ai-dev

[warm tiger]

Good luck with that, I’m already having trouble setting a proper context formatter for a non-runtime dialog writer (usual issues with too much exposition, low understanding of relationships etc, and thats with models as big as GPT5.2)

[full robin]

Thanks. Yeah, I guess I'll need some luck if I'm to make it work reliably. I had some success with my previous version, but making it work like a coding agent seems to pose a whole different challenge.

[mellow osprey]

Hey AI-Dev folks! 👋

I’ve been experimenting with different AI models as game masters, but I kept running into the issue of exceeding the context window. That meant NPCs would have “bleed” and things would get mixed up over time. I tried this across a few different AIs—I won’t name names, but let’s just say a few well-known ones—and it got me thinking about a more reliable way to keep the world state consistent.

So, drawing on my old database programming days (way back in my twenties!), I decided to write everything back into a relational database. This way, the AI acts as the game master but always pulls each scene’s state from the database like a file cabinet. When you enter the library, the NPCs have the correct memories and relationships for that scene. When you move to a new area, the updated memories get saved back to the database, and the next scene loads accurately.

Basically, I’m using a Python CLI (command-line interface) to grab all the info from the database and feed it back to the AI for each new area, so the AI can run the game smoothly without losing any context. Thanks for letting me share this in more detail!

[full robin]

I'd recommend checking out SillyTavern if you haven't already. Although it has only rudimentary extensions for the long-term memory feature, it has a pretty complex RAG system, such as 'World Info'.

[mellow osprey]

Thx

[wind quartz]

deafknightjr thanked mysticfall

[full robin]

I managed to make the new Response Client work with OpenRouter/MS Agent Framework. I'm not sure if it's better or how the developer message works when the history grows, though.

[full robin]

I hit an unexpected roadblock - the indefinitely running agent model worked relatively well, but many APIs like chat store or reducer in MS Agent Framework seem to get triggered when a (non-tool) response is received.

Maybe such a model isn't how an agent system is supposed to be used. It's not a big problem, however, since the other refactoring results like using a conversational history instead of an one-shot prompt are still valid.

I'm thinking about instructing the agent to use responses as a way to keep its internal thread of thoughts/plans, similar to reasoning responses but for itself rather than the user.

Previously, I implemented a checklist and memo system for such a purpose as tools, but maybe a short plain-text responses could work just as well.

[mellow osprey]

Hey — just wanted to follow up now that I’ve had some time to experiment more.

You were right that SillyTavern wasn’t an exact fit — I treated it more as a reference point, like you suggested. I spent about 10–12 hours poking at it and it helped clarify the core issue I was running into.

What I kept hitting across different AI GM setups was context window bleed over time — NPC memories drifting, relationships getting mixed, scenes losing continuity. That’s what pushed me back toward something closer to my old database days.

What I’m building now is a Hero System–faithful setup where the AI acts as GM, but all world state lives in a relational database. The AI doesn’t “remember” the world — each scene pulls state from the DB (NPCs, relationships, perception, etc.), runs the adjudication, then writes the results back before the next scene loads.

I’m using a Python CLI to fetch only the scene-relevant data and feed it to the AI, so it can GM cleanly without accumulating drift. It’s kind of like a hard-state version of what you mentioned with RAG for world lore — but with strict persistence and rule enforcement instead of soft recall.

SillyTavern was useful to look at, but it confirmed I’m less in the RP space and more in the simulation + adjudication space. Just wanted to share where that exploration landed. Appreciate the earlier nudge — it helped narrow things down.

[full robin]

Yeah, ST is focused on RP, and an ideal RAG system for RP could be quite different from the one for a simulation type game, as you noted.

Still, the important question could be what information to pull than from where. If the information is simple enough, you get get away with dumping everything into the context. But you'll see problems like memory loss or mixed up relationship like you said, if you put too much info into the context.

ST tries to mitigate this by implemeting a very complex conditions & rules to determine what information to pull from the World Info. It doesn't matter how you store the source - RDBMS, SQLite, VectorDB (it may matter if you thinking about using an embedding), etc. - but how to select the essential informations out of it does.

Another thing to consider could be adopting an agentic architecture. Letting the agent dynamically pull relevant information could be much more efficient than populating a prompt from a RAG store before sending it to LLM.

Also, you can try things like building an agent workflow, if latency isn't too much an issue, to make a 'guardrail' agent can verify the validity of what the GM agent responded, for example.

But probably there's no silver bullet in such a matter, and each game may have different requirements. After all, there haven't been enough number of people using LLMs/AI agents in their games yet to establish best practices of such a usage.

[mellow osprey]

Yeah, that all makes sense — and I agree there’s probably no silver bullet here.

Where I ended up landing is that for my use case, the selection problem is more important than the storage problem, like you said. I’m less worried about whether the source is RDBMS vs vector DB and more about being explicit about what the GM is allowed to see at a given moment.

That’s why I’ve been leaning toward a very explicit “scene snapshot + recent events” model instead of broad RAG pulls. The idea is that the GM agent never has to infer relevance from a large lore store — the engine tells it exactly who’s present, what they know, what just happened, and what rules apply, and everything else is effectively out of scope for that turn.

I’ve thought about agentic approaches too (guardrail / verifier agents in particular), and I may experiment with that later, but right now I’m trying to keep the first version as deterministic as possible so I can tell whether errors are coming from model behavior or state selection. Once that’s solid, adding verification layers feels safer.

Totally agree this space is still early and that different games will want very different tradeoffs. I’m mostly trying to get something that behaves like a by-the-book tabletop GM — slow, constrained, sometimes boring — and see how far that paradigm can be pushed.

Appreciate the thoughts, though. They’ve helped me frame the problem more clearly.

[full robin]

Good luck, and please consider sharing what you find when you make progress. 🙂

By the way, I'm in the middle of refactoring my project which used the “scene snapshot + recent events model" into a "minimal context + agent pull model" at the moment.

The previous model worked well. So I know it can be a valid approach.

[mellow osprey]

Current rudimentary UI to test and debug my Social Engine. It does not have the questions I asked you will have to infer. Working on it. The scene is Location ID 7 in a table which is Rebecca Law's back yard pool area. The images on top are the NPCs present and I play Steve Savage.

[full robin]

That looks promising 🙂

By the way, I stumbled upon this the other day, while browing the SillyTavern subreddit:https://github.com/SpicyMarinara/rpg-companion-sillytavern

Haven't tried it myself but the screenshot looked interesting.

[mellow osprey]

Thanks I will check it out tomorrow evening.

[wind quartz]

deafknightjr thanked mysticfall

[full robin]

https://coral79.github.io/frankenmotion/

[full robin]

Now that we have a practical way of making an NPC speak naturally, and reactively without scripts, I think what I linked above could be the other half of the puzzle that would give us human-like autonomous NPCs in games.

The link is just a paper so it may need some time before we see a usable implementation, however.

[Coming Soon] We will release the full code, pre-trained models, and the Frankenstein Dataset.

On a side note, I think it might make sense to make a special kind of PC peripherial just for running AI models.

I'm already running a TTS, STT, and lipsync generator locally in my game, so my graphics card might lack VRAM to run yet another AI model.

[tiny hawk]

That's AI structure for My Romance game

I know it's hard

But there's nothing I can do

[compact wedge]

how do you make this chart?

[tiny hawk]

I first made in rough paper then AI generated

[compact wedge]

ah that makes what ever is going on after the candidate actions make sense

[tiny hawk]

Yeah

Since it's R&D stuff it should be hard

[compact wedge]

I should do stuff like this

especially if im going to try and make increasingly more Game Ai-heavy games

[tiny hawk]

Yeah, you should try it If you’re aiming to make more ai heavy games anyway experimenting like this now will help a lot Even rough tests can teach you what works and what doesnt

[full robin]

I need something like that! I had some of those parts as rough ideas in my head, but seeing it in a chart made it much clearer 🙂

My version would be much simpler and some parts wouldn't be applicable in my project. But that was enough to clear up my ideas.

On a side note, I find it interesting your version also starts from a "perception -> observation" loop, which is exactly how I implemented mine.

I thought the abstraction wouldn't be too common, since I suspected modelling sensory data would be too fine-grained for most other games.

Relevant parts:
https://github.com/mysticfall/alleycat/blob/main/src/Ai/Mind.cs
https://github.com/mysticfall/alleycat/blob/main/src/Ai/Agent/AgenticMind.cs

I did some experiments about having dedicated planning and decision making stages, by the way, which didn't go well.

They work relatively well in what they were supposed to, but were too slow in a realtime interaction.

So I consolidated everything into a single agent and in the process of testing it to see if it'd be capable of making reasonable decisions.

I was thinking about reviving the planning part as a parallel loop if needed as a compromise.

I can't do any AI stuff until an issue I reported for Language-Ext gets resolved, though.

[full robin]

https://github.com/QwenLM/Qwen3-TTS

Looks interesting, especially its voice designing and natural language instruction features.

[mellow osprey]

I am working on a Social Engine and your chart inspired me to compare my flow.

[tiny hawk]

tbf thats chart was basically brain architecture , Then i just modified it to an AI simulator 😛

[tiny hawk]

yeahh, thats basically tradoff i ran into to, i did some test with split decision -> plainning stages and the behaviour was cleaner, but the letency cost made it slower in Real time
then what better worked for me is kinda hybrid like
-> keep a fast reactive loop every tick (utility scoring + pick next action)
-> run a planner at lower frequency / in parallel (every N ticks or on big events)
or -> treat planner output as a proposal that can be cancelled replanned if the world changes

[full robin]

Yeah, I think the split model is the way to go. I simply can't think of any other model that would work in this context.

[devout geyser]

Hey folks, I am currently working on a NavMesh navigation system. Currently, I am using Recast to generate and query my NavMesh. My game has strategy / building builder elements where the NavMesh needs to be updated (using the re-tile functionality with Recast to handle this at runtime). Because this needs to be somewhat real time I turn down the tile to terrain resolution (aka the sample rate of the base terrain along with the marching cube size of the mapper). This results in a NavMesh that is great on flattish terrain, but adding any hills and the NavMesh pokes through the terrain.

My solution:

• I periodically project the navigating Agent's world position back to the NavMesh. This requires a query of the NavMesh with a specified search bounds.
• I navigate the agents based on the NavMesh, where I need to query agent and target positions from world coordinate to NavMesh coordinate (projection).

The new problem:

• In general this is something like (simplified examples) TerrainPos: (0, 1, 0), SearchExtents (1, 1, 1) - but periodically given the error from terrain to the NavMesh we exceed the search extents (y) coordinate.

Literally typing this all out I think the solution is to not project the agent positions from the NavMesh at all and treat that as the source of truth. Only in Rendering do I project the agent back to the terrain which can be as simple as a heightmap lookup and/or a raycast into structures simplified collider (example walking on a bridge). Welp, too late to turn back now. I might as well ask it as a question, would this be how you solve the problem (use NavMesh position as the source of truth of agent positions and project/raycast onto terrain/detail meshes at render time)?

I should also mention that find nearest position on navmesh using Recast is kind of expensive when I do it per agent per-frame (source+destination). Man that was a lot of words and I think I have my solution. Thanks

[alpine edge]

How long until someone tries vibing pathfinding by feeding the graph to an LLM?

[full robin]

I have no plan of doing something like that, but I'm thinking about feeding an LLM a top-view of the navigable are and character positions to let it determine the coordinates to walk to for the character it controls.

[quick pelican]

Dont these benchmarks show model to model improvement?

Where did the claim that "ai is already better than most humans" come from?

This isnt really supporting your claim

[ashen plaza]

holy crap we have an ai friendly space?

[agile depot]

sorta this room was for ai-dev

i mean if you look at the description its meant for using ai for generative diolog etc not for general ai dicussion of coding

there should be a agentic coding room

[ashen plaza]

So my panmathos, an ai controlled universal engine fit here? lol

[agile depot]

yeah

[agile depot]

look at where humans typically rank when triying to do the same tests

[ashen plaza]

do you trust a review from pro gemini model, or pro chat gpt 5.3 inspecting code and determining functionality without context? say, would you believe a review of a project folder to be factual, (function-structure-capability?)

[hasty horizon]

have you tried claude code yet?

i ask because you seem to be all-in on codex

[agile depot]

i tired it a little but couldn't really do much w/o paying for the full plan and i already got openai

i didnt see it too be much better..

I think Codex is much better at staying precisely aligned with my style

I give it detailed plan docs and it loves it.

[tender schooner]

i personally use deepseek whenever im doing sanity checking because it's insanely cheap

[ashen plaza]

I want panmathos on that list some day..

[ashen plaza]

Whats funny is when i started panmathos (was dataforge), allot of what i "envisioned" i was told was impossible lol.. and i did it anyway.

[full robin]

I'm glad to see active discussions on this channel. I wish I could talk more on stuff I do with AI than on AI itself (e.g. ethics), but probably it'll take a month or so before I can revisit the AI part of my game project.

On a side note, I've been rather slow to adopt code agents, although I've long been using an AI assistant that shipped with the IDE.

In a month or so, I'll start a pilot project at work which I'll use to evaluate the best agent option for the team and establish related conventions.

For now, I'm leaning towards IntelliJ ACP + OpenCode + OpenRouter, for maximum flexibility & integration.

My focus will be creating a setup (e.g. agent policies / instructions) that'd allow the team to reliably run agents to implement new features and fix bugs.

[ashen plaza]

Panmathos will allow me to have "smart npcs", ones that can remember and dynamically change over time. My mmo is designed to be an evolving world with an environmental pushback, i needed a complex list of capability's, way more than i alone can easily create and work with, so panmathos IS my team lol, it my own solution to generativ ai flaws. My goal is i design the astetics, anatomy, environment, world atmosphere, just once, and the rest of the world is generated by my ai in my image, although in this process i have discovered a few things, found additional "good use" panmathos could be for me, now its something special lol.

[full robin]

@ashen plaza Do you have a Github repo for Panmathos? It sounds interesting. 🙂

[ashen plaza]

No i have not made it public in any way, it owns a hard drive on my pc and i have 30+ previous builds sitting on my primary hard drive lol, been allot of walls and new ideas along the way. i like to solve problems i encounter, not skirt arround them.

leyoki will be, thats the game side of things, and if panmathos really is everything i hope, ill package and build modules for other things, book writing, record keeping,ect.

[ashen plaza]

so far ive been met with mostly negative reactions about my project. so i never ended up working with anyone, hence another reason to build my own team lol.

[tiny hawk]

That sounds ambioutius project, don’t burn yourself out ship it step by step

[ashen plaza]

when i feel the burn, i switch to a smaller game or smaller project for a bit, usually that project involves a small element of my bigger design, so i can take a moment to test a smaller system.

Ran an inspection through cluade, its been the nicest lol

i feel like im close, the hardest systems i have already designed and implemented sucessfully, a few "cant be done" things lol

[ashen plaza]

Currently upgrading to work with 3d lol

[full robin]

Not gamedev related, but I just managed to complete a very tedious task using an agent, which would have taken me at least couple of hours to finish:

It only took a few minutes to write the prompt and the result seems to be ok (I'm still validating it).

[tiny hawk]

OpenAI says GPT 5.2 helped scientists discover soemthing new in theoretical physics, it found that a partical interaction people thought could not happen maybe its a special case, but the result is published as preprint and still needs revieve

its currently unconfirmed in the scientific sense yet

[prisma acorn]

Chat is this Ai?

[wicked igloo]

So uhm, I created a platform where you can configure an AI and github access. From there you can chat with AI, explore a git and create tasks of what you want to do. Then you can ask the AI to spin up containers wherein AI will implement the actual tasks and create a PR. You can give it feedback, sometimes it can ask u questions, etc. With every PR it also automatically give you a preview url where you can see the app in the web. All secured per user (even the preview urls), AIs nicely contained and safe and no code is pushed without human reviews. Devops is all done by the platform, fronted/api/databases/links/plugins/configuration/env-vars/secerts/dns, you name it, all handled (and/or configurable).

With this platform you can pretty much run a dev team.

[full robin]

Sounds useful. But don't many MCPs exist for similar tasks? What does your platform do differently from them? (When I searched for "github" on mcpmarket.com, it returned 5988 results.)

And how modular is it? Many dev teams mix and match different platforms/tools for Devops. So, it'd be better if it supports different platforms other than Github.

My company uses Jira + self-hosted Gitlab + Jenkins + ArgosCD + K8S on Naver Cloud. But the previous company used Jira + Bitbucket + EKS on AWS, for example.

I still think such a platform could be useful. But I'd be interested more if you can provide some information about how it's different from the rest. Maybe some Github project page for your project?

[zinc raft]

What?

[full robin]

I think my project isn't a good candidate for using a code agent, but I decided to give it a try.

As a first step, I'm going to ask an agent for a code review of the current codebase:

It did a fairly extensive code review, but I feel some of it was rather superficial. The conclusion it reached was more or less spot-on, however:

(Except for 2 and 4, which are due to misunderstanding of the architecture.)

Probably I better focus on 1 for now, and have a more in-depth discussion with AI.

I don't know how I can standardise between Eff and IO - I don't think it's even possible. So I'll skip it for now.

As for 5, it's more or less due to my laziness, but I found this project that I can use for that purpose the other day:
https://github.com/chickensoft-games/GodotNodeInterfaces

[full robin]

It's more of a programming discussion than that about AI, but I think it can be an interesting case of using an AI agent for a non-trivial refactoring & code reviewing process.

The AI understood the core issue I explained perfectly, but I think it didn't get the "static constructor" idea as well.

It understood what I meant by a "static constructor" correctly (which was impressive, since it's not a commonly used term), but I don't think its reasoning is sound.

[full robin]

Still doing the code reviewing session with the AI:

[wicked igloo]

What I'm trying to make is the following:

• Allow users to code with AI in the cloud with their git repo (think claude code GUI with any model in your browser)
• Provide quick previews (URL accessible builds) of PRs
• Provide environments like DEV and PRD with domains set up
• Provide a Kanban like workflow and overview of tasks
• By standardizing most of Devops through a giant complex yaml file that supports everything from VMs, docker, k8s and more. (which an AI can easily create)
• By integrating most planning tools (git, jira, etc) with webhooks to update the tasks overview
• If everything works well together, I've a product that essentially provides a small team of AI software devs that can ship a product from scratch to PRD (targeted for small companies). And if needed, a dev can jump in anywhere in the flow to have a human in the loop.

[full robin]

Interesting... but what benefits you expect from using your platform instead of, say, some agent orchestration tool (e.g. Oh My OpenCode) with (MCP) tools for integrating existing infra (e.g. Jira, BitBucket, AWS, etc.)?

[wicked igloo]

The benefit for you is that it's all in one place, which is only a small leverage

But the benefit for small companies, especially small non-tech companies, this can be huge

You're a software dev, if this product is trying to do most of your job while you can do it yourself, it won't be of much benefit for you

[full robin]

I guess, some small companies building a DevOps infra from scratch might be interested in using such an one-stop solution. Still, I'd feel concerned about vendor lockin.

But yeah, maybe it could be because I'm looking at it from a software dev's point of view, instead of from that of some non-tech oriented person searching for a carefree way of deploying their service.

Anyway, good luck with your product 🙂

[wicked igloo]

Yeah, I don't vendor lock though (it would be smart, but I don't). You've access to your git at any time and can do whatever you want with it. You don't even have to use my env deployments if you just want PRs and previews.

[full robin]

By vendor lock in, I meant that there are a lot of choices for each component in DevOps infra, like AWS vs Azure vs GCP, for example.

[wicked igloo]

And you can still use those

as the yaml file supports external services

The yaml file can have a record of external service of AWS S3 buckets which requires Config X and Secret Y. The platform sees this and provides config and secret options to be filled in for each env.

[full robin]

Well, I defer my judgement until I see it in action (or at least documentation) then. 😅

[wicked igloo]

Fair, just hope competition won't kill me in the short run

[alpine edge]

so you're making an ai-first PaaS

doesn't Vercel have AI features though?

[wicked igloo]

Vercel does, and very good ones. But on techstack it's limited to NextJS and postgres, which is very icky to me.

Something as simple as, "Make me an excel tool to automatically do this with AI" (a very common business ask right now), it would probably get stuck on what to do with the AI part or implement it very badly.

[median pawn]

Hi 👋 ,

Repost of https://www.linkedin.com/posts/mathieun_machine-learning-roundtables-gdc26-activity-7431604766406967297-QXyj, feel free to send a connection request.

As part of the GDC2026 Machine Learning Summit, I am co-organizing two roundtables.

To make them practical, lively, and grounded in what actually shipped, we put together this small survey: https://forms.gle/jPxvfh6Xw6z6Hza26. You can contribute to shape the discussion and receive a written summary of the roundtables even if you don't attend GDC.

RT1 - AI/ML Runtime Techniques (rendering, compression/decompression, behaviour, animation, bot, …) is Tuesday, March 10, 2026 at 4:30 PM (PT). This is about what made it on screen, what survived real constraints, and what got killed along the way. Not theory. Real systems in real games.

RT2 - ML in Games Development and Operations (tools, genAI, vibe coding, GaaS, DevOps) is Thursday, March 12, 2026 at 1:50 PM (PT). This is about what actually stuck in practice over the past year, what collapsed at scale, what changed how teams ship, and what got quietly rolled back.

[full robin]

I tried OpenClaw yesterday and I was seriously impressed, and that was even with so many bugs and missing documentation/features that I had to fight against to get it working.

As a test-drive, I created an agent which can serve as my "friend" on Discord with shared interest on dark fantasies I have.

I gave it a role to chat with me on this theme and build a lorebook from what we talk. Whenever the agent find something interesting, it creates a Markdown file for the subject, and periodically edit and reorganise the file structure to make it look like a local Wiki.

It worked like a charm, and I really enjoyed talking about the topic which was a bonus. I'll probably add things like a custom command to export the current lorebook to a PDF and post on Discord, and vibecode some tool to export it to SillyTavern's world info format.

If it can work so well for such an unusual task like that, I can imagine how it can be made useful for many other more practical purposes.

[full robin]

I'll probably create another agent which can monitor Gitlab for MRs and notify me after doing a code summary and preliminary review.

I'd just add a review bot normally, but the new company is using a free self-hosted version of Gitlab, which is managed by a different team.

I usually do 3 or so code reviews a day, which can be time consuming when they are not trivial changes.

[full robin]

Added another agent and successfully made it checks for new merge requests that need my approval every 30 minutes during work hours/weekdays.

I gave the agent a personality, so now I'm working in a roleplay setting. 😅

[tiny hawk]

<@&133522354419662848> <@&1224208037892591666>

[wicked igloo]

I've decided to scale down my idea of creating a AI devops platform to targeting small non tech businesses

Now I've a platform where someone can ask an AI to build a website. You start a project, it provides a basic first index.html page. You go to your website, only for the admin (the website creator) there is an AI chat tool to talk with. It has the ability to write full html web apps, it shows the changes right on the website for you. You can select html elements to guide it more specifically if you want. You can undo/redo/reset/save changes. The platform provides services like Auth, User&Role management, Stripe for billing, Sending mails, linking you domain and storing data per user/website. The AI knows about all these features and can hook it all up himself.

[full robin]

That sounds more concrete and reasonable in scope to me. Good luck with your service. 🙂

[wicked igloo]

I've started like this weekend

And it's already blowing my mind

of how good this works

Can't put this online yet because it's not security tight yet

[carmine burrow]

Here please don't hate too much uch

[elder hazel]

looks fine

[silent vale]

Nice

[carmine burrow]

Yes wanted to ask if they look decent or not I did try to make it consistent in other poses but still beginner

[wicked igloo]

The AI website builder was having trouble with images. Like giving them a name that does match the image content and then used for wrong placements on the website.

So I gave it the ability to describe and understand any image and now it has no problems at all.

[celest ravine]

Yep, sounds like a issue i had during the making of my own AI as well 😭

but i also fixed it with the same method.

[agile depot]

FYI...

[tiny hawk]

https://youtu.be/RJyPVLMyyuA?si=AkHY5DI5yKdZrP92

Found this accidentally

[full robin]

Looks like I'll finally have an opportunity to try agent-based dev on a new project.

So far, I've only used coding agents to refactor existing projects. I'll be using this opportunity for establishing the project wide conventions/practices regarding agent use, and see how far I can automate the development in my team.

I'll likely to start with an IntelliJ + OpenCode + OpenRouter setup for now.

[wicked igloo]

I think this channel came from AI behavior trees and simple game agent trainings. Anyhow, we took it over with LLMs

[full robin]

Half way done... I made a system that allows you to generate requirements from a JIRA issue and Figma design layers, and an OpenAPI specification from them. It also creates a design guideline "skill" from associated Confluence pages, so coding agents can use them to work on each project following team conventions.

Now I need to create an agent that can find differences between source Figma designs and the current page under development. I suspect it might be a challenge since the current Figma documents weren't designed to support such automation. Figma provides many such features, but our designer didn't seem to know that.

[wicked igloo]

What is the purpose of the requirements and OpenAPI spec though? Shouldn't those be clear before creating the JIRA issue and Figma designs? Seems backwards ...

[full robin]

Usually things like an OpenAPI spec or detailed requirements come after opening a ticket (JIRA issue, in my case).

And in real projects, it's often the case the development has to begin with merely a few lines of description and a Figma document due to a pressing schedule.

As for OpenAPI, many projects generate the spec from existing code, not vice versa They use things like Java annotations on Spring Boot controllers for that.

It's a valid case, but not desirable if you're going for a more spec-driven approach.

What I'm aiming at is keeping the most high-level artifacts like a JIRA ticket as the "source of truth", and establish an semi-autonoumous, but iterative workflow to generate concrete implementation out of them.

Anyway, I've made so far as to generate a full React/Redux Toolkit/RTK application with full unit tests that look reasonably similar to the design document.

Looks like I'll need a couple of days more to refine the process to make it usable for production.

I won't be able to do any gamedev this weekend because of it, though. 😦 I really hate when I have to work through a weekend, but at least I'll be doing something interesting this time.

[wicked igloo]

But, if the JIRA tickets are the source of truth, they need to be monitored and carefully crafted to have a desired outcome

And JIRA issues from users are usually not a "good" source of truth

[full robin]

In my project, the tickets aren't coming from end users but they are managed internally by the team.

[wicked igloo]

So then the requirements are probably present before creating the issue

[full robin]

Not really. Say, if we decide some new feature is needed, we just open a ticket for it. And we refine it by adding requirements or design documents, using tools like Confluence or Figma.

I believe that's the most common way of managing projects using a task system like JIRA.

[wicked igloo]

Why not just put an AI inbetween or next to your ticket system

Why put the AI behind ur ticket system

[full robin]

To minimise the cost of manually writing specs and code?

[wicked igloo]

but why write the spec after the ticket

Like code by AI sure.

[full robin]

As I said, tickets typically precede detailed requirements, like they do in popular methodologies like Scrum or Kanban.

[wicked igloo]

I would make AI tools that help write/maintain the JIRA ticket AND the spec. And then let code AI figure it out from there.

[full robin]

It's only natural to think of doing something before having a detailed plan for that.

[wicked igloo]

Sure, but if the ticket is the source of truth, it should also be the spec

And should co-exist and both be created by the AI

[full robin]

That's not incompatible with what I said. In fact, they often write specs in Confluence and link them in the JIRA issue.

But typically, tickets come first because they are often created without having a detailed implementation plan. They evolve as people spend time refining the details, and when they become concrete enough to be implemented, they are moved up the backlog, then included in an active sprint.

[wicked igloo]

Yes, so when do you then use it to write spec?

Right before AI dev or just during the ticket creation/refining

[full robin]

In this particular case, I only had a Figma documents with some comments, and a few lines of description in the ticket. For a coding agent to work properly, you need a more concrete plan than that. So, the specs must be written, and it doesn't matter if a human or AI does it.

In that workflow I'm making, there's an OpenCode command to write specs as local Markdown documents, and I instructed it to use all documents linked to the JIRA issue as sources.

It's because it'd be very inefficient and error prone to tell an AI agent to read JIRA/Confluence/Figma everytime.

You can, of course, use AI to write detailed specs on Confluence or even create Figma designs. But it'd much better to have them as local files in a structured format also.

[wicked igloo]

Sure, the markdown artifact is useful, I won't deny that. I'm just saying, it would be even better if the artifact is generated next to the JIRA ticket, not "only" when passing the ticket to AI dev.

It's just a perspective of the UX.
Instead of User > JIRA > SPEC > Dev you would have User > JIRA + SPEC > Dev.

[full robin]

For now, I plan to commit those artifacts along with the prompts and configs for AI.

As I mentioned earlier, tickets usually come before specs in a vague form. It's because tickets are created whenever people come across an idea, like in a meeting. You don't usually write detailed requirement documents on the spot.

[wicked igloo]

It's fine man, my point isn't coming across and I'm over it. Github is releasing this AI flow anyway so whatevs

[full robin]

Workflows can differ between projects and teams. Although what I mentioned above is the most common practice.

[wicked igloo]

It's not, because you see maintaining and refining a Task in the same step as Dev.

[full robin]

Have you worked in a Scrum, or other Agile project?

[wicked igloo]

Yes I have

I've worked with JIRA, Azure Devops, hell even simple github and trello kanbans

From small to large teams

[full robin]

Then you must be familiar with the concept of a backlog or sprint. When you open a ticket, it goes to the bottom of the backlog by default, because it's usually vague and lack detailed requirements needed for implementing it.

[wicked igloo]

Yes, that's the initial start of a ticket, sure

[full robin]

Then people progressively refine it and move it up the backlog until it becomes detailed enough to be included in a sprint plan.

[wicked igloo]

Yes, we get in meetings, create designs, define features AND write spec over time of the existance of the ticket BEFORE it goes to DEV.

[full robin]

Yeah, "over time" is the keyword here, that is after the ticket, not alongside it.

[wicked igloo]

The ticket stille exists, doesn't it?

[full robin]

Yeah, it does.

[wicked igloo]

So it's not AFTER the ticket, it's DURING

[full robin]

I didn't think you understood me as saying we are doing something like deleting the ticket once requirements are ready.

[wicked igloo]

I didn't think you understood me as saying bro, I don't know what you're saying right now.

What I understood from you is that you're making something that once a ticket is ready for DEV, the flow runs, writes the spec and does the dev with that spec.

[full robin]

Then I guess that was where the confusion came from. We use Confluence where people write requirements for tickets by hands. My system isn't meant to eliminate Confluence entirely. In reality, some tickets have more detailed requirements attached to them than others, and some of them are even entirely missing them. All I did was creating an AI instruction to collect whatever is available at the time of running to compile them into something a coding agent can work with, by filling the holes and reorganising the structure.

[wicked igloo]

Yeah, and I'm arguing, while that's great to have a tool to fill the missing pieces. The tool would be more useful if it could do both fill the missing pieces before the coding agent gets it AND help someone write spec alongside refining the ticket before the ticket is considered ready for dev.

Instead of a tool that only fills in the gaps after the ticket has moved to dev, it would now be a tool that exists next to the process of refining a ticket.

[full robin]

It'd be nice if the generated documents could be automatically fed back to the source system, but I'm afraid it will complicate the things too much for me. At least, my system can synchronise the source if it got updated after the generation of the specs is done.

[wicked igloo]

Well it would be an artifact tool or whatever that hosts the documents and adds links to the docs in the ticket. I can imagine if you want to be JIRA native, you would have to write a jira plugin which is an headache yeh.

[full robin]

Atlassian added their own AI bot recently. I'm not sure if it also helps users writing documents though. Personally, I'm not entirely convinced by the idea of making the same tool to serve as a writing assistant for humans and a compilation util for coding agents. They look similar, but have a slightly different set of requirements. The Markdown specs, for example, aim to be a concise instruction for agents (to save tokens and improve prompt adherence), while those for humans can be more verbose, and can contain additional images or elaborations that would only confuse AIs.

[wicked igloo]

Well, we devs are moving towards a world where we'll be prompting AI the best way we can. And if you box off the artifacts to only the prompt generation for the code agent, you might be missing out on stuff. While if as a team you can be generating these artifacts during/after meetings and review them before handing it to a coding agent.

It indeed would have to be clear that these are artifacts for AI, they need to be concise, token efficient etc.

Prompting skills will be a skill/job in the future.

[full robin]

I do agree that it would be ideal to have AI and humans work together using the same workflow. But in a meanwhile, I don't need a writing assistant as much as I need a spec compilation tool, and I have only a limited time to build such a system (which is why I have to work through this weekend 😉 ).

[full robin]

Now it can correctly delegate to a subagent to open a page in a browser and compare it with the Figma design source to find visual discrepancies.

And it can feed the information back to the main agent so that it can ask the coder agent to fix the issues:

[full robin]

The back and forth between the "tester" and "coder" agent is working pretty well now:

[wicked igloo]

What model are u using?

[full robin]

Mostly GLM5. I used Codex for complex tasks, and Qwen 3.5 for visual testing.

[zinc onyx]

hello

[zinc onyx]

Guys.. I'm tinking of launching an Ai LLM API for $10/month unlimited usage

what you think will anyone wanna try it?

[peak totem]

Whats the opinion on antigravity’s agent for game dev. I made a game in jsut an hour without touching my keyboard whatsoever. 1 prompt in chatgpt asking for a detailled prompt for the software. It did everything. Assets to index.html

[wicked igloo]

You're usually limited to static or simple non animated 2D assets. And art style are what the game feel like a game, so yeah. So while antigravity or any other AI coding tool can program a game pretty easily, it can't "make" a game.

[wicked igloo]

An API is useless if it doesn't show stats of the AI model. Is it in the same ballpark with GPT 5.3 or Claude Opus 4.5? If not, you're competing with cheap open source models.

[full robin]

I'd consider subscribing if the service is comparable to OpenRouter, but I don't think it'd be easy to offer as wide a variety of models at similar prices as they do.

As for making games with AI agents, the issue with using them for anything related to art (which includes games) is not what it can generate but how much control you have.

AI can be a legitimate tool if you can establish a workflow that gives you enough control to steer the output to what you want in an iterative way. Otherwise, it'll just generate slops.

And it can be challenging to do that because games are difficult to exactly describe or verify by prompting alone.

It's far better, IMO, to use AI agents to enhance the existing pipeline (for which they can excel) rather than trying to "vibe code" the whole thing at one go.

[hexed sedge]

yo what do you guys think about an ai that create ais for exactly what you ask him for
and i know there is already ais that does that but thay keep crying about laws

[full robin]

What do you mean by "AI that creates AIs for exactly what you ask it for"?

Maybe an example would clarify what you mean?

[tiny hawk]

Probably he is talking about AI systems that generate other AI models,
like Neural Architecture Search (NAS) or AutoML systems things

[hexed sedge]

it generate ais by researching by it self and create it for you
if you haven't notice i like automating stuff

[full robin]

By "generate ais", do you mean things like training a model?

Or do you mean things like generating prompts, skills, mcps, etc.?

[hexed sedge]

training a model

[full robin]

That certainly piqued my interest then. Can you elaborate on typical use cases you have in mind?

[hexed sedge]

i just want to reach a point that i can automate everything 😁
there is no big thought behind what am making just breaking through the limit of what i can do with less time

[full robin]

Honestly speaking, I can't find anything interesting to add, in that case. Good luck with whatever you're trying to do, however.

[hexed sedge]

i've been stuck here any tips

I've changed the code for the vision_manager a thousand time and the same problem keep showing

[hexed sedge]

fixed it

[hexed sedge]

Hay is there a way to get open ai API for free or something else works the same or should I make more template's

[agile depot]

My bulletproof Agents file. refined over a year of wok with Open AI CODEX.

[full robin]

Try one of the model providers. OpenRouter, for example, usually offer (often less powerful) models for free.

[wind quartz]

mysticfall thanked ttv_qsilvaq

[full robin]

Dang... I didn't think it'd make me "thank" the spammer 🤦 Please, delete the message as well.

[zinc raft]

<@&1224208037892591666>

[earnest vapor]

has anyone been playing around with Trellis2 yet?

that looks awesome

apparently it can generate things inside transparent containers - like fish in a bowl of water

[full robin]

I haven't spent much time on 3D-gen stuff yet. I only used Sam once to create a mocap from a video. I already have a good 2D AI workflow, so I'm planning to use it as a basis for generating 3D models, using things like SAM, Hunyuan, or Trellis.

[earnest vapor]

getting hunyuan to work locally was fucking PAIN

it's good tho

to an extent

not as in it creates useful 3d models- but if you know what you're doing, you can make them useful

[full robin]

I can use a Runpod instance for such tasks 🙂

[earnest vapor]

in that case go with trellis... hunyuan works on consumer gpus, my 3060 runs it perfectly well

trellis needs 24gb - which is i think 4090 territory? 5090?

may aswell rent a gpu then

[full robin]

That sounds nice 🙂 But I can easily run them on a cloud when needed.

It's not that expensive.

[earnest vapor]

no? how much are we talking for let's say 2 minutes of generation time?

[full robin]

Let me check.

[earnest vapor]

o shit

that is quite affordable

huh

i may tinker with that

[earnest vapor]

prepare with an asset list, spin it up, let it run through all the generations quickly, and then not use it for a while again

[full robin]

This is the current price table for 24GB+ machines on Runpod.

Yeah they are pretty affordable for such a task.

[earnest vapor]

might be worth a consideration to even go serverless if all im doing is just quickly generating 20 3d models and then spending a day post processing them

[full robin]

I have a serverless setup for a work project. But it's more suited for automated stuff, like running a render farm.

I'm running Blender on a headless mode with that setup.

For generating a batch of 3D models occasionally, I think just burning credits for a spot instance would suffice.

On a side note, I'm cautiously positive about those 3D models although I haven't used them much yet. From what I've seen they seem to do a decent job at image-to-3D tasks. Combining with the 2D workflow I have already, and with a help from a retopo tool, I feel I'd be able to create usable models as I intend them to be.

The only part that I'm unsure of is texturing. But I can do it myself with Substance/Blender if needed.

[hexed sedge]

Yo I got into AI 7 months ago I was just a programmer got in game dev but blender was kind of impossible for me so I decided to make an AI that can do anything in blender I tried bpy and it broke everytime so I decided to make it work by seeing the screen and simulate a mouse and keyboard, got any tips.

I am trying a lot of things in ai
Mostly trying to automate things and it kinda of easy

[earnest vapor]

well sorta.
the stuff they spit out can't be really used as-is, but if you know what you're doing you can always make it work and use it as a good starting point

[earnest vapor]

generally blender and ai REALLY don't get along, because the software has changed dramatically over the years, and all datasets are poisoned with outdated, wrong and unworkable information

[full robin]

I expect I'd have to do some retopo, but it's still better than doing it from a scratch 🙂

[earnest vapor]

always depends on what the endresult should be

i found 3d ai very useful when trying to achieve a ps2 style look

[full robin]

On a side note, what I really need is a texture generator that can create seamless, photorealistic skin maps for a given UV map, like that from MakeHuman models.

But maybe I'm approaching it wrong when we have DLSS5 now...

[earnest vapor]

that's probably not happening any time soon, but skin is mostly a shader issue anyways, you can get away with almost no texture in many artstyles when making skin

[full robin]

I briefly explored options for making a photorealistic character before AI took off. And I was apalled by the amount of work it'd take to create texture maps like those used by the Digital Emily project.

Like you said, it wouldn't be an issue for stylised models, but when you're aiming for photorealism, it's really difficult to do it by hand.

[earnest vapor]

ye it's always a question of how much you're willing to scope down.

anything photorealistic will always take an enormous amount of time, no matter how many tools you layer ontop of it

even if we said DLSS5 is perfect and makes faces look photoreal-

it'll look creepy as shit, because the animation isn't photorealistic, which leads to a massive disconnect and dives right down the uncanny valley

so now we gotta ai the animations too, which will screw with gameplay, and so and so fourth

[full robin]

That's why I'm interested in the prospect that AI might generate such textures some day.

Before AI, people used things like 3D scans which only covers a diffusion map that usually has shades and highlights embedded in them.

[earnest vapor]

maybe in 10 years

[full robin]

I had briefly tested a mobile video -> animation pipeline and it worked decently.

And I ran it locally on my 4070Ti desktop.

So, I can basically "act" out any motion I need in front of my tablet camera, which I can easily convert into animations that I can import into Blender for cleaning up.

The only complaint I had was I wish there was a way I could turn them into an IK animations.

But I guess it's a limitation in mocap in general, but game studios use mocap all the time. So I guess it's not really needed to make good game animations.

[earnest vapor]

oh sure you can do that, but there's a difference between "I can get an animation to work"

and

"my character behaves photorealistically"

as in-
there are minute little differences in his walking depending on if he wears heavy shoes or light ones.
or
when he gets into his car, all of his clothes actually rub on the fabric, and bunch up properly, as he searches for his keys in his pocket

[full robin]

I can imagine how things like subtle facial animations can make a night and day difference in that matter.

[earnest vapor]

it comes down to miniscule details that, if they're incorrect, will look creepy if the still image has been made technically photoreal by ai

making something photorealistic is just a ridiculous amount of effort, and always has and always will

[full robin]

I tend to agree... but as someone who's working on a VR NSFW game, I have to set my goal high in that regard - at least I hope I won't be too picky with details when I play a lewd game 😅

[earnest vapor]

you work on a nsfw game and you try to go photorealistic?

that seems very counter productive...

[full robin]

You can always dream. And as for productivity, I guess that's why we're talking about it in this channel. 😉

I feel we're slowly collecting all the necessary puzzle pieces to achieve that, with things like Godot's Human Shaders, Nvidia's fork of Godot with RTX & DLSS 5 support, and so on.

The one glaring missing piece I see is textures. To be specific, a way of easily generate a variant of them without having to painstakingly create them from photo scans by hands.

By the way, this was a test I did with a purchased model to see how much fidelity I could get out of it in a realtime environment.

I think that's an achievable goal if Godot, DLSS, AI, etc. keep moving on their present direction for a few more years.

On a side note, I also managed to hook that model into an Audio2Face setup, which generates a life-like facial expressions & lipsync animations procedurally.

I had to abandon it because Nvidia decided to close up their platform, and I hated Python too much. I used an open-source alternative called Neurosync for my game. But somehow the project maintainer suddenly disappeared, so I'll probably have to look for an alternative again.

[mint shoal]

example codex prompt after running a codex prompt and upgrading it with gpt pro 5.4

so how is everyone

[pale light]

are you working on anything?

[mint shoal]

a roguelike game with 2d, isometric and 3d game modes that are interchangable

[full robin]

I'm making finishing touches for the "spec-driven agentic workflow" that I developed using OpenCode:

I successfully tested it to create a frontend and a backend project. I'm planning to reimplement the legacy modules belonging to the project one by one using this setup.

[tiny hawk]

How much RAM you need to run a model like that locally?

[full robin]

I'm just using Codex/OpenRouter/Z.AI etc. now, although test to see the feasibility of running it locally is planned.

[tiny hawk]

Hmm so how reliable it is? Does it actually produce usable code or need fixing?

[full robin]

Depends. What I've seen so far, it exceeded my expectation, but it's partly because it's optimised for certain patterns specific to my project.

The generated code is pretty reliable since I tried my best to add a lot of guardrails. The backend is a pretty simple CRUD server, but it has 198 unit tests, and AI automatically runs hundreds of static analysis checks.

After I demoed the project to other team members, one of them said the open source version of jOOq doesn't support Oracle. So I just told it to "replace the persistent layer with a MyBatis-based implementation" and it did it without any error while I had a cup of coffee.

Later I found a few code design issues which I addressed with additional prompts, but it got the functionality perfectly fine at a first try.

[pale light]

Codex 5.4 is great

This is running in Chrome

custom engine from scratch using Codex

uses planck physics, polyanya for pathing, PIXI for draw calls
local detour, game loop, controls, and the ORCA/RVO2 written by Codex

[lavish ruin]

oh nice

[mint shoal]

if you use codex long enough you get

one roguelike 2d-isometric and 3d modes which can be switched around easy

i use gpt 5.4 pro to enhance codex suggesitions, and then feed it back into codex 5.4 mini ( and it can spit out a full 7000 LOC new feature across 40 files in about an hour) and one shot it

[wicked igloo]

who made the art tho?

[earnest flint]

has someone documented their gamedev journey relying on ai ?

i saw bunch of posts recently where people made games as hobby for themselves or their children and they became popular

i haven't done any game dev in a while. But since the last time, there has been a huge shift in terms of AI generated code and all these agents

I wanted to know, to what extent are we relying on AI ? coz i have my reservations against completely handing over control to it

[full robin]

I'm seriously considering to start my project over to try an AI-driven workflow, after having a success establishing it at work. I have been very sceptical about using it for gamedev, and I still feel that way. But having so little free time and experiencing what AI agents can do, I feel compelled to give it a try at least.

I expect a lot of roadblocks along the way, and I'll probably spend more time building the harness than making content at first. But if I commit to this path, I will share the workflow as an open source project.

[pale light]

what do you mean harness

I don't know what that means

[pale light]

this is like dozens of folders and a hundred files and 10's of thousands of lines of code and I did it with codex

Codex 5.4 IDE extension is really good

[full robin]

If you want to go beyond one-shot generating simple projects, you'll need more than just prompting.

And games have additional difficulties, especially if you want to build non-trivial 3D games using AI.

[pale light]

the video I showed is a from scratch engine I made using ai and I wrote almost no code

[full robin]

I can see that. But it's because the game is simple 😉

[pale light]

this is the level editor

this is the delauney triangle navmesh decomposed into convex polygons using polyanya pathing

[full robin]

Well, try telling it to build a VR sandbox game with MakeHuman characters with proper IK set up, for starters 😛

[pale light]

these are the ORCA/RVO2 algorithm reciprocal avoidance overlays to prevent ship to ship collisions
|

it has recoil, reticle bloom, and inaccuracy

it bakes different sized nav meshes per ship size

[full robin]

Let me put you into a perspective. I tested my workflow to build two "simple" non-game projects last week at work.

They involve some 20K LOC in 200 files with 300+ unit tests.

And my plan is to apply that standard to rewrite other 4-50 subprojects.

[pale light]

when you destroy ships depending on where the damage came from they undergo voronoi decomposition into physics controlled pieces

[full robin]

It's just that we have different ideas of what is "simple" or "complex".

[pale light]

lol ok buddy

good night

[full robin]

And even for actually simple projects, it's always good to setup some workflow which can orchestrate agents and let them incrementally discover contexts as needed.

Otherwise, you'll see them perform increasingly poorly as your project gets larger, and your token consumption will skyrocket.

Also, even the best models can and will make mistakes. You need a way to automatically validate their output in one way or another, which can be challenging in game projects.

[pale light]

My advice to you if you want to elevate from toy projects is to focus your effort on finishing one project instead of spreading thin on 50 subprojects. That is a common mistake newbies make 😉

[full robin]

That's sound advice. The only problem is that they are not toy projects, but a product we sold to many customers including banks and governmental departments. And I am not a newbie but someone who's been programming, likely long before you were even born. 😉

[pale light]

I noticed several times it took over a minute for you to only type 1-2 sentences. If your words per minute is only like 20, that can be a huge bottleneck to succeeding in your journey learning how to be a programmer.

My advice to you is you should should use a typing practice site, for example https://www.typingclub.com/ 😉

you've been typing for like 3 minutes

definitely use the site

your WPM must be like 10

I'm not going to read anything you write now so don't bother

I'm leaving the channel, have a good one

[full robin]

I'm sorry I have hurt your ego so bad that you had to resort to such a childish attempt to insult me (which wasn't my intention).

Just 2 things: 1) English isn't my native language, and 2) it has nothing to do with AI-dev.

If you want to start a drama, at least try to do it without going offtopic.

[earnest vapor]

@pale light

how is the ai better with bigger projects tho

are you actually just vibe coding stuff, or coding yourself and the ai is just kinda helping along?

[queen spade]

Took yer jobs

[pale light]

The different pieces of a project represent solutions to all of the problems the project is meant to solve

So if I have a command "send a unit to point B"
then that causes the strategic planner to do A*, then produce waypoints, then another system creates manuevers based on those waypoints

The folder structure, imports, comments, and existing code examples in context are a solid description of 'how this repo solves problem xyz'

The next time you ask the AI to make a change, the pattern of what is already there makes it far more deterministic in how it will solve the problem

If I ask the AI to
"make me a new ship that looks like the gunship, it has 4 turrets, they fire missiles that travel for 0.5 seconds then activate thrust, they seek their own targets, they explode when within 200px of their target, and when they detonate they create a flak explosion with shrapnel"

My game has:
-Hardpoints
-Gimbals
-Weapons
-Projectiles
-HomingCalculators
-HomingTargets
-ThrusterDelays
-InitialVelocityOnLaunch
-LaunchGates
-ProjectileFuses
-OnHitEffects
-OnDeathEffects

So creating that new ship is a matter of instantiating the set of components together in the right combination

More content is just authoring within the existing systems instead of having to write new systems

My experience has been that once I get the systems right, the combinatorics of those systems becomes easier and faster

the AI will do a decent job making that new ship and new behavior, then I open up my ship editor and see how the specific positions and angles of those weapons could not possibly have been placed well by the AI (the AI can't really see or play my game, but it can make up some initial numbers)

so then I go into the ship editor and just fine tune the angles, positions, ranges, cooldowns etc.

if you make sure each piece has a small number of responsibilities, and you ask for a change, the AI only needs to look at a small number of pieces to make changes so the context doesn't really creep upward that much

I had 3 agents working on the same repo at the same time, one working on the smoke system, one working on a custom ORCA pathing for large ships with compound discs, and another doing Tauri level authoring at the same time

the 'contracts' and tests for each piece are like promises that each piece will take certain things, and provide certain things, so each piece of the project can black box its internals and be trusted to adhere to the contract

so multiple pieces can change simultaneously, and you can refine the internals of each piece, given that the 'contract' is not broken

the shape of what is passed in and the shape of what is returned must stay the same

the quality of that result can improve as long as the 'shape' (like the waypoints are a series of x/y coordinates)

[earnest vapor]

so

[pale light]

even the context does not actually grow necessarily when the project is huge

[earnest vapor]

is it 100% vibe coded or do you actually know what's in your code?

[pale light]

everything is abstractions, so if you have methods that solve the problem, the AI only needs to chain your existing methods together instead of write the internals

I can actually see that my AI doesn't need more context

[earnest vapor]

no i get how programming and how ais work, and i get how agents work too, but it's just in my experience - if something fucks up, or is not buggy, but "human buggy" - as in "the ships rotation doesn't FEEL right" - the ai can't really help you

[pale light]

(it says how much context it is using)

[earnest vapor]

sure the ai can implement A*

no shit

[pale light]

I know you know how software works I'm just explaining why it seems to me that AI actually seems to get a little better once the project is bigger

[earnest vapor]

but if fighting your npcs isn't fun because of the way the gun mechanics play with the movement-
how are you gonna fix that using an ai?

[pale light]

I was surprised how it started fitting together and being more consistent and faster when the code was larger

[earnest vapor]

the ai can't actually play your game and experience it...

[pale light]

I spent like an hour carefully tuning the algorithm for how the spread growth worked

I'm a QA person/playtester

I definitely have to tune everything myself

in the gif you can see my crosshari growing and the shots getting less accurate

the AI's algorithm for that was terrible

I wrote my own

[earnest vapor]

okay, from the lil video you posted, you're making a space game, im a lil triangle spaceship, and i shoot other spaceships.

cool.
Now let's say you want to make an enemy that's a suicide bomber.

[pale light]

my bang bang face mouse algorithm was hand written too

[earnest vapor]

and the AI didn't bother making it all object oriented / component based, to account for the fact you may wanna make different enemy types

is it gonna refactor everything, or just tack on another hardcoded enemy

suicide bomber enemy needs a completely different ai, but still use the same A*

but now the goal changes every frame

which means prediction, recalculation, maybe even chunking, etc.

and i mean sure, you can tell that to the AI, and it might work-

but how do you know

it jsut feels like you're pulling a lever on a slotmachine, hoping it spits out a winner

[pale light]

the ManueverSystem requires intentions and locations
it takes those intentions and locations and then outputs specific desired manuevers like
"face left and move upwards"

the FlightController takes the ManueverSystem's output intentions and does math to figure out with the given ship what the right actuation of thrusters are to cause that manuever to happen

If I wanted to make a suicide bomber I'd create a system that produces intentions and locations and set those on the ship, which would get picked up by the maneuver system and the rest would be history

There is a system for producing explosions, shrapnel, flashes

I might give the ship a weapon that fires a projectile using the existing system that has
lifetime: 0
onDeath: explosion, friendly fire on

The suicide bomber AI system that gives its manuevers would tell it to fire its weapon when it is within X range of an enemy

There would be an event sent to the event bus that an explosion happebed
the sound system listens for the event and plays a sound
the flash system listens for the event and shows the flash

but I coudl talk to the AI for 30 minutes with these suggestions and it would come up with a better design than I could probably

[earnest vapor]

yes i know how one would make a suicide bomber

the problem is, if you've told the ai to setup your game NOT factoring in the possibility of evental suicide bomber enemies-

[pale light]

no I'm saying thats what the AI would code in my game

[earnest vapor]

it'd have to refactor half your ai & enemy code

[pale light]

and those systems are in the game already

[earnest vapor]

i think you don't get what i'm saying

but i also guess that's kinda the point. People who let ai do all the things for them, don't really need to actually know what's in their codebase

and if it works - cool.
but if it doesn't - time for a new project i guess...

[pale light]

are you saying you think that my
Ship - Hardpoint - Gimbal - Weapon - Projectile

Component system with
Auto Aim Controllers
Direct Player Control
Firing Gates
Cooldowns
Fire Rates

You think its not modular and reusable?

will you watch the video

[earnest vapor]

no, i think that your A* functions, your pathfinding components and your general enemy ai classes are probably not architecturally sound, and you may be taking on alot of technical debt

i don't care about a random float that defines your cooldown on your weapons

i care about the moment where you say
"alright, i wanna have 400 spaceships, fight 400 spaceships from team 2, without bunching up in the middle, and they all need to have dogfighting ai, while choosing their opponents intelligently"

THAT'S when ai is probably gonna shit the bed

because that's when it goes beyond boilerplate

[pale light]

I guess we'll find out in a few weeks

[earnest vapor]

good luck

[pale light]

thats running on my laptop

(this is from 9 years ago)
If I run into issues in any particular area and the AI fails I'll just fix the code

I studied physics in college

[full robin]

I'm glad that I decided to start my project over to switch to an agentic/spec-driven workflow, although it meant abandoning everything I've done since last summer.

Before the change, I could only work on my game project on weekends. But now, I'm running three agents concurrently on a Monday, two for my game project, and one for my work. 🙂

[earnest vapor]

what do you guys use for agent vibe coding btw?

[full robin]

I'm using OpenCode. And you can see my current setup from here: https://github.com/mysticfall/alleycat

Note that it's very incomplete now, since the project is literally two days old. I'm trying to establish a similar workflow I made for my work projects.

[hasty horizon]

me too

okay so, judgement free space lol

I'm trying to make a game with 100% LLM code just to see how it does

so far I'm extremely impressed.. progress is going faster than I'd expect, 80% of my time is just writing planning documents.. the code is mostly good and works, but requires help/bending into the right direction often

[pale light]

do you have anything to show?

I've been working on a game using AI as well

[hasty horizon]

I don't want to share it yet

[full robin]

I think I have a working environment for making 2D UIs. I just started working on creating a 3D workflow:

Again, I'm happy that now I can work on my game project on weekdays. Today, I'm running 2 agents for work and one for gamedev.

[full robin]

It's going good so far. Made a test scenario in which the agent loads up a character and visually verify the scene from multiple angles:

[full robin]

Started working on an actual feature (an IK setup for neck-spine bones):

The agent failed to use a proper framing for screenshots, however, which I need to fix.

[full robin]

Fixed 🙂

[pale light]

rts controls
ORCA vs terrain
rudimentary first AI pass
moving stations where ship AI understands how to avoid it

[pale light]

normal map shader support

[full robin]

I refined the visual workflow. And these are what the agent created with the tools I provided it to implement a neck-spine IK node:

[compact wedge]

man, I love normal mapping

never done it myself but I love looking at it

[hidden rune]

if you have a game you like and a youtube channel with gameplay videos, here's my bot https://github.com/eternalyze1/youtube_imitation_learning_game_bot

[solemn walrus]

Hey 👋

[full robin]

<@&133522354419662848> Could you clean up the spam? Thanks!

[earnest vapor]

cleaned

[full robin]

Thanks! 🙂

[wind quartz]

mysticfall thanked hallojoby

[eager grove]

Hey all, anybody following the DLSS 5 situation? When I heard how hard people were coming down on the tech I felt like it was a ridiculous idea that it could only make things photorealistic. I did some experiments with Comfy Cloud and ControlNet to look at what such a technology might be able to do if directed differently. I wrote an article, but just scroll through if you only want to see the images!
https://aitalesfromthefield.substack.com/p/nvidias-dlss-5-controversy

[pale light]

I wrote my thoughts, then read your article, and I think we're basically saying the exact same thing.

Jensen has admitted it is basically just image to image ai image gen (it takes in input image, prompt, and outputs an ai generated image). There is no reason why that should be limited to forcing faces toward photoreal, but it does come with the 'training data common denominator' problem where its going to have a mind of its own that isn't the artist's because of the training.

Prompt + Image = Output

I can take a person and say "draw them in ghibli style" and then the output can look like ghibli style but with the AI generated feel. Do that quickly and with stability over time, and you have DLSS5

I think its exactly as good or bad as ai image + prompt to image AI generation is, which depends on the user, but it more often than not gravitates toward bad unless the person using it has style and sense.

If you take low poly game with good gameplay and say "make it look like photorealistic clay stop motion" and it does, that is kinda cool and can be creative

If you take a decent looking game and ask it to "make graphics better" and it looks like ai image to image pushing training data's real human faces onto games it looks sloppy and lazy and terrible

If a game is made from scratch with using DLSS5 in mind, you could think of trying to make the visuals as creating the right prompting information to intentionally let DLSS5 'finish' the look for you and then have DLSS5 do something big and creative with it, like "make the whole world look like photorealistic realistic yarn"

[full robin]

Personally, I've long believed that 3D will survive in the age of AI, but its role would shift from rendering to providing a "source-of-truth" about geometries and style guidelines.

[eager grove]

I think we are pretty aligned. I also like your examples. Regarding style text prompting, I did give that a try, but I found that the other factors (img2img, depth map, edge maps) constrain the output too tightly for the prompt to have much effect. That's why I went with LoRAs, as they have a deeper effect. Also note that the graphics systems provides lots of buffers and reprojection, so it can be both more stable and semantically correct than a pure img2img diffusion process.

[eager grove]

Yes, I definitely think that's a possibility. My question is how much can generated content be anchored to the world? Will clothes switch from having zippers to buttons just because you turn around? If that's an issue, can we either make the generation more predictable, or record it to keep it consistent temporally?

[full robin]

That's why we need 3D model as a "source of truth". I think we'll still create zippers on clothe models in that case, but instead of spending hours on making zippers as a 3D mesh, simply drawing a rough shape of zippers in the texture will suffice to prevent such issues from happening.

[eager grove]

Let me rephrase. Unless someone finds a different approach there will always be a gap between the literal geometry and the AI completed frames. This could lead to inconsistency. Perhaps the new discipline will be working out the minimum polygon/texture budget required for various things such that the inevitable differences are indistinguishable for most players.

[cobalt fjord]

anyone know about advanced AI for survival horror games much like that of Nun Massacre (unity, puppet combo game) or Aka Manto (unity also, chillas art game)?

i mean like AI that will hunt you down until you quit the game and probably hunt you down in the real world too

every little audio cue matters, the flick of the lighter, your feet padding on the creaky wooden floor, the door's frail wood structure screeching as it opens, the list goes on and on

ambushes, playing with it's food like a cat hunting it's prey and waiting for just the exact moment to strike and end your run then and there

and if it sees you, know that you are already dead. you never stood a chance. by the time that you realize your stealth has been compromised, you're already being torn to shreds by a knife forged long before your time.

Hiding is a temporary fix to an inevitable problem. Eventually, no matter how long you stay in a locker or under a bed, it will find you.

One way or another, it will weaponize your paranoia and fear against you in such a way that you spend nights trying to just comprehend.

so like anyone know how I can begin doing something like this

very passionate about this project i've been cooking up

had the idea on the back burner for a while but i figured it was time to focus on it and begin writing a story and lore

[eager grove]

I'm not an expert per se. You could certainly have something like behavior trees for the enemy's internal state. Objects that emit signals like sounds and enemy sensors to detect those signals. Most games cheat a bit. I believe Alien Isolation semi scripts when the alien turns up and doesn't bother with pathfinding when the alien's in the vents. Maybe people have been using generative ai too? If so I'm out of that loop.

[cobalt fjord]

idk

not much of an AI guru either

I think the most AI's done for me is motivate me to keep playing Aka Manto

[eager grove]

Depending on your starting point I'd suggest checking out the AI and Games channel, it discusses how the AI in a lot of popular games works. https://m.youtube.com/channel/UCov_51F0betb6hJ6Gumxg3Q

[onyx ermine]

Hello
anyone can help me understand the reason why we use Bias in Single Layer Perceptron (Neural Networks)?

I understand its to offset from the origin

But why are we moving it away from origin?

[onyx ermine]

I think I got it, as I understand its to move it closer to the threshold, or else all values will stay around origin and we can't a simulate real world scenario

[earnest vapor]

it's very early days on that stuff

but i can't see actual photorealism work without useage of ai

there's just too much detail that needs to be handcrafted, you gotta put some sort of ai in there somewhere if you wanna bridge the uncanny valley

[deft mesa]

https://streamable.com/kkxgnx

[cold merlin]

"All of those embodied agents are seat opportunities," Jha said, envisioning organizations with more agents than humans — each effectively a user that must pay for a software license, or "seat" in industry lingo. https://www.businessinsider.com/microsoft-executive-suggests-ai-agents-buy-software-licenses-seats-2026-4?utm_source=reddit.com

[urban parrot]

what's up

[deft mesa]

ai generated web page, with ai playing soccer lol made ai with ai

[full robin]

Stumbled upon a video that provides a good overview on where AI-3D workflow is at currenty: https://www.youtube.com/watch?v=PyxnLyRfKFo

[minor warren]

As i didn't saw any assets of Stalin so i created wolfenstein 3d esque screen for ussr campaign

[full robin]

https://x.com/TencentAI_News/status/2044604715376546006?s=20

I mentioned of NVidia's Lyra today on another channel, which looks nice but is proprietary. And now we have an open source alternative that we can freely use.

From the video, it looks like there's still a gap in quality between the two. But I'm glad that we at least have an open-source option for a world model now.

[pale light]

I vibe coded this procedural blended leg animation system by describing a bunch of algorithms to the AI

[agile depot]

this works.

[pale light]

I'm not sure I know of any large bodies of work that I could even in theory have AI just go on a big loop and do autonomously for long durations right now.

I just talk to the AI like its a person and have conversations about how to get what I want accomplished, form requirements, get detailed, form an implementation plan, then have it follow it to completion.

[full robin]

I just briefly tested NVidia's AI animation generation model, Kimodo, and it was not really usable for me. The tech looks promising but it may need more time before becoming production ready.

[silent vale]

Hmm this is nice only time i give ai the credit for that

[agile depot]

<@&133522354419662848> <@&1224208037892591666> spammer multiple channls

[full robin]

I think that's a sensible way of developing any complex project, not just games, especially when using AI agents.

And that largely coincides with my own method, building a project using components, each having its own specification, tests, integration tests, and a test scene.

Probably the only part I may have a different opinion could be that about distributing LOC evenly across the project.

[agile depot]

well it was more about how many LOC written per turn. each turn it was giveng commads to split those lines into seperate files.

basically maximize your output so each run is maximum context length

[full robin]

I think imposing a rule on LOC might have an unexpected impact. And it's better to manage context using things like a compress/compact plugin anyway.

[pale light]

I think 90% of the context being used can be the thought the ai is having about your task, I've been able to fill 256k context with a question about 100 LOC because the amount of research and thought required to understand and solve it was big

You could have a file 50x longer that uses less context

[pale light]

ai produced procedural normal mapped 2d sprite atlases workshop

so I can make rotating 2d stations/structures with doodads and details that get lit as they rotate just from code

and the smoke has a fluid simulation now

[abstract moat]

are there any ai game engines yet or?

[static phoenix]

Probably not soon

[abstract moat]

ah

[static phoenix]

Seen 1, but it's very basic

[abstract moat]

ooh?

i mean ue and unity have ai right?

[tiny hawk]

@ashen plaza is making one

[ashen plaza]

Panmathos is my project. Simply put its a middle man that learns and enforces governance and determinism, able to build and expand its own tools.

[cold merlin]

https://www.youtube.com/watch?v=sDMQlC3QIds

[dim summit]

used ai quite a bit for a separate tool to bake lighting information onto sprites for my project. The results are not too bad although there is only self shadowing and fairly low detail shadows. but happy w/ the results regardless.

[pale light]

nice! I'm baking normal maps procedurally for my game too

trying to see if I can make good low latency multiplayer with prediction using webtransport protocol that became standard in March 2026, its brand new and seems really good

[tiny hawk]

https://copy.fail/

Co founded by AI btw

[dim summit]

the lighting looks great.

[normal tundra]

First 😛

[zealous flame]

so what about that ai

if (seeEnemy) shoot() amirite

[elfin vortex]

Hell yeah

No I want Wu Dao 2.0 API

someone hack me a copy I’ll pay $999

I intended for mostly ML topics applied to game dev, I think there’s a ton of untapped potentials

From every aspect, graphics, AI, neutral language, image synthesis

[fathom temple]

writeAICodeForMe();

[elfin vortex]

Get copilot

[river saffron]

I'm gonna need this channel really soon

Might need to either implement custom a* or just hack my way through with waypoints

[glacial eagle]

tbh I don't know how copilot is any useful

you have to describe ur problem in detail and you have to check the answer given by the AI

[elfin vortex]

no

that's definitely not the only use case

it's very powerful at pattern completion

[glacial eagle]

I could interperet that as multiple things

do you mean when you intend to create a for loop it will automatically suggest one for you?

or like a getter

[elfin vortex]

For example I typed the up case

it's really good at always knowing the "rest" of what you're trying to do

[glacial eagle]

what am I supposed to look at

the cases to the right?

[elfin vortex]

left side is code right side suggestions

I type the UP case

[glacial eagle]

wait what really

[elfin vortex]

it suggests the stuff to the right

let me find another

[glacial eagle]

if you make ur "UP" case it can automatically detect all other directions?

that's pretty cool

[elfin vortex]

yeah

and a lot of the times, when I do cout

it will fill it with every I need let me try to find

like this, but sometimes on much bigger scale depending on context

[glacial eagle]

it suggested prints?

I dunno, I always have in mind exactly what I want to type out and if shit appears or behaves not exactly how I type it confuses me

[elfin vortex]

correct answer is

  glVertexAttribPointer(1, 2, GL_FLOAT, GL_FALSE, vertexBytes, (const void *)12);
  glVertexAttribPointer(2, 3, GL_FLOAT, GL_FALSE, vertexBytes, (const void *)20);
  glVertexAttribPointer(3, 3, GL_FLOAT, GL_FALSE, vertexBytes, (const void *)32);
  glVertexAttribPointer(4, 3, GL_FLOAT, GL_FALSE, vertexBytes, (const void *)44);
  glVertexAttribPointer(5, 1, GL_FLOAT, GL_FALSE, vertexBytes, (const void *)56);
  glVertexAttribPointer(6, 1, GL_FLOAT, GL_FALSE, vertexBytes, (const void *)60);
  glVertexAttribPointer(7, 1, GL_FLOAT, GL_FALSE, vertexBytes, (const void *)64);
  glVertexAttribPointer(8, 2, GL_FLOAT, GL_FALSE, vertexBytes, (const void *)68);

looks like it got it right

somewhere else in the file I have the vertex layout

and it got it

[upper dirge]

@elfin vortex this VS* or VSC?

[elfin vortex]

tf is Vzs

VSCode yes

[upper dirge]

oh

Your theme looks like VS

@elfin vortex how often does copilot annoy you?

[elfin vortex]

like 10% maybe, since that and intellisense are both tab completion

but definitely more help overall than hinderance

[upper dirge]

Ooo

[glacial storm]

lol

[elfin vortex]

Still looking for wu dao 2.0

Does anyone have any cool copilot moments

[placid garden]

"--- BEGIN PRIVATE KEY ---"

[elfin vortex]

Lmfao

[elfin vortex]

Lol

Vscode, the next generation calculator

[radiant island]

i mean it’s not far off

[patent blade]

you gotta carry the // themperor

[placid garden]

Ah thats my mistake

[elfin vortex]

It’s been more than a week since I joined waitlist

Shame on you OpenAI

Don’t even deserve that name

Should be illegal

[zealous flame]

you need to be on a waitlist for something called OpenX?

wild

[elfin vortex]

EXACTLY SOMEONE SUE THEM

[patent blade]

Open™️AI

[true frost]

I have to figure out some enemy AI for an action platformer, something more substantial/dynamic than "approach player and attack on a delay"

Anyone have any good reading for this sort of thing?

Or general advice?

Espcially for bosses, how to design satisfying attack patterns

[patent blade]

this is something you're gonna have to spend some time on

i had the same problem

(different type of game)

i recommend breaking down things your enemies can do into 'atoms' and then it may be easier to decide how to combine them

e.g. enemies can shoot, walk, charge towards, flee, shoot at, jump, crouch, wait, dodge, shield, maybe do some action with other allies, etc

if you generify it like this, you can also use this to build friendly AI too

@true frost ^

[dusk plinth]

This is exactly the space I'm in right now, and while I don't have good reading (though I would be happy to chitchat about it). My thinking is you really have to have some feature of the game that makes it so that enemies have something to do that isn't just mob the player and attack.

There are lots of games where they have alarms and what not for the AI to trip or whatever, but if the AI has some set of conditions under which they logically want to run away, life gets significantly more interesting.

[dusk plinth]

Maybe rename it dev-ML then?

I've been puzzling over good uses of ML in gamedev. The first real potench I think would be for something like automatically generating different 'Pokemon-like' creatures.

Imagine a Gen III spoof, but you use CLIP or some such to make every pokemon a unique snowflake.

That is ML based culling and refining of procedurally generated content so that that creature you have is truly unique. It's yours, you own it.

[wispy silo]

I have a weird pathfinding problem. When my enemy approach my player at a certain distance they stop chasing him.Basically they dont follow the path and they stand still.After my player run away from them they chase again. I have no idea how to fix it.

[elfin vortex]

[gleaming coral]

soooo just gonna get started with ai in unreal tomorrow, any good tuts?

pls @ me

[glacial eagle]

@gleaming coral I know just how to solve your problem. You want a tutorial? These people will give you a tutorial #unreal-tutorials

[elfin vortex]

Got lucky at work, given some free time to do R&D for machine learning applied to our use cases heh

was always trying to break into that tech stack but don't have personal time

right now looking at some resnet models

[elfin vortex]

looking into YOLOR right now. The github implementation https://github.com/WongKinYiu/yolor seems to suggests training on a large dataset (coco2017) myself instead of using a pretrained model of this implementation, is that typical?

[elfin vortex]

nvm I'm dumb and didn't read everything

[craggy pawn]

Kk

[tight storm]

@kind night

Ты обиделся?

Ну пиздец

[dusk plinth]

What are you planning on using this for?

[elfin vortex]

NDA :p

[dusk plinth]

https://tenor.com/view/frodo-keep-your-secrets-alright-then-lotr-gif-12907121

Anyway.

The biggest uses of ML in game (indie) development I would think would be either for better balancing or generation of levels, or image generation for content.

For AAA development I would imagine it'll completely overtake existing methods for post-processing.

What do you think?

[elfin vortex]

yeah what I'm doing rn is not related to game dev, but my next personal project is

and yeah, all those things are potential applications I want to try

in addition to quest/dialogue generation

and even ai/animation

[dusk plinth]

I sat down for a bit and tried to get torch to work with Godot.

[elfin vortex]

lol

[dusk plinth]

Since Godot is C++ backend and Torch has a C++ implimentation

Mostly got stuck on the Godot end, actually

So I thought I might wait for Godot 4 and then take another swing at it.

[elfin vortex]

I'm debating whether to keep all this as api or packaged with game

I wonder how big gpt 3 is memory wise

[dusk plinth]

Not hart to estimate

*hard

it's between 1 and 10 B weights

each weight is I think a 16 or 32 bit integer

• excuse me,

I'm off by an order of magnitude

[elfin vortex]

here's your 400GB game, enjoy

[dusk plinth]

GPT3 is 175 B

[elfin vortex]

oh boy

[dusk plinth]

So try 4 TB

[elfin vortex]

here you go, 4TB update just dropped

so yeah, probably kept as api

[dusk plinth]

Clearly the implementation should be handled before release

There is a subset of ML attempting to optimize networks that operate on a single GPU

[dusk plinth]

This is the reason why all ML engineer jobs nowadays are asking for AWS or other cloud computing experience.

[ruby basin]

WOW

[distant onyx]

I want to develop AI and robots which will be best programming language?

[upper dirge]

@next ice

[elfin vortex]

Most AI stuff is python

[elfin vortex]

Geez

Was training my YOLOR model today

Ran out of gpu memory on my 3080 using default configurations

[distant onyx]

Well someone said C programming covers most of how computer works is it rughtz

?

[upper dirge]

Nope

[distant onyx]

Which programming language covers the most of computer's basic operations

?

[glacial eagle]

uh?

any of them?...

[river saffron]

assembly

gives you the most access

[glacial eagle]

it gives you direct access, but you can compile abstract code to assembly

which effectively provides the same access

[distant onyx]

tochka vs quantumedbox

[glacial eagle]

shit vs shit*

[distant onyx]

C vs python which i should choose?

[glacial eagle]

depends on your goals

[river saffron]

if you want to do low level, which is what it sounds like, then C

c is the lowest level language outside of assembly, python is (one of) the highest

[glacial eagle]

or pick something else that provides the same system level access :p

[river saffron]

i wonder if theres an actual c alternative out there

[distant onyx]

I want to learn basics of computer, little bit of games like modding minecraft or smth , i want to do some AI stuff later

[river saffron]

speed wise

[river saffron]

basics of computing = c without a doubt

[glacial eagle]

of course there are

[river saffron]

no popular ones

not even rust comes to c level

[glacial eagle]

rust often outperforms C/C++ because of its memory model

[distant onyx]

Hows TS

[glacial eagle]

you can prove by it some low level optimizations

[river saffron]

if you want to learn the basics of computing, then bad

if you want to learn programming, then fine

[distant onyx]

Want to do ai stuff for that i need python and to get basics of computing down i need C right?

[glacial eagle]

https://benchmarksgame-team.pages.debian.net/benchmarksgame/fastest/rust.html

pretty close

[distant onyx]

Which topics of maths are important for AI and machine learning

[river saffron]

sorted by how "close to the metal" they are: Assembly > C/Rust/CPP > Java/C#/R > Python/JS/Ruby

technically you can do ai with any language

python is just the popular one

[distant onyx]

So now i should go with C in order to get the basics down?

Cuz i think a strong foundation is important

[glacial eagle]

python is built on top of C infrastructure, i learned them both side by side and it was quite good experience

[river saffron]

i started with python, went down to c#, went down to cpp, currently thinking about going down to c/asm

no, im not a masochist

[glacial eagle]

yes you are :p

[distant onyx]

If i start wirh python and then go to C, will it be fine?

[glacial eagle]

yeah

[distant onyx]

So i am gonna start witb python again

I started programming then started C without hsving a basic understanding of python and now i dont know what i know

What im gonna do , plz correct if i say anything wrong.
Python , then C

[glacial eagle]

nothing wrong with any of ways

but imo this particular one makes sense

[distant onyx]

Could u elaborate?

[glacial eagle]

in my eyes in gives you understanding on how to design stuff via python and how systems are implemented via c
just by learning pure python, for example, you can have no idea about cost of certain operations

[distant onyx]

So i shoukd learn C

[glacial eagle]

it's handy, but isn't necessary

decide by yourself

[distant onyx]

I choosee u C !!!

[crisp blaze]

Ai and robotics are different fields that go well together.

There are a lot of things you need to learn before you start choosing a language such as Math, Probability and Statistics, Calculus, Geometry and a lot more.

Now of course you can practice the above with some coding. I would recommend Python cause it is a good choice for Ai (and Robotics tbf)

But there are a lot more options such as C++, R, matlbab and more.

Skipping assembly would be really bad so at some point you will need to learn Assembly too, if you wanna learn Robotics.

I would recommend starting with C, it will be hard but then moving to Python is somewhat of a joke. It will also help you with assembly.

[distant onyx]

Hmmm.....thusm

[river saffron]

couldnt robotics just be done with c these days

[glacial eagle]

very few things are done with c nowdays :p

[river saffron]

even fewer are done with assembly

[crisp blaze]

Of course, you can choose a lot of languages.
And in fact C is a nice language for that

But you know whatever you are more comfortable with and doesn't limit you will work too

[glacial eagle]

C is really nice language imo, but it's just not that easy to develop software with it as you need to make almost everything yourself

[crisp blaze]

Yeah you won't need C to develop Software
You can just use Lua or other languages that will suite you better

[glacial eagle]

yep

[crisp blaze]

And no one said that using C means your program will be fast

You need to really know C to have such a gain

So for software I wouldn't choose C

But for backend I would, like a game engine, robot's logic (in our case) etc

[glacial eagle]

yeah, it makes sense in that regard

tho something like rust or cpp still are better for backend/engine usage imo

[crisp blaze]

Haven't used rust, but for C++ yeah that's true.
It can save you from a lot of things
And for game engines its somewhat the standard

I recommend creating a Physics Framework/Library with C
It is fun and you learn a ton

[glacial eagle]

i tried making C engine backend for Python once, but in the end i spent so much time making UTF8 lib/various container types/debugging it all

and with rust or cpp i would not even needed to care about it at all lol

[distant onyx]

U guys know any good book or couse to srart C programming?

[crisp blaze]

Of course, there are a lot of good books but you can't beat the one the only: The C Programming Language book

And it's actually free

http://math.ecnu.edu.cn/~jypan/Teaching/ParaComp/books/The C Programming Language 2nd.pdf

[patent blade]

learn c++ not c

[glacial eagle]

learn c not c++

:p

[patent blade]

false

[crisp blaze]

Learning C will not only help you with C++
But in general is a much simpler language to learn than C++

[zealous flame]

ehhhh, it depends how you define "simpler"

[patent blade]

c++ will teach you c

and c++

and c++ is far more widely used

[zealous flame]

no it wont? C programming is vastly different to c++

[glacial eagle]

and has infinitely more shit to learn

[patent blade]

i mean techniques are different

but there's a reason c++ exists

[glacial eagle]

to annoy :p

[patent blade]

it is very easy to make bugs in c

[crisp blaze]

Well, you don't get to use C++ while learning it for quite a while

You start with the basics and then you move on to smart pointers etc

So why not start with C?

[zealous flame]

if you want to ultimately code in c++, learn c++. if you want to code in C, learn C

[glacial eagle]

so is true with c++ tho, mostly

[zealous flame]

they are different languages with concepts that dont map toe ach other

[patent blade]

also don't do c with classes

[glacial eagle]

you can't stop me

[zealous flame]

#define class struct where is your god now

[crisp blaze]

Learning raw pointers, stack, heap, data types, operations, structs, data structures and many more
Are all basic things that you must learn

It's fun to use std::string sure, but if you don't first experiment with pointers to use as strings then you won't learn how the string behaves in the first place.

Will C++ learn you that? Yeah
Do you need to start learning C++ for that? No

Starting with C is basically learning C++ without the coolest features of C++

So I don't see a reason not to start with C

[glacial eagle]

private is no more

[patent blade]

its not just the coolest features of c++

[zealous flame]

because c++ has so many fundamental semantics that doesnt exist in c??

[crisp blaze]

Then moving on to C++ will be easier

[patent blade]

its a proper object lifetime model, type safe generics

type safety in general

[zealous flame]

you can also learn about pointers, stack ,heap, with ASM. so why learn C when you can learn ASM?

moving onto C will be easier because ...

[glacial eagle]

#based-chat

[crisp blaze]

Again, you won't use any of the C++ features until you know the basics

And if you don't know the basics why use C++ features in the first place.

[zealous flame]

so why learn C when you can learn ASM instead?

[patent blade]

i mean classes are pretty basic to c++

[zealous flame]

i'll cut to the chase: I am trying to show you why your argument to learning C before C++ is foolish: there are many semantics in C that make learning ASM almost useless as to how there are many semantics in c++ that make learning C almost useless

because I can map any argument you have into learning ASM before C, which is obviously a fools errand

[crisp blaze]

We can do this forever though, you can take it even more back than Assembly, that doesn't make any difference to the argument.

C++ wasn't made to make you not learn C. Nor to replace it. You can't do that. Why?
Cause C is much simpler than C++. C++ has so many things going for it that not only you need years to learn them but you need double the years to be even considered good in C++.

OOP is cool, smart pointers are useful, STD lib is powerful.
Classes are a basic thing to any language that supports OOP. That doesn't make it a must to learn right away. If you don't know how memory is allocated when declaring a simple int variable what's the point of learning OOP.

And what's the point of picking C or C++ if you don't wanna learn those things from the very beginning.

It's the same with JS and TS. Ts has really nice features for JS but you need to know JS to use TS efficiently

[zealous flame]

so why not learn ASM instead of C? give me a reason

[patent blade]

knowledge is power but you do not need to know js to use ts efficiently

[zealous flame]

ASM is simpler than C as well

[patent blade]

but you could make that argument for anything down the stack

https://www.cs.utah.edu/~elb/folklore/mel.html

[crisp blaze]

Are you actually kidding me?

C is known as an English version of Assembly (human readable form of Asm)

[zealous flame]

okay

and?

its simpler

[patent blade]

c is not an english version of assembly

[zealous flame]

C is a wrapper around ASM the same way C++ is a wrapper around C

[patent blade]

and most c programmers probably don't konw exactly how c maps to asm/machine code

[glacial eagle]

this thread is strange

[crisp blaze]

If you think that, what do you have to say about C++ users?

[patent blade]

the same thing

[crisp blaze]

I don't think that there is a point of continuing this.
Especially on this channel

[zealous flame]

okay

[distant onyx]

Thank u for free link , books are very costly sometimes

[wind quartz]

physics makes sense thanked Depression

[elfin vortex]

https://github.com/WongKinYiu/yolor I was able to extend this model and add another image class, and was able to train with my custom data, however when I run the inference script against the output weights, all the previous image classes don't show up anymore, anyone know?

[quartz field]

that was a great read lmao

thanks

[elfin vortex]

Progress after first 2 weeks of ML R&D for my work to detect empty shelves in supermarkets, trained on ~500 labels over 300 images for half an hour using YOLOR

[torn panther]

That's pretty damn solid. What GPU?

^ ping me too if anyone answers

Are you like, 100% new to it or?

https://www.cs.drexel.edu/~santi/teaching/2014/CS387-680/schedule.html @glacial eagle

Go through some of those slides

Game AI is more of an art than a science

[prime moss]

stupid question: is this channel also applies to player made ai ?
where you'll train ai to take control of a game

[patent blade]

sure why not

[elfin vortex]

yes

[glacial eagle]

how into AI in C++

[not unreal]

[patent blade]

What?

What kind of ai, what kind of game?

What platform?

[glacial eagle]

just an ai that would learn

yknow like those btree neurobrain ai's

platform? linux

[patent blade]

Find a library

[glacial eagle]

wow no shit

which lib u recommend

[patent blade]

I don't have any recommendations

What have you found?

[glacial eagle]

nothing

¯_(ツ)_/¯

[elfin vortex]

why c++

almost everything python

[grizzled marsh]

Hey, I have always been fascinated with AI, and have some interesting ideas, that if they work, could really be promising. Could someone DM me so I could ask some questions in order to see if my idea is do-able?

[patent blade]

ask them here

[spark vigil]

Fascinating

[snow cave]

i guess thats why he chose that name... he wants to keep it a se(r)cret xd

[hoary egret]

What are some of your guys' favorite AI patterns?

[patent blade]

what kind of game?

[hoary egret]

I guess I mean in general. I really like utility tables. I'm curious about stuff I probably haven't heard of

[patent blade]

never heard of that

also do you mean game AI or machine learning AI?

[hoary egret]

game ai 😛

i've thought about machine learning in games though

[patent blade]

i use a behavior/task based system in my game

[hoary egret]

i like the idea of generational ai

[patent blade]

AI's have a list of behaviors they can perform, which is selected is based on 'senses' (high/low health, enemies nearby, etc).

[hoary egret]

cull/reproduce some coefficients over the course of a game

[patent blade]

behaviors are composed of a list of tasks that run in sequence

tasks are individual actions (run here, pick a target, shoot at target, etc)

[hoary egret]

behavior tree?

[patent blade]

sort of but not really

[hoary egret]

kind of a behavior table it looks like

sounds actually similar to a utility table. utility tables you have rows of normalized considerations, agent picks the highest scoring behavior

[patent blade]

sounds like its probably that then

[hoary egret]

with a different scoring strategy

[patent blade]

i modelled it after things like half life and destiny

senses while quick to check, aren't really efficient to evaluate though so i need to come up with a better design

[last sequoia]

Those were made with AI, the AI makes a pretty decent concept artist, although it doesn't seem to handle creating character sheets or different angles of a model very well, it tends to create each view a little differently as it goes, it gets too creative and carried away 😆

if anyone is interested in things like that just say so and I can link some references of online tools and resources to use. I don't think it could replace a person yet, but it can definitely be used to assist or generate quick references to build from

[last sequoia]

Just look up algorithms related to what you're referring to, you can usually find links to source docs and project downloads
Gpt3 is open source
Using python a lot of ai is easy and quick to set up and understand, even if it's not as efficient at running, it'll get you going quicker
Lots of other algorithms are open source, remember, open source means you can find the entire algorithm for free and download it to try yourself
Lots of government websites have tons of databases, like the census has a massive labeled text database you can use to train an ai to identify written letters and numbers
There's also lots of free to use ai api's that people have made that you can use to interact with their machine learning algorithm they have set up, like deepsleep or nightcafe
Nightcafe uses a model called VQGAN+CLIP, its great at interpretation and abstract, and that's the type of model I used to generate those images, as well as my logo

[glacial eagle]

okey first i can't glue python all over the codebase i already have Lua glue all over the place and the code between c++ <-> lua is horrible

There's also lots of free to use ai api's that people have made that you can use to interact with their machine learning algorithm they have set up, like deepsleep or nightcafe

these is what am i interested on. which is best to work with?

[last sequoia]

I'd recommend trying nightcafe first, their api is very easy to use and you can use it a few times for free each day, but paying for tokens to use in it isn't expensive, basically they provide the brute force and hardware so you can use the ai even if your computer is a potato, and the tokens are just priced so that they can keep the lights on
About $15 will let you generate 100 images

[glacial eagle]

i can't pay for some random ai

isn't there something like c++ tensorflow

i just have a bunch of inputs (the game is server<->client) and i want to interwire it to some brain that can make reasonable outputs

[last sequoia]

You can download those models and train them yourself, like I said, the model they use in particular is called VQGAN+CLIP
that means it's 2 different ai algorithms wired together to accomplish a goal

[glacial eagle]

no but it has to be offline

cuz game is singleplayer (yeah server<-> client can be done locally for SP), so i have to like put a bot to replace human interaction

[last sequoia]

Downloading the training database would let you train offline, or you can train online
Once the ai is trained its just a program, it can be ran offline then

[glacial eagle]

???

[last sequoia]

Oh, a machine learning algorithm probably wouldn't be good to plug directly into a game for client side, they take up a lot of processing power. You'd need a very good rig to pull it off, high end computer at least
I doubt even the new consoles could run anything more than a basic machine learning algorithm while running a game

[glacial eagle]

there is no training database?? the inputs are basically parts of the game (i.e no images/letters, just raw data like economic charts and etc)

[last sequoia]

You could have a simple one that's not to complex which uses the player as its training set and evolves over time, that would be doable

[glacial eagle]

yeah i want that

[last sequoia]

Or uses changes in game state as the training data

[glacial eagle]

so how do i do that

do i just like make a random "if else if else" then put some neurons and some weights and change stuff?

[last sequoia]

Well you'd need to probably use pre defined arrays as the data structure, since that would be one that you can update the data in during runtime, and the O factor is an O 1 rate. However arrays inside arrays would work better and give you more flexibility. I'd probably recommend a 5 dimensional array. That's an array inside an array 5 times over. It could hold tremendous data while keeping it all very organized and quick to access for performance. Then you just need to identify the behaviors you need the ai to execute, find a way to streamline them as much as possible, and make it loop regularly to update the data it runs on

I would recommend coding in cpp(c++) for processing speed

[glacial eagle]

so basically a big state machinery

[last sequoia]

Make sure your regression models are fine tuned to the data you're analyzing as well

I would recommend a 3 dimensional regression for better flexibility and still somewhat cheap performance

[glacial eagle]

why do i want a 3 dimensional array

is there going to be 3*3*3 "neurons"?

[last sequoia]

So it can run the ai goal in parallel with analyzing the player behavior and generate a single path to follow both

[glacial eagle]

oh i gotcha, so like an ant-type algorithm

the problem is that there is no single "player" on the game (even in SP), it's either all AI's, all humans or mix of both (and nobody can predict what will happen in the next "move", since a mix of rng and determinism and more rng)

[last sequoia]

Yea, like a very simple ant algorithm

[last sequoia]

Then analyze the game state not the players

[glacial eagle]

yeah but then how do i do that

do i just make the big array with double everywhere and change super specific places and take paths according to those weights?

[last sequoia]

Think of it in terms of chess.
No matter how players are controlling the pieces, the board states are the same either way. The ai just needs to identify patterns in that board state to move towards a goal

[glacial eagle]

yeah but in chess you use the min max algorithm

[last sequoia]

That's also part of the reason for the 5 dimensional array
You could have one of the early arrays represent the players in the game and the sub arrays represent the variables they control

[last sequoia]

I was just meaning theoretically
The chess ai doesn't know how many people are controlling the board pieces it's against, as long as they follow the same rules, it doesn't care

In other words just design the data in a way where the players are a hidden layer not an output layer

[glacial eagle]

hmmmmmmmmm

[last sequoia]

Its helpful to remember the ai has no idea what a player is or a person is or any of that. They're just a bit of data the same as all the other data
All it needs to do is identify useful patterns that bring it closer to its goal

[glacial eagle]

yeah but then i just need to make a btree holding doubles and i give a random number to each neuron at each generation depending on how well they performed on their outputs

[last sequoia]

Just don't update too frequently, it'll get expensive quick

And step out the calculations
This will keep the performance cost more stable so you're not getting lag spikes

[glacial eagle]

yeah but the problem with this is that it just sounds like tomfoolery and not like a real AI

that is why im asking if there is anything like tensforflow for AIs or such

[last sequoia]

Most ai is basically tomfoolery 🤣

[glacial eagle]

:P

[last sequoia]

You won't be able to run a full ml program
There's not many ways I can think of to dynamically compile the code real time, you'll have to use some effect of Tom foolery

[glacial eagle]

guess i will just stay with random else-ifs

[last sequoia]

I don't think there's actually any way to dynamically compile the code during runtime
There's a few ways to fake it but they're expensive

[glacial eagle]

you can recompile code in runtime with ring 3 privileges

[last sequoia]

But you can do a ton with a multi dimensional array and regression matrix

Then it's just a matter of plugging all that data into your behavior tree

[glacial eagle]

k

[last sequoia]

I would probably recommend a puppet system as well
Have a bunch of AI controlled my a master puppeteer ai

Several AAA games use a form of puppet system if I'm not mistaken

[last sequoia]

Did any of that help?

[last sequoia]

Also, I thought ring priveledges were for antivirus and security checks and stuff. Would that even be usable for an ai during runtime?

[glacial eagle]

Rings are usable with the usage of a driver which is not isolated on userspace

if the OS runs a driver under userspace you're kinda fucked

otherwise you can hack-in and start skewing the page tables to mark certain regions as code (and more importantly, allow to modify code sections), this should allow for self modifying code to run on a kernel device driver

since Windows is a hybrid architectural kernal the drivers can be in userspace or in the kernel address space, the latter being preffered over the former due to the aforementioned reasons, kernel space allows for execution of code not present on the initial loading

i don't know how to get a devide driver to kernel space however, since the kernal is very strict on what can and cannot go (i believe that it has to be a certified driver before being allowed to run), this is to allow Windows to be more stable because a malicious kernal driver could just nuke the entire space out of orbit and nobody wants that

tl;dr too much shit for little gain, and won't work on linux and i use linux so fuck it

[last sequoia]

Interesting
I never thought about that approach

[glacial eagle]

it's perhaps the most "it just works" approach ever :P

[last sequoia]

That's game dev in a nutshell

I also prefer linux, having a virtual machine is a godsend lol

Especially when you accidentally write code that wants to devour all your data to the kernel level indiscriminately 🤣🤣 gotta love dev/null for getting rid of trash

[glacial eagle]

Hewwo

Anyone has tips for where to begin in studying AI? -

ʕ •ᴥ•ʔ

[patent blade]

what kind of AI?

[last sequoia]

Google free beginner machine learning projects

You'll get plenty of results
There's even many YouTube videos
Jabrils has good ones

And there's always crash course or freecodecamp
Those are decent resources, crash course is good entry level to intermediate info and freecodecamp is great for college level knowledge for free

[glacial eagle]

Sure, thanks

[steep sand]

Good eveNing everyone

Any tips on separating render logic from ai update logic?

Not sure if this is the right place to ask

[patent blade]

um... separate them?

have an Update() function and a Render() at the most basic level

[steep sand]

To my surprise, it was simpler than I thought

Thanks for the insight

[torn panther]

running inference is pretty cheap actually...

[rugged zodiac]

What datastructure would you usually use for a 2d grid-based system with NPC path-finding? My implementation would require really big grid and a lot of path finding on it, i'm actually new to programming so my only ideas are 2d array or graph, and for both i'm kinda struggling in performant* path finding.

[elfin vortex]

If grid is sparse or have very large data structures for entities use 2D maps of pointers or references to some dynamically allocated object. If dense or each grid’s object’s memory footprint is similar use 2D array of objects

It’ll depend a bit on your choice of language too

[rugged zodiac]

And for path finding on this grid, how would you check for closest enemy to move to for example?

Atm I'm considering to map all the NPC's positions to an array or something and then try to find the closest one in it, but since i'm planning to have really lots of NPC's it kinda sounds bad. Or maybe simple vision system where the NPC's would trace in all directions until they touch an obstacle, but that may be bad at performance i suppose cuz i could have really long ranges without any obstacles on my grid.

[elfin vortex]

if you're okay with c++, the classroom way of doing this would be everything inherits from something like GridEntity, each with its own tick() or doSomething(), where an Enemy that inherits from GridEntity can be performing pathfinding to the player (doesn't have to be every tick to save some performance), Space can also inherit from GridEntity and just have nothing in the tick function. This way you can loop through the whole grid every tick and call tick() on all the GridEntity and it would be a normal approach (you can instead store all the enemies in a separate array to call their tick() each frame, which removes the need for inheritance and heavy usage of OOP). If your grid is very large, you'll definitely want some kind of acceleration data structure to do pathfinding with (normal non grid games can break the world down to rooms, and have doors and windows or covers as connection points that eventually lead to the player) (in 2D grid you can break it down into 8x8 or 16x16 or some sort of cell, and do pathfinding from the enemy cell to the player cell, then do path finding from the enemy to the closest cell border and then player to the closest cell border to complete the path). You'll almost always be using the A* algorithm (google this) rather than some sort of look in all directions (which only gives you 4 or 8 directions and wouldn't suffice for good path finding). As you are implementing this, go for the most straighforward solution and don't worry about performance. Usually you'll want to profile if the game is running slow and fix bottlenecks instead of spending lots of energy trying to make every little piece very performant and make the code hard to read in the meantime

[rugged zodiac]

Thanks! Especially for pointing out A* algorithm.

[wind quartz]

GamerZero thanked Mudkip Hacker

[torn panther]

look into nav meshes also

[patent blade]

a grid is fine

but pathfinding is expensive

you want to avoid doing it as much as possible

and use things like flow maps if you have lots of objects pathfinding

[solar pivot]

hey so...

I can't seem to google my way into finding if this has been done, but are there any ais that manage other ais

like, this AI is good at understanding languge, this one pictures, etc. couldn't you just network them and have a hierarchy to mange them to build an AGI that way?

[patent blade]

are you talking about like machine learning AIs?

[solar pivot]

yes

[elfin vortex]

I think connecting existing models won’t get super far as far as AGI, it would only be as good as the sum of its part not greater. It needs some new unsupervised learning breakthrough

[hushed bone]

figured this would be the best place to ask for help, but I'm having an issue with boids. They all seem to have an issue related to cohesion where they all head south

I'm fairly certain it's cohesion related, because if I remove cohesion from the "equation" the issue is gone

[dry pike]

Maybe someone can point me in the right direction. I'm contributing to an existing open source AI system for a Trading Card Game. The system is set up in a way where I just hook into events, and then I make the AI consider what to do based on the current game state. The high level of that is, I just create an "executor" which is just a struct that ties a function (returns bool shouldExecute), card type, and action type together. Every time there's a new phase, the system just attempts to run every register executor sequentially.

As I'm building these AI scripts though, I'm noticing that most of my functions are just going to end up being gigantic if-else statements. Especially for every game-state condition I need to check. Is this common in AI? Or is this a sign that I need to try a different approach?

I've thought about somehow serializing some sort of configuration that defines card interactions, but that might cause more problems than not. I'm almost thinking that my two options are, machine learning, or giant if-else statements.

[patent blade]

do you have examples of some of these if elses?

this is what decision trees solve generally