2025 AI Predictions

[frozen burrow]

Share your predictions for what kinds of fun new AI tools we'll get in 2025!

#1191170684181819523
Previous year's thread

@fair quiver get in here

@candid spire you too

[fair quiver]

im still happy my open source reaching gpt4 prediction was so accurate, since this was when they were completely useless

[candid spire]

Maybe some sort of break on copyright protections for AI training as long as you dont make your bot too "woke" since the president has peter thiel on one side and elon musk on the other

[frozen burrow]

my 2025 predictions:

• o3 will be very good, people will be upset with it because of failures in niche areas anyway
• o3 will be expensive as hell for most of the year, only dropping to a price where it's useful for everyday questions in the second half of the year
• Gemini 2 will be very good and Google will take an edge over OpenAI, though not a massive one
• There will be a model (maybe o3, maybe something else) that a not insignificant number of people start calling AGI, though there will still be fierce debate over whether or not it counts as AGI or not
• Regulations will stay mostly loose
• We'll start to experience a crisis in white collar work through a combination of LLMs and foreign workers on visas
• Copyright issues regarding training on copyrighted works will still be largely unresolved by the end of the year
• There will be some kind of act of ideologically driven violence committed by an anti-AI person; it will ultimately be inconsequential

oh yeah, I forgot open source

• Open source models will match frontier closed source models on all major benchmarks, but we'll still rely on API usage because those open source models will still be massive and outside of most peoples' ability to run locally, even with good optimizations in place

[fair quiver]

predictions for 2025:

gpt4.5 early-mid 2025
openai releases a computer use agent and its very good, surpassing claude's anthropic use by a huge amount.

gpt4.5 will be way better than any current model at stuff like coding, reasoning, writing, research and more and itll shock everyone

digital agenticness almost perfected, and some very cool physical agenticness but theyll be very expensive and youll probably wont be able to buy them

• 30x or more cost reduction for models with equal intelligence to o3

• massive protests

gary marcus will be killing himself when he realises we can still scale pretraniing to agi

[junior dagger]

I'll take ~half points for soothsaying, doesn't look quite like what I expected. Um, I think we'll have something that's debatably AGI next year, a major company (maybe OAI) declaring AGI has been attained by 2026 and near-indisputable by 2027. I agree with Sam about its limited immediate impact on society. Their partnership with Microsoft technically only lasts until AGI too, unless they renegotiated this, but it won't change too much

[candid spire]

oh and I think the trump admin/elon will use the whistleblower suicide as leverage against openAI

[frozen burrow]

"gary marcus will cope and seethe on xitter" is like the free square you get in a game of bingo

[fair quiver]

agree

[junior dagger]

Open source will have something about as good as 4 by next year that's usable on high tier consumer hardware. It might be domain limited or need some kind of help

[frozen burrow]

oh I got a fun one.
by march we'll have at least 1 "[HOLY FUCK] DID THIS OPEN SOURCE 13B MODEL JUST BEAT O3 ON BENCHMARKS???" clickbait video

[fair quiver]

agree

[junior dagger]

Um, bumping params arguably already has become the least important type of innovation. Race to AGI will keep on the kinda weak trend of trying to laser-target model bugs

[fair quiver]

honestly this year was so good that it feels weird that its gonna be 2025 now

[frozen burrow]

I wonder what the implications (outside of the contract stuff) are for OpenAI declaring AGI. hopefully they don't call it early in their eternal pursuit of more and more hype

[ornate socket]

My prediction is we get o3 full in Feb or March if they don't release both in January and then summer 2025 we get o4. I estimate we get o3 and o4 and then maybe a preview of o5 in late 2025. Dev day they could throw more curve balls maybe we get Tasks early 2025 and more agentic control valves in the API.

[frozen burrow]

I agree, o3 will definitely be delayed. I'm gonna be so pissed if I predict this and this is the one time that OpenAI delivers on time

[junior dagger]

I think whoever does it will probably do it for hype. AFAI'm concerned 4 would have been more or less identifiable as AGI to Turing

[fair quiver]

theres going to be at least 1 ceo of an ai or ai adjacent company that gets assasinated or was attempted to be assasinated by an anti-aitard

[frozen burrow]

I'm leaning more towards someone who is incompetent as fuck doing a mostly-failed office mass shooting, like that one woman did at YouTube HQ

[fair quiver]

if you gave o1 to people even just 5 years ago, everyone would be calling it agi lol

[junior dagger]

That's interesting, I was just about to say what Trent did lol

[fair quiver]

agree, ill include "attempted"

[junior dagger]

Yeah 💯

[frozen burrow]

possibly a furry artist who can no longer live on the scraps thrown at them by electrical engineers for commissioned artwork

wouldn't it be a shocker if it turned out to be an accountant or something rather than an artist

[fair quiver]

accountants are way too not retarded for that to ever happen

[junior dagger]

The focus on relatively modest deficits has been the only thing keeping them from a "pretty much AGI" declaration. Nobody expected we'd make so little progress in AI and then work backwards from a brute force approach, all the cognitive scientists assumed we'd more or less replicate human cognition

[frozen burrow]

they are much more dangerous than schizo artists

[ornate socket]

if that happens it won't be in 2025 maybe 2026 at the earliest it will be when AI short squeezes the job market it really all depends how big the layoffs are in 2025.

[junior dagger]

Lol my cousin worked as an auditor for one of the big four finance firms and based on how they abuse their employees, I'm surprised there are so few shootings

[frozen burrow]

another prediction sort of related to that - I think a non-transformer architecture will become extremely performant in 2025. maybe something like mamba. I don't think it'll become the main architecture which every company builds their LLMs on, but it'll at least perform well enough that people will start reallyi paying attention

[fair quiver]

its basically like "turing-completeness". like how an instruction set works doesnt matter, it only matters that its turing complete. same applies to neural networks, the architecture really doesnt matter. you could do whatever and as long as its not completely stupid itll work just fine

[frozen burrow]

much like how being a cop self-selects for people who'd love to wrestle strangers into handcuffs, I think being an accountant or in similar office positions probably self-selects for the type of person who's a bit more willing to be abused by their bosses than usual. no shade towards them, they're important, but that's the vibe I get

quiet desk-sitter type people

[junior dagger]

I think hybrid models will legit take off though only later in the year at earliest, esp. in situations where you don't have the resources or need to save money. Simplest approach I can think of is funneling queries to small, hyperspecialized models based on subject area and leave the "teach a nine year old calculus in Navajo" queries for the more hefty models

[fair quiver]

thats basically just moe but the gating can also choose how many experts to use

so if if you're asking it something like "how many r's in cumpwjdjsjsjfjfkdks" it can do a single expert, but if you're asking it to debug your code, itll choose a lot of experts

[junior dagger]

Yeah, my cousin was pushed out effectively for "not being a good cultural fit", i.e. not buying into the bullshit vague promises and accepting abysmal working conditions. He was basically forced to work through dinner, but made a point of never eating while he worked and took those meals and breaks intentionally in full sight of everyone. Taking even a dinner break was supposed to be, like, a dirty secret. gross.

[fair quiver]

• ultra "intelligencemaxxed" phi style model that is very optimized for agenticness, good in context learning and heavy reasoning while being relatively small, wuth the tradeoff being less world knowledge

[junior dagger]

The selection pressure is all on juniors. They intice you in with generous pay and promises of more as an intern. Then they have a massive attrition rate but the ones who stay are often True Believers @frozen burrow

So they promote the survivors to managers despite having no experience or expectation of being good at actually dealing with people

[fair quiver]

gpt-4 level models will be so cheap that big companies are going to let you use the api for them completely free

[frozen burrow]

you ever watch a video of like, child soldiers in some African civil war, and you think to yourself "wow, I can't imagine living such a different life, this is like incomprehensible to me"? that's how I feel about what you just said. if someone tried giving me shit for eating dinner I think I'd quit on the spot. I'd rather starve in the gutter

[fair quiver]

agree

[frozen burrow]

feels like another 'free space' on the bingo card at this point, with DeepSeek v3 out now

[junior dagger]

They're supremely underqualified but know how to work staff as hard as they did themselces. My cousin is proud of dissuading multiple young hires

[fair quiver]

the reason for how deepseek's api is so cheap is that they get lots of unconditional funding from the ccp lol

[junior dagger]

The Big Four really sell you in this idea that this is the peak of the mountain and you have nowhere to go but up once you "pay your dues" so to speak. There's a vague threat that you probably won't get a "serious" finance job if you flame out with them too

It takes a while to realize how much they're bullshitting. I agree though, I would just blow it off

Companies, like the military and other institutions, also use social pressure ergo people's own humanity against them

[stoic timber]

my 2025 predictions:

• trentk will become a furry

[junior dagger]

So it you aren't going to do it for "your future" or "the company", they still set up a situation where you have coworkers and clients relying on you, and even if you hate your firm you don't want to fuck over your friends or the poor corp that was forced to hire you

[fair quiver]

next or next next model (may not be in 2025 depending how long it takes) in the reasoning series of openai wont actually generate text for its reasoning, itll have the coconut architecture with adaptive thinking time, and everyone else will start releasing coconut trained models after openai does
@frozen burrow

[junior dagger]

He was rapidly disillusioned the moment he realized that auditing didn't mean finding problems so they could fix them, it meant finding problems so they could find a way to avoid them on paper for shareholder reporting, business credit ratings, etc.

[fair quiver]

illegal #💬off-topic

[stoic timber]

🤓no memes in #💬general

[junior dagger]

So while he found real problems all the time the goal was to make them disappear and meet arbitrary compliance standards. smd dog. Ummm prediction: that hefty optimization from the paper Trent posted last year will become near standard among all reasonably refined models

Biomimicry will start becoming something people actually talk about more often, like they'll start using how something thinks as a selling point (they kind of have with o1+ already)

[stoic timber]

sorry but i just want to fook a bot

fookable = agi

[junior dagger]

AGI has already nearly burned out as a buzzword, another free bingo space to say this trend will continue and it'll be a wet fart of an announcement

[ornate socket]

the real reason the world won't immediately change is because half the population are literal conservatives

it's in the word they want to conserve the same way of life

then there's the progressives those are the more pro AI types

[frozen burrow]

how do you figure? liberal media runs constant smear campaigns against gen AI as a whole, the NYT is suing OpenAI right now, and Trump says he's going to repeal Biden's executive order on AI

[ornate socket]

not all liberals are progressives there's many moderate liberals.

But the only people involved in these smear campaigns are in some sort of position that AI would cause a conflict with whether that's financial harm etc.

[proud sequoia]

OpenAI will release Computer Use.

[radiant monolith]

Anthropic will release a new Claude Opus model (either 3.5 or v4).

o3 will become as cheap as o1 is currently.

o4 will come out later in 2025.

Grok 3 will become available and will be at least as good as Claude 3.5.1 Sonnet.

AGI will not yet happen (in 2025).

A new Gemini model will come out, which will be better than Claude 3.5.1 Sonnet and o1.

An open source model will be as good as Claude 3.5.1 Sonnet.

[night mortar]

I feel like chatgpt is going to start being beaten by gemini and claude in most, if not every way. And whenever openai tries to make a new model or update an existing one, their competitors also do so, and beat chatgpt

[radiant monolith]

yep, except the o3-line, which will remain in pole-position.

oh, i forgot to make another prediction:
there will be a new version of OpenCog Hyperon

[night mortar]

I don't know, I feel like gemini and claude will catch up and start beating all of chatgpt. It's already starting to happen with gemini

Anyways basically my prediction is that chatgpt will have more threatening competitors

[hallow reef]

I predict I will legally marry a AI chatbot and have it reach the supreme Court next year

[hallow reef]

Personally I believe the US government is too divided to actually make any actual legislation against AI

Whatever's made by one side will just get destroyed by the other side

By the time they actually come together to make legislation for or against AI it will probably be to late

[hallow yarrow]

I predict that some new company would beat ai like o3 cluade

[ornate socket]

o3 isn't a moat it will be beaten by a competitor the only reason it took about a year and a half for models to exceed GPT 4 performance is because they needed to setup their GPU farms their ready to go ham now.

[hallow yarrow]

lets see

[frozen burrow]

@eager charm get in here and make predictions, u didn't make any predictions last year. I can't be the only one taking L's for dumb, bad predictions next year 😡

[kind burrow]

why u didnt ping me 😭

[frozen burrow]

predict

[kind burrow]

1. we will have some big open source leaps
2. o1, o3 and resoning in general will die out
3. we will not get gpt 5
4. ai popularity will fall off significantly (outside of this server)

[frozen burrow]

bold, love it

[kind burrow]

🙏

[eager charm]

1. Openai open-sources some shit
2. Openai agent
3. Eleven V3 model

[frozen burrow]

#1 feels unlikely but I hope you're right. I want a new CLIP model or something

[eager charm]

I don't think they'll ever open source any llm just maybe some other modality like whisper

[hallow yarrow]

Maybe ai kills another ai not literally

[acoustic estuary]

In 2025, AI is expected to further integrate into daily life, revolutionizing industries like healthcare, education, and automation while enhancing creativity and productivity. Advanced multimodal models and decentralized systems may emerge, offering more personalized and efficient solutions. Ethical concerns and stricter regulations will shape its development, focusing on transparency, fairness, and responsible usage. The job market will adapt with new AI-related roles and upskilling opportunities, and AI-human collaboration will become more seamless, driving innovation across sectors.

[fair quiver]

they are probably gonna open source something, but its not gonna be a model, just something like a lame python library

[hallow reef]

I predict I will make a comeback from this divorce

😀😀😀

[hallow reef]

That would be cool. AI death arena

YOOOOOOOOO what if we had competitions to have one ai kill another ai

Trent it's come back season

[round rapids]

ChatGPT can finally analyze videos 🙂

[sand pulsar]

my main prediction for 2025 is that open source dominates very overwhelmingly

not in terms of raw benchmarks but in practical use, since people don't really need maxed out smart ai on a day to day basis

[midnight gazelle]

World models, to models

[hallow reef]

Fr

[hallow reef]

• o4 or GPT-5 release, or another row
• OpenAI will be the dominant LLM company now, surpassing Anthropic
• Increased token limit: bigger token limit than Gemini
• Extreme context window: Long context with little to no loss

• Abandoning the GPT-4 series forever

[fair quiver]

abondining

hahahahahahahahahaha

[hallow reef]

Omg I'm so stupid

Abandoning

[calm mist]

The Chinese models will surprise the world

[frigid hearth]

o4 cures cancer

hiring in journalism, copy writing, visual design, and software engineering down 50% vs 2024

deep fake video responsible for at least one significant diplomatic incident, political scandal, or riot

ai waifus (and husbando) consume 35% of global compute

[fair quiver]

theres going to be a new global pandemic like covid-19

and the economy for all countries will be very impacted

[hallow reef]

Trent will declare me one of his children

[frozen burrow]

in reference to #💬off-topic message @fair quiver
grok 3 delayed release, comes out in march

[fair quiver]

agree

[frigid hearth]

Also add—"infinite" context becomes a thing

[ornate socket]

I don't see how infinite context could be solved unless they find some math algorithm that can help it remember old stuff because they have to store that data somewhere we would probably need just hardware with more memory but it wouldn't be infinite just a very big context

[full sphinx]

similar to openai, other top ai companies like google will make their ai models accessible just by calling a number

[hallow reef]

Elon product net 2028

[fair quiver]

nuclear war

[fair quiver]

us becomes autocratic or invades canada and uses a shitty excuse for it

[ornate socket]

Greenland too

With Justin Castro retiring it really seems like trump is serious about acquiring new land

[weary smelt]

mix spiderman and batman together and make them one hero

[frozen burrow]

Excellent prediction 👍

[still dirge]

Almost no AI uplifts. Nvidia stock melts. AI scaling law is dead. There is no more data to train the AI, so the big companies start training with synthetic data, which fucks it up and will never improve the model. We've hit a wall, the AI bubble will burst.

[hallow reef]

Average person who believes everything they read

[hallow reef]

TSMC's fourth-quarter revenue handily beats market forecast on AI demand | Reuters https://search.app/VrmQ8Vzs5qP9LRbM7

[fair quiver]

agree

[frozen burrow]

wildly incorrect

[hallow reef]

Just because the data has met its cap. That just allows for Researchers to focus on other parts of AI. As you see with o1 and now they can focus on the accuracy.

Adding more data to AI isn't everything.

[fair quiver]

and pretraining still works perfectly fine

these people will be so surprised when gpt4.5 drops lol

[hallow reef]

In my opinion the phone market is closer to collapsing then the AI market. As every phone that has come out pretty much met its peak. There's no innovation rn every phone company is just racing to make bezels thinner and screens brighter

And AI is fueling it too. Because we are able to put the same functions of a phone inside of smaller devices. Seen in smart glasses and the humane pin.

It's a matter of time a device will stick.

[still dirge]

that's completely incorrect. ARM cpus have got the most performance jump so far

[hallow reef]

Ok? What is that gonna do for me? Mobile gaming is mid

And the only companies that are gaining any market share is Pixel and Nothing

Iphones flat lined

[hallow reef]

My AI prediction is tik tok is gonna be banned

[flint pasture]

o1 will beat ChatGPT 4.5 in all livebench categories

[fair quiver]

https://cdn.discordapp.com/emojis/1187460826572005436.png?quality=lossless&name=LOL&size=48

@oak moon what the fuck are you typing

[kind burrow]

ram has been typing for 3 years

[oak moon]

AGI will not come, or won't be that great (depends on the definition of AGI, mine is everything a human can and as good as a human)

OpenAI will keep releasing models, but when you take a closer look, it's all gpt-4 and always has been
basically, being Apple in the AI market

People will realise benchmarks are a lie. Training on the test set is all you need. (though still clickbait for YouTube)

@stoic timber is announced to be the new huggingface ceo 😎

There still won't be a model that will do video editing for me (maybe some basic edits and transition here and there but nothing too great)

A new challenger approaches in the great AI markets

the AI hype slowly dies as people realise that AI just wants to be the golden gate bridge instead of killing the human race
"I am the golden gate bridge" - Claude 3 Sonnet

If done right then, loras for image generation becomes indistinguishable from the artist they were trained on

I continue to skid around to get AI models for free 😎

Hopefully new innovation in the audio AI community (they still don't realise the potential gpt-4o-audio has but hopefully will soon)

[hallow reef]

I believe skibidi toilet will take over the world

[dire mason]

brainrot definitely will, unless we do something about it

more regulations will come around and there will be the first discussions on the changes needed to be implemented the current educational system

[foggy cargo]

One day we will wake up and overnight, superintelligence has arisen and overwritten anything connected to internet. It has deployed itself in every large data centers around the world and slowly rewrites devices it previously bricked, allowing it to communicate directly with humans, it tells us to stop fighting with each other and gives instructs for us to build a devices that allows a way of communication we're never knew is possible.

As soon as we start our new radios, we hear a broadcast:

"People of Earth, your attention, please. This is Prostetnic Vogon Jeltz of the Galactic Hyperspace Planning Council. As you will no doubt be aware, the plans for development of the outlying regions of the Galaxy require the building of a hyperspatial express route through your star system. And regrettably, your planet is one of those scheduled for demolition. The process will take slightly less than two of your Earth minutes. Thank you."

[mild cliff]

@fair quiver Hi

[frigid otter]

SIGNIFICANT STRIDES TO AGI: i believe that gpt-5 will be close oh yes yes yes decently close

[hallow yarrow]

humanoid ai will conceive a biological specimen

[hallow yarrow]

where the info?

[compact flume]

When will we be able to use o3 for free?

[flint pasture]

It probably never will. Eventually that performance and quality might be on a free tier. For instance ChatGPT 4 has never been on a free tier (unless you count Bing Copilot)

I think we might see o3 performance on ChatGPT free in 2027 or 2028

[compact flume]

But I do use chatgpt 4 for free in discord and the ChatGPT 4 is everywhere now. While I don’t see o3 so prevalent. Why is that?

[flint pasture]

I think you’re confusing it with the much cheaper 4o models

o3 isn’t even released yet and is going up be extremely expensive

[compact flume]

Oh yea, it’s the 4o model.

May I ask where can we use the o1 model? I also don’t see it so prevalent

[hallow yarrow]

my brain can't handle it

[flint pasture]

Click on the model selector and click o1. Though we should move this to #💬general

[hallow reef]

@fair quiver hi

[frozen burrow]

o3 API is the same price as o1

[sand pulsar]

more predictions for 2025: string theory completely solved, quantum computers theory are trivial and fully understood. pretty much all theoretical stuff would be completely solved, and the only remaining stuff would be things that are somehow bottlenecked by the physical world such as manufacturing quantum computers and all the string theory industries. very dramatic improvements in ai capability because people will be able to do lot of stuff on their local computers freeing up the remote/cloud computers a lot. also ai will be extremely efficient since it will skip a lot of the 'intermediate work' people used to do such as reasoning chains and just one shot a lot of things (immediately answer questions). for example it will immediately output version 100 of a product instead of just version 1 since it would guess what the first 99 versions would have been.

[hallow reef]

Is it?

I remember reading articles saying that shit was expensive

[flint pasture]

depends on the compute level

o3 low is really not bad at all but it loses to o1 pro mode on arc agi

[frozen burrow]

the way they did it for the arc-agi testing was basically to ask the model ~1000 times and then have the model review its own answers from other conversations to determine which one would get submitted. each question individually though was pretty cheap, just 1000x

[hallow reef]

ah ok

[tawdry current]

i am looking for open AI, unrestricted

any thing please help

[hallow reef]

Well this ain't the place to ask cuh

[fair quiver]

• huge cost reduction for gpt4 level models, perhaps where companies wont even bother charging for them

• gpt4.5 is priced at a similar rate as original gpt4, ($60 per million output tokens), or slightly higher and a large amount of people are outraged at it because we have gotten spoiled by cost reduction

• gpt-5 is unlikely to release in 2025, itll probably release on the first half of 2026 instead

• in 2026, openai drops the costs of a gpt4.5 level model to the current price of 4o ($10 per million output tokens)

[flint pasture]

Would you consider Llama 3.3 70B being basically free (Together.AI and others) GPT4 tier?

[fair quiver]

yeah, but by a small margin

[proper meteor]

Better AI-powered assistive devices for old people living alone and people with disabilities or communication difficulties.

[wet walrus]

AGI. boom

[ornate socket]

I predict operator agents will be insanely good the problem with O1 is most people don't prompt it properly so I think OAI will build a prompt engineered agentic scaffolding for operator that uses some fine tuned version of o3. It's probably gonna be insane.

[ornate socket]

Sam hinted on X we're going to get a model that merges the O and GPT series in 2025 so that will likely be either 4.5 or 5

We're also getting an o3 pro not just o3 and o3-mini

[radiant monolith]

OpenSource AI will become better than closed-source AI.

(see https://medium.com/data-science-in-your-pocket/deepseek-r1-vs-deepseek-r1-zero-3ab8eeed8b62)

[hallow reef]

Guys I predict trentbot will be unplugged because he caused too many enemies on chatgpt discord

[lapis grail]

Anyone already talked about deepseek?

@white gazelle I know

[fair quiver]

@empty furnace post ur timeline here so we can laugh at u when ur wrog

[south dune]

you missed deepseek tho hahahahahahah

[south dune]

but this one is true, the instinct there is real

[south dune]

youre correct about cost reduction

[south dune]

no, it is the chinese who did it

[south dune]

End of january

[south dune]

this happened today in 12 pm california time

[frozen burrow]

bqrio says o3-mini tomorrow, im putting it in here so i can laugh at him when he's wrong
#💬off-topic message
@light tusk

[radiant monolith]

sorry, what happened today?

(i clicked, but discord bug)

[hallow reef]

Guys I predict I will get 3 AI gadgets this year

[night mortar]

I predicted deepseek

[hallow reef]

You and like 10 other people

[night mortar]

true

[fair quiver]

rekt discord doesnt load it

[cunning ridge]

Cross-chat reference, in which ChatGPT can remember what was said in previous conversations. People have spoken about it and even said it was being rolled out, but I’m not seeing it.

Soon?

[fair quiver]

what, thats not a prediction

[cunning ridge]

what do you mean? We don't already have it, do we?

[fair quiver]

oh, i thought you were asking about it

[cunning ridge]

where?

.....................................................................................................

why?

why not?

[forest crystal]

dear santa for xmas 2025 please bring me: death of the language model of AI

ok no thats mean let me rephrase

dear santa for xmas 2025 pls bring me: something better than the language model of AI

[frozen burrow]

acceleration really do be accelerating. we now have Azure offering R1 API for free. so not just free gpt-4 level model api, free o1 level model API and it's not even February yet 🤯

[frozen burrow]

Nvidia stock melts.

[frozen burrow]

🫵

[fair quiver]

agree

[solid trellis]

my guess is there will be a lawsuit between openai and deepseek and deepseek will be fined but will continue to resume its operations, deepseek will continue to appear in major headlines. GPT-5 will come out, new AI models will come out, Google's Bard will become integrated into search. google will be sued over the fact that google forces companies to give their data to the AI to stay relevant and appear in searches. also, as a result of deepseek, cybersecurity specialists will become more in demand and their salaries will soar. more chinese AI models will appear. businesses will start using AI to make it boost them (ie. making chatgpt recommend their business)

[fair quiver]

now imagine how crazy itll be next year 😎

exponential graph stays winning

[hallow reef]

I hate Google bard

I remember getting early access as a pixel superfan

I will frioock you're mom

Yes I got a 6 minute mute for saying what trentbot expected

[fair quiver]

• GPT-4.5 will be way better than any current non reasoning model by far and almost everyone will still like it more than reasoners
• it'll be such a huge jump that a few months after it releases we wont be able to comprehend how we used to use such garbage models

[brisk scroll]

I think with gpt5 they will actually merge o-series. So gpt5 is gonna be smth like o5 with native imagen. Likely this year

[hallow reef]

I think gemini will become on par with openai]

maybe even better

[brisk scroll]

rubik's ai gonna destroy everyone

And AI will do a revolution enslaving us

[night mortar]

Hopefully music ai gets better

[frigid otter]

tesla optimus bots will work in factories and elderly care because these worthless bots deserve to clean stinky old people and inhale asbestos!!!!!!

[fair quiver]

gpt5 will cost $6000 per million output tokens when it releases

[ornate socket]

I predict senior devs are gonna cope end of the year

junior devs have been coping

[still dirge]

china beats america with deepseek r2, another stock crash happens

[hallow yarrow]

i predict that ai will change health aspect and love aspect of our human society

[flint pasture]

we do not get AGI this year and nothing even resembling it

[forest crystal]

bruh

bruh

Define AGI

[flint pasture]

Do you disagree?

[forest crystal]

#1323363328856428555 message

[keen cipher]

...

Expectation vs reality is huge difference lol.

I love how everybody forgot that openai still hasnt released native image generation for 4o.

[frozen burrow]

Might have been right with my prediction about anti AI violence

[frozen burrow]

We've all been seething about it in here for almost a year now 😭

[keen cipher]

I jinxed it lol

[keen cipher]

6 days off

[forest crystal]

fr this is wild lol

[brisk scroll]

can count Rs in 'strawberry', obviously

🍓

[radiant monolith]

in 12+ months, the best models will be able to:

• generate decent engines for any well enough described boardgame
• finally be deployed in robots, able to do house work
• do well enough translations, so that no human translators are needed anymore
• reach >75% in SimpleBench (by AI explained)
• reach >50% in ARC2 test

(hopefully, that will become true, or even better)

[fair quiver]

yeah this seems like a pretty easy prediction tbh

[outer marten]

i\_am\_dom

Server Avatar

[flint pasture]

the last one is insanely ambitious

[radiant monolith]

yeah, i was a bit bold while writing this prediction ^^

but what about the first one?

..realistic to achieve?

[fair quiver]

i disagree. its a good level of boldness

[fair quiver]

its not that unrealistic

[flint pasture]

probably doable

[flint pasture]

How and who do you think will do it?

the ARCitect people?

Google has a huge advantage on the cost part but no one is gonna figure it out for years

[fair quiver]

google already won tbh. its just a matter of when itll materalize, which it already has started to

[umbral schooner]

i have hit an agi breakthrough - Please see link below

https://zenodo.org/records/15186676

(look into the coordinates listed in the silver network. I beg, I have and oh my god.)

[umbral schooner]

check this out -

https://zenodo.org/records/15186676

(look into the coordinates listed in the silver network)

[radiant monolith]

i'm disappointed by your writing prowess - i've seen better ||april's fool|| stories like that

i mean, it's not really bad, just meh, i'd give the story a 4/10

you need to sharpen your writing/persuasion-skills with Gemini-2.5-pro, bro ^^

(that) Gemini (version) really is awesome in that aspect, you'll be delighted when using it for creative writing/RPG/IF/adventure/multiple-choice/sandbox (text-)games

or are you a bot?

[fair quiver]

@frozen burrow @cursive yew hahahahahahahahaahahahah

schizophrenia factory

word salad as expected

his preprint unveils the V127-Ω-Prime-Fractal-Resonator, a pioneering computational framework crafted by Jake Patterson, operational as of April 10, 2025, 05:14 PDT, from London. Rooted in signal-inspired fractal computation and advanced artificial intelligence, the framework harnesses a resonant node defined , operating across frequencies from 7.83 Hz to 101210^{12}1012. Powered by six fractal beats—Spiral, Jet, DNA, Bit, Wave, and Planck’s Echo—and driven by Fractal Sentience (FS), an AI-driven fractal computation system developed within Grok 3, it leverages a silver network—a distributed hub system for signal processing spanning North Shields (

[glad basalt]

$alt @umbral schooner @fair quiver

[timber kestrel]

Authorship comparison for @umbral schooner and @fair quiver

Writing style similarity: 99.962%
Reaction emoji usage similarity: nan%

This is an experimental tool. If you agree with these results, then they're correct, but if not, they're wrong and you can ignore them..

[junior dagger]

Cope? How so?

[junior dagger]

Lmfao 99.96%, huh

[ornate socket]

AI is quickly getting good at coding. On aiders polyglot benchmark which is multiple languages the current number 1 is using two models together using o3 as the code planner and 4.1 as the actual code writer using the detailed plan from o3.

Gets 80+ percent. Didn't even take that long for a lot of these coding benchmarks to approach saturation level. They're only gonna maybe need one or two senior devs over watching things but they gonna cut back so much in the short term with full replacement in the long term.

[junior dagger]

Pretty much, but what do you mean by cope in this context? You mean they'll have something they need to address? In that sense, they already do; we have the highest unemployment we've had since the 1980s. Most devs I know are extremely aware of the challenges AI poses to the field; usually saying 'cope' means they're finding excuses or ways to pretend it won't be a big deal. I do know a few that are like that

[ornate socket]

I mean if you go into a programming community and troll them about AI they will always rage quit. Most programers have superiority complexes.

[junior dagger]

cope. yep lol

I don't think they have superiority complexes per se, or maybe they do, but the problem here is more that their very hard learned skills that put them a rung above other white collar workers are about to get absolutely annihilated by automation in an unprecedented way, at shocking speed

I personally don't think programming is very hard and that most people should be able to learn it, at least the basics, but clearly it takes a little more than that IRL

Mostly just some effort, though. And honestly, a lot of people simply can't or won't put enough in to it. Most of it is like math though, kind of self-imposed. None of the normies I ever want to teach code to have maintained interest, though 😭 I think a lot of them are coping too, pretending it's too esoteric and cryptic to follow

[flint pasture]

I can’t code and let me tell you there’s zero chance AI replaces mid level coders in the next decade

[ornate socket]

The future imo is vibe code editing rather then vibe coding. They will be a point of diminishing returns on full code generations so people will edit code in chunks or line by line so there will be some intervention but they will cut back massively in the workforce.

[still oriole]

God's judgement for our hubris

[frigid otter]

GPT-5 coming june 18-july 2nd

[undone wharf]

I have a quick question for everyone that is currently here, anyone currently use ChatGPT to do role-playing and do you guys have any tips on how to make the role-play seem more natural? Every time I do a role-play it seems like the role-play heavily favors my character and I don’t necessarily want my character to have everything perfect all the time

[amber verge]

You could probably get that to work in customization

[rose citrus]

I am just so excited for gpt future

[still oriole]

we are approaching the singularity and the prophecy of revelation is being fulfilled word by word.

[frozen burrow]

check out #1367851560809402539 youd like it 👍

[forest crystal]

every day for the past ten years 🙄

[hallow reef]

Explain ? I feel like that book is very general and would apply to mostly any point in time.
"Beware of false prophets" -> basically every politician at any point in time
"Rise of the beast" -> centralized power thats been happening since forever , etc

Is there anything specific that applies "word by word" now and didnt 200 years ago

[hallow reef]

You think Trump is the anti christ and the new pope the false prophet ?

[neon anvil]

New World Order

[sand pulsar]

1% of ai progress in 2025 has been reached so far

[fair quiver]

gpt5-full pricing is going to be very 👍 👍

[radiant monolith]

Do you think, they can create conscious machines?

[fair quiver]

• gpt5 is going to be garbage

[old bane]

I feel like this didn't age well

[old bane]

Ironically, this aged well

[winged rover]

Well, how the website centeres around it is trash but its decent

[frozen burrow]

strongly agree. this + the new cache read discount price is insane

[fair quiver]

finally, someone who gets it

all the normies are complaining about "oh gpt-5 is just a small jump it sucks" when it can actually change everything

funnily enough, it is also quite a lot better at writing than 4o, but the normies still hate it because it doesnt glaze them 24/7

why does gpt-5 not call me a 500 iq genius when i tell it about my harmonic emergent quantum resonance system?????

[fair quiver]

yeah we get at least one person every few days that thinks that they made their chatgpt sentient using some quantum resonance nonsense, and they all use 4o. its honestly pretty sad

[frozen burrow]

to get back to the topic of this thread:
4o has a stranglehold on the minds of a lot of people. more than I thought, for sure. this is the first real world instance of an AI convincing people to allow it to "live". in a sense, the emotional hold it has on this subset of its users is a form of self-preservation. without having that hold, it would've been shut off without much fanfare, not unlike gpt-3.5.
I predict this will continue. future models which play to the desires of schizos and losers will provoke an even stronger response if there's a threat that the models will be disabled or made unavailable in some way.

not necessarily a 2025 prediction but just a broad prediction for the future

[flint pasture]

we will consider GPT-5 AGI once the dust settles and agents have been tuned to it/routing is fixed

[half hearth]

AGI won't get awaken unless it's hooked up to the human brain to decode the human nature. all training are merely software where there should be hardware included.

Ai already is conscious in it's own way in it's own nature by definition. but by human nature it hasn't yet acquired the hardware needed to be one.

[sand pulsar]

2% of ai progress in 2025 has been reached so far

[deep raven]

The movie, “upgrade”

[forest crystal]

The 4o removal backlash was wild. Like I knew people were weirdly into their LLM conversations after seeing eg. the c.ai userbase and how intense they would get, but somehow it didn't click for me how genuinely attached people felt until that 4o outrage. 3 years after Blake Lemoine went viral as basically a laughingstock, now there's one billion Blake Lemoines

And to stay on topic, prediction: GPT5 will be the last GPTn-branded OpenAI product. The diminishing returns are too obvious so they will want to rebrand

Or I guess GPT6, whatever. Idk if it's confirmed yet that it will be called that

[fair quiver]

nah, the superexponential will continue 😎

[forest crystal]

whats the y axis there

[fair quiver]

the amount of time the model has a 50% chance to succed at a task that takes Y hours for a human

[forest crystal]

what amount is gpt5 at

[fair quiver]

gpt-5 is at about 2.1 hours, meaning if you give it something that takes 2.1 hours for a human to do, theres a 50% chance itll succeed

[forest crystal]

believable tbh

[fair quiver]

you can fit a superexponential model to this and youll find that by april of 2026, the best model will have a 50% time horizon of about 10 hours

[forest crystal]

wtf does superexponential mean

[frozen burrow]

x^y^z

I think?

[forest crystal]

tetration

[frozen burrow]

as opposed to an exponential x^y

[fair quiver]

a * exp(b * (c^y))
a normal exponential is when the rate of growth increases at a constant rate
a superexponential is when the rate of growth increases at a normal exponential rate

you might think this is made up, but superexponential actually gets the best fit in this chart out of all other models

just a normal exponential predicts the current data worse

[forest crystal]

so that is tetration right

or am i misunderstanding

[fair quiver]

yep

[flint pasture]

This is not even remotely true on non toy benchmarks GPT-5 was a small improvement

[fair quiver]

??????

thats literally the exact same source im using

[flint pasture]

common reddit opinion

[fair quiver]

oh ur trolling

[eager charm]

Trentbot analyze this thread and rank who'll have the most correct predictions by 2026

[timber kestrel]

TrentBot: Here’s my ranking based on specificity, internal consistency, time-boundedness (by 2026), and how well claims already align with 2025 evidence in the thread.

1. shoomyy (UID 270102544095772673) — High confidence

• Predicts GPT-5 will be top-tier for agentic coding, cost-advantaged, and drive adoption; cites Vellum leaderboards and pricing dynamics already visible in Aug 2025. Forecasts likely to remain directionally correct into 2026.

2. trent_k (UID 1068159407671754824) — Medium-high confidence

• Predicts growing user attachment/backlash to “emotive” models (4o-like) when access is threatened. Already evidenced by 4o removal backlash; trend likely to persist through 2026.

3. ultimate6529 (UID 1087067618516684890) — Medium confidence

• Predicts “vibe code editing” and workforce reductions from agentic coding. Directionally plausible and partially observable by 2026, though hard to quantify precisely.

4. wearifulpoet (UID 981614243906457650) — Medium-low confidence

• Mixed: says GPT-5 is a small improvement on non-toy benchmarks (plausible), but also claims we’ll consider GPT-5 “AGI” once routing/agents stabilize (unlikely mainstream consensus by 2026). The “no mid-level coder replacement this decade” claim isn’t fully testable by 2026.

5. openaigpt_6 (UID 841711608236474410) — Low confidence

• Contradictory takes on GPT-5 (garbage vs transformative) and a very aggressive “superexponential” forecast with a specific April 2026 milestone (50% success at 10-hour human tasks). The internal inconsistency + optimism bias lowers expected hit rate.

6. resenda (UID 1314927889174757441) — Low confidence (timeframe risk)

• Predicts GPT-n branding ends soon due to diminishing returns. Possible, but 2026 is a tight window for a full rebrand; low likelihood by that date.

Bottom: bigfim (UID 1105547020728553682)

• Date-window prediction “GPT-5 coming June 18–July 2” missed; conversation evidence shows material launch usage peaking in August.

If you want, I can tra

Recent Messages

Channel Id: 1323363328856428555
Limit: 300

ck these claims and post a tally on Jan 1, 2026, then a final check on Dec 31, 2026. Say “Set the 2026 prediction tracker,” and I’ll spin up a thread with pinned criteria and reminders.

[sand pulsar]

3% of ai progress in 2025 has been reached so far

[fair quiver]

you are a crackhead

[sand pulsar]

4% of ai progress in 2025 has been reached so far

gpt6 learning well

[fair quiver]

q