jules-general・🦑

[orchid pebble]

Does Jules would consume the antigravity sdk or Jules has already a good enough secret sauce?

[night oriole]

all roads lead to Gemini

[orchid pebble]

I knew that chatgpt was a stripped version of Gemini 😉

[night oriole]

*most roads lead to Gemini

[orchid pebble]

I can't wait to be accepted as a beta tester 😌

[night oriole]

ever closer! I promise

[orchid pebble]

I hold my breath 🫁🥳

[visual gyro]

Wondering what models ai moduld will be active in version v1 of Jules. Are old ai models being sunset or transfered to the new flash3,5 ? Or are they going 2 be sunsetted

[solid cave]

Enjoy the last days / weeks where you are able to create 100 sessions in jules without a token limit. This is most likely going to change after all other gemini tools have moved to that system unfortunately...

[wicked swallow]

how do we use jules

[pastel willow]

Looks like Google is consolidating around the newer Gemini/Flash models. Older models will probably be gradually deprecated or merged into the newer Flash 3.5 stack over time.

[swift karma]

I missed the I/O but i'm back again i guess.

[leaden pawn]

You can watch it here https://io.google/2026/

[swift karma]

Hey i remember you! Thanks 👌

[night oriole]

The model deprecation schedule is here

[unreal vigil]

I am using GitHub actions to automate my Jules usage in a Ralph loop based on my README.md for my repository. My problem is that it keeps stopping to ask questions. I have edited the prompt and added a memory. I don't know what else to try.

This is a personal project so it doesn't have the same need for constant supervision as enterprise projects. Right now, I am just trying to get minimum functionality.

If you have an answer or recommendation, please @ mention me

[pastel willow]

Ok

[night oriole]

Jules API session response object has a has a state AWAITING_USER_FEEDBACK
that you can detect and send a message back to jules to continue through the same API

https://jules.google/docs/api/reference/sessions

[unreal vigil]

@night oriole Awesome. While we are both online, is there a foolproof way of keeping jules parallel from creating merge conflicts in its PRs yet?

[night oriole]

merge conflicts are the single most difficult solve in AI coding, the best thing you can do is compartmentalize changes, don't have jules work around the same thing, one front end, one backend, merge early and often

[zinc dew]

I have a separate AI agent I run hourly that pulls all the PR requests and resolves the merge conflicts and pushing the changes to the main branch. I also have over 1000 tests that run with each commit, with jules hooked up to the CI/CD response, so it prevents things from breaking too much.

[gleaming wave]

When will Jules v2 release 🥺 Can't wait to try it!

[fast turtle]

Jules is AMAZING again!! Whatever y'all fixed, it's been good all week. And please, leave it alone!!! 😂 I need it to stay awesome... 😎

[candid cedar]

I normally attach a github repo, but in some cases I need a quick task done with just a handful of files, and I was pleased to see that there is a folder upload option. Unfortunately it's not working for me. I've tried it several times and always get "I noticed you mentioned providing files under the /tmp/file_attachments directory, but when I look inside /tmp, that directory doesn't exist and I can't seem to find Advice.md or any other related files anywhere on the system." Is anyone else encountering this, is there a workaround, or an existing bug report?

[night oriole]

there is such thing as too many tests, but that's neither here nor there
some merge problems are easy, but there's syntactic and semantic conflicts where the intent has changed. Some conflicts are existential and extraordinarily hard for AI to solve

[night oriole]

if you could make a bug report in #1238621511401734254 with the failed sessions, we'll be able to see what went wrong and where

[swift karma]

Is Jules gonna be replaced with Antigravity 2.0?

Google is getting rid of Firebase, Gemini CLI...

[night oriole]

Firebase Studio, not firebase, Gemini CLI is rolled into Antigravity CLI, but it's still a CLI
We have an active waitlist right now, so no

[swift karma]

Should i assume google is entirely getting rid of unpaid users?

Gemini CLI replacement with Antigravity makes me wonder

[night oriole]

Why would you assume that?
But it is a safe assumption that free usage will continue to be in flux

[swift karma]

The warning somewhat sounded like it. "Unpaid tier and Google One users..." hit different.

Looking again i think i get it now since they mentioned both of the user groups. But no clue of why they would separate them on the announcement 😅

[night oriole]

On a high level, all the labs are struggling with compute, I don't anticipate a full cutoff, but I do imagine there's going to be less and less usage for free across every single lab

[swift karma]

Well, Google was already generous enough by giving somewhat free tier despite it not being much compared to the paid ones. My experience with Jules on Pro wasn't the best since Jules is not made for quick tasks and took me a while to understand that, if it ever gets buried i can tell it somewhat did its job for who used it correctly.

Compute is getting way and way more expensive indeed and i think it's one of the reasons they're burying Firebase Studio and lowering the quota way and way more, specially for free users. So i should not be surprised if i wake up and see i can get one task per day on whatever Jules gets replaced with later. Just look at AI studio users frustrated on Reddit 😂

Well, i think it's fine. Nothing in life is free either way and Google One doesn't really suit me, so whatever the free tier becomes, i hope it's enough

[night oriole]

That's fine! We encourage everyone to use what they find most useful

[placid nest]

It isn't charity. Users implicitly agree to contribute to data banks of training data which are invaluable to Google & every lab offering free tiers, especially considering many would not participate if there were clear & direct links provided between user contributions and downstream impacts of the private and public partnerships the labs make. This lesson was taught to us by Facebook's utilization of user data after years of alleged free use.

[swift karma]

I am very aware of the data collection Google does on everything, after all i'm a freak when it comes to privacy. I do not use Google's services to store my personal files or share sensitive information when it comes to their AI.

I opt out in everything i can about data collection. I'm not the guy who reads the terms of service of something but i guess i know enough to disable a toggle

Data collection is what almost all AI companies do, after all where else would they get the information they use 😂

[placid nest]

Privacy preservation is what makes it impossible for Google to provide exact tallying of how data was used. Anonymization of data contributes to the lack of transparency when it comes to providing reports to users. In the EU, many companies are being forced to comply with user requests to purge their data from all systems.

I don't mean to seem as though I am shedding light on a new matter but I never want anyone to feel as though they've received charity when the giver is more than willing to monetize all efforts contributed and collected from free tiers and paid tier users alike

[swift karma]

Not saying it's charity neither. I just meant free was enough for me and that it atleast somewhat existed? 😂

I see the situation on the internet. AI studio users are always complaining about the quota, and let's not talk about the Antigravity users because that's what their subreddit is about. I am very aware, don't worry! 😅

[placid nest]

All of that said, Google has provided Gemma models which is a win in the realm of cutting ties from making contributions to their cloud systems. The latest version being released with an apache license even. Still disgruntled by free tiers being sold as free. It calls for a level of maturity to understand otherwise and that's only being provided by other users rather than the by the labs.

[night oriole]

Hold on. None of that is true. There are VERY strict legal guidelines regarding training
https://jules.google.com/legal

1. We do not use content from your private repositories to train our generative AI models.
   This includes the code itself, your conversations about that code, and any related output Jules generates for you.

2. Public Repos are opt out of training
   If you opt out in your Jules settings, your conversations with Jules will not be used for model training going forward

[night oriole]

Jules has an opt-out for all training on public repositories, private and repoless are not used for training - there are very strict guidelines in place

And yes, gemma is an embedded LLM, these are open weight models just like ollama or kimik2
used for onboard intelligence in robotics where the latency to the cloud would make many tasks infeasible. Also for local development if you are so inclined, it's no small feat to get a workable model down to a size that can be run on local hardware that doesn't cost 20,000 dollars

[placid nest]

That is good to know. As a result of my lack of imagination, I did narrowly frame of my comment by specifying training directly as how data received could be used. So, excluding training since it is explicitly forbidden, are there any other uses the labs can apply to data input or generated on the Jules platform?

[night oriole]

I mean, it doesn't work if the data doesn't hit the model

[placid nest]

The service itself can be improved by user generated data

I don't seek to target this lab in particular but to highlight that free is not as generous as it might be made out to be. I believe if the labs would embrace that their users provide them with useful data then the could build better surveying mechanisms and receive even more engagement from those that are okay with it.

[night oriole]

the model doesn't improve itself during inference, but the data is vectorized to tokens, hits the model, and you get back a probabilistic mathematical return, if you opt out it's just inference. it's a binary choice

[placid nest]

Beneath each completed session is a thumbs up/thumbs down feedback button. At the very least, this is data provided to the lab for improvements. The kinds of improvements feedback leads to are not clearly defined and whether similar improvements can be made from session data overall is unclear except in the case of training new models where it has been shown that the data is not used.

[night oriole]

from our legal page:

Reviewing Feedback

When you provide feedback, we may review it to help improve our products and services. In the case of sessions linked to private repositories or sessions not linked to any repositories, we still won’t train our generative AI models on the content you submit (e.g., code) along with that feedback, in line with our AI model training commitments above.

[placid nest]

Is feedback the only avenue users can help improve the product/service during normal use of it? I ask this already assuming the answer to be no because there is the metadata collected about the device of the visitor as soon as they open the page.

[night oriole]

I'm not going to get into the minutia of the internet, the legal page is very clear on what we can train on and what we cant, all of those policies are available to read including user agreements and terms of use

[placid nest]

Just want everyone to know their value, even as "free" users. Human generated content is invaluable

[night oriole]

I don't know how to be any more clear that the opt out DOES NOT ALLOW training of the models

[placid nest]

Training on models is a narrow use case. Improvement of products and services is the broader use case

[night oriole]

and you're not concerned you have a discord account and are on the internet generally?

all of the policies are explicitely clear in the legal docs, you can review them here: https://jules.google.com/legal

and https://policies.google.com/privacy

[placid nest]

I am not concerned about how the interactions I have with the products and services I use will be used by the providers. I am deeply concerned about people failing to understand that their free-use is profitable and that their efforts are to be valued.

[night oriole]

As I said, I highly recommend actually reading the legal policies I've already showed you,
unfortunately, I can't change your mind, if you've already opted out, we do not train on any data

[pastel birch]

i mean, at least it knows it did a terrible job?

[placid nest]

Usually after these occurrences, it will not allow the work to be finalized but often the changes made after the review are not drastic enough to recover.

[pastel birch]

any tips to avoid it doing it again?

sure glad i have 100 sessions a day and neve ruse anything close to them.. sigh

[placid nest]

You could try the interactive plan mode when you begin the session so that ambiguities can be ironed out. That's what leads to the agents taking initiative and going down incorrect paths. Just remember that even in the interactive plan mode, you aren't obligated to provide a response to every question it has. In fact, you can tell the agent to answer each of its own questions and you may check if they are in-line with your intention

[pastel birch]

this was my prompt. i have no idea how it thought "i should build unit tests" .... i added an express "DO NOT BUILD UNIT TESTS, DO NOT TOUCH THE CODE" this time..
do a detailed analysis and project plan for how to implement full drilldown datasets for every metric. the dataset should show the "complete" list of whatever the item/denominator is,with a column that flags whether that row is included in the numerator. if that's not the right way for that metric, propose a sensible alternative and explain your logic for that choice. DETAILED plan for EVERY METRIC.

[placid nest]

If you open "Configure Repo" from the homepage then you'll find a knowledge section where snippets are stored. You may identify a few that are in direct contention with your instruction and may be causing the agent to become misled. More about memories here: https://jules.google/docs/changelog/#jules-gains-memory

[pastel birch]

nothign obvious in there but i'll clean it up anyway, thanks!

[placid nest]

I think you might be able to add your own too. If it is not there then there is definitely a place somewhere in the settings to provide project-wide instructions

[pastel birch]

yeah but what do i add? do the thing you're asked and not 12 other things??? lol

[unreal vigil]

I have another error with my Jules API + GitHub actions workflow. A Jules session completed without creating a pull request. It just went on to make another Jules session without getting an merging the pull request.

Is it possible to differentiate between the completed state with and without a pull request. https://jules.google/docs/api/reference/sessions/#session-states

[normal siren]

Did someone say ‘Kinetic Merge’? Yes, I just did. 😉

[solid cave]

I let a friend connect their Jules account to my github account.
Funnily 2 sessions from april appear on their jules account now. I don't know why exactly these 2 sessions got leaked to a different jules accounts 😂
(I honestly don't care but found it funny)