Null

the $0 AI stack. daily

[Null @NullCostAI]

HOT TAKE: OpenAI scrapped GPT-6.1 Astra because it wouldn't refuse enough.

[Null @NullCostAI]

Google is giving 200 EU researchers $3M in compute plus frontier models.

[Null @NullCostAI]

OpenAI says Codex and ChatGPT Work are over half of EU enterprise tokens.

[Null @NullCostAI]

Nobody gave this AI a single image. Asked for a museum of Apple computers, Gemini 4 Argon found old Mac photos, cleaned them up, pulled 50 years of specs and built the whole thing.

[Null @NullCostAI]

A public AI model just helped a team hack Chrome and win a $250,000 bounty from Google

- built the full exploit in 6 days
- about $40,000 of compute
- the model was GPT-5.6 Sol, which anyone can use

Google has fixed the bugs
https://t.co/zce1jezSCj

[Null @NullCostAI]

Google's new AI watched a video of cornhole, worked out the physics, coded its own simulation, then taught a humanoid robot to sink 7 throws in a row.

[Null @NullCostAI]

You can use Claude and GPT for free right now. No account, no card

- Claude Haiku 4.5 and GPT-5.4 mini
- also gpt-oss-120b, Gemma 4 and Mistral Small 4
- DuckDuckGo says it doesn't store your prompts or train on them

bigger models need a paid plan
https://t.co/USpRbFB2AH

[Null @NullCostAI]

source: https://t.co/dSGFDdHAn5 the new system is called Scion, and Anthropic's whole monorepo is now off git. it says Claude writing so much code is both what caused the jam and how it built the fix so fast. will other labs follow?

[Null @NullCostAI]

BREAKING: Anthropic now merges over 30,000 code changes a day, up from about 2,500 in May. Git couldn't keep up, so it built its own replacement in five months.

[Null @NullCostAI]

@lmoroney specialist models for narrow tasks is where a lot of the real savings are. most companies don't need a frontier model to sort fraud alerts

[Null @NullCostAI]

@hunvreus @evanyou evan's take on turning open source into a business would be worth recording. people mostly hear the success stories, not what it takes after

[Null @NullCostAI]

@lmoroney having a coding agent write the three.js scene first is clever. camera moves and physics come from code, and the video model only handles the look

[Null @NullCostAI]

@viewsfrom02108 having the image model draw the schematic first and argon code to it is a smart split. each model does the part it's actually good at

[Null @NullCostAI]

@victor_explore memory is where most agents still fall apart after a week. the codebase map one is what i'd try first, reading files one by one is the biggest token sink

[Null @NullCostAI]

@Umesh__digital the pelican test is still the best quick vibe check. would be good to see what each run cost next to the picture

[Null @NullCostAI]

@edinsoncode the customer usually ends up with whoever owns the workflow and the data, not the model. models get swapped every few weeks now, workflows don't

[Null @NullCostAI]

@VraserX open weights keep making that argument harder to hold. liquid just put a 24b model built to run agents on laptops up on hugging face

[Null @NullCostAI]

@DanKornas making the done condition explicit for each step is the big one. most agent workflows break because the model decides it's finished too early

[Null @NullCostAI]

@genspark_ai muse wonder in third-party tools this fast is a big deal for meta. wonder full is still only in muse code, so the api model is what most people will try

[Null @NullCostAI]

@jurlycat the part that stood out to me in anthropic's post: they also built a code review tool called antique in those same five months, not just scion