Executive Summary
The AI landscape saw a massive shakeup over the last 24 hours, headlined by OpenAI’s announcement of an analytical proof for the Navier-Stokes Millennium Prize Problem, produced by an internal model group. The newly spotlighted GPT-6 "Astra" model demonstrated unprecedented agentic capabilities, autonomously completing video games like Portal and generating complex 3D environments in Blender and Three.js. Meanwhile, the developer tooling ecosystem rapidly adapted to these advanced agents, with new local LLM frameworks, multi-agent orchestration templates, and streamlined CLI tools hitting the market. Hardware and local compute setups also saw significant discussion, as developers optimize for on-prem AI inference and training.
Key Events
- OpenAI announces an analytical proof and Lean formalization for the Navier-Stokes Millennium Prize Problem, with Sam Altman detailing the behind-the-scenes coordination and rivalry with Anthropic. → link
- GPT-6 Astra autonomously beats the game Portal in 24 hours for only $571 in token costs, showcasing deep computer-use capabilities. → link
- Astra demonstrates advanced 3D modeling and game development, building a playable Catan game in three.js and generating Blender assets directly into Roblox. → link
- HuggingChat releases "ML Intern", a feature that connects to the HuggingFace ecosystem to autonomously run evals, fine-tune datasets, and deploy Spaces. → link
- Amp CLI removes message queueing for agents to allow faster, more natural parallel communication, while expanding memory capacity to 44GB for its "Orbs" compute instances. → link
- Tinygrad suggests GitHub should charge $20 per account creation to stem the tide of AI-generated spam Pull Requests. → link
Analysis
The last 24 hours highlight a massive leap in both AI reasoning and practical agentic execution. The Navier-Stokes proof validates AI's utility in deep theoretical mathematics, while the "Astra" model's ability to autonomously play, build, and navigate complex 3D environments signals that computer-use agents are crossing the threshold from experimental to genuinely productive. We are seeing a parallel surge in developer tooling designed to manage these hyper-capable agents, such as multi-agent "chief of staff" templates and local inference optimizations. Moving forward, watch for increased friction in open-source contribution pipelines due to AI spam, and expect rapid iteration in orchestration tools as developers figure out how to chain these new agent capabilities together.
Tweet Feed
AI Research & Breakthroughs
@sama · 2026-09-08T18:02
I spent much of the weekend talking with the team who did this work. Seb--and everyone else--acted with integrity and generosity throughout.
Initially we believed the other team had also solved the problem. We wanted to collaborate and do a joint release.
When we learned that they had Euler but not Navier-Stokes, we offered to let them go first, to suggest that they should be the ones to get the prize, and optionally for Tristan to be the lead author on a rewrite of the OpenAI proof. We felt it was challenging to offer the same to Levent (an Anthropic employee), who was not willing to talk or coordinate with us anyway. We were open to other solutions.
We would have greatly preferred coordination. We did not rush to publish even though the other team wasn't communicating with us. The team threatened us with unfounded accusations of plagarism.
Now that we can see their work, the approaches appear to be different. It is also worth noting that our latest model can solve many, many other math problems.
It is true that we tried this because there were rumors on the internet last week that Anthropic's models had solved a millennium problem and we were curious if ours could do it too. → tweet link
@gdb · 2026-09-08T17:30
Announcing a solution to the Navier-Stokes problem, one of the seven Millenium Prize problem.
Represents a major milestone for both AI and mathematics, as well as pointing to an upcoming renaissance in scientific discovery.
Hoping that the knowledge created by this model can help the community better understand non-linear PDEs, which are used to model much of the physical world. → tweet link
@sama · 2026-09-08T17:46
RT @OpenAI: This model represents a step-function improvement on many benchmarks, and its training is ongoing.
Our internal model group ar… → tweet link
@jezell · 2026-09-08T17:43
RT @OpenAI: The group produced an analytical proof and Lean formalization that via Navier-Stokes dynamics a fluid can develop a singularity… → tweet link
@jezell · 2026-09-08T06:24
RT @imjustnewatai: OpenAI’s internal model ‘Bel’ has solved navier-stokes.
tristan buckmaster says he was told it produced a roughly 100-p… → tweet link
@LinusEkenstam · 2026-09-08T11:33
The future will have it all, and this is highly disturbing.
This is for educational purposes only. I do not endorse the creation of these deepfakes, but we should all have a conversation about this.
I see no clear pathway of this getting stopped.
Not by laws, or by filters, content blocks etc, because the tools to create these are so simple to download onto your own computer, that you would have to remove peoples computers.
Secondly: you could automate the generations of these clips (clip length is not an issue) and you could spit out and distribute thousands or tens of thousands of clips everywhere. Causing real harm. Because there is no "forbidden" words or actions in these videos. → tweet link
GPT-6 Astra Demonstrations
@jezell · 2026-09-08T06:14
RT @tomshardware: OpenAI’s GPT-6 Astra model autonomously completes Portal in 24 hours — feat cost just $571 in tokens https://t.co/m9SEkyN… → tweet link
@MengTo · 2026-09-08T11:49
I asked Astra to create a Catan game in three.js and it came out incredibly well.
I played it, and the animations make it so much fun. The instructions and info cards make the game easy to learn. It even has AI players and trading.
Astra used another game I had started building as a reference. The more games I build, the more skills and assets I have for the next one. It can remix the UI, 3D models, pages, and sounds instead of rebuilding everything.
Now I want to see how many other board games I can make from this foundation. → tweet link
@KingBootoshi · 2026-09-08T04:02
FULL ASTRA 6 BLENDER CREATION + ANIMATION + IMPORT INTO ROBLOX + SCRIPTING ONE SHOT
GG
I prompted "design a slime monster in blender then import it into our roblox game and make it an actual interactable enemy" https://t.co/HfIY5e5QqV → tweet link
@MilksandMatcha · 2026-09-07T19:07
6.837 (Computer Graphics) was the hardest class I took at MIT.
There is a ridiculous amount of code and math around Bézier patches, surface normals, etc. just to make a teapot look like ateapot.
We had several PSETs around the Utah teapot, which in college always took me until the last hour to figure out.
Astra made this interactive version in 1 min 43 seconds.
One must wonder how my college years would have been different if Astra had existed. Would I have graduated smarter or dumber... → tweet link
@jxnlco · 2026-09-08T18:34
RT @kylejeong: GPT-6 Astra broke the internet with its ability to use blender and generate 3D worlds, all a byproduct of it's new computer… → tweet link
@jxnlco · 2026-09-07T21:06
RT @Apocriton: GPT-6 Astraが独力でRimWorldをクリアしました！与えた指示は、「全て自分で判断し、ゲームをクリアせよ」だけ。 GPT-6はネット検索やComputer Use、外部記憶ファイルの作成・参照などを駆使し、連続襲撃に勝ち、全員を生きて惑… → tweet link
@jezell · 2026-09-07T19:21
RT @Kalshi: JUST IN: GPT-6 Astra successfully completes all 48 levels of the “I’m Not a Robot” challenge https://t.co/2Tn7VfKAAR → tweet link
@gdb · 2026-09-07T20:11
Astra for spectrogram -> sound identification → tweet link
@gdb · 2026-09-07T19:22
Astra for visualizing how a drug works: → tweet link
@badlogicgames · 2026-09-08T11:12
astra, when you give it a subagent tool.
wtf is this model doing. https://t.co/NnQshqP9EB → tweet link
@jezell · 2026-09-07T20:06
RT @onihachi: GPT-6 AstraでのBlenderモデリングが、劇流行りしてますが、そもそもmegascansのアセットだけを使って森を作ってもらえば、すぐフォトリアルになるんじゃないかと思ってやってみました。結構良い結果になりました。参照で渡したのは添付の画… → tweet link
@steipete · 2026-09-07T22:19
RT @gregpr07: Astra mogs Fable 5.1 on browser use 😮
77% Astra 56% Fable 5.1 Fable just refuses too much https://t.co/hsEK0zcBSm → tweet link
Developer Tools & Agents
@sqs · 2026-09-08T09:10
Queueing your messages to the agent made sense for easily distractible models.
Now? Amp sends your message to the agent without waiting.
It's way faster and more natural. Plus, more information makes the agent do better.
https://t.co/vRt8Lerp2s https://t.co/y0hfFTRJXg → tweet link
@sqs · 2026-09-07T23:33
Orbs, now with up to 44GB memory.
a1.3xlarge at https://t.co/lakQcGkPq4. Gigawatt and Enterprise only for now, due to capacity constraints. https://t.co/DxJhdIQ8dB → tweet link
@kunchenguid · 2026-09-07T21:36
sharing my Grok @Bot setup as a template you can one-click install
you might have already heard of the "chief of staff" pattern that's being widely adopted
firstmate is THE chief of staff - it's probably the project that pushed the furthest on this pattern
it spins up other bots to get work done and actively improves their memory and instructions so they do better and better over time
it sets up a local sqlite db to manage tasks, delegation, results, and decisions, so persistent records are never lost
it uses cursor cloud agents for coding tasks, and grok bot for non-coding, to balance the resource and quota usage and avoid scalability bottlenecks
the most common feedback from people who have adopted firstmate is "there's no way to go back" :) → tweet link
@victormustar · 2026-09-08T17:35
cool new HuggingChat feature: "ML Intern"
basically it connects HuggingChat to the complete HF ecosystem 🥳
you toggle it, ask for an eval, a dataset, a finetune, a paper repro, whatever, and it goes and does it: research, training on hf jobs if needed, can ship Spaces too :) → tweet link
@ivanfioravanti · 2026-09-08T15:19
DwarfStar Qwen 3.8 Flash Next update focused on 64GB usage and more: - Tools calling fixed in some specific use cases (@anemll you were right during your initial tests) - Prefill/Decode performance improvement (chart below from M3 Ultra) - Memory usage: PLE was always prefaulted, while on 64GB machine is now demand-paged greatly reducing memory pressure. Now 128K context is feasible 💪 Here DS4_QWEN4_PLE_EVICT_TOKENS=1024 can be used to keep ~480 MiB only in memory to have even larger contexts, but there is a performance hit, because every 1024 tokens generated, sidecar's pages are dropped.
I've just merged another PR that should give an additional boost in performance. Testing now, tomorrow we'll see 🤞 → tweet link
@ivanfioravanti · 2026-09-08T13:46
deepseek-v4.1-flash I'm really curious to understand the architecture behind the scenes. 🤔 Model is really fast as you can see in the video! I've got an average decoding speed of around 350 t/s.
BTW it is clearly experimental and: - currently, the model is under heavy load and failing quite often. - it thinks like there is not tomorrow - it keeps failing editing file - after 4M tokens my frogger game is still in progress - it can't see images out of the box on DeepSeek Harness, you have to add these config lines → tweet link
@thdxr · 2026-09-08T17:52
so that experimental rift thing we built which uses copy on write to save space ended up being used by a lot of people
now can integrate that directly → tweet link
@victormustar · 2026-09-08T12:38
RT @adaption_ai: Bringing AutoScientist to 15+ million @huggingface researchers and builders. 🤗
AutoScientist automates model training. Sp… → tweet link
@steipete · 2026-09-07T22:35
Did a PR to one of our upstream projects and they requested some minor changes.
What’s even the point with this workflow? You already wrote the prompt, why make me ping my agent again so your agent then merges? → tweet link
@jezell · 2026-09-08T16:10
RT @derrickcchoi: Great blog post by @Dimillian about how he used Astra to build a house in Blender, then to Unreal Engine 5.
He reviewed… → tweet link
Hardware & Local Compute
@MilksandMatcha · 2026-09-08T17:36
To understand why Cerebras built a very, very big chip, it really comes down to one number: the energy it takes to move a bit of data.
Off-chip: ~1–10 pJ/bit On-chip: ~0.1 pJ/bit
Moving data on-chip can take up to ~100× less energy than moving it off-chip. At AI scale, where you’re constantly moving trillions of bits, that difference becomes enormous.
Cerebras is not big for size's sake. By keeping compute and memory close together, you get minimal data movement. → tweet link
@sudoingX · 2026-09-08T11:05
day 2 in the new place and the floor looks like a hardware crime scene.
the old 6k thb studio room was full and bad ac. this one has a proper power grid and 2 ac cold enough to keep silicon happy at 2am. that changes the math. i can add machines now without the walls sweating.
by moving to this place i am buying the ceiling on what i can run at once. every box on this floor is a parallel experiment waiting for a shelf.
tonight the desk goes up. tomorrow the lab starts producing. i can already feel the angles this room is going to give me.
from a trolley bag to this in 6 years. the boxes just keep getting heavier. → tweet link
@sudoingX · 2026-09-08T04:19
building on windows in 2026 is doing the sport on hard mode for no reason. just come to linux anon. → tweet link
@jezell · 2026-09-07T19:47
DGX set up last week. Asus NUC with 128 GB RAM came yesterday. Mac studio on order. That will add Intel, ARM, and Mac to the local build cluster. Now just need to build a little device farm on top. → tweet link
@FrameworkPuter · 2026-09-08T17:04
We have a positive update today on LPCAMM2 pricing! We were able to source a limited quantity at lower cost, and are retroactively reducing prices for orders that shipped at higher price, as well as offering the lower price to orders through Batch 10. → tweet link
@ivanfioravanti · 2026-09-08T10:57
M2 Ultra 192GB is still a great machine!!! → tweet link
@TheAhmadOsman · 2026-09-08T04:58
At some point you gotta ask yourself if these kinds of clusters are the most economically efficient setup for the quality and quantity of tokens produced
I personally wouldn’t spend my money on this, but to each their own → tweet link
Software Engineering & Open Source
@tinygrad · 2026-09-08T16:25
GitHub should charge $20 to create an account. This is no barrier for any real software engineer, and it would prevent all the AI spam PRs from the brand new accounts. To anyone doing this, stop wasting money on tokens, it takes 10 seconds for me to close the PR and ban you. → tweet link
@MatejKnopp · 2026-09-08T11:32
Please somebody explain this shit to me. I manually quit every single application. I specifically left this checkbox unchecked. And yet after restart macOS will open every fucking application under the sun. What the fuck are they doing in Cupertino. This is so infuriating. https://t.co/9MnA536gNR → tweet link
@jezell · 2026-09-08T17:14
Can confirm EROFS is great. https://t.co/Iwj6PafA9t → tweet link
@jezell · 2026-09-07T19:00
Great software requires iteration with users. LLMs make the iteration faster, but they are not the users. → tweet link
@levelsio · 2026-09-08T18:25
Thanks everyone for recommending me to buy an LG TV two years ago but especially everyone for recommending me to NEVER ever connect it to the internet
But instead to get an Apple TV and connect that with HDMI to the LG TV
I knew LG TVs were full of spyware but it's even worse than everyone thought!
You're 100% safe though if you never connect it to the internet → tweet link