Editor’s Note: Traditionally the Agentics newsletter has focused on longform content about how to think about agents, from suggestions on how to implement workflows to case studies to thoughts on AI config.
We also get a lot of requests for takes on various hype-y things, and most of those don’t warrant a full response. Still, we often want to address what’s going on and provide a few lines of opinion where relevant, so we are starting a hopefully-weekly-but-probably-every-other-week series of posts that is solely focused on things happening in the AI world.
Agentics NYC August Flagship: Slides and Recordings
Thank you everyone who came out to the August Agentics NYC Flagship event! We had 1400 people sign up and over 300 people pack into a Times Square nightclub to talk about AI agents. Recording links:
Inference Designed for Long Horizon Agents — Jack O’Brien, Subconscious
Unbundling the Goose SDK — Jack Amadeo, Block
Modal Training Gym SDK — Joy Liu, Modal
Software Factories and How to Build Them — Amol Kapoor, Nori Agentic
One AI Harness: A Workspace for Everyone — Danny Avila, Librechat
Reducing Our On-Call Burden with Cloud Agents — Jeff Silver, Opus
The New Software Crisis — Rohan Gupta, Telos
More pictures, slides, and recordings of the event can be found at the Agentics NYC event page. Subscribe on Youtube to get all of the Agentics recordings as soon as they go up, and keep an eye on the luma for more events!
AI Takes
Navier-Stokes controversy. This is definitely going in a longer post, but I have to mention it here. The Navier-Stokes equations model fluid dynamics. One important open question in the math world is whether these equations fall apart at some extremes. This is a very important question because we use these equations to do things like design planes and map how blood vessels work. So important that this was selected as one of 7 millennium problems at the turn of a century. It even has a $1m prize pool attached.
A few days ago, OpenAI announced that it had used an internal model and literal millions of dollars of compute to solve Navier-Stokes.1 This was immediately mired in controversy. Turns out another two mathematicians — Tristan Buckmaster and Levant Alpoge — were working on this problem for the last year. They also happened to use ChatGPT quite a bit in the process. These mathematicians were very close to their own solution. Rumor got out that they were close, it reached OpenAI, and OpenAI then burned all of those tokens over the course of two weeks to try and scoop the result.
This raises a question: did OpenAI look at Tristan’s previous work?
Tristan’s write up is not flattering for OpenAI:
Levent had been told by Sebastien “very little human input” had been used. This turned out not to be true. Over the course of the call, as members of their team sent Sebastien corrections and details over their internal chat, it emerged that an entire team had been working on the problem…
I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project. I was told the model did not look up user data. I asked again, about training, and I did not get an answer….
Sebastien twice asserted that he wanted Levent removed from authorship, and said it would all be simple if only it were not the case that, and it was so annoying that, Levent works at Anthropic…
I said that if OpenAI released its result in the way proposed I would go public with what happened. The reply was, “Why would you ruin your career?” I replied that I am an academic, and asked why he thought going public would ruin my career. The reply was, “If you don’t want me to be nice, then I don’t have to be nice.”
But OpenAI’s write up isn’t that flattering for OpenAI either!
After the completion of our full project and Lean verification (on September 6th), believing from the rumor they also had a solution of Navier–Stokes, we reached out to them to offer a concurrent release of our result and to recognize their priority in a joint announcement.
…
While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.
They probably did not directly feed Tristan’s and Levant’s sessions into their model. But of course they trained on that session data! Of course they did! Why would they not? Frankly, it’s crazy that OpenAI is even hedging with ‘while unlikely’. Data is such a scarce resource, everyone is training on every bit of data they can get, every company that sells data to these companies gets literal billions of dollars over night. We’re supposed to believe that OpenAI decided that this particular session from two extremely talented mathematicians was somehow pruned from the training data, at the exact time that OpenAI was trying to make their models better at math?
Bluntly, unless you have a very explicit no-train contract with the frontier labs, you should assume that all of your data is going right into the model training. Which means that anything that you feed into the model now will potentially be available in a few months to your competitors, just by asking.
I have to imagine that this makes cloud inference even less appealing for large enterprises, especially in finance. JaneStreet is publicly buying billions in datacenter capacity, that’s not for Anthropic models!
While we are on the subject of Millennium problems and OpenAI, rumors are swirling that they ALSO solved the Hodge conjecture, another of the Millennium problems. Big if true.
Related: Anthropic released a formalization of Fermat’s Last Theorem. Less drama here.
GPT6 Astra. Speaking of OpenAI, they released a model recently. It’s very good, and has replaced Claude Opus 4.8 as my daily driver.
Astra is state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work. Astra saturates FrontierMath Tier 4 with a 98% score, having already helped solve long-standing open problems in mathematics. Astra also saturates ARC-AGI-3 with a 99.9% score and ExploitBench with a 100% score. It also sets a new frontier on computer and browser use, handling the most demanding professional work with unmatched speed, accuracy, and judgment.
This thing is very good at UI/UX and design. Some of the demos floating around the web are incredible. That said, the “100% on exploit bench” bit is like the part of the horror movie where they have the news person foreshadowing the zombie outbreak on the TV.
Interestingly, I find Astra quite a bit cheaper than Claude Fable/Mythos, and much more capable generally. I still don’t really trust it for more targeted work — it is too ambitious and will change too many things — but I like it a lot for larger refactors or product features that have pretty solid detailed specs.
Also, a request: so far, most of the people using Astra are AI people. I would love it if some game designers picked up Astra to, well, make games! I think there’s a ton of potential here, but most of the demos I’ve seen thus far are uninspired.
Gemini models. The Google Gemini team has been on a tear with releases:
Building on the momentum of 3.7 Flash from three weeks ago and marking our third Flash release in only six weeks, today we’re introducing Gemini 3.8, our best reasoning and coding model yet, at the same speed and low cost of 3.7. Gemini 3.8 introduces 2 variants:
Gemini 3.8 Flash: our most intelligent workhorse model, delivering significant improvements from 3.7 Flash across software engineering, agentic tasks, and critical, multi-step reasoning in specialized domains. It is available at the same introductory price as 3.7 Flash at $0.75 per million input tokens and $3.75 per million output tokens.
Gemini 3.8 Flash Cyber: our most capable cybersecurity model with frontier-level performance in vulnerability detection and automated patching, available to trusted defenders through our new Fairwind Program.
That said, these are still flash models. We haven’t seen a new Gemini Pro model in some time now. I think the current running theory is that something got botched and they have to do another pre-training run. I also kinda suspect that Google has decided to shift focus from using compute to train models to using compute to…sell to other people trying to train models.2
I really have a soft spot for the Gemini models though. I don’t use them very much, but thus far my favorite model generation was when Gemini 2.5 was on top.
Somewhat related: Google has put a lot of effort into their Omni world models. I was pretty bullish on those efforts, because it seemed to play to Google’s unique strengths: they have way more video data than everyone else, and way more compute to run that data. But GPT Astra seems to have a very compelling world model (in the 3D physics sense). Not sure what to make of that.
Meta muse. Meta released muse.ai, a personal AI assistant to compete with the likes of GrokBot, Town, Instinct, vellum, Orchid, etc.
First, I just really don’t know how to feel about the personal AI assistant space. I think it…kinda misunderstands how consumers work? Like I wrote in my Silksong review, the average consumer doesn’t really care all that much about time efficiency. Companies care about time efficiency, because…well, they’re paying you. I think consumers actually like wasting time, which is why the gaming industry exists.
Second, have you noticed how the moment any startup has a decent implementation of a product UX, there are like 20 other implementations of the same concept? muse.ai has “A dedicated, secure computer with its own browser,” is “designed around the way you already communicate,” and “connects to your apps and builds its own tools.” Meanwhile GrokBot has “A computer of its own,” that you can “message like a teammate…from your phone or desktop,” and “uses tools you already use.” Wow software really is just infinitely copyable huh? There’s no product differentiation left!
Agentics is the study of how to use and reason about agents. If you are an expert in coding agents, or interested in learning more about agents, join our community slack. Check out our agent learning hub.
If you are trying to set up cloud agents or build a software factory, we can help! We build white-labeled custom cloud agents infrastructure. Check out noriagentic.com for more.
300 billion tokens, to be precise, which is $15m at public pricing.
Which is also why Jeff Dean et al. ended up leaving the company not that long ago.
"Bluntly, unless you have a very explicit no-train contract with the frontier labs..." aren't all commercial tiers with Anthropic no-train? And even individual users can toggle off training? Do you think Anthropic isn't respecting that?