You know that feeling when you open your feed and every other headline sounds like it was written by someone who just watched a sci-fi movie? That was this week. OpenAI launched GPT-6 Astra, then casually dropped a claim about making progress on a Millennium Prize Problem — you know, the million-dollar math questions that have been sitting unsolved for decades. And somewhere in the middle of all that, Meta shipped an agent that will happily shop and email for you, in a private VM so it can't peek at your life.
Let's slow this down, because there's a lot to unpack and a fair bit of it deserves some side-eye.
The 88-hour math machine
The Navier-Stokes story is the one I keep chewing on. OpenAI published a formal write-up saying an internal system produced an analytical proof that 3D incompressible fluid dynamics can develop a finite-time singularity. The effort ran with something like 10,000 concurrent agents for roughly 88 hours, and then GPT-6 Astra spent about 17 more hours turning the argument into a Lean formalization — meaning a computer checked every logical step.
Here's the honest catch: the model that actually produced the proof was an internal system that OpenAI says was significantly more capable than the Astra they're selling you. So no, you can't go reproduce this in ChatGPT this afternoon. And the company also acknowledges concurrent related work from Tristan Buckmaster at NYU and Levent Alpöge at Anthropic, so the "we got there first" part is genuinely messy. Math priority disputes are ugly enough when it's two humans; throw in a lab-scale swarm of agents and a rival lab's model and it gets very tangled.
Still, the workflow detail is the part that matters more than the prize. A research pipeline where one very large internal model generates the argument and a shipping model formalizes it — that division of labor is new, and it's going to show up in a lot of labs before long.
Claude did Fermat's Last Theorem in 11 days
Right on cue, Anthropic-adjacent work made the rounds: an AI formalized and verified Fermat's Last Theorem in 11 days. The careful phrasing matters here — it didn't rediscover the proof, it made the existing one machine-checkable. Which is exactly the kind of task that used to take human mathematicians years of grinding.
I have a soft spot for this because I spent a miserable semester in a formal verification course. The idea that a model can now chew through a proof that size and have Lean certify it... honestly, it makes me feel both impressed and slightly obsolete. It's also a reminder that "AI does math" is not the same as "AI discovers math" — the discovery part is still the hard, slow, human-flavored thing. For now.
Meta's Muse wants your inbox
Meta unveiled Muse, its big agentic push. It drafts and sends emails, hunts for deals, plans purchases — and it runs inside a personalized virtual machine with encrypted connections back to your devices. A separate agent called Sentinel is supposed to make sure Muse doesn't do anything you haven't approved. Later this year there's a "Confidential VM" variant that even Meta claims it can't see into.
To be fair, the privacy scaffolding is more thoughtful than most agent launches. But here's my skepticism: every big lab is shipping the same "agent that does tasks for you" pitch, and the hard part was never the demo — it's what happens when the agent hits an ambiguous situation mid-task, or a website changes its layout, or your inbox contains something you'd rather it not "learn" from. I'll believe the VM story when I see an independent audit, not a blog post. Signups are open in the US for iOS and Android, so at least we'll get real-world reports fast.
The local AI box that's actually a workstation
Minisforum showed off the MS-S1 Max-P495 at IFA: AMD's Ryzen AI Max+ Pro 495 with a whopping 192GB of unified memory in a mini PC. That's a genuinely usable local LLM rig for people who want to run bigger models without renting GPUs or handing their prompts to a cloud.
Don't get me wrong — 192GB is a lot of memory and the price of that configuration will make your eyes water. And unified memory bandwidth is not the same as a data-center GPU stack; you'll run 70B-class models at conversational speed, not training anything serious. But for a certain kind of tinkerer, the appeal is obvious: private, always-on, your own weights, no per-token meter running in the background.
Along the same silicon vein, SEMIFIVE started mass production of HyperAccel's "Bertha" LLM inference accelerator on Samsung's 4nm node — a big-die chip over 500mm² aimed at data-center inference. The interesting bit isn't the chip itself, it's that custom ASICs for inference are moving from "announced" to "in a fab, in volume." Inference is where the real cost lives once models mature, and that's exactly where the industry is now pointing its money.
The smart people are nervous
OpenAI's chief scientist wrote an essay making the rounds: essentially, nobody is ready for how much smarter these systems are getting, and the pace is outrunning our institutions' ability to absorb it. He notably stopped short of calling for a slowdown, arguing instead for steering automated research toward safety.
It's an odd thing to read the same week a lab claims a Millennium Prize Problem. The stock market felt it too — S&P 500 software names sold off on AI worries. And in a completely different corner, McKinsey reports farmers are adopting AI faster than any other tech they track, mostly for day-to-day decisions. Which is probably the most quietly useful AI story of the week: no proofs, no agents, just people using a tool to make better calls about crops and weather.
My own take after this week: the distance between "model that can check a proof" and "system you actually trust with your workflow" is still enormous. The demos are getting jaw-dropping, but the boring parts — pricing transparency, auditability, failure modes when the agent does something dumb — are moving way slower. Keep an eye on the math papers and the VM audits, and maybe hold off on letting any agent loose on your primary inbox just yet.
Speaking of reference material, if you're in the middle of comparing specs for a local AI box or just want to double-check what a chip actually supports, I've been leaning on Engineering Reference to keep my notes straight. Handy for exactly this kind of week.
Top comments (0)