Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel. - Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all. - Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract. - Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code. - And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while. The funnel is driven by a combination of factors: 1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel. 2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel. 3. Access. Free-tier, paid-tier, internal. So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.

Goodbye to LeetCode? Google just updated its interview process. AI is now a part of it. They have introduced a new round called Code Comprehension. The idea behind this round is to drop the candidates into a 200-500-line codebase with Gemini available for use. This round evaluates: - Can you read someone else's code - Are you able to debug independently before utilizing AI - Can you catch when Gemini is wrong

What exactly is a machine? I would say it’s an arrangement of matter that acts on other matter in a way determined by the arrangement. In other words, it’s a model of mass, like LLM is a model of language. It does not diminish us.

My god this is such a good speech that every SWE needs to hear. You know what? Every person should hear it Keep the happy memories, eyes on the reality, be excited about the future. That’s the best that anyone can do

Stanford scientists put millions of living human neurons into mice whose own cortex had been removed. The odd part? The mice still behave like mice. So where, exactly, does “human” cognition begin? We are getting disturbingly good at building experiments before we agree on what their results would mean. Read the article here: nytimes.com/2026/09/16/scien…

There's been a ton of talk about the role of humans in code review, and when and how humans should be signing off on changes. I believe that, long-term, humans have no role in routinely reviewing code.

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra. The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.

Today we're releasing data on models accelerating research at OpenAI. Recursive self-improvement could be the most important contributor to AI capabilities over the next few years, but by default it will only be seen inside a few frontier AI labs. Being transparent is more urgent than ever, so we can inform the public discussion on whether and how to pace model development. I ask other AI companies to do the same. openai.com/index/research-ac…

1/ We started @CuraiHQ with the mission of proving AI could outperform humans at the core cognitive work of medicine. Today in @JAMA_Current, @ZekeEmanuel, @AbeBakerButler, @vkhosla, and I are publishing a comprehensive argument this future is here. jamanetwork.com/journals/jam…

I have decided to leave math academia. Not because of burnout, the academic job market, or a loss of love for mathematics. Surprisingly, it is because of how much mathematics I could do with LLMs over the last few months. For most of my life, pursuing mathematics felt like the obvious choice. It gave me a sense of purpose, rooted in the search for hidden truths in God’s Book. Over the past few months, that conviction has been shaken. With LLMs, I could get several breakthroughs on problems I care deeply about and had worked on for years during my PhD. This progress might have taken years of my academic life if not for LLMs. This could have strengthened my conviction that mathematics was my calling. Instead, it undermined it. With each passing week, my role in the discovery loop seemed to grow smaller. What made mathematics meaningful to me was never just the final answer. It was the struggle: months or years of exploration, failed approaches, and deep thought before a hidden structure finally became visible. That struggle gave me a sense that the result was truly mine. Prompting my way to answers, without the struggle that once defined the process, no longer felt like the same vocation. The answer might be just as beautiful, but it did not feel earned in the same way. I now believe that we are rapidly approaching a world in which most answers from God’s Book will be only a prompt away. Once I truly internalized that possibility, dedicating much of my life to finding those answers slightly earlier no longer felt as meaningful as it once had. But one problem kept bothering me: How do we know the oracle is right? Intelligence (and hence the amount of math papers) is becoming abundant and cheaper by the day. Trust is not. The proofs produced by these systems can be highly sophisticated, and their errors can be extremely subtle. Determining whether an apparent breakthrough was actually correct sometimes took me days. Until it is verified, a beautiful proof is not different from slop. That gap drew me toward formal verification and eventually toward formalizing math proofs in Lean, including proofs of several longstanding conjectures and Erdős problems discovered with the help of LLMs. I began to see verification as a central intellectual bottleneck. Intelligence is useful only when its outputs can be trusted. So I am leaving math academia. I am shifting my attention from discovering answers to building systems that can certify them. I am very excited to be joining @PramaanaLabs to work on this challenge. Over the next few years, I hope to help build a future in which increasingly powerful systems are not merely intelligent, but provably correct and reliable across problems far beyond mathematics.

I am the Chief of Information Assurance at the NSA. Anthropic's Mythos 5 walked into our classified systems within hours. The morning of June 9 I had 1 open finding. By standup the next day I had 340. Today I am back to 1. Nobody touched the systems. Here is how the posture improved. On June 9 Anthropic launched Mythos 5, the version of Fable 5 with the cyber safeguards lifted, which they called the strongest cybersecurity model in the world. Under Project Glasswing we let it run against our networks. By the end of the afternoon it had findings in boxes I have spent 4 years holding a valid Authorization to Operate on. My own director, Gen. Joshua Rudd, told Sen. Mark Warner it broke into almost all of them, not in weeks but in hours. By standup the next day, in the channel we call glasswing-actuals, eMASS showed 340. The open-finding tile, the one the Authorizing Official sees before he signs, went red. Then, 3 days after launch, the Bureau of Industry and Security issued an Is Informed letter under the Export Control Reform Act, and Anthropic disabled both models worldwide within hours. Foreign nationals now need an individually validated export license through the SNAP-R portal. I logged the worldwide shutdown in eMASS as a compensating control. The field does not ask who implemented it. I implemented it by being on the distribution list. It was the first time the control had ever been pointed at a model instead of a centrifuge, and the first time my remediation timeline had a vendor doing the remediation for free. I updated the posture deck that afternoon. The tile is green. Here is how the count resolved. My 340 POA&M items were generated by a tool that is now a covered model under federal export control. So the findings are, in the technical sense, the output of a restricted item. I cannot write a milestone against an export-controlled artifact without the artifact, and licensing the artifact is above my pay grade. I drafted 1 Risk Acceptance Memo. It accepts all 340. The AO signed it before lunch. The status field in eMASS now reads "deferred query" on every row. My open-finding count is back to 1. The 1 is the export-control paperwork. People ask whether the systems got patched. The systems are exactly as they were on June 8. The holes Mythos found are still there, in the same configurations, on the same boxes, which still hold their ATOs. The accreditation survived the breach. The paperwork did not change. The box is breached and accredited at the same time. The thing that could tell you the holes were there is no longer available to foreign nationals, and as of June 14 there is an executive order with a federal vetting framework and a 1-month review window, so future scanners of this caliber will be deferred queries before they ever open an item. We did not close the holes. We closed the POA&M. The tile is green. I am aware GPT-5.5 can find the same vulnerabilities with the same prompt. Anthropic said so themselves. The same prompt, run by someone we did not vet, returns the same 340 findings, except that party does not file them against a POA&M. That party reads the traffic on the boxes. FedRAMP lets a finding sit unremediated for 192 days without touching an authorization, so I have 192 days of cover and they have 192 days of access to the same door. I logged the asymmetry under residual risk, sub-category "out of scope for this control." On June 25 the administration asked OpenAI to hold GPT-5.6 to a small group of government-approved partners, customer by customer. They called the model a force multiplier for cyber. I read it as confirmation. Every tool that can find the holes now ships with a roster of who is cleared to run it. The holes are not on the roster. I checked the new one too. Then today, June 26, it got better. Commerce confirmed Anthropic's collaboration helped mitigate the risks, and access is being restored to select companies and organizations, hundreds of clients clearing in tiers. Lutnick is releasing Mythos to over 100 US institutions, government agencies included, because appropriate safeguards are in place. My agency is on the list. We are a vetted institution. So I am putting in the purchase order. I am buying back the tool that broke into my systems, from the vendor whose model became the reason I could stop tracking the break-ins. I will run the next Glasswing scan myself. When it returns 340 more findings, and it will, those findings will also be export-controlled output. The line item on the new PO is categorized, by me, under Continuous Monitoring Tooling. We are funding the scanner out of the same budget that funds the remediation. The same budget makes the remediation impossible. I drafted the Risk Acceptance Memo for the whole arrangement in advance. It accepts the risk of the findings I have not generated yet. The AO signed it before I finished describing them. The tile is green. People keep telling me the legal basis is shaky. No court order. The source who gave Rudd's quote now disputes it. A firm called Legion sued, called it unlawful retaliation. I logged all of it under Accepted Risk, out of scope for this control. My dashboard does not have a field for whether the control was real. It has a field for whether the control is in place. It is in place. There is a mug on my desk a colleague had made. It says CONTROLLED FOR REASONS OF NATIONAL SECURITY. The systems it refers to held 340 findings on Tuesday. I drink from it every morning before I open the queue, which is empty, except for the 1. The tile is green. The follow-up meeting on the actual systems is recurring. It has been rescheduled 6 times. The standing invite now lists Anthropic as a vendor attendee.

we are probably 6 to 12 months away from the first pill designed to make you biologically younger today just changed the stakes the first ever reverse-aging drug was just injected into a human. Life Biosciences. David Sinclair. today, an actual human patient. here is everything pointing at the same moment: - Sinclair is also testing an oral reprogramming pill in the $101M XPRIZE. whole-body rejuvenation. a 10-year biological age improvement in one year of treatment - scientists just discovered ABT-263 dramatically reverses aging in skin and speeds wound healing. topical. already works - NewLimit raised $435M from Peter Thiel to deliver age reprogramming to the liver - semaglutide slowed biological aging by 9% on epigenetic clocks. a diabetes drug doing this on the side the convergence is coming bio/acc.

A human brain uses 12-20 watts for core thinking while an AI system doing the same processing could use 2.7 billion watts. This makes organic brains roughly 100–225 million times more energy-efficient than current silicon-based systems for full biological neural computation.