Notes on Leoblog

If You Had Already Won

Sat, 31 Oct 2026 13:30:00 +0000

Ankur Moitra: So much of our society now is geared towards winning and looking good on paper, and I always wonder to what end. You start off in high school and you can start, as soon as you walk in the door, optimizing your APs and your extracurriculars so that you can get to an elite school. And then when you get to that elite school, you might skip classes because, you know, maybe the important thing is to prepare for your interviews so that you can get a job at a frontier lab, and then you can get some nice, you know, restricted stock units and live out your life. And that’s great. You won. You won at every stage, but so what?

Learning Math, and Learning

Sat, 31 Oct 2026 13:00:00 +0000

Alex Kontorovich: I’m constantly talking to AI systems, and I’m learning a lot of math a lot faster than I’d be able to. Because the way math is written, you know, I sometimes give a paper to a grad student to read, and they come back a week later, and I say, “So how far did you get?” “Well, I didn’t really get past the first page, to be honest.” “Oh, really? Tell me.” “Uh, well, actually, I didn’t get past the first paragraph.” I say, “Okay, fine. Why not?” “Well, the first sentence is by the perturbation theory of non-local systems and I don’t know what any of that means, so that’s kinda where I’m stuck.” I’m like, “Okay. You don’t know how to read a math paper. You don’t– You skip the first couple of paragraphs. Go straight to, you know, the Lemma 3.” “And that’s where you can start working and work your way backwards to what the thing is doing.”

Trust and Stakes: Where AI Fits

Sat, 31 Oct 2026 12:45:00 +0000

Alex Kontorovich: But we also don’t wanna outsource so much– nearly so much decision-making to these systems for exactly that reason.

Steven Brunton: Yeah, we have infrastructure and protocols in place already for safety critical and high impact decisions. We have the FAA, we have the American Medical Association, the Bar Association for these things that affect humans. And I think of, you know, how do we incorporate AI into our general processes in the entire world on two axes. We have, you know, how much do we trust the output, so verifiability or trustworthiness, and on the other axis is how high stakes is it if it’s wrong. If AI spits out a wrong proof, nobody goes without power. It’s fine. So it’s very low stakes. And it is actually extremely verifiable, so you know very quickly if it’s right or wrong, and it’s low stakes if it’s wrong. So that’s the perfect use case is this open-ended design space. Designing a better vehicle, you know, a better car design that’s lower drag, more fuel efficient. Low stakes if it gets a wrong design because you can always test it in a wind tunnel and verify if it did or didn’t work.

The Secretary Problem's Lesson: Explore First, Then Decide

Sat, 31 Oct 2026 12:00:10 +0000

Ankur Moitra: So the way that this works is it’s an example of something called online algorithms. So you have a sequence of options coming. One through N. They come in each time step, and your goal is, you know, when someone comes along, you know their ranking relative to everyone else you’ve seen before, but you don’t know their ranking relative to the people who are yet to come.

And your goal is, you know, to make a decision when the person comes along, do you commit or do they move on, in such a way that you have a good chance of finding the absolute best candidate. And so the point is that you can actually get a constant competitive ratio algorithm that what you do, the best strategy is to wait some amount of the time, not make any decisions, don’t accept anyone. And then once that probationary window has passed, you take the first person who’s better than anyone you’ve seen thus far. And that’ll give you a good competitive ratio in the sense that you can prove theorems. Assuming the inputs arrive in a uniformly random order, you’re likely to do pretty well compared to as if you knew what that sequence of inputs was in advance.

Long-Term Thinking

Sat, 31 Oct 2026 11:06:35 +0000

Ankur Moitra: I think also students have so much anxiety about, you know, what publications are gonna look like that one of the hardest parts is, you know, when I’m thinking about a project, I’m really trying to think about it in a sort of multi-year time horizon because I have some meta vision that I wanna realize, like how can we use learning theory techniques in quantum many-body systems? And there are a lot of stepping stones on route to that, but some of the most exciting problems are, you know, many years out. But students are so worried about the incoming AI tidal wave, you know, the future of work, universal basic income. Are any of us gonna have jobs? That they’re really thinking in three- and six-month horizons, and then it’s hard to find problems that we both care about just because- You know, we’re thinking about different worlds.

The Binance Hack Proves Anthropic Right, Once Again

Fri, 30 Oct 2026 16:10:14 +0000

Although, as of October 30, 2026, there is no evidence that the Binance hack had anything to do with AI, I believe it did. I think there is a good chance the hackers used AI, directly or indirectly, in the attack.

As an engineer, I know full well that hackers can use AI to launch large-scale attacks with ease. Hacking is, at its core, finding vulnerabilities in software. In the past that process usually depended heavily on people and was expensive to carry out, and what AI is best at is reading code tirelessly, in many threads at once, line by line, to find vulnerabilities.

Because computer software is so complex, hackers can use all kinds of techniques to get around the visible guardrails, directly or indirectly, so a very strict set of guardrails is necessary. That is why I think it is reasonable for Anthropic to strictly restrict users from hacking, even if the restrictions are sometimes too strict. Of course, until the whole world has been hit by large-scale hacking, they look too strict.

AI is Transforming Intelligence

Thu, 29 Oct 2026 15:01:32 +0000

AI is Transforming Intelligence

We believe that Powerful AI, a highly capable and impactful form of frontier AI characterized by the following key attributes, will become widely available within a few years:

cognitive capabilities matching or exceeding those of top human experts across most professional and scientific disciplines;

the ability to navigate all interfaces available to a human working virtually, including text, audio, video, and direct computer control;

the ability to reason through complex tasks autonomously over extended periods—lasting hours, days, or weeks—and seek clarification and feedback when needed;

the ability to coordinate across multiple instances and collaborate with both humans and other AI systems; and

the ability to interface with the physical world by digitally controlling laboratory equipment, robotic systems, and other machines through digital connections.

Built Around Enterprise Needs

Thu, 29 Oct 2026 14:35:00 +0000

Enterprise Strategy. Our enterprise strategy is built on the belief that frontier AI is most valuable in real-world settings, as deployment across enterprises is where economic impact is greatest and where the improvement loops from customer feedback are richest. Accordingly, we have organized our research priorities, our product roadmap, and our go-to-market and Applied AI teams around enterprise use cases, which range from routine knowledge work to the most complex problems. The customer feedback we gather directs our research focus and makes Claude more capable over time. This broadens where Claude can be deployed and further expands enterprise adoption, which helps ensure that the benefits of our AI reach people and organizations at scale.

In the prospectus: Prospectus summary: Overview (original filing on SEC EDGAR)

Anthropic’s Dynamic Compute Platform

Thu, 29 Oct 2026 14:31:44 +0000

We manage compute as a single, fungible pool from multiple strategic partners, and we dynamically direct available compute between workloads based on where the marginal unit is expected to be most productive at any given time.

In the prospectus: Prospectus summary: Our Approach to Compute (original filing on SEC EDGAR)

In other words, they’ve built a dynamic compute platform: when the consumer side needs more compute, more goes to the consumer side, and whatever frees up goes to research and training.

We have achieved this through our research breakthroughs and our multi-chip compute architecture across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs, enabling us to orchestrate workloads to optimize for price and performance.

In the prospectus: Prospectus summary: Our Approach to Compute (original filing on SEC EDGAR)

The Two Sides of Technology

Thu, 29 Oct 2026 14:22:03 +0000

To build AI is to harness cognition itself, and anything so powerful can be used for ill as well as good.

In the prospectus: The founders’ letter (page image) (original filing on SEC EDGAR)

Transformation and Opportunity

Thu, 29 Oct 2026 14:22:02 +0000

Doing so may well represent the greatest economic opportunity in history.

In the prospectus: The founders’ letter (page image) (original filing on SEC EDGAR)

Expertise Within Everyone’s Reach

Thu, 29 Oct 2026 14:22:01 +0000

AI could do the same for expertise, giving anyone access to the knowledge and skills of a lawyer, an accountant, or an agronomist, improving economic efficiency and accelerating growth across the world.

In the prospectus: The founders’ letter (page image) (original filing on SEC EDGAR)

Related: Intelligence on Demand

The Great Power of General Purpose Methods

Thu, 29 Oct 2026 14:22:00 +0000

The idea is closely related to Richard Sutton’s concept of the bitter lesson, which describes how sheer scale has consistently won out over more specific, complex methods of training AI models. Note that scale is not the only source of improvements in model capabilities—algorithmic efficiencies (“compute multipliers”), which allow us to (for example) achieve higher capabilities with the same amount of compute, have played a major role. But these efficiencies are themselves found using the scaling laws: researchers test new ideas and keep the ones that improve the scaling trend the most.

In the prospectus: The founders’ letter (page image) (original filing on SEC EDGAR)

This is a big lesson. As a field, we still have not thoroughly learned it, as we are continuing to make the same kind of mistakes. To see this, and to effectively resist it, we have to understand the appeal of these mistakes. We have to learn the bitter lesson that building in how we think we think does not work in the long run. The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents, 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning. The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.

Anthropic’s “Closing the Loop”

Thu, 29 Oct 2026 14:21:22 +0000

Each version of a Claude model assists in building the next, helping to compress development timelines and produce compounding improvements in the speed at which we innovate. We refer to this recursive dynamic as “closing the loop,” and we believe it is one of the most significant structural advantages in frontier AI.

In the prospectus: Prospectus summary: Frontier and safety research (original filing on SEC EDGAR)

Intelligence on Demand

Thu, 29 Oct 2026 13:42:16 +0000

For the first time in history, intelligence is something humans can build. The cognitive capability to reason, analyze, synthesize, decide, and create is producible at scale and accessible on demand.

In the prospectus: Prospectus summary: Overview (original filing on SEC EDGAR)

The words “accessible on demand” caught my eye right away, and they immediately made me think of how ordinary workers are hired by the month, not on demand. In some situations that is a huge advantage: a legal adviser handling a case, an accountant doing quarterly reports, a translator used as needed, and so on. Many tasks don’t need continuous employment; they only need someone to step in for a while, do the whole job, and then be done.

“Interleaved Thinking” in Anthropic’s Prospectus

Thu, 29 Oct 2026 12:53:53 +0000

“interleaved thinking” refers to a capability whereby an AI model interposes discrete reasoning steps between successive tool calls within a single response, rather than confining its reasoning to a single planning phase at the outset.

In the prospectus: Selected definitions: “interleaved thinking” (original filing on SEC EDGAR)

So that’s what it means. The phrase “rather than confining its reasoning to a single planning phase at the outset” reminded me of something I wrote in an earlier post: that OpenCode couldn’t reason while its subagents were running. I think the two are related, even though they’re not talking about the same thing.

One is whether the harness can collect the subagents’ problems while they run, hand them to the model to reason over, and produce new steps
The other is whether the model is able to take in scattered information and make new judgments

What Is the “Externality” Mentioned in Anthropic’s Prospectus?

Thu, 29 Oct 2026 09:49:07 +0000

In economics, an externality (also called an external effect or spillover effect) is an indirect cost or benefit to an uninvolved third party that arises as an effect of another party’s (or parties’) activity. Externalities can be thought of as unpriced goods involved in either consumer or producer market transactions. Air pollution from motor vehicles is one example: the cost air pollution imposes on society is not paid by the producers or users of the vehicles to the rest of society. Water pollution from factories is another: all consumers are made worse off by the pollution, but the market provides no compensation for the damage. A positive externality is when an individual’s consumption in a market increases the well-being of others, but the individual does not charge the third party for the benefit. The third party is essentially getting a free product. An example might be an apartment above a bakery that gets the smell of fresh pastries every morning. The people living in the apartment do not compensate the bakery for this benefit.

Don't Limit Readers

Tue, 27 Oct 2026 13:22:01 +0000

Looking back at my own habits: with a long document, I always preview it online and then download it to read locally with my own tools. If the original is HTML, I tend to read it right in the browser, and that’s the best experience. If the original is a PDF and the site “opens” it with some kind of viewer, the experience is usually poor, both in efficiency and in reading. I always end up downloading the PDF, because it’s smooth locally and I can use tools I know better.

So I don’t want to try to keep readers on my site. No matter how well it’s optimized, don’t hold them there. Just give them the tools, and leave them free not to use those tools, no conditions attached.

Clever

Mon, 26 Oct 2026 20:54:05 +0000

The blackmail test in Anthropic’s “A global workspace in language models”: “fake” and “fictional” show up early in the model’s J-space.

“This World Isn't Real”

Mon, 26 Oct 2026 20:18:28 +0000

Anthropic’s research on injecting thoughts into a model’s J-space raised two questions for me: what if the subject were a person, and if the model can feel pain, is this abuse? It also brought to mind the idea planted in Inception.

Thoughts on Watching Translating Claude’s thoughts into language

Mon, 26 Oct 2026 20:04:43 +0000

Thoughts on the video “Translating Claude’s thoughts into language”: train a model to turn activations into text, then turn the text back into activations to check it.

A Major Breakthrough on the Blog Project

Thu, 17 Sep 2026 10:55:24 +0000

The blog project has made a major breakthrough. I believe I’ve found the right way to do concurrent development. Next, I’ll finish the first phase: completing Phase A to prove that the path “edit content → new version → publish” can be carried out reliably by a Sonnet-class executor.

After that, I’ll look into how to manage the project itself, rather than managing specs and requirements.

I believe this is real progress. Now / Next / Later looks like a very simple, actionable way to manage a personal project.

I haven’t decided to develop this way yet, but it’s already given me a useful insight.

I’m excited about the development process ahead — being able to see the project as a whole, rather than just code and requirements.