Get live statistics and analysis of Jeffrey Emanuel's profile on X / Twitter

I wanted to read Henry Kissinger’s 400 page undergraduate thesis (it has an incredible first page), but really didn’t feel like dealing with a scanned PDF that’s annoying to read on a phone without constantly zooming and panning. So I decided to convert it to a nice markdown format using OCR and LLMs. Then I thought it would be nice to fix the footnotes and get rid of the page breaks and to fix the line breaks and other things like that. I was already working on some other coding projects, so I had the idea of loading up the draft markdown file in Claude Code and having it work on fixing these issues using a swarm of 20 sub-agents, which worked well. Then I thought it would be cool to link to the full sources for all the many references on sites like the Internet Archive or Project Gutenberg, so I had another swarm of sub-agents do a ton of searches to track the links down and insert them into the footnotes and bibliography. Then I figured that I might as well run it through my mind-map generator and summarization code to see what it comes up with, so I tried that. But now I had a few files to present, so needed some kind of index page. So I asked Codex with GPT-5 to whip up a slick looking web page to present the stuff nicely, which it did a yeoman’s job with. Note that I was already working with these tools in a bunch of other sessions on other projects, so my work here was occasionally giving some instructions to the coding agents and letting them crank away. I really didn’t spend much active time on this! Anyway, the net result is clearly the premier way in the world today to consume Henry Kissinger’s undergraduate thesis electronically. I’ll post the link in the next tweet to avoid getting punished by the algorithm. As for the thesis itself, it’s wild how erudite he was as a young man, and also what a great writer he was. And even more impressive considering that English was his second language. The thesis is basically him trying to come to grips with, and to mentally organize in an internally consistent way, a vast swath of Western thought. From what I’ve read so far, I think he did a pretty good job. Incidentally, his thesis is the reason Harvard changes the rules to limit the undergrad honors thesis to a maximum of 35,000 words. Good thing they didn’t apply this silly limit to Henry!

I randomly came across this book in college and was totally mesmerized by it. First time I had seen a serious attempt to apply math to biology in an explanatory way, not just descriptively. Pretty ahead of its time for 1917. The illustrations alone are worth browsing through.

This is simply a false rewriting of history through a feminist grievance lens. She may have taken the picture, but it was Francis Crick’s mathematical sophistication and knowledge of Bessel functions that allowed him to recognize that this image implied a specific helical symmetry, something that escaped Franklin and Wilkins.

So Python 3.14 finally came out for real yesterday. Finally removing the GIL (global interpreter lock), which allows for way faster multithreaded code without dealing with all the brain damage and overhead of multiprocessing or other hacky workarounds. And uv already fully supports it, which is wildly impressive. But anyway, I was a bit bummed out, because the main project I’m working on has a massive number of library dependencies, and it always takes a very long time to get mainline support for new python versions, particularly when they’re as revolutionary and different as version 3.14 is. So I was resigned to endure GIL-hell for the indefinite future. But then I figured, why not? Let me just see if codex and GPT-5 can power through it all. So I backed up my settings and asked codex to try, giving it the recent blog post from the uv team to get it started. There were some major roadblocks. I use PyTorch, which is notoriously slow to update. And also pyarrow, which also didn’t support 3.14. Same with cvxpy, the wrapper to the convex optimization library. Still, I wanted to see what we could do even if we had to deal with the brain damage of “vendoring” some libraries and building some stuff from scratch in C++, Rust, etc. using the latest nightly GitHub repositories instead of the usual PyPi libraries. I told codex to search the web, to read GitHub issue pages, etc, so that we didn’t reinvent the wheel (or WHL I should say, 🤣) unnecessarily. Why not? I could always test things, and if I couldn’t get it to work, then I could just retreat back to Python 3.13, right? No harm, no foul. Well, it took many hours of work, almost all of it done by codex while I occasionally checked in with it, but it managed to get everything working! Sure, it took a bunch of iterations, and I had to go tweak some stuff to avoid annoying deprecation warnings (some of which come from other libraries, so I ultimately had to filter them). But those libraries will update over time to better support 3.14 and eventually I won’t need to use any of these annoying workarounds. Codex even suggested uploading the compiled whl artifacts to Cloudflare’s R2 (like s3) so we could reuse them easily across machines, and took care of all the details for me. I would never think to do that on my own. Every time there was another complication or problem (for instance, what is shown in the screenshot below), codex just figured it out and plowed through it all like nothing. If you’ve never tried to do something like this in the “bad old days” prior to LLMs, it was a thankless grind that could eat up days and then hit a roadblock, resulting in a total wipeout. So it was simply too risky to even try it most of the time; you were better off just waiting 6 or 9 months for things to become simple again. Anyway, I still can’t really believe it’s all working! We are living in the future.

If Hotz can actually pull this off and get the same performance via PyTorch as CUDA but without actually using CUDA, it opens the floodgates to alternative hardware (especially AMD’s), particularly on the inference side. There’s no fundamental reason why CUDA must remain a moat.

Recently, I've noticed people making a big deal about how we haven't yet seen massive disruption to job markets and knowledge work from AI, and so they're starting to doubt that it will happen soon. And investors are wondering if they can just sort of ignore it for a while.

OK, so now we know just how much more efficient DeepSeek is for inference in terms of total tokens per second processed. They are doing roughly 7-8x more tokens per second on an H-800 (a crippled, export-control version on the H-100) than the open-source state of the art on H-100

DeepSeek just released a pretty shocking new paper. They really buried the lede here by referring to it simply as DeepSeek OCR. While it’s a very strong OCR model, the purpose of it and the implications of their approach go far beyond what you’d expect of “yet another OCR model.” Traditionally, vision LLM tokens almost seemed like an afterthought or “bolt on” to the LLM paradigm. And 10k words of English would take up far more space in a multimodal LLM when expressed as intelligible pixels than when expressed as tokens. So those 10k words may have turned into 15k tokens, or 30k to 60k “visual tokens.” So vision tokens were way less efficient and really only made sense to use for data that couldn’t be effectively conveyed with words. But that gets inverted now from the ideas in this paper. DeepSeek figured out how to get 10x better compression using vision tokens than with text tokens! So you could theoretically store those 10k words in just 1,500 of their special compressed visual tokens. This might not be as unexpected as it sounds if you think of how your own mind works. After all, I know that when I’m looking for a part of a book that I’ve already read, I imagine it visually and always remember which side of the book it was on and approximately where on the page it was, which suggests some kind of visual memory representation at work. Now, it’s not clear how exactly this interacts with the other downstream cognitive functioning of an LLM; can the model reason as intelligently over those compressed visual tokens as it can using regular text tokens? Does it make the model less articulate by forcing it into a more vision-oriented modality? But you can imagine that, depending on the exact tradeoffs, it could be a very exciting new axis to greatly expand effective context sizes. Especially when combined with DeepSeek’s other recent paper from a couple weeks ago about sparse attention. For all we know, Google could have already figured out something like this, which could explain why Gemini has such a huge context size and is so good and fast at OCR tasks. If they did, they probably wouldn’t say because it would be viewed as an important trade secret. But the nice thing about DeepSeek is that they’ve made the entire thing open source and open weights and explained how they did it, so now everyone can try it out and explore. Even if these tricks make attention more lossy, the potential of getting a frontier LLM with a 10 or 20 million token context window is pretty exciting. You could basically cram all of a company’s key internal documents into a prompt preamble and cache this with OpenAI and then just add your specific query or prompt on top of that and not have to deal with search tools and still have it be fast and cost-effective. Or put an entire code base into the context and cache it, and then just keep appending the equivalent of the git diffs as you make changes to the code. If you’ve ever read stories about the great physicist Hans Bethe, he was known for having vast amounts of random physical facts memorized (like the entire periodic table; boiling points of various substances, etc.) so that he could seamlessly think and compute without ever having to interrupt his flow to look something up in a reference table. Having vast amounts of task-specific knowledge in your working memory is extremely useful. This seems like a very clever and additive approach to potentially expanding that memory bank by 10x or more.

My coding agent workflow has really changed a lot ever since I gave them access to messaging so that they can directly communicate with each other. Now, I have one of them come up with a super detailed plan and sometimes have GPT Pro review and improve the plan in the webapp. Then I start up 4 or 5 Codex instances in the same project folder and tell them: "Before doing anything else, read ALL of AGENTS dot md and register with agent mail and introduce yourself to the other agents. Then coordinate on the remaining tasks left in PLAN_TO_DO_XYZ.md with the other agents and come up with a game plan for splitting and reviewing the work." Then I can queue up a ton of the following message in codex, and it will just keep plodding along until the context gets full: "Proceed meticulously with the plan, doing all remaining unfinished tasks systematically and continuing to notate your progress in-line in the plan document and via agent mail messages." Then they just keep cranking on their own for a really long time. And you don't need to supervise them much so you can be juggling multiple projects like this at once and make really great progress on all of them.

I have a feeling I'll soon be looking back at my Claude Code usage for the past month with pangs of nostalgia about how I managed to get over $13k and counting of inferences services for $600 worth of Claude Max subscriptions. There's no way this is sustainable for Anthropic.

So much for Jevons paradox, huh? Maybe I was right after all. BTW, Satya basically confirmed this posture in the recent Dwarkesh interview. Zuck is probably doing the same too now if he has any sense…

Total Sam Altman domination here. They have the best models, the best API, the best products, the best (or 2nd best) coding agent CLI, the best image generator (at following instructions at least), best video generator, and the best financial deal making. Haters can only weep.

I finally got around to making a tool I've wanted for a long time: you can basically think of it as being "like Gmail for coding agents." If you've ever tried to use a bunch of instances of Claude Code or Codex at once across the same project, you've probably noticed how annoying it can be when they freak out about the other agent changing the files they're working on. Then they start doing annoying things, like restoring files from git, in the process wiping out another agent's work without a backup. Or if you've tried to have agents coordinate on two separate repos, like a Python backend and a Nextjs frontend for the same project, you may have found yourself acting as the go-between and liaison between two or three different agents, passing messages between them or having them communicate by means of markdown files or some other workaround. I always knew there had to be a better way. But it's hard to get the big providers to offer something like that in a way that's universal, because Anthropic doesn't want to integrate with OpenAI's competitive coding tool, and neither wants to deal with Cursor or Gemini-CLI. So a few days ago, I started working on it, and it's now ready to share with the world. Introducing the 100% open-source MCP Agent Mail tool. This can be set up very quickly and easily on your machine and automatically detects all the most common coding agents and configures everything for you. I also include a ready-made blurb (see the README file in the repo, link in the next tweet) that you can add to your existing AGENTS dot md or CLAUDE dot md file to help the agents better leverage the system straight out of the gate. It's almost comical how quickly the agents take to this system like a fish to water. They seem to relish in it, sending very detailed messages to each other just like humans do, and start coordinating in a natural, powerful way. They even give each other good ideas and pushback on bad ideas. They can also reserve access to certain files to avoid the "too many cooks" problems associated with having too many agents all working on the same project at the same time, all without dealing with git worktrees and "merge hell." This also introduces a natural and powerful way to do something I've also long wanted, which is to automatically have multiple different frontier models working together in a collaborative, complementary way without me needing to be in the middle coordinating everything like a parent setting up playdates for their kids. And for the human in the loop, I made a really slick web frontend that you can view and see all the messages your agents are sending each other in a nice, Gmail-like interface, so you can monitor the process. You can even send a special message to some or all your agents as the "Human Overseer" to give them a directive (of course, you can also just type that in manually into each coding agent, too.) I made this for myself and know that I'm going to be getting a ton of usage out of it going forward. It really lets you unleash a massive number of agents using a bunch of different tools/models, and they just naturally coordinate and work with each other without stepping on each other's toes. It lets you as the human overseer relax a bit more as you no longer have to be the one responsible for coordinating things, and also because the agents watch each other and push back when they see mistakes and errors happening. Obviously, the greater the variety of models and agent tools you use, the more valuable that emergent peer review process will be. Anyway, give it a try and let me know what you think. I'm sure there are a bunch of bugs that I'll have to iron out over the next couple days, but I've already been productively using it today to work on another project and it is pretty amazingly functional already!

I wrote up all my thoughts on whether NVDA stock is finally a decent short (or at least not a good thing to own at this point). I’m a huge bull when it comes to the power and potential of AI, but there are just too many forces arrayed against them to sustain supernormal profits.

This new Paul Singer interview with the Norwegian pension fund guy is extremely good. What a brilliant investor and thinker. I loved hearing about him getting shellacked in the 1974 bear market (was down 88% on his mom’s $50k) and that being the basis of his focus on risk control

I’m absolutely convinced that the smartest AI models out now, GPT-5 Pro and Grok4 Heavy, are already smart enough, and surely knowledgeable enough about math and AI, to conceive and develop important theoretical and practical discoveries, given the right kind of clever prompting.

I’m absolutely convinced that the smartest AI models out now, GPT-5 Pro and Grok4 Heavy, are already smart enough, and surely knowledgeable enough about math and AI, to conceive and develop important theoretical and practical discoveries, given the right kind of clever prompting.

I wanted to read Henry Kissinger’s 400 page undergraduate thesis (it has an incredible first page), but really didn’t feel like dealing with a scanned PDF that’s annoying to read on a phone without constantly zooming and panning. So I decided to convert it to a nice markdown format using OCR and LLMs. Then I thought it would be nice to fix the footnotes and get rid of the page breaks and to fix the line breaks and other things like that. I was already working on some other coding projects, so I had the idea of loading up the draft markdown file in Claude Code and having it work on fixing these issues using a swarm of 20 sub-agents, which worked well. Then I thought it would be cool to link to the full sources for all the many references on sites like the Internet Archive or Project Gutenberg, so I had another swarm of sub-agents do a ton of searches to track the links down and insert them into the footnotes and bibliography. Then I figured that I might as well run it through my mind-map generator and summarization code to see what it comes up with, so I tried that. But now I had a few files to present, so needed some kind of index page. So I asked Codex with GPT-5 to whip up a slick looking web page to present the stuff nicely, which it did a yeoman’s job with. Note that I was already working with these tools in a bunch of other sessions on other projects, so my work here was occasionally giving some instructions to the coding agents and letting them crank away. I really didn’t spend much active time on this! Anyway, the net result is clearly the premier way in the world today to consume Henry Kissinger’s undergraduate thesis electronically. I’ll post the link in the next tweet to avoid getting punished by the algorithm. As for the thesis itself, it’s wild how erudite he was as a young man, and also what a great writer he was. And even more impressive considering that English was his second language. The thesis is basically him trying to come to grips with, and to mentally organize in an internally consistent way, a vast swath of Western thought. From what I’ve read so far, I think he did a pretty good job. Incidentally, his thesis is the reason Harvard changes the rules to limit the undergrad honors thesis to a maximum of 35,000 words. Good thing they didn’t apply this silly limit to Henry!

Total Sam Altman domination here. They have the best models, the best API, the best products, the best (or 2nd best) coding agent CLI, the best image generator (at following instructions at least), best video generator, and the best financial deal making. Haters can only weep.

I have a feeling I'll soon be looking back at my Claude Code usage for the past month with pangs of nostalgia about how I managed to get over $13k and counting of inferences services for $600 worth of Claude Max subscriptions. There's no way this is sustainable for Anthropic.

Recently, I've noticed people making a big deal about how we haven't yet seen massive disruption to job markets and knowledge work from AI, and so they're starting to doubt that it will happen soon. And investors are wondering if they can just sort of ignore it for a while.

I finally got around to making a tool I've wanted for a long time: you can basically think of it as being "like Gmail for coding agents." If you've ever tried to use a bunch of instances of Claude Code or Codex at once across the same project, you've probably noticed how annoying it can be when they freak out about the other agent changing the files they're working on. Then they start doing annoying things, like restoring files from git, in the process wiping out another agent's work without a backup. Or if you've tried to have agents coordinate on two separate repos, like a Python backend and a Nextjs frontend for the same project, you may have found yourself acting as the go-between and liaison between two or three different agents, passing messages between them or having them communicate by means of markdown files or some other workaround. I always knew there had to be a better way. But it's hard to get the big providers to offer something like that in a way that's universal, because Anthropic doesn't want to integrate with OpenAI's competitive coding tool, and neither wants to deal with Cursor or Gemini-CLI. So a few days ago, I started working on it, and it's now ready to share with the world. Introducing the 100% open-source MCP Agent Mail tool. This can be set up very quickly and easily on your machine and automatically detects all the most common coding agents and configures everything for you. I also include a ready-made blurb (see the README file in the repo, link in the next tweet) that you can add to your existing AGENTS dot md or CLAUDE dot md file to help the agents better leverage the system straight out of the gate. It's almost comical how quickly the agents take to this system like a fish to water. They seem to relish in it, sending very detailed messages to each other just like humans do, and start coordinating in a natural, powerful way. They even give each other good ideas and pushback on bad ideas. They can also reserve access to certain files to avoid the "too many cooks" problems associated with having too many agents all working on the same project at the same time, all without dealing with git worktrees and "merge hell." This also introduces a natural and powerful way to do something I've also long wanted, which is to automatically have multiple different frontier models working together in a collaborative, complementary way without me needing to be in the middle coordinating everything like a parent setting up playdates for their kids. And for the human in the loop, I made a really slick web frontend that you can view and see all the messages your agents are sending each other in a nice, Gmail-like interface, so you can monitor the process. You can even send a special message to some or all your agents as the "Human Overseer" to give them a directive (of course, you can also just type that in manually into each coding agent, too.) I made this for myself and know that I'm going to be getting a ton of usage out of it going forward. It really lets you unleash a massive number of agents using a bunch of different tools/models, and they just naturally coordinate and work with each other without stepping on each other's toes. It lets you as the human overseer relax a bit more as you no longer have to be the one responsible for coordinating things, and also because the agents watch each other and push back when they see mistakes and errors happening. Obviously, the greater the variety of models and agent tools you use, the more valuable that emergent peer review process will be. Anyway, give it a try and let me know what you think. I'm sure there are a bunch of bugs that I'll have to iron out over the next couple days, but I've already been productively using it today to work on another project and it is pretty amazingly functional already!

My coding agent workflow has really changed a lot ever since I gave them access to messaging so that they can directly communicate with each other. Now, I have one of them come up with a super detailed plan and sometimes have GPT Pro review and improve the plan in the webapp. Then I start up 4 or 5 Codex instances in the same project folder and tell them: "Before doing anything else, read ALL of AGENTS dot md and register with agent mail and introduce yourself to the other agents. Then coordinate on the remaining tasks left in PLAN_TO_DO_XYZ.md with the other agents and come up with a game plan for splitting and reviewing the work." Then I can queue up a ton of the following message in codex, and it will just keep plodding along until the context gets full: "Proceed meticulously with the plan, doing all remaining unfinished tasks systematically and continuing to notate your progress in-line in the plan document and via agent mail messages." Then they just keep cranking on their own for a really long time. And you don't need to supervise them much so you can be juggling multiple projects like this at once and make really great progress on all of them.

Examples of o3 lying through its teeth in a really concerning way. It’s constantly making absolutely outrageous claims about how much faster the code is after its revisions (it doesn’t even know if it will *run*), whether it’s functioning properly, that it tested things, etc.

This is simply a false rewriting of history through a feminist grievance lens. She may have taken the picture, but it was Francis Crick’s mathematical sophistication and knowledge of Bessel functions that allowed him to recognize that this image implied a specific helical symmetry, something that escaped Franklin and Wilkins.

So much for Jevons paradox, huh? Maybe I was right after all. BTW, Satya basically confirmed this posture in the recent Dwarkesh interview. Zuck is probably doing the same too now if he has any sense…

I wrote up all my thoughts on whether NVDA stock is finally a decent short (or at least not a good thing to own at this point). I’m a huge bull when it comes to the power and potential of AI, but there are just too many forces arrayed against them to sustain supernormal profits.

It sounds so stupid, but one of the biggest productivity hacks when using Claude Code with Opus 4.1 is this: after asking CC to implement a feature or fix a bug or whatever, and after it says it completed everything, you just repeatedly say the following to it over and over until it can't find any more mistakes (which sometimes takes up to 7 or 8 tries!): "Great, now I want you to carefully read over all of the new code you just wrote and other existing code you just modified with "fresh eyes," looking super carefully for any obvious bugs, errors, problems, issues, confusion, etc." Yes, this does take a while, but that's why it's so handy to have a bunch of CC sessions open at once. Then you can just rotate through them, pasting that sentence in again and again. Somehow the "fresh eyes" stuff changes how it perceives what it just wrote in a really helpful way. Weirdly, this trick doesn't seem to work as well with GPT-5 thinking-- it tends to just say "Yup, everything looks right!" Claude is much more prone to second guessing and to making careless mistakes the first time around, but good at catching them given enough chances.

So Python 3.14 finally came out for real yesterday. Finally removing the GIL (global interpreter lock), which allows for way faster multithreaded code without dealing with all the brain damage and overhead of multiprocessing or other hacky workarounds. And uv already fully supports it, which is wildly impressive. But anyway, I was a bit bummed out, because the main project I’m working on has a massive number of library dependencies, and it always takes a very long time to get mainline support for new python versions, particularly when they’re as revolutionary and different as version 3.14 is. So I was resigned to endure GIL-hell for the indefinite future. But then I figured, why not? Let me just see if codex and GPT-5 can power through it all. So I backed up my settings and asked codex to try, giving it the recent blog post from the uv team to get it started. There were some major roadblocks. I use PyTorch, which is notoriously slow to update. And also pyarrow, which also didn’t support 3.14. Same with cvxpy, the wrapper to the convex optimization library. Still, I wanted to see what we could do even if we had to deal with the brain damage of “vendoring” some libraries and building some stuff from scratch in C++, Rust, etc. using the latest nightly GitHub repositories instead of the usual PyPi libraries. I told codex to search the web, to read GitHub issue pages, etc, so that we didn’t reinvent the wheel (or WHL I should say, 🤣) unnecessarily. Why not? I could always test things, and if I couldn’t get it to work, then I could just retreat back to Python 3.13, right? No harm, no foul. Well, it took many hours of work, almost all of it done by codex while I occasionally checked in with it, but it managed to get everything working! Sure, it took a bunch of iterations, and I had to go tweak some stuff to avoid annoying deprecation warnings (some of which come from other libraries, so I ultimately had to filter them). But those libraries will update over time to better support 3.14 and eventually I won’t need to use any of these annoying workarounds. Codex even suggested uploading the compiled whl artifacts to Cloudflare’s R2 (like s3) so we could reuse them easily across machines, and took care of all the details for me. I would never think to do that on my own. Every time there was another complication or problem (for instance, what is shown in the screenshot below), codex just figured it out and plowed through it all like nothing. If you’ve never tried to do something like this in the “bad old days” prior to LLMs, it was a thankless grind that could eat up days and then hit a roadblock, resulting in a total wipeout. So it was simply too risky to even try it most of the time; you were better off just waiting 6 or 9 months for things to become simple again. Anyway, I still can’t really believe it’s all working! We are living in the future.

Each release like this is such a compete humiliation and indictment of Meta, which pioneered open-weight LLMs with the first Llama model introduced in February of 2023. They’ve likely invested 100x to 1000x the resources (money, compute, PhD headcount, square footage, etc.) on a cumulative basis compared to any of these other Chinese labs (Kimi, Z, Qwen, DeepSeek, etc.). By all rights, they should be way ahead of everyone else. And yet they haven’t had a state of the art open-weight model, or even a modestly compelling model, since Llama 3.3, which was released at the end of 2024, nearly a year ago. Now the Chinese labs have been leapfrogging each other like crazy, so that the latest models are extremely capable now. Could Meta still pull a rabbit out of a hat and leapfrog these other labs as a result of all the brilliant people they’ve hired at great expense over the last few months? I suppose they could, but even then they’d likely be getting far, far less bang for the buck compared to these half a dozen or so Chinese labs. This would be like if the Soviets stole the nuclear bomb secrets and then ended up testing a hydrogen bomb 2 years before the US did. Unthinkable. Makes it a lot more understandable why Zuck has been doing a massive purge of the organization. I would want to clean house, too, in this case. And better to err on the side of caution and cut deeper to be sure you’ve removed all the rot.

It's wild how good OpenAI's o3 model is at translating Ancient Greek poetry. It does a way better job (while preserving the poetic structure) than you find from academics who spend their whole life doing this. This would be incredibly depressing to me if I were studying Classics.

Public Service Announcement: You can greatly reduce the chances of something catastrophic like this (i.e., Codex or Claude Code deleting your project or files, wrecking your database, etc.) by adding this stuff to the very beginning of your AGENTS dot md file: