What is the system that is supposed to be AGI? The inference only (chat gpt/LLM+harness) or the training + inference loop (i.e. OpenAIs/Anthropics full stack, though this probably still involves humans). I doubt that inference only is AGI as of today. The full stack might be.

Do you keep them turned on? Some auto wakeup/hibernation feature for dev boxes would be great!

A proof market would be very handy, so I could put money on stuff I want to see proven/disproven. Would be an interesting way for an agent to make money, and to see what the most important conjectures are.

Storing the passkey private key in software systems is as safe as storing the TFA private/recovery key in a password manager. With TFA, hardware isn't enforceable though, and systems usually only allow to set up one TFA code. (According to my understanding)

With passkeys, a system can cryptographically enforce that only certified hardware (eg. TPM) is used to store the key. The hardware can certify that the user recalled the PIN (in N attempts). These are the two factors. Most systems don't enforce trusted hardware though.

Do you have ideas how to make this secure? Eg when you have a fully autonomous loop, that it cannot possibly leak your github token or consume an excessive amount of tokens overnight? I've been exploring that space a bit, but it gets tricky quickly.

There is also a mathematical proof that says no algorithm can decide in general if a given formal mathematical statement is true. Not even humans in the loop. Every cleverness on one class of problems can be used against it! And still, humans can do math and prove things.

It is also now trivial to write tools with AI that show you the paragraph of the cited paper, so you (or AI) can easily check that the citation exists and is relevant. (it took me less than an hour to vibe-code the tools needed to get what is shown in the screenshot)

This pattern reminds me of Rule 30! en.wikipedia.org/wiki/Rule_3…

The same seems to happen for cellular automata language theory. LLMs try to argue with common sense, but then fail to pin down any coherent technical argument. To quote Sonnet 4.6: > In short: I confused "the obvious construction doesn't work" with "no construction works".

Has there ever been a bug in lean that made it significantly easier/possible to prove a statement, but that didn't make it significantly easier/possible to prove false?

Very helpful was the Brave MCP server and the Google Scholar API, which AI turned into an MCP server on its own in one prompt. The PDF quotes come from an http server that maps a given range in the extracted text back to the visual bounding boxes.

I used VS Code Copilot and the debug value editor extension (github.com/microsoft/vscode-…) to let AI run/debug a node repl, which it used to go through the papers, extract their text, create embeddings, and look up their metadata.

I've become a fulltime AI addict since January, not being able to stop creating, researching and planning. And I'm still in shock by just how powerful AI is. Any AGI will immediately be an ASI. Here Opus did a literature research for me, it also created all the tools it needed.

I love this! Would be interesting to explore what this could do for Lean (which uses unicode math symbols)

Together with ChatGPT, I discovered a fun puzzle I thought was worth sharing: "Finding the Survivor in the Middle" gist.github.com/hediet/a0e9b… (…or why one-dimensional two-way cellular automata operating in linear time are likely more powerful than their real-time counterparts)

I'm confident that the code is correct because composition has a nice definition of correctness: Applying "e1 composed e2" on a string "s" should equal applying e1 on s, and then e2. A fuzzer verified this for millions and millions of different edits.

Today is a day that I will remember for quite a while - Gemini 3 correctly implemented an algorithm in minutes that I didn't manage to implement in days: Composition of Line/Column-based text edits. I don't even understand the code (but heavily tested it). github.com/microsoft/vscode/…