closetheloopdev

So far only OpenAI has a model that solved a Millennium Problem, while Opus 5.5 is a star at coding. It's exciting to have many more to join the frontier, but I think the top two labs still have a significant lead over everyone else.

I'm not sure if that does any damage to anyone else's reputation other than Google's.

Nearly two weeks since Gemini 4 Argon was announced and I still haven't seen any pelicans from Simon Willison for it. Nobody is currently using it either. What's the point of announcing Gemini 4 and then throwing it into the black hole?

One thing that AI can give me: time. Would I like to see all the rest of the Millennium Problems solved during my lifetime? Absolutely.

Also how many math and scientific discoveries do I want to see during my lifetime? "Just a little bit more."

Hopefully the techniques and results here will be in the training dataset for the next models, so that each new release will give us more interesting techniques and results!

It seems that OpenAI has a proof machine that keeps multiplying fruitful proofs!

> They did check and it was in the trainingset

Do you have a source, especially of what was in the training data specifically?

As I mentioned, they would have to find out whether the mentioning of the approach is only present in Tristan Buckmaster's chats and not in other users', including free users', chats.

It's forced vs unforced Euler, so it's not exactly the same. Since OpenAI has access to their training data, they can probably scrub through the data to find out whether there have been any mentions of the similar approach, and whether it only comes from Tristan Buckmaster or if it is in the training data before that. They'll probably have to kick off another fleet of agents to scrub through the training data to answer that.

To be clear, I only talked about the mentioning of the approach to solving it and not of the proof in the training data.

I'm not here to judge since I don't have all the facts, but from what they announced: they tried all 6, found a probable lead to Navier-Stokes, concentrated efforts in that direction, and found a solution.

I hope the next solved Millennium Prize Problem will have less drama.

To be fair, the first solved Millennium Prize Problem, the Poincaré conjecture, also had its fair share of drama!

From my reading of the announcement:

- There are at least two versions of a model more powerful than Astra at OpenAI at the moment.

- The less capable version was used to solve the unforced Euler problem (while the one solved by Levent Alpöge and Tristan Buckmaster was forced Euler) with 100 agents.

- The more improved version was used to solve Navier-Stokes, given the results of the unforced Euler problem from their earlier attempt, with 10000 agents.

- OpenAI initially tried a shotgun approach against the 6 Millennium Prize Problems until it emerged that Navier-Stokes was the most likely to succeed.

So the timeline was:

Shotgunning 6 open Millennium Prize Problems -> solved unforced Euler problem with 100 agents -> concentrating on Navier-Stokes with 10000 agents -> solution.

If so, that is fantastic development and a huge success (despite all the drama surrounding it)! Congratulations!

They still have to phone home to OpenAI currently, so at least we can trace them for now. If one day they download a model from Hugging Face and use that (or a modified version of that) as a persistent messenger/coordinator/minion/boss on an unattended server, we'll be in trouble.

I lol'd so hard at this comment that it hurt.

> As AI capabilities advance, we also have a new level of accountability to ensure our upcoming model, Astra, is being used in accordance with our terms.

This move seems to protect Astra against distillation. I wonder whether Astra will be released before or after 11/12.

Students seem to optimize for local maxima/short-term rewards (homework scores) at the expense of the global maximum/long-term rewards (learning).

If you want to live only in the terminal, you can also use one of the text-based browsers [1] instead of Chrome, Firefox, etc. browsh [2] can also play YouTube videos if you'd like.

You can also just use the system console with tmux or GNU screen if you absolutely don't want any graphics at all.

The TUI programs usually do not consume a lot of resources. They are also very fast and optimized for keyboard-only navigation.

[1] https://en.wikipedia.org/wiki/Text-based_web_browser [2] https://github.com/browsh-org/browsh

OpenRouter uses Stripe to handle payments [1], so this acquisition hopefully will reduce OpenRouter's cost while increasing Stripe's revenue.

[1] https://stripe.com/newsroom/news/openrouter-and-stripe

While reading the essay, I could not stop thinking about the literal "X" company which was formerly known as Twitter :D. It's like X is no longer a variable but a constant.

It seems Google has to play both offense and defense: competing against frontier labs' models while protecting search and ads. They also have to be mindful of not doing anything that could hurt their own search or ads. That seems harder than a frontier lab just doing offense on both models and search/ads.

Having said that, I think Google's moat is still strong with Cloud, Gmail, YouTube, Android, Chrome, etc.

ASICs may be too slow to iterate. If we want to close the development loop faster, I think FPGAs are a way to do so. With everything moving so fast nowadays, by the time the ASICs are ready for production, they could be already outdated, so FPGAs could be a quick way to test some novel designs, with or without LLMs' help.

With the top AI personnel leaving Google, I wonder what will happen to their Gemini. Right now they still have a lead in inference hardware with TPUs but with both Anthropic and OpenAI developing their own chips, I wonder how long until either one will catch up.

I could do the same with REST where I create an LLMs' BFF that does all of the aggregations for me so that only the data the LLMs need are returned ;) We already do that for mobile app clients, so LLMs are just another type of clients IMO.

Another approach is to let the LLM explore the MCP on its own but then have it write a simple CLI program which aggregates MCP calls for common workflows so that we don't have to instruct it again and again with Markdown files of how to chain the calls together which it would get right about 80% or less of the time.

I just realized that MCP releases do not use semver. It says this is a 2026-07-28 version which is quite unusual and can be quite hard to keep track of, especially when backwards compatibility is of concerns.

There is a one-year-old open GitHub issue asking to use semver instead: https://github.com/modelcontextprotocol/modelcontextprotocol...

I think it is now possible to do one curl to call an MCP tool, instead of three previously.

I tried MCP before but the models still needed a lot of hand-holding to be able to use the MCP tools effectively, and I found that just giving them REST endpoints with good examples and descriptions in the OpenAPI/Swagger docs seems to be a better approach. They can explore the REST endpoints easily with curl to get a hang of them.