jph00

Latent Space just released an interview about Solveit: https://www.youtube.com/watch?v=01ybLOH1fnU

no definitely not the case.

Absolutely right! Although I've found AI great for learning. It doesn't write my code for me, or my prose, or my devops scripts, but it's part of the process of me learning as I write them.

Thanks Erik - I've added that video to the article now.! :)

No, the first example shown doesn't use AI at all. The AI is an optional helper in the process, if needed/wanted.

The course is about a methodology, not a product. It's the ideas Eric Ries and I have been working on for decades. 5 weeks is a crash course that can only touch on the ideas. And it covers learning data structures and algorithms, foundations of web programming, system administration, startup creating, and much more.

It's really a rapid "how do to <x> the solveit way" for a variety of x. Each of those x is likely to become a full course in the future.

(We actually built the tool for ourselves, and only decided to make it publicly available when we realised how much it's helping us. We're a PBC so our mission is not entirely financial. We're not trying to compete with existing tools, but provide an alternative direction.)

(For those that don't recognise the name, Eric is the creator of The Lean Startup, and also founded the Long Term Stock Exchange. He's the co-founder of Answer.AI, which has built the solveit platform, and which fast.ai is now part of.)

It's both! We built a new platform designed to be a good place for using the approach we've designed. So it's a course, plus access to the platform.

Sorry it is a bit confusing tbh!

As the post mentions, a year ago we did a trial of it, and have been working with that group of 1000 users since then.

The course is about a methodology, not a product. It's the ideas Eric Ries and I have been working on for decades. 5 weeks is a crash course that can only touch on the ideas. And it covers learning data structures and algorithms, foundations of web programming, system administration, startup creating, and much more.

It's really a rapid "how do to <x> the solveit way" for a variety of x. Each of those x is likely to become a full course in the future.

Hi all - Jeremy Howard from Answer.AI here. Really excited to share with you all what we've learned over the past year about how to work with AI in a way that's entirely human-centered.

Whilst most folks seem focused on how to remove humans from the loop entirely and make AIs do all the work, we've concentrated 100% on how to make humans part of the loop in a way that makes us more and more capable and engaged.

I've enjoyed building and using our tool, "solveit", for the last year, and do basically all my coding, writing, reading, research, etc in it nowadays. I use small fast iterative steps and work to maximize my learning at each step.

Yes, since it supports Starlette/ASGI, FastHTML should work just fine.

That's not widely true. E.g the GPT 4 tech report pointed out nearly all their experiments were done on models 1000x smaller than the final model.

On this bench it underperforms vs glm-4.5-air, which is an MoE with fewer total params but more active params.

As the article notes, hypochondria is not an officially designated term.

Historically a great many diseases that medical research had not understood, have been widely assumed to be imaginary.

For instance: “fakers disease” which you’ll now know by its medical term: “multiple sclerosis”.

"Rank" means the number of dimensions of an array.

So "rank polymorphism" means being able to write expressions that work correctly regardless of how many dims the arrays involved have.

For example, in numpy you can write a function that handles both lists and matrices automatically, by taking advantage of broadcasting. (The J language takes this idea a lot further -- numpy is a fairly minimal implementation of the idea.)

You're welcome!

It's intentional -- sometimes you can get it to start spitting out its system prompts, but shortly after it does, a monitoring program cancels the output in the middle. It also blocks tricks like base64.

Asking in base64 doesn't work either - grok 4 heavy blocks that too. They seems to have a filter model that tests inputs and outputs to monitor for possible prompt leaks.

That isn't synced with what's in prod. E.g the system prompt changes that xai said were made during the "mechahitler" phase did not show up in that repo.

This seems similar to the situation where x.com claimed that their ML algo was in github, but it turned out to be some subset of it that was frozen in time and not synced to what's used in prod.

Label Studio is great, but by trying to cover so many use cases, it becomes pretty complex.

I've found it's often easier to just whip up something for my specific needs, when I need it.

From the RFC:

""" A well-known URI is a URI [RFC3986] whose path component begins with the characters "/.well-known/", and whose scheme is "HTTP", "HTTPS", or another scheme that has explicitly been specified to use well- known URIs.

Applications that wish to mint new well-known URIs MUST register them, following the procedures in Section 5.1. """

There was no "large group of very smart people" behind llms.txt. It was just me. And I'm very familiar with the registry, and it doesn't work for this particular case IMO (although other folks are welcome to register it if they feel otherwise, of course).

You can't just put things there any time you want - the RFC requires that they go through a registration process.

Having said that, this won't work for llms.txt, since in the next version of the proposal they'll be allowed at any level of the path, not only the root.

That's not what llms.txt is. You can just use a regular markdown URL or similar for that.

llms.txt is a description for an LLM of how to find the information on your site needed for an LLM to use your product or service effectively.

The next version of the llms.txt proposal will allow an llms.txt file to be added at any level of a path, which isn't compatible with /.well-known.

(I'm the creator of the llms.txt proposal.)

The paper link on that site doesn't work -- here's a working link:

https://arxiv.org/abs/2506.06105

But that’s the point! It’s just html!

But since the direct 1:1 mapping to html returns a python object, you can build up rich component hierarchies using plain python, mapping to plain html. And thanks to MonsterUI it looks pretty good too.

You seem to be arguing that more complexity, moving further from the foundations of the web, is a good thing. The fasthtml philosophy is the opposite — sticking close to the fundamentals is good.

That book is amazing! An absolute gem showing a wonderful way to think about web programming.

This isn't that kind of tool at all -- it's quite the opposite actually. It's a very thin layer on HTTP and HTML. MonsterUI uses tailwind, and you can directly use tailwind everywhere yourself.

I'm the creator of FastHTML -- I see quite a few of the comments so far are about FastHTML, rather than MonsterUI, so I guess I should provide some background.

FastHTML moves back closer to the foundations of the web. The functional components are a direct 1-1 mapping of m-expressions to HTML tags (i.e children are positional parameters, attributes are named parameters). There's nothing new to learn there if you know HTML already. I've been using functional components for web apps for >25 years, including building Fastmail with them -- they work very nicely for me at least, and are standard in most functional programming communities. I like being able to treat my components directly as plain objects directly in python (including using the python debugger, profiler, etc, and using IDE standard python navigation), rather than having to switch to a separate templating language with its own way of doing things.

MonsterUI is mainly a FastHTML wrapper around FrankenUI -- a pure js/css/html version of shadcn. I like using native components like this, instead of needing special react/vue/svelte/etc framework libs. FastHTML's functional approach make server-rendered components feel a lot like jsx in react (jsx behind the scenes is a wrapper over a similar data structure to FastHTML's, but js doesn't natively provide the needed m-expression syntax, hence the need for jsx).

For newer devs that never really learned native HTTP/HTML ways of developing for the web, and have lived in the react/nextjs/etc world, FastHTML will seem very odd. But for anyone that grew up with Perl, PHP, etc, writing plain server handlers that deliver HTML directly to the server, FastHTML will feel quite natural. They are two different approaches, each with their pros and cons. At Answer.AI we've written a lot of useful apps in FastHTML now, including SSE and websockets, rich client interactions, etc, and it's been working nicely for us.