benob

I don't think this would qualify as clean room (the Library was involved in learning to generate programs as a whole). However, it should be possible to remove the library from the OLMO training data and retrain it from scratch.

But what about training without having seen any human written program? Coul a model learn from randomly generated programs?

What about doing that with movies and music?

Also, don't forget to install our local ad blocking LLM. Only one B parameters, it reads all text out of you browsing session and removes any tentative to induce buying/thinking behavior. Best in town...

I had that in mind too. What if you handcraft a subnetwork with (some subset of) Turing machine capability? Do those kinds of circuits emerge naturally during training? Can reasoning use them for complex computation?

What is the need for dependencies when you can code them from scratch?

Same with asbestos, I mean what could go wrong in 20 years?

You could totally make a believable timing generation model from a few (hundreds) recordings of human writing. Detecting AI is hard...

Give me self hosting: LLM generates compiler which compiles LLM training and inference suite, which then generates compiler which...

Why don't LLMs directly generate machine code?

I really like the "file lens" example:

> “Focus on…” would allow the user to specify what they're interested in changing and present only files and lines of code related to their specified interest.

> “Edit as…” would allow the user to edit the file or selected code as if it were a different programming language or file format.

What local models shine as local assistants? Is there an effort to evaluate the compromise between compute/memory and local models that can support this use case? What kind of hardware do you need to not feel like playing with a useless shiny toy?

Too bad it doesn't have zooms echo cancelation / isolation

Maybe it will also change the whole publication as evaluation of science.

An LLM targeting language and no token efficiency?

It was over when they named the company Open AI

Is there a pip installable version?

What, no reference to quantum or crypto?

Will there ever be a llama12? Is it going to go the yolo route?

How was the attack detected in the first place?

Was it this one? https://infini-gram.io/

What does it mean nowadays to start from scratch? At least in the open scene, most of the post-training data is generated by other LLMs.

If caching is bounded in time, can't you use other fingerprinting methods to seal the gaps?

Why doesn't this apply to any kind of cached content?

Google's move is very good for the web. By pushing app makers away from walled platforms, you turn them to standardized, open ones such as the web.

> Bring Your Own Language

Few-shot new languages is going to be a game changer for linguists

Not tested on that particular model, but the idea has been flying around for some time: https://arxiv.org/abs/2509.04166v1

It comes with a basic interpreter, but the thriving community develops plenty of stuff: python, lua, forth, lisp... have nice ports which you can play with on device. There also is a library of software developed for rp2040 which have been ported, such as a mac classic emulator. If you feel that a micro controller is too low power, you can plug in a luckfox lyra which runs a proper linux with 128M of RAM.

I have been having a lot of fun with PicoCalc. It's not targeted at end users but is fun for developers alike who want a taste of developing things from first principles. More than anything it can live independently from your other devices.

This is very lenient French: "fetchez le dico"

Audio tokenization consumes at least 4x tokens versus text. So there is an efficiency problem to start with. Then is there enough audio data to train a LLM from scratch?