Dos desvíos

@StephenLCasper @GoodfireAI "In the past few years, several papers have demoed this kind of thing" would you providr some examples?

@GoodfireAI, I think this hype-milling verges on dishonesty. I believe that this paper has the potential to do big disservice to its readers, particularly less experienced ones who are newer to interp. Nothing new was accomplished here, and it wasn’t done in a useful way. This project just used interpretability methods as a circuitous way of contriving the rediscovery of predictive features in data sets, like sequence length. This project validated its interpretations about the salience of features by validating them as predictive features within a test set. But if that is what we treat as the ground truth, there’s no point to the use of interp tools. This is not a proof of concept for a repeatable recipe for scientific discovery as the post and thread claim. In order to show that these tools are valuable, you need to show that you can use them to discover something that wouldn’t be trivial to discover just by looking at the datasets. In the past few years, several papers have demoed this kind of thing. But this paper is not one of them. When you limit yourself to a hammer, everything looks like a nail. Especially when you’re also selling that hammer. In 2023, I told the GoodFire founder that I think a venture-capital-backed, for-profit interpretability research startup was the last thing that the epistemics of the interpretability community needs. I think this is still true and that GoodFire is establishing a pattern of grift.

Esto de los LLM se nos fue absolutamente de las manos

Si las black box más black box de la historia se terminan explicando a sí mismas en lenguaje natural va a ser un giro totalmente poético

Un efecto colateral MUY positivo del research en interpretability es la cantidad de material didáctico de buena calidad que generó

@nickhjiang Thx for ur answer! For that purpose, I could use LDA or any other topic modeling technique, can't I?

@dosdesvios Great question! The advantage of these labels is that you don't need to pre-define them, meaning that you can find insights about your data without any priors.

New work! What if we used sparse autoencoders to analyze data, not models—where SAE latents act as a large set of data labels 🏷️? We find that SAEs beat baselines on 4 data analysis tasks and uncover surprising, qualitative insights about models (e.g. Grok-4, OpenAI) from data.

@ChrisGPotts Love these videos! Please keep uploading practice runs!

Here is the full talk: youtu.be/iaDwT-bDL4A?si…

I've posted my practice run of this talk on YouTube (link just below). This clip gives the core argument:

Cursor es el pináculo de la civilización.

@gptcrosa Jaajja atroden crosa

Que hermoso ver la ley de alquileres en nyc va a ser hermoso lo que odio esa ciudad sobre valorada es tremendo

@stanfordnlp @tomchen0 Interesting! Will you record it?

Hi everyone! We're looking forward to the first NLP Seminar of the year! For this week's seminar, we are excited to host Tong Chen (@tomchen0) from University of Washington! If you are interested in attending remotely, please fill out the form below: forms.gle/E1iL719njyG1Nf…

@simonw This would explain why they usually don't come up with deep or new relations, the same way an encyclopedia stores a lot of knowledge but isn't able to rearrange it. They lack the big picture

An LLM is a lossy encyclopedia simonwillison.net/2025/Aug/29/lo…

La historia del NLP puede rastrearse en las notas a las sucesivas ediciones de esta biblia hermosa

Relegaron naive bayes al apéndice :(

No uso métodos anticuados, hago NLP ecológico.

Me parece fascinante que este experimento sea replicable en español entrenando vectores de 50 dimensiones con 400MB de Wikipedia.

Franco Moretti es tanto mejor que el promedio de los investigadores en digital humanities porque él llega a las dh como una necesidad más que como un arbitrario punto de partida.

Los LLMs "resolvieron" muchos problemas del NLP, con un costo energético inédito y en gran medida obligándonos a usar modelos PRIVADOS! Los métodos "clásicos" son baratos, mejores con el medio ambiente y mucho más respetuosos de la privacidad. Y esto no va a cambiar...

El affaire Grok está despojando a los LLMs de su ilusorio halo de neutralidad. Eso es más bueno que malo.

Muchas de las visiones apocalípticas sobre el futuro de los LLMs asumen que *alguien* les va a dar la potestad para tomar decisiones fundamentales. Pero darles esta potestad a los LLMs no sería menos absurdo que dárselas a un perro, o a un algoritmo que genera números al azar

@yoavgo What is it built on then, in your opinion?

"Modern ML is built on Linear Algebra". lol no its not.