Barak

Meat computer seeking to understand the universe @GoodfireAI

[Barak @barakwidawsky]

@varunneal You found my model at a very Chinese time of its thinking trajectory

[Barak @barakwidawsky]

@AgustinLebron3 I’m not sure it’s RLHF

I suspect it may be a form of in-line thinking/planning

[Barak @barakwidawsky]

@tszzl More, but alignment will become harder as people increasingly lean on the models

[Barak @barakwidawsky]

@samzliu I’m not sure we’re ready for the unironic “companies are conscious” conversation

[Barak @barakwidawsky]

San Francisco’s top 10 GORGEOUS parks to check slack in🧵👇

[Barak @barakwidawsky]

@_arohan_ Can’t wait for kernels that prefetch from magnetic tape

[Barak @barakwidawsky]

I believe OpenAI and Anthropic get quite a bit of usage in the broader economy 🙂. It’s certainly possible to claim that you’d expect more usage at the claimed capabilities but that’s a different story

This has not materialized for Meta or SpaceX, and SpaceX is certainly reselling the compute (so it’s still getting used)

To your original point models themselves may not end up being that competitive! But open model inference usage just keeps going up, hence my statement that compute itself is a big advantage

[Barak @barakwidawsky]

@menhguin Claude Mythos -> Chronicle -> Akashic Records

[Barak @barakwidawsky]

@deanwball To be understood is to become stale

[Barak @barakwidawsky]

@snowclipsed Same

[Barak @barakwidawsky]

RLHF is forever going to be a moving target. As soon as people start picking up on patterns that a model adopted because people genuinely prefer it (and then identifying those patterns as being "AI slop"), their preferences change

"Delve", em dash, warm colors in image gen....

[Barak @barakwidawsky]

@stochasticchasm Ayoooo!???!

I got to try this

[Barak @barakwidawsky]

Hey Claude please raise my series A make no mistakes

[will brown @willccbb]

we raised $130M @ $1B for our series A

>$100M run rate

we’re just getting started

[Barak @barakwidawsky]

If you are a feature hiding inside model activations, we will find you 🫵

[Goodfire @GoodfireAI]

If models think in shapes, our tools should too.

Our latest research: Block-Sparse Featurizers (BSFs), a new way to find concepts in model activations - using multidimensional “blocks” instead of single directions. (1/9)

[Barak @barakwidawsky]

@tszzl I agree “tool AI” is a losing concept, but I am optimistic that “collaborator AI” is possible

[Barak @barakwidawsky]

@jeffreyhuber @willccbb Policy update is limited

The SDPO paper demonstrates some evidence for this

[Barak @barakwidawsky]

@giffmana @manuelcecchett

[Barak @barakwidawsky]

@giffmana @manuelcecchett If the image is just the pale blue screen, then no I don’t see it there

When I’ve made diagrams with boxes, the box interior doesn’t usually have the noise but the area around the boundaries does

I’m assuming it’s a generation artifact

[Barak @barakwidawsky]

@giffmana @manuelcecchett I don’t know if it’s a watermark or a generation artifact, but *every* ChatGPT-image-2 generation has a really strong (and subjectively unpleasant) speckle pattern over it

It was worse in the initial release but it’s still there

[Barak @barakwidawsky]

@GergelyOrosz The machine culture will have great influence

[Barak @barakwidawsky]

@tenderizzation Even more! For the same selected tokens there’s different compression methods