Lucas

ML engineer

[Lucas @block_lucas]

And Pi is a great harness being consistently better and cheaper. But being able to swap harnesses seems to be the biggest moat

[Lucas @block_lucas]

The databricks swe internal benchmark is really interesting.
It really shows the potential of open models like GLM 5.2
A cheaper model is not always a cheaper choice, sonnet ends up being more expensive than Opus per task.
https://t.co/jrOvHLFgYH

[Lucas @block_lucas]

I like the idea put forth in the economist that a lot of the hate towards datacenters is due to them representing the embodiment of AI rather than nimbyism

[Lucas @block_lucas]

Small things like these is such a great short book, Claire Keegan is so good

[Lucas @block_lucas]

I had never seen an approach like this to explaining positional encodings, pretty cool
https://t.co/lHYSAJ3DTf

[Lucas @block_lucas]

parakeet-tdt is too good, nvidia really cooked

[DSPy @DSPyOSS]

indeed it's all just signatures (specs), modules ("harnesses", "inference scaling"), and optimizers (learning algorithms for prompts, weights, and hyperparameters)

can you imagine what would happen if there was a framework that divvied this up in 2022 and is still growing?

[Lucas @block_lucas]

@sailaunderscore Such a weird take. Scenic is not the same as safe. And there is many examples of how Maps already biases who gets business, this would just make it worse (for instance https://t.co/LtqxEWn15C)

[Lucas @block_lucas]

Had to get thinking with Claude in London
Anthropic really winning the aesthetics