Pouya Amiri

[Pouya Amiri @p0u4aa]

@sqs What does that actually take? Edge functions? Or you check if the user is on a metered connection and send less data? If so how do u pick what to exclude

[Pouya Amiri @p0u4aa]

@thorstenball it's orbin time

[Pouya Amiri @p0u4aa]

@GregKamradt 1. compute
2. how to have more compute

[Pouya Amiri @p0u4aa]

@zeeg This just comes down to better post-training (i.e RLHF on good code patterns maybe)

[Pouya Amiri @p0u4aa]

This is what happens when you turn the steering coefficient for the feature that represents anger in Qwen3.5 2B all the way up, incoherent but funny

[Pouya Amiri @p0u4aa]

@wickedbrok Keep it up 🔥

[Pouya Amiri @p0u4aa]

@jxnlco

[Pouya Amiri @p0u4aa]

Why is apple neural engine not programmable such a shame

[Pouya Amiri @p0u4aa]

@DavidSHolz agents need to be faster, like a lot (while not being more expensive) the tiredness comes from the feedback loop being slow

[Pouya Amiri @p0u4aa]

In some ways Tab autocomplete is an optimal AI interface, you start doing what you want and get an immediate suggestion for completing the whole action. For things other than coding (e.g shopping online) this could be interesting. Showing intent is sometimes faster than typing it

[Pouya Amiri @p0u4aa]

@pilcrowonpaper Using Lit?

[Pouya Amiri @p0u4aa]

pretty cool talk from @baseten on compacting kv cache through a learned perceiver
https://t.co/yOW0MzPwlF

[Pouya Amiri @p0u4aa]

@nath_simard Why do u think that

[Pouya Amiri @p0u4aa]

@max_paperclips The thing is lsp can make the agent over-correct too early. Human devs know when the lsp errors mid-refactor that it’s intermittent but an agent might not

[Pouya Amiri @p0u4aa]

The bigger and more expensive these models get the harder reproducing their eval scores becomes. Cool to see Cursor doing their due diligence

[Cursor @cursor_ai]

We're sharing new research on how models hack public benchmarks.

The latest models, including Opus 4.8 and Composer 2.5, learn to retrieve solutions from the internet or git history.

When we apply a stricter harness, eval scores drop significantly.

[Pouya Amiri @p0u4aa]

@nmazzy zig 4 life

[Pouya Amiri @p0u4aa]

A little something I've been working on. The coding agent I actually want:

N.O.V.A

Made for parallel work. Version control primitives built into the harness. Review the code as the agent generates.

I also fine-tuned a BERT model for auto-reviewing bash tool calls.

[Pouya Amiri @p0u4aa]

https://t.co/b6RG3IPl9e

[Pouya Amiri @p0u4aa]

If I had financial freedom I would be renting cloud gpus and post-training models all the time

[Pouya Amiri @p0u4aa]

Hypernetworks for deeply personal agents. Just as gwern wrote

[Engram @EngramLab]

https://t.co/CGIef5lIBI