Phil Hauser

AI Engineer, Linux and opensource enthusiast, IPI Member, Linkin Park fan

[Phil Hauser @pjhauser1]

Prefix reuse in numbers: a 60k-token prompt that only costs the delta - 2.2s prefill where a cold run paid 45s. Cache-hit accounting belongs in your product metrics

[Phil Hauser @pjhauser1]

@MiaAI_lab What's amazing is that that hasn't been enabled by cheaper faster hardware, the opposite is in fact happening, it's instead been enabled by people like @ashxhart and projects like #tensorfold

[Kamil Baranek @kamilbaranek_]

Claude just finished what Claude 41 mins ago said would take 2 weeks to implement.

[DHH @dhh]

My hunch is that people don't bother spending their tokens on projects that are openly hostile or restrictive towards AI contributions. Nor according to where people who say things like "slop tokens" think they should go. https://t.co/leQZ6qnb6t

[Phil Hauser @pjhauser1]

Maybe this is the turning point for the snobbier FOSS libraries that boycott AI driven development? Then again... Elitist hackers usually know better innit

[Ayaan 🐧 @twtayaan]

Linus Torvalds is using Claude Code to make the Linux kernel faster.

He’s published a development branch targeting a bottleneck when many threads make small reads from the same file.

Claude Code helped write the code, create test programs, review the changes and draft the patch description, all under Linus’s direction.

He ran the measurements on his own machine, with significant improvements in the targeted workload.

These patches aren’t merged yet, and a place in Linux 7.4 is still unconfirmed.

Source: https://t.co/9KMVbqHBCS

[Phil Hauser @pjhauser1]

Anyone had a play with amd for AI? All I see and hear is Mac and NVidia

[DHH @dhh]

It's very important that we don't concede open source to socialists and communists. Capitalism is the greatest engine of growth and prosperity the world has ever seen. Open source is a great technique for aiding and accelerating it.

[Phil Hauser @pjhauser1]

MTP speculative decoding on a desktop GPU: 20-24 of 24 drafts accepted. The acceptance rate is the whole story - speculation without accuracy is just latency

[Mia @MiaAI_lab]

https://t.co/5gH1aoMj1z

[Phil Hauser @pjhauser1]

Prompting is the new programming language

Agent control planes is the new dev ops

Agent harnesses are the new IDE's

APIs are the new APIs

[Phil Hauser @pjhauser1]

Holy moly

[Ash Hart @ashxhart]

TensorFold is going exospheric, literally. 🚀

We’ve been invited to test it on Jetson AGX for potential use aboard satellites.

Ground testing comes first, but the thought of something I built running above Earth is a little surreal.

What is actually happening 🤯

[Phil Hauser @pjhauser1]

KV snapshot banks: precompute whole-context states to disk, resume long-running sessions in seconds. The line between a cache and long-term memory is thinning

[Phil Hauser @pjhauser1]

A 4B student with one epoch of LoRA beats prompt-engineering the same task. One variable, one base, 34k training pairs - small models with good data go far

[Phil Hauser @pjhauser1]

@sudoingX There are so many options now for extremely capable local agents its amazing! #tensorflow unlocked my local agent setups for sure!

[Phil Hauser @pjhauser1]

@ashxhart https://t.co/yUR6Qj4VeI

[Phil Hauser @pjhauser1]

I saw people trying to cram Qwen 3.6 35B A3B onto an 8GB card and wondered if I could get it working with #tensorfold ... Suffice it to say not only does it work, but I actually get a usable TPS on my laptop 🤯 @ashxhart bringing local AI to the masses

[Phil Hauser @pjhauser1]

@ashxhart This is with a 32k context window btw

[Phil Hauser @pjhauser1]

@ashxhart Experts in VRAM and the rest in RAM

[Phil Hauser @pjhauser1]

MTP speculative decoding on a desktop GPU: 20-24 of 24 drafts accepted. The acceptance rate is the whole story - speculation without accuracy is just latency

[Phil Hauser @pjhauser1]

@axhiii88547 Big big fan of @cloudflare here