Tristan T

MTIA kernels @meta | ai and audio in free time | 🇳🇱

[Tristan T @trirpi]

Created deep dive into Anthropic’s performance take-home test—the custom VLIW SIMD simulator, the tree-traversal hashing challenge, and how Claude Opus 4.5 outperformed most humans. Link in comments

[Anthropic @AnthropicAI]

New on the Anthropic Engineering Blog: We give prospective performance engineering candidates a notoriously difficult take-home exam. It worked well—until Opus 4.5 beat it.

Here's how we designed (and redesigned) it: https://t.co/3RZVyhpVij

[Tristan T @trirpi]

@nic_carter fixed it

[Tristan T @trirpi]

@rssmrm startup idea 💡

[Tristan T @trirpi]

@tensarahq maybe average out multiple benchmark runs? I'm getting latencies ranging from 40us to 70us on identical code.

[Tristan T @trirpi]

if your openclaw is disagreeing with you, you should delete the "you're allowed to disagree" rule from the SOUL.md.

so much nicer now my bot is a pushover

[Mario Zechner @badlogicgames]

linkedin is the real moltbook.

[Tristan T @trirpi]

what happened to mamba and state space models?

[Tristan T @trirpi]

@ElbaSatGuy What's the latency on this thing? How does this compare to CPU?

Pretty cool to see some GPU audio processing, just curious on the usefulness.

[Tristan T @trirpi]

lol can someone tell me what’s a KB cache

[Y Combinator @ycombinator]

Inference Chips for Agent Workflows
@sdianahu

Most AI chips are designed for "prompt in, response out." Agents don't work that way. They loop, branch, and hold context across dozens of steps, and current GPUs hit 30–40% utilization as a result.

That gap is where purpose-built silicon wins.

[Tristan T @trirpi]

Got like 11 dependabot alerts for some simple webproject with 87-line backend. They were not a concern for my project, but still insane how many security issues in pretty popular node packages

Most of the issues were part of `qs`, the string parser used by express.js

[Tristan T @trirpi]

claude keeps calling every bug a smoking gun 🧐

[Tristan T @trirpi]

This is quite interesting.

Basically, mentions we should also compare companies based on price/book value. This is the ratio of a companies profit divided by the cost of its assets.

Intel is at 3.6 p/b while TSMC is at 10.6.

[Irrational Analysis @insane_analyst]

Intel is still cheap on price/book value. https://t.co/6Ey3s3yRZE

[Tristan T @trirpi]

Source: https://t.co/ysZZq2QBkK.

[Tristan T @trirpi]

One unique feature of jax is it's syntax to describe sharded matrices. You define a mesh, and can then use the device to specify how to shard over it, there's a ton of possiblities on how to shard as shown in the picture.

[François Chollet @fchollet]

When looking at deep learning profiles, one of the most obvious tells between a mediocre and great candidate is whether they list PyTorch or JAX.

[Tristan T @trirpi]

10/ However, HFT has not caught up fully on these prediction markets, and arbitrage opportunities are way more plentiful there, so I'm sure a few opportunities like this exist there. Good luck!

[Tristan T @trirpi]

Apparently arbitrage opportunities on polymarket had a duration of ~12s in 2024 and currently still for ~2s in 2026 with 73% captured by <100ms bots.

[Tristan T @trirpi]

9/ Now, before you ask openclaw to find these opportunities, please realize that this technique has long been optimized and turned into a latency game by high frequency trading companies running it in nanoseconds on FPGAs with more specialized versions techniques.