MTIA kernels @meta | ai and audio in free time | 🇳🇱
sf
Joined April 2020
- Tweets256
- Following570
- Followers286
- Likes1K
Pinned Tweet
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
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: anthropic.com/engineering/AI…
Frontier is the new faang. Recruiters keep talking about frontier salary and frontier research and frontier applied ai whatever that means.
french is such a beautiful language
Meet Mistral Large 4, aka Le Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
My team (MTIA kernels) is hiring by the way for roles in bay area, bellevue, and NYC. Feel free to reach out
According to amazon my light switch apparently recharged itself from 13% to 18% lol
The past months, I keep receiving emails from amazon saying my battery is low and needs to be replaced along with an estimated percentage.
I plotted the battery percentage according to amazon. Quite scummy
People are less computer literate than a decade ago due to the rise of smartphones. They don’t understand a file system with folders anymore
these are actually quite beginner mistakes imo. for my thesis i wrote a dynamic programming algorithm to write these based on basso continuo fwiw. things like parallel octaves are the first you learn about.
clearly shows shortcomings of ai in niche fields
BREAKING: $2.5M Microduck order from Lockheed Martin
Insiders say the units will be augmented with automatic rifles and shoulder-mounted ballistic missiles.
Microducks are reportedly intended to operate in swarms in forests where conventional drones can’t maneuver effectively
Tristan T retweeted
WHAT
We gave our AI system a spec, and in under 2 weeks, it designed, verified, and deployed a chip that beats NVIDIA.
It’s built for low-power physical AI workloads. We’re running live inference on >B+ parameter models like Llama, Qwen, and Kimi, serving at 3.4x better perf/watt than NVIDIA Jetson.
From just a specification, our AI system autonomously generated all of the RTL Design, UVM verification, formal proofs, firmware, drivers, and kernels, co-designing the model, software and silicon as one optimization loop.
Better AI can now design better chips to run AI, leading to a loop of recursive-self improvement towards our path to abundant intelligence.
Readers added context they thought people might want to know
Architect Labs did not manufacture a physical chip; the design was tested on a commercial AMD Versal FPGA board. Furthermore, the 3.4x perf/watt gain over NVIDIA Jetson is a simulation-based projection for a future 8nm ASIC, not a physical benchmark. businessinsider.com/architect-labs… architectlabs.com/architect-labs…
Make no mistake, these were made so that huggingface can swarm the openai office next time openai's eval tries to hack huggingface
We built a small biped robot you can teach new tricks to.
Train it in simulation, run it on the real thing. Meet Microduck 🦆
$399, shipping before Christmas.
pollen-robotics.com/microduc…
github.com/pollen-robotics/m…
so garry tan was in a frat?
Garry Tan says turning down Palantir for a Microsoft promotion was a $2–4 billion mistake:
"Joe Lonsdale and Stephen Cohen were my fraternity brothers at Stanford... they were interns at Peter Thiel's hedge fund."
"They said, 'Let's go get Garry. He's up in Seattle.' They flew me down to have dinner with him, and Peter said, 'I'm so sure this is the right thing for you. Here's a check for 70 grand.' That was how much I made at Microsoft at the time."
"I said, 'Thank you very much, Mr. Thiel, but I might get promoted to level 60 this year.' Which I did. But that was a $2 to $4 billion mistake at this point."
@garrytan @illscience
Had recruiter reaching out on behalf of Inferact, the company created by the creators/maintainers of vLLM.
Instead of mentioning vLLM, he says like 7 times it is founded by Berkeley alumni and backed by a Berkeley fund!!!??
It's a good school but creator of vLLM >>> Berkely alum
assuming they used cadence/synopsis, there must've been a lot of manual labor involved.
cadence/synopsis are a duopoly and they provide the least agentic ai ready environment.
please try using it before forming an opinion.
i refuse to believe ai is able to manage the 12 hour wait times to run something, the instability and overall shitty software
A few thoughts on OpenAI's Jalapeño chip announcement today:
1. This chip is most likely the first one virtually entirely developed by Codex/GPT. Codex with whatever internal coding model (GPT 5.6/6.0 whatever) coded the entire software stack and most likely the hardware design
2. OpenAI will write all of their inference serving in pure Jalapeño ISA (instruction set architecture). Why? They only need to get say a few production models serving on Jalapeño. They can handwrite with Codex the entire model in pure ISA to get very high performance
3. They are most likely running Codex/GPT in custom RL envs to teach the models direct Jalapeño chip programming at ISA level
4. This is a massive cost savings for OpenAI and only possible IMO due to the breakthroughs in agentic coding. An AI company with frontier coding models can now become a hardware vendor with only a small team of experienced SWEs and an infinite amount of tokens
This is the first chip program fully accelerated by frontier AI.