Piotr Mazurek (in SF ๐บ๐ธ at last ๐ค)
@tugot17
966 Following 4.7K Followers
might be a good moment to re-read "Why are embeddings so cheap?" on Tensoreconomics;
we discuss embeddings but Jev is largely the same compute-bound pattern
the value of 4090-like chips increases substantially if jev-style models gain popularity
Show more
jev is the top model at 1k-10k context on openrouter!
my new favourite past time is rage-baiting every European ๐ช๐บI meet here that: "being in Europe is merely existing, being in SF is living" and showing them these
the reason you will have to care about heterogenous compute is that as we get closer to RSI you will be just priced out of accessing the leading edge chips by the labs; there is no upper cap on the demand here
Show more
$1B makes sense. 70% on training is logical
> DeepSeek is using NVIDIA gaming GPUs to run inference on smaller models
this is news to me but tbh is an obvious move. Imagine how much work you could speed up just with Jev-style small models, or a custom-made Qwen 35Bโฆ
I want it
Show more
after text now we are making vision faster as well ๐ท ๐ซก
probably the first speculator for small VLMs if I'm not mistaken
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
๐งต
Show more
this and producing a trillion tokens ๐ซก๐
At long last, the interactive DeepSeek-v4.1-Flash diagram is up ~
when you make a release you need to mention the Amdahl's law, i don't make the rules
the number of scams waiting for you when trying to rent a flat in SF is just insane ๐
Next ๐ฆ in making ๐ซก
Last Thursday I officially founded Roboteur GmbH at the notary's office. The journey begins! ๐ฆพ๐ช๐บ
Still canโt comprehend that seeing self driving cars means Iโm home ๐
The countdown to FARR San Francisco and SF Tech Week is On!
Weโre excited to reveal the next four brilliant minds joining the FARR lineup, coming together on one intimate stage in San Francisco!
@steipete created OpenClaw, the breakout open-source AI agent platform focused on autonomous workflows and personal AI assistants. In 2026, he joined OpenAI to help shape the future of agentic AI systems and intelligent personal assistants. Before his work in AI, he founded PSPDFKit, a document technology company trusted by thousands of developers and enterprises worldwide.
@sam_debrouwer is Co-Founder and CEO of XY AI Labs. She co-founded doc ai, which was acquired by Sharecare, and now builds agentic AI for healthcare operations: the systems that decide whether clinicians spend their day on care or on paperwork. She also co-curated and produced TED AI, San Francisco.
@eshear is the Co-founder and CEO of Softmax, an independent research lab building a massively multiplayer benchmark for social intelligence and alignment. Before Softmax, he co-founded Twitch, was a partner at Y Combinator, and briefly served as interim CEO of OpenAI.
@ramin_m_h is co-founder and CEO of Liquid AI, a device-native foundation model company, and an AI Scientist at MIT CSAIL. He is known for advancing efficient, adaptive AI systems for real-world autonomy, device intelligence, and robotics, and his work on Liquid Neural Networks has earned international acclaim.
October 5, San Francisco. Limited seats โ reserve yours soon.
RSVP here ๐
#FarrSummit# #AI# #AIAgents# #FoundationModels# #AIAlignment# #HealthcareAI# #SFTechWeek# #Openai# #AIScientist# #Sanfrancisco# #Frontierai# #Autonomy# #Research# #Robotics#
Show more
wtf since when a 5090 costs $7k๐ญ
FlashAttention on an actual napkin: very Boston-coded experience with @vtabbott_
yes you do need to ... in the same fashion at it was impossible to build ChatGPT without access to the model weights and training recipes
edge compute but at scale ๐ซก
The era of physical labor abundance is about to be unleashed and you are laughing?
when you see chinese humanoid companies scaling production before real demand is there, remember what happened last time they tried the same playbook
"look at the data" has to be the highest alpha advice for working on LLMs;
I'm currently working on optimizing LFM VL models and I realized I could improve my intuitions on vision datasets;
a thread where I go through the popular vision benchmarks, 1/x
Show more
