minhash (@minhash) header
minhash (@minhash) profile picture

Research Researcher @FeynAI (YC X25), prev startups, @oracle @IITGuwahati | I use nixos, btw

San Francisco, CA
Joined September 2020
Photo by minhash (@minhash)Photo by minhash (@minhash)Photo by minhash (@minhash)Photo by minhash (@minhash)Photo by minhash (@minhash)Photo by minhash (@minhash)Photo by minhash (@minhash)Photo by minhash (@minhash)Photo by minhash (@minhash)Photo by minhash (@minhash)
i generally don't think there's much to "learn" > get the GPUs (the hardest part) > set up vllm or sglang (they have recipes) > set up a few replicas on the same node > set up a cache aware router (the one from sglang should do) > set up a cache like lmcache so your replicas can see cache hits when they are under load > set up a gateway to manage replicas that are far apart because you run on heterogenous compute > enjoy see? very straightforward
Everybody I see learning Inference Engineering. Inference Engineering might be one of the most important jobs in AI right now. One 2026 industry analysis reported Inference Engineer job postings up 163% YoY. And current roles with disclosed salaries average around $250K/year, with frontier labs going much higher. If you're working in AI and not learning inference engineering yet, this is a good time to start.
13
31
3
608
46,170
minhash retweeted
What is the best LLM for maximizing frequency of a Chip Design? We measured performance of 8 different LLMs - giving them 20 designs and real EDA tools to see which models are the best. My feed is full of people saying LLMs can do chip design. Lots of anecdotes, very little data. I wanted to see how LLMs perform in a real-world chip design task. What we found: - GPT-6.2 Astra scored highest (75/100): closest to the best result across the designs, and the only agent to improve all 20. - Claude Fable 5.2 raised Fmax the most, +10.4% on average. - Claude Opus 5.5 made MemPool 50% faster, the biggest single win. - Half of all candidates failed their equivalence proof. - Iteration pays. In nearly half the sessions that found a gain, the best result came after the fifth attempt, and GPT-6.2 Astra was often still improving when it ran out. Links to full write-ups below!
5
4
7
330
/goals are all you need
Can AI really do your job? We put 20 AI models in a work simulation, asking them to read their environment and do work without being asked. The latest models did 20% of planned work well enough to accept unchanged. In January, this number was 7%.
1
36
minhash retweeted
Can AI really do your job? We put 20 AI models in a work simulation, asking them to read their environment and do work without being asked. The latest models did 20% of planned work well enough to accept unchanged. In January, this number was 7%.
1
1
2
9
229
i wake up to my habibi cooking me models
I get 4 amazing hours of sleep ✨ Someone has to wake @minhash in the morning 🔪
2
3
204
You can literally get a production model trained in less than 24hrs and less than $1000 in most cases, especially if your provider has thought about cost and speed (like us!) Come make production models with us, usefeyn.com
This simple idea (use the largest model at the highest reasoning when developing, once done - build the specialized targeted model for production) is gradually percolating through the industry. I am surprised the adoption is not higher, though - constantly see people doing literally the reverse...
4
205
who's building cloud computers for general consumer? think RAM + GPU on demand to run games, adobe, some chrome tabs etc.
62
rl gets 20% because rollouts are forward-only decode, which is bandwidth bound, and 4-bit weights are half the bytes. pretraining never gets that discount
unavailable
3
163
probably 98% cache tokens
Fireworks serves >40T tokens a day of "specialized models." I'd define this as any post-trained model for a customer.
91
minhash retweeted
Introducing Cloudy, a research tool for long-running experiments in the cloud. Your experiments move between CPUs and GPUs as needed, with terabytes of storage each. Pause anytime and resume right where you left off. cloudy.so
26
21
4
216
15,988
I think electron might be why we need 32gigs of RAM minimum in 2026
Electron is great software and it is the best option for many interactive desktop applications
5
312
cope. as long as you can make bad behavior impossible when working with agents, you should. they will do more than save you time and money. languages, frameworks, libraries. they are very much still needed for the next few years.
rust will die soon. it was just cope for humans writing bad code
1
160
minhash retweeted
Our mission with Keiki is to introduce a new medium your business can increase its revenue from. Today, we're excited to launch Apple Pay on iMessage for all agents hosted on Keiki, powered by @stripe . Apple Pay increases conversion by 22.3% - per Stripe.
7
3
2
40
3,408
wait till you find out the worth of work a single gpu (B300) can churn out with a open-source model silicon >>> carbon
a single GPU now makes more than the federal minimum wage of $7.25/hr fucked up to say, but isn't that a sign of compute being underpriced? one B300 produces much more value than a minimum wage human ☹️
7
350
the interesting line isn't GLM vs claude it's that the gains showed up once tokens stopped being rationed turns out the bottleneck was engineers hoarding context like it was gpu quota
BREAKING: Revolut moved ~2,000 engineers onto self-hosted GLM for coding and pushed Claude Code and Codex to the exception list. Per David Abitbol, who leads AI ops there: 56 Nvidia B200s, a blended 30 cents per million tokens, roughly a tenth of what they paid Anthropic and OpenAI. Engineering lead time improved 70%. Here's the part that matters: the gains showed up when tokens stopped being rationed. Engineers quit saving AI for the hard tickets and pointed it at everything. Frontier models become the exception, open weights become the default. That's a pricing problem for every lab heading to IPO. h/t @OliverMolander Would you move your team off frontier APIs if the bill dropped 90%?
99
nixos but /nix/store is one sqlite file reproducible AND faster someone stop me =)
It’s so funny that SQLite is genuinely faster than most traditional filesystems. I wish there were more experimental operating systems using dbs for the backend. Not many people playing around in this space...
1
4
253
"a paragraph, a block of code and a table don't need to be split the same way" putting this on a billboard on 101 =)
A paragraph, a block of code and a table don’t need to be split the same way. Chonkie gives you different ways to split documents for RAG. You can divide text by token count, sentences or meaning, with separate methods for code and Markdown tables. You can also combine the steps in one pipeline: split the document, add overlap between chunks, create embeddings and send them to a vector database. For example, start with larger Markdown sections, then divide those into smaller chunks based on meaning. It works as a Python library or a self-hosted API, and you can install only the components you need. MIT licensed.
2
223
minhash retweeted
A paragraph, a block of code and a table don’t need to be split the same way. Chonkie gives you different ways to split documents for RAG. You can divide text by token count, sentences or meaning, with separate methods for code and Markdown tables. You can also combine the steps in one pipeline: split the document, add overlap between chunks, create embeddings and send them to a vector database. For example, start with larger Markdown sections, then divide those into smaller chunks based on meaning. It works as a Python library or a self-hosted API, and you can install only the components you need. MIT licensed.
2
2
1
6
427
this seems cool but i bet schmidhuber did it in the 80s =)
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
3
175
This is massive! Happy for @jeffreyhuber and co. this is a really good acq for Replit and excited for what they do together ...but I imagine it would be insanely hard for the transition of folks to other services (read: turbopuffer) migrations are annoying, with or without AI
Chroma is joining Replit! Together, we will build the context layer for the self-driving company. trychroma.com/blog/chroma-jo…
1
9
435
can i sell hand written code on etsy? i don't know what else to do with it
1
11
374