
it's an incredible time in tech rn, everyone trying out crazy shit and twenty different permutations of a thing popping up and dying in an afternoon. We're right in the middle of the kitchen and stuff is flying everywhere

Inference, Kubernetes, reliability, and the messy reality of running AI in prod. Built some popular OSS things (llmfit more recently) axjns.dev

it's an incredible time in tech rn, everyone trying out crazy shit and twenty different permutations of a thing popping up and dying in an afternoon. We're right in the middle of the kitchen and stuff is flying everywhere

dangerous rhetoric, I believe there is absolute necessity for advancement of frontier AI within the UK. This is as important as critical national infrastructure
UK does not need its own frontier AI model, says AI minister bit.ly/4zxiGuS

so i'm joining a new company in 7 days, so that's pretty exciting

let's be honest, most people won't apply a mote of judgment to what is shown, and reduce the human experience to finger muscle spasms
Today we launch Sensei, the first generative experience across all your devices. Get Sensei: kosuke.ai Ship wherever you are: Mobile, Watch, iPad, Glasses, Web, Desktop (for now). We made execution a taste game, in a Tinder-like experience. Let the agent do the work. You just swipe right if you like the output, left if you don’t. We believe in a world in which the laptop is dead. Superintelligence is coming. Software and execution are going to zero. Models will do the work more and better than you. It makes no sense to spend 8 hours a day in front of a computer. It’s time to ship from the swimming pool, or between one match and the next 🎾

was halfway through talking tlb misses and hypervisor page table walks with fable and it ended the chat on me. i'd told it upfront this was just to get my own thinking going. if a model won't touch nested paging it's not much use for systems work

A film you've watched more than six times with a gif. No Star Trek, Star Wars, Lord of the Rings, Marvel, Disney or Ghibli.


this has such a great old school energy to it
I never thought I needed to know the answer to the question "How many low lvl hunters does it take to kill a devilsaur." but now I do.

llama.cpp going 1.0 feels overdue. it's been quietly holding up local ai for years, semver just makes it official
llama.cpp v1.0.0 is now available huggingface.co/blog/ggml-org…

Most of it ends up as heat. a sliver ends up as thought. The miracle isn't the efficiency, it's that we keep building it anyway imo

think about the human effort to mine lithium, drill for oil, design and prototype HVAC, rack sleds, invent network protocols.. when the output is so incredibly ineffective for that investment, especially since the human brain does this for 20w

I spend a lot of time thinking about energised megawatt conversion into real world output. And when you stare really hard at that calculation the lossyness of the conversion is staggering

Second hand GPU market is an absolute blood bath. I'm struggling to equip the new llmfit lab without having to sell a kidney

this is fun! to clarify; it can work. the M5 Ultra Studio peaks at 385W, so 440W off the pack after the inverter, and 750Wh e-bike battery gives about 1.7h per Mac. so it works, you just need a lot of batteries!
It’s live. First portable AI cluster in the world. Thanks @fdotinc for the space


Congrats @MiaAI_lab I'm really blown away to be in such good company with llmfit. The collective good is surging ahead here
I'm deeply grateful to @HRF for the grant to support the work I do. The mission was, is, and will always be the same: Pushing local AI forward. I will do everything in my power to use this grant the best way possible to advance my mission, put out great local AI recipes, and build the best open-source tools I can. Thank you! hrf.org/latest/hrfs-ai-for-i…