building software, rethinking systems, and enjoying life in chicago
Chicago, IL
Joined May 2025
- Tweets874
- Following175
- Followers64
- Likes3.2K
Pinned Tweet
Writing bug reports to OSS projects shouldn't be quick.
Maintainers are heroes for donating the amount of time they do to their craft, and you should allocate a modicum of yours to save as much of theirs as possible.
Anything less comes off as obnoxious and entitled, IMO.
Touch on a laptop is a legitimate "Pro" feature IMO. Being able to use an iOS simulator like an actual phone is lowkey super useful for developers.
i have no idea what "argon" means relative to "carbon" when it comes to gemini 4.
there is zero intuition around these names, unlike haiku/sonnet/opus/fable & luna/terra/sol/astra
this absolutely mogs Genie 4 in actual day-to-day usefulness
stands to have a vastly larger positive impact on the lives of actual people
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
Generative AI slop movies and games is far and away the least compelling usecase for a World Model.
@Google what are we DOING
Bleeding edge R&D to create a "product" nobody in their right mind would ever actually want to use.
henry retweeted
Today we're introducing Genie 4 — our newest interactive world model from @GoogleDeepMind 🌐
Much more value to be extracted from the things you're NOT doing than from the things you are. The latter is finite, but the former is infinite.
So much is possible now.
It's essential to re-evaluate your priors.
God I am so hyped for overalls.dev
More excited for this than GTAVI
😎 overalls.dev now supports pausing/resuming firecracker machines.
here's how it works:
if the machine is 'pinned' it shows up on the top of your dashboard and is prevented from being deleted for suspended.
if the machine is created via /threads (ie. orb workflow) then the machine is automatically suspended after 4 hours of inacitivity from tool call. a new prompt/toolcall will unpause a machine.
if a network request (ie. vscode web, ssh or shared webapp) arrives it and the machine is paused then it will automatically unpause/resume.
if there is no capacity on the lilly node (the bare metal machine) then it will >automatically relocate the machine< to another lilly node with capacity
machines can now be cloned (ie. forked) with their live memory state instead of being turned off (loss of memory state).
of course you can pause/suspend/clone directly via ssh
all of the above rules hold true even if the machine needs to be relocated to another lilly node with capacity.
tldr; cpu/memory is expensive right now and now resource mgmt is built in as a platform primitive that you can self-host on your own infra 😎
henry retweeted
😎 overalls.dev now supports pausing/resuming firecracker machines.
here's how it works:
if the machine is 'pinned' it shows up on the top of your dashboard and is prevented from being deleted for suspended.
if the machine is created via /threads (ie. orb workflow) then the machine is automatically suspended after 4 hours of inacitivity from tool call. a new prompt/toolcall will unpause a machine.
if a network request (ie. vscode web, ssh or shared webapp) arrives it and the machine is paused then it will automatically unpause/resume.
if there is no capacity on the lilly node (the bare metal machine) then it will >automatically relocate the machine< to another lilly node with capacity
machines can now be cloned (ie. forked) with their live memory state instead of being turned off (loss of memory state).
of course you can pause/suspend/clone directly via ssh
all of the above rules hold true even if the machine needs to be relocated to another lilly node with capacity.
tldr; cpu/memory is expensive right now and now resource mgmt is built in as a platform primitive that you can self-host on your own infra 😎
henry retweeted
Following Ukrainian strikes on Russian data centers:
Tucker Carlson's views dropped 71%
Douglas Macgregor's views dropped 62%
TCN views dropped 34%
...or even otp-js 😱
OK being able to stick an entire runtime and agent development environment in a V8 isolate is dope.
@elixirlang folks if TypeScript (or any language) is to "win" in the AI era, it'll be because of stuff like this.
Legit think we might need an otp-rs...
youtu.be/ND_kRgU_WlY
OK being able to stick an entire runtime and agent development environment in a V8 isolate is dope.
@elixirlang folks if TypeScript (or any language) is to "win" in the AI era, it'll be because of stuff like this.
Legit think we might need an otp-rs...
youtu.be/ND_kRgU_WlY
Announcing tsc-rs (aka ts-rust), my complete rewrite of the Typescript compiler, type checker, and LSP, all in Rust.
This is a project I've had agents working on for 5 months now. I burned ~$400k of Codex tokens and got nowhere. Burned ~$20k of Opus and got there in 2 weeks.
I have not read a single line of the code.
tsc-rs is an open source drop-in replacement for tsc, and it's available now.
