Chief Gearhead of WizOps (https://nitter.cf/t.co/K90uAzoJgi)
Redwood City, CA
Joined March 2007
- Tweets27.5K
- Following541
- Followers1.5K
- Likes8.6K
Testing agent-desktop. It's basically scraping the structured map of the desktop spanning apps and components intended for accessibility. It's snapshot could have sensitive info buried in JSON but should be fine if you're a boring person.
Most of that number is Nemotron stream audio encoder sidecar weights. Real weight reduction is much lower but quantization does its job. The new version taps directly into primary ASR instead of the sidecar, creating complications, ofc. 🤞
Don Park retweeted
Come test drive all these models in @Flux!
We’re big believers in giving our customers access to frontier intelligence and the freedom to switch between models without having to switch their entire agent environment.
Different models have different strengths, and those strengths change constantly
Our job is to give them the best possible hardware engineering environment, regardless of which model is behind the wheel.
Today, we're releasing @hardwarebench, the first hardware engineering benchmark built in partnership with the world's most advanced hardware companies.
Measurement has always been core to our work at @arenaphysica: the very first version of Atlas shipped with our internal framework for evaluating frontier models on real hardware engineering tasks.
HW-Bench evaluates frontier model performance on production hardware data and captures the full hardware context across tools and file formats. Each task is graded against rubrics and golden answers reviewed by subject matter experts at our partner companies, who build some of the most advanced hardware in the world, from the chips behind modern supercomputing to the autonomous systems and vessels redefining defense at sea and beyond (more coming soon).
We'll be building towards jobs that span disciplines, like designing, manufacturing, and testing a 140 GHz, 512-element D-band phased array. To start, see how 15 frontier models performed at designing a 2.45 GHz antenna at hw-bench.com.
I built an experimental WebSockets-over-P2P tunnel using iroh yesterday. It was usable but shelved because tunnelling cost in performance. Revived it after realizing I could just use the WebSockets API and replace the rest with QUIC + HTTP/3 WebTransports compatibility layer.
Turns out it's better to hack specific WebSockets libraries directly. For the experiment, I hacked axum to cut through the protocol to send and receive more efficiently and benefits of QUIC's multiplexing drips down.
Yes. AI can change individual weights, its own or another's, as easily as it can change an array item or we can choose to stir our brain matter with a hot-iron any time. Crazy, huh?
Dragology is a TypeScript library that applies constraints declaratively to draggable SVG objects.
Drag-and-drop makes the digital graspable, but building it well is fiddly enough that many don’t bother. Declarative Dragging (with Josh Horowitz & Jeff Heer, #uist2026 best paper award), lets you make an element draggable just by listing the states dragging it can reach.
Wrote a radio jockey agent to use as radio stations I can call. Available radio stations in an OPML resource exposed via MCP, like how WebMCP works. This is actually TopicRadio idea reprojected using a phone metaphor. Radio and TV stations, phone numbers, doorlock, no difference.
Maybe I'll add an agent that talks like the passive-aggressive elevator at Restaurant at the End of the Universe when someone calls: "Hello, I am to be your elevator for this trip to the basement of your choice."
Now imagine seeing `AROUSED: 89%` blinking at the corner of your smartglass while talking to your customer.
Our new Audio Decision model is INSANE. This is the fastest we've ever shipped since I joined and it might be our finest work. The research team at @hume_ai is at the top of their game.
What's weird is that these mood-sensing intimate talks feels creepy while spying on people you're talking by monitoring and inferring intents and emotions does not despite it being a power move, at least for sales folks.
Sample (incomplete) turn annotations for Open Yap 1k dataset made with Tyger and small GRU models from Pipecat.
huggingface.co/datasets/donp…
Distill Tyger duplex decisions into a shareable small-model dataset - behind-the-scene:
github.com/donpark/idfon/iss…
Weight-Space Programming
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.…