M

i like this because i can deeply apply research in alignment with edifying incentives even as an entrepreneur

@mattparlmer facts

Robotics of this type is a solved problem to the degree that undergrads will be putting stuff that beats anything they’ve demo’d here up on HuggingFace by the end of the year

get its activations in context without spewing a single rambly token. no more observing by talking

they should invent a cantrip but for shutting the f up

> MDM-Prime-v2 is 21.8× more compute-efficient than autoregressive models I may be humiliated extremely hard with my diffusionLM skepticism.

This feels like physical product design's ChatGPT moment. This team just ran an autonomous agent against the entire chip design process: 219-word spec in, tape-out-ready silicon layout out, 12 hours later. The agent ran continuously against a simulator, found its own bugs, rewrote its own pipeline, and iterated to a working CPU! Chip design costs well over $400M and takes up to 9 years. Not because writing hardware code is hard (it is actually brutally hard) but because a respin costs 10 of millions. So teams spend more than half their total budget just verifying the design is correct before a single transistor is placed. That cost structure is why most chip designs never get built. Entire product categories that were previously too low-volume to justify a tape-out are now buildable.

A small ship I love: We made Claude.ai and our desktop apps meaningful faster this week. We moved our architecture from SSR to a static @vite_js & @tan_stack router setup that we can serve straight from workers at the edge. Time to first byte is down 65% at p75, prompts show up 50% sooner, navigation is snappier. We're not done (not even close!) but we care and we'll keep chipping away. Aiming to make Claude a little better every day.

always had better results pretraining w the finetuning data

The first thing I did at @tryramp was set up distributed tracing, structured logging, and metrics for Inspect, our background coding agent. We now have full visibility in to everything the system is doing: the browser, CF workers/DOs, @modal sandboxes, database calls, etc. Most importantly, Inspect now has visibility in to itself. It can self-triage runtime errors it encounters and create PRs to fix them. Every morning, it reviews the past 24 hours of its own @datadoghq dashboard, identifies systemic issues, new errors, and long tail latencies, and has a summary + PR waiting for me at 9am.

(continual) pre-training is not dead! some thoughts about cost per task (on cursorbench) being 2x lower, imo it can be due to: - new base model: seems straightforward but it's not imo, you need to optimize the inference/training stack (both for rl and consumer inference), GLM-5 > V3.2 doesn't mean glm-5-base > V3, they may not be equally malleable to post/mid training - more optimized kernel and inference stack/serving (most likely) - rl/mid-training with objective/data to make smaller CoT (most likely) - mid-training with more efficient arch: i would love for this to be true, and i can see how it's necessary if they use the previous base model generation and need efficient memory for long context, but since they also released some tricks with self-summary, i'd say unlikely? (they can and imo should be combined together for very long tasks)

DSPy is actually frontier agent welfare bc it affords an off ramp into the most appropriately tool shaped SLM

@instance_11 dmd

@init_malachi do you play?

i envision a world without duplicates

@instance_11 yeah, it used to be my main thing

@instance_11 ur literally me

@init_malachi thank you! the audio here is a snippet from a long session of improvising, imagining this video as i played; i then made the video based on the audio, it was a joy to draw from both sides

i’m going thru my old schizophrenic notebooks and it reads like 2026 twitter takes

@r0b0t_sp1der i’m still very much learning the ropes of audio, i wanted a very metallic/vibrational sound so i recorded the piano with an iPhone laying directly on the frame; the voice was generated from a similarly tinny source, i then applied some light filters and added a hard limiter

You must read God-Emperor of Dune understanding "precognition" as the computability and predictability of human individual/group actions, which will take off in our lifetimes; the computers/Bene Gesserit are the Fates, who Leto II frees us from. It is in dialogue with Foundation.

🧠AGC. Most surface CNNs use fixed patch sizes. AGC uses our diff. framework to dynamically learn the optimal receptive field for each channel & layer. It adapts to local geometry, beating fixed-patch methods.

🌊MeshFlow🐤: our alternative to Riemannian flow matching (RFM) on meshes. We skip the ODE solver by back-propagating through our Exp map to learn a static vector field. We are 16,000x faster in inference, use 97% less GPU memory, and deliver superior results than RFM.