AI Builders Digest

Diffusion will win on inference economics, not on raw intelligence

Stefano Ermon's bet is that autoregressive models have a structural problem nobody can engineer away: generating one token at a time is memory bound and maps badly to GPUs, while diffusion processes many tokens at once and looks at inference like it looks at training. "The bitter lesson is that the more parallel solution is the one that is eventually going to win." Inception's Mercury models now match the speed-optimized tiers from frontier labs on benchmarks while running significantly faster, and voice agent company Open Call moved off Cerebras custom silicon to get comparable speed on ordinary NVIDIA GPUs. He estimates 20 to 30 percent of workloads today are latency bound, and flags a second advantage: because diffusion refines coarse to fine, you can score and steer a partial output instead of waiting for the finished object.

• #hardware
• #open-source
• #products

Rauch: the real slop risk is that reading itself gets discounted

"Reject non-understanding." Rauch argues slop grenades are not just code and PRs, and that the actual damage is people giving up on reading at all, worn down by low-quality unverified AI prose thrown at them constantly. His example is a thread going viral that credits a performance win to a compiler change, when the PR description itself has the AI saying the gain came from changed algorithms and data structures and that a similar win was available anyway. What he wants instead is AI in the service of understanding the universe and enhancing human cognition.

• #developer-tools
• #culture

Tan's current bug-fix loop: Capy running GStack autoplan on prod issues

Tan says his favorite way to fix bugs now is capy.ai with GStack /autoplan pointed at a live production issue, running GPT-6 at medium reasoning. Away from tooling he is pushing Garry's List house parties as the answer to people who get mad on social media and have no idea what to do next, with education and merit as the organizing hook. Also: be chalant, don't LARP.

• #agents
• #developer-tools

Claude limits went from barely usable to basically unlimited overnight

Yang noticed Claude's usage limits flip from barely usable to effectively unlimited with no explanation for the change. He is also building a conversational Japanese tutor on Google's new Gemini audio API, structured as 10 lessons with 10 phrases each, and tried Antigravity for the first time to do it, with no obvious place to send feedback.

• #products
• #developer-tools

Plato's Protagoras, novelized by a human and filmed by Opus 5.5

Shipper wrote a novelization of Plato's Protagoras, then handed it to Opus 5.5 to turn into a movie, releasing scene one and scene two back to back. The interesting part is the pipeline: a human does the adaptation work, the model does the cinematography.

• #products
• #creative-tools

One year of making videos with Claude Code, and the iteration cost collapsed

Thariq marks a year since one of the first posts about using Claude Code to generate video. The ones he made back then each took a long time to get right, iterating with the model and pointing out details that were wrong over and over. His point is the gap between that grind and what the same workflow produces now.

• #products
• #creative-tools