Meta’s Muse agent runs in a cloud VM and uses a computer

#1 ▶️

Meta is pivoting again... everything you missed from Connect 2026

Fireship

#2 ▶️

What product looks like when coding is solved | Geoff Charles (Ramp CPO)

Lennys Podcast

#3 ▶️

How to scale intent, quality, and artistry with Al | Katie Dill (Stripe)

Lennys Podcast

#4 𝕏

Santiago demonstrated prompting Codex to use Hyper3D’s Rodin MCP to generate three different 3D car models, then having GPT-6 write a racing game with them. He said the process required no code, though that claim was not independently verified.

#5 𝕏

Jeff Dean commented that Waymo’s safety data was improving: the latest comparison covered 270M miles and showed a 20X better rate of crashes with serious injury than human drivers, versus 13X better over 170M miles in Mar 2026 and 10X better in an earlier comparison.

Also covered by: @Jeff Dean

#6 𝕏

Sebastian Raschka shared the fifth installment of “Reasoning from scratch,” covering log-probability scoring, token probabilities in PyTorch, numerical stability, and self-refinement through critiques and revised answers. It also includes MATH-500 evaluation results, takeaways, and next steps.

Also covered by: @Sebastian Raschka

#7 𝕏

DeepLearning.AI shared Andrew Ng’s explanation of why AI engineering tactics should adapt to a project’s stage for speed and reliability. The letter in The Batch covers evaluation pipelines and metrics, scalable software architecture, and product feedback loops.

#8 ▶️

Claude Opus 5.5 Is About to DOMINATE Kalshi & Polymarket

All About AI

#9 𝕏

Jason Zhou demonstrated a dynamic 15-second motion graphics video he said he generated in one shot from Opus 5.5 using a résumé-showreel prompt. Surprised by the result, he said he planned to investigate what Anthropic had put into it.

#10 𝕏

Garry Tan shared his “favorite way to fix bugs”: using @capydotai with GStack /autoplan on a production issue and GPT-6 medium reasoning.

#11 𝕏

Santiago commented that Reef’s /reefine command lets builders refine agent behavior through instructions and works with any agent harness using its adapter. He suggested—explicitly as an assumption—that the adapter could evolve over time and be personalized.

#12 in

Udi Menkes recapped Claire Vo’s “The Last Roadmap,” warning that AI-enabled velocity—approximately 3x as many PRs and 40 Grok bots—can create backlog, parity, and ship-and-forget traps. He advises AI product builders to pursue three to five long-term bets over one to two years, using AI to test evidence quickly while protecting customer trust.

#13 𝕏

Guillermo Rauch warned that persistent low-quality, unverified AI prose could cause people to discount reading. He cited a viral thread attributing a performance gain to a compiler change and said the referenced PR description instead attributed it to changes in algorithms and data structures. He expressed a desire for AI to advance understanding and enhance human cognition and creativity.

#14 𝕏

Boris Cherny expressed anticipation about what people will build, commenting on a ClaudeDevs post referencing a new plugin portal, MCP/skills for Claude, and 110x usage growth.

#15 𝕏

Yann LeCun commented that neural nets already support protein-conformation prediction, materials-property prediction, driving assistance, and medical-image analysis, with some pouring chemicals in “breakers.” He emphasized that none of these applications are based on LLMs.

#16 𝕏

Yann LeCun argued that human-level performance is not the same as inventing new solutions to new problems. Modern AI assistants can achieve human- or superhuman performance on some tasks, he said, but largely by learning from human-produced solutions through imitation learning, RLHF, or other methods that compile existing knowledge.

#17 𝕏

There's An AI For That demonstrated a motion graphics video it says was produced by Opus 5.5 from a prompt requesting a resume-style showreel explaining how prompt injections work.