AI Builders
Digest

The Takeaway: Speed in AI inference is now the dominant factor, enabled by massive specialized chips that outperform GPUs for real-time agentic work. Cerebras CEO Andrew Feldman, who built the world's largest wafer-scale chip (58x larger than a GPU), explains how big chips solve memory movement bottlenecks in inference. Unlike GPUs relying on slow HBM memory, Cerebras uses SRAM on a dinner-plate-sized chip for 2500x faster weight movement. This enables blisteringly fast tokens, transforming AI UX from waiting to real-time. Key insight: AI's 'broadband moment' like Netflix shifting from DVDs to streaming. Feldman notes three hidden bottlenecks (HBM, CoWoS, 3nm) that Cerebras avoids, plus surging CPU demand from agents. On CUDA: 'It takes eight keystrokes to move from a GPU to us.' The $20B+ OpenAI deal underscores massive demand. Quote: 'For AI, bigger chips are undoubtedly the best way to go.'

Anthropic's Claude Blog announces Artifacts in Claude Code: sessions now generate live, self-updating web pages like PR walkthroughs, dashboards, incident timelines, and release checklists. Built from full session context (codebase, tools, conversation), they refresh in place for teams. Private by default, with org controls. Practical for debugging, reviews, and collaboration without extra infrastructure. Available in beta for Team/Enterprise.

Replit CEO Amjad Masad highlights autoscale deployments down 80% in cost. His chess autoresearch agent earned a 'PhD in modern LLM finetuning.' Viktor used Replit to disrupt agencies, then built an autonomous agency via MCP agent loop. Strong momentum in developer platforms and agentic workflows.

Box CEO Aaron Levie argues AI amplifies experts who steer agents and incorporate outputs into high-quality work. Specialization remains crucial as tools raise market expectations. Non-experts produce slop; experts achieve far more.

Vercel CEO Guillermo Rauch announces Python code now starts 2x faster automatically. AI Gateway sees continued rapid enhancements from the team.

Claude adds support for Opus and Sonnet in Voice, tool access mid-conversation, and more languages (Spanish, French, Hindi, Japanese). Rolling out on mobile, desktop, web.

Swyx has been dogfooding an agentic GitHub clone with built-in CI/CD via Workers. More features coming; hiring for swyx inc to influence roadmap.

Swyx praises Poolside AI for openness: excellent small coding model beating others, full eval dataset published across benchmarks with multiple runs.

Peter Yang excited about multiple ChatGPT Voice threads for team-like collaboration. Shares before/after examples and suggestions for notifications and pronunciation.

Garry Tan pushes to build housing in SF and repeal CEQA. Emphasizes importance of open weight models.

Madhu Guru: Great leaders understand the jagged frontier of their people like AI models. Discusses security challenges with infinite agents inheriting permissions.

Thibault Sottiaux from OpenAI teases cool tech and shares Jarvis-like Voice experiences in desktop app. Multiple posts on impactful work.