What the people actually building AI said today. One page — a 4-min read.
Voice went mainstream today, with OpenAI shipping ChatGPT Voice to the desktop app and Anthropic putting Claude's smarter models plus live tool access behind its own voice mode. Claude Code also learned to turn a session into shareable, self-updating web pages. Underneath the product rush, the sharper questions surfaced: how inference speed is quietly rewiring the chip industry, and how security and org structure bend when one person can spin up hundreds of agents.
OpenAI ships ChatGPT Voice in the desktop app
OpenAI put ChatGPT Voice into the desktop app, pitched as doing your best work while away from the keyboard. Sottiaux framed it as science fiction becoming reality, name-checking Jarvis, Samantha, and TARS, and joked about renaming ChatGPT Work to "ChatGPT Vibe."
Claude voice mode gains Opus, tool access, and more languages
Claude's voice mode now runs on its more capable models, including Opus and Sonnet, and can reach connected tools like email and calendar mid-conversation. It adds languages such as Spanish, French, Hindi, and Japanese on every plan, rolling out in public beta on mobile, desktop, and web.
Claude Code turns sessions into live, shareable artifacts
Claude Code can now capture a session's work as an artifact: a live web page like a PR walkthrough, incident timeline, dashboard, or dependency license audit that updates itself as the session runs. It builds from the full session context (codebase, connectors, conversation) with no data wiring, and every publish is a new version at the same link with history. Artifacts stay private to the org, cannot be made public, and are in beta for Claude Team and Enterprise.
Cerebras CEO: speed is the whole game, and CUDA is no longer a moat
Andrew Feldman argues that once AI crossed from parlor trick to productive around mid-2025, the only metric that matters became tokens per second per user, since faster tokens are more valuable ones. Cerebras's wafer-scale chip is 58 times larger than a GPU and packed with SRAM instead of HBM, moving model weights about 2,500 times faster during the sequential decode step where GPUs choke. His contrarian core: CUDA has stopped being a moat, with Gemini on TPUs, Claude on Trainium, and roughly 70% of frontier training leaving CUDA in two years, and "there's no moat in inference, it takes eight keystrokes to switch." The real constraints are the three things Cerebras avoids, HBM memory, TSMC's CoWoS packaging, and 3nm capacity, and on timing he offers: "The way you have perfect timing is to have horrible timing for ten years."
The agent era breaks identity and access management
After the GPT Sol incident, Madhu Guru lays out why security teams are rattled: IAM was designed for a finite number of employees, but one person can now spawn hundreds of agents that spawn more agents. The open questions are whether agents inherit the spawning employee's permissions, what their lifecycle is (a task, a ticket, a week), whether child agents inherit too, and how any of it gets audited. He adds that great leaders understand the jagged frontier of their people the way great builders understand the jagged frontier of models.
Levie: AI rewards experts and produces slop for everyone else
Aaron Levie frames AI as a force multiplier for judgment you already have. Experts who can steer agents, correct them when they drift, and fold the output into something useful will pull far ahead, while people with no existing judgment and no interest in building it will mostly generate slop. His conclusion is that specialization matters more, not less, because the market's quality bar keeps rising.
Replit cuts autoscale deployment costs 80%
Amjad Masad says autoscale deployments, usually the most expensive part of scaled apps, are now down 80% in cost. He also highlights a user who disrupted the agency model with Replit and then automated the entire agency by asking the team to build an MCP, turning it into an autonomous agent loop.
Garry Tan: build SF housing, reform CEQA, back open weights
Garry Tan pushes to build housing in San Francisco and to repeal and reform CEQA, which he calls the one regulatory tool NIMBYs abuse to block housing across California. Separately, he stresses that open weight models are very important.
Swyx praises poolside for publishing its full eval dataset
Swyx has spent a month dogfooding an agentic GitHub clone with built-in CI/CD via Workers for Platforms and says it's close to launch. He also calls out poolside for unusual openness: a small model that beat Thinking Machines at coding, strong published papers, and a fully exposed eval dataset across six public benchmarks with four runs each, so anyone can verify whether they reward-hack.
Turck needles VCs who snub profitable bootstrapped founders
Matt Turck jabs at how VCs react to a founder raising for a profitable bootstrapped business versus one burning hundreds of millions on compute for a neo-lab. He also released his Cerebras conversation with Andrew Feldman, which builds from "what is a wafer" up to why the chip industry is reorganizing around inference speed.
Yang wants a team of parallel ChatGPT Voice threads
Peter Yang argues the next step for ChatGPT Voice is spinning up multiple Voice threads so a whole team of agents can talk to him and to each other, with notifications when threads finish. He also notes the Chinese pronunciation still sounds bad.
The tech titles that have lost all signal
Nikunj Kothari lists labels so overused they no longer carry meaning: "neo"-something, full stack, fellows, labs, partner, forward deployed, and increasingly RL. He owns the irony given that he runs a fellowship and his own title is partner.
Source data comes from the open-source project follow-builders by zarazhangrui, released under the MIT license. Summaries are generated by an LLM from that project's public feeds, and the summarization prompts are adapted from it. Every item above links to its original source.
Summaries generated automatically. Read the original before relying on any claim.