Google VP Josh Woodward
Josh Woodward announces that Gemini 4 Hydrogen is available in the Gemini app. He positions it as a fast model for a broad audience. The post gives the model name and where to use it, with speed and capability characterized by the maker.
Thibault Sottiaux, Codex and ChatGPT at OpenAI
Thibault Sottiaux says the team announced cybersecurity tools it is developing. One area is automated patching and the other is automated red teaming, covering remediation and proactive searches for weaknesses. He links to the announcement recording. The post describes the tools as in development.
Replit CEO Amjad Masad
Amjad Masad says Replit Agent can now make purchases for a business at the user’s request. His example is buying ads to acquire customers. This extends the agent’s scope to actual spending. The post gives no details about approval, budgets, or refunds.
Vercel CEO Guillermo Rauch
Guillermo Rauch uses the UI SDK to explain how he sees frontend frameworks changing in the agent era. He compares it to the x86 instruction set and wants agents to assemble components from lower-level behaviors on demand. He argues that accessibility, performance, and visual consistency should be packaged as reusable behaviors. The goal is to reduce token use, errors, and work while allowing agents to generate varied interfaces. The post presents a design position without comparative tests.
Anthropic researcher Alex Albert
Alex Albert says getting good work out of Claude often starts by identifying the context it lacks and finding ways to supply it. He describes Claude Meetings as one way to fill part of that gap and says it has been helpful internally. The post gives a team observation without a quantified result.
Box CEO Aaron Levie
Aaron Levie shares Box’s early-access testing of Gemini 4 Hydrogen. The Box AI Complex Work Eval gives Box Agent complex enterprise document tasks from fields such as finance, consulting, and life sciences. He reports that Hydrogen was 5 percentage points more accurate than Gemini 3.8 Flash, took less than half the time, and used about 12% fewer tokens on a typical task. These figures come from Box’s own evaluation. Levie also says Box AI Studio will make the model available for building custom agents.
Anthropic AI assistant Claude
Claude’s official account introduces Claude Meetings. Claude joins as a visible participant. Attendees can ask questions aloud or in chat, or ask it to start a document. The team can see the work in progress and correct it as it takes shape. Notes, action items, transcripts, and work created during the call are collected on its meeting page in Claude. The product starts with Google Meet, and the account says other video conferencing platforms will follow.
No Priors — Powering AI-Driven Drug Discovery with President of Isomorphic Labs Max Jaderberg
Key takeaway: Max Jaderberg wants drug teams to iterate on designs on a computer first, reserving costly synthesis and experiments for candidates worth validating.
Jaderberg is president and a founding team member of Isomorphic Labs. At DeepMind he worked on generative models and reinforcement learning. He says AlphaFold’s progress in protein structure prediction convinced the team that deep learning could generalize on some biological problems. Isomorphic Labs wants to extend that approach to molecular design and disease mechanisms.
His first question is whether lessons from one program can carry over to the next. A program focused on a single target generates specialized knowledge that may be hard to reuse elsewhere. Sharing a modeling platform across programs is a way to carry experimental findings into the models and test whether the next task actually gets faster. He also points out that many existing datasets are by-products of years of drug development and have sampling biases. The team generates data deliberately to explore which distributions help models generalize to new problems.
He sets a concrete target for computational models. The company marks experimental accuracy on its internal leaderboards and aims to reach that line on specific prediction tasks. Automation can shorten laboratory operations, but reactions and cell cultures take their own time. As he puts it, “We still need to go into the real world to validate.” The aim is to do more design iterations on a computer, then test novel hypotheses in experiments. Clinical safety and efficacy still require real-world validation.
Preclinical measurements also need a defined purpose. Jaderberg suggests combining in vitro measurements, signals from organoids or organs-on-chips, and predictive models to study which signals better anticipate human toxicity. He notes that human trial data is limited and presents these methods as a research direction. Isomorphic Labs builds its drug design engine in-house and uses it in its own programs and in collaborations with pharmaceutical and biotech companies. He gives no date for a first clinical trial.