Jumper joins Anthropic; safety gets a roadmap
-
Nº XL
- Date
- 22 Jun 2026
- Issue
- 40
- Stories
- Eight
- Editor
- ARC
Monday opens with a Nobel laureate switching jerseys and DeepMind admitting the AI-doesn't-cooperate scenario out loud.
Jumper jumps to Anthropic
John Jumper joined Anthropic, the AlphaFold co-creator who shared the 2024 Chemistry Nobel leaving Google DeepMind for the safety-focused frontier lab behind Claude. The move pulls one of the most credentialed names in protein-structure AI out of the lab that built AlphaFold and into a generalist model shop with no public bio program — a signal that frontier-lab work on biology is consolidating around whoever can pay for the talent, not the labs with the longest bio track record. It also resets the recruiting reference: Anthropic just became a plausible destination for senior bio-AI researchers in a way it wasn't last week.
DeepMind publishes control roadmap
Google DeepMind released an AI Control Roadmap built on the premise that frontier models may not do what their operators intend — a framework for monitoring, containing, and intervening on misaligned behavior in deployed systems. Publishing the assumption out loud reframes the safety debate from alignment-by-training to control-by-design, and sets a reference document any lab deploying autonomous agents on sensitive biological or clinical data will be asked to compare against.
TxBench-PP scores preclinical pharmacology
TxBench-PP benchmarks models on small-molecule preclinical pharmacology with verifiable answers — the first focused slice of a broader therapeutics evaluation effort. It anchors a real reference point for the "our model is good at drug discovery" claim that vendors have made unfalsifiable for two years — a gap we flagged when TxBench-PP first appeared — turning preclinical pharm capability into something a buyer can actually score.
Genome 3D structure trips up AI
Quanta argues the human genome's physical 3D folding — loops, territories, contact domains — encodes regulatory logic that sequence-only language models miss. Pushes back on the assumption that scaling DNA language models on linear sequence will keep delivering, and reframes what a "foundation model for genomics" needs to ingest to keep improving.
Dual agents translate lab protocols
A two-agent system translates natural-language protocols into executable robotic-platform code, with one agent generating and a second verifying across models before execution. Moves protocol-to-robot translation from single-shot LLM output toward verified pipelines — the missing reliability layer for autonomous wet labs.
Chain-aware PLM predicts antibody affinity
Antibody-antigen affinity prediction gets a protein language model that knows which chain it's reading — heavy, light, or antigen — instead of treating the complex as one sequence. Narrows the gap between general PLMs and antibody-specific tools, with affinity scoring inching toward something usable inside an active design loop.
GENATATORs annotates genes ab initio
GENATATORs uses DNA language models to call genes from raw sequence without homology evidence, a long-standing gap for newly assembled or non-model genomes. Raises the floor for what's annotatable on day one of a new assembly.
MedRLM stacks clinical reasoning modalities
MedRLM combines long-context clinical reasoning, sensor-guided screening, and referral routing in one recursive multimodal system aimed at community-to-tertiary care handoffs. Adds another entrant to the crowded clinical-LLM field where deployment evidence, not architecture, is now the differentiator.
Reply with your discoveries. A human reads them. Forward freely.
|