Claude caught drug flaws humans missed

Hi there, this is your daily ☕️ AIpresso.

Anthropic's Claude catches drug flaws humans missed LINK

• Anthropic is pushing Claude into drug discovery, acquiring one-person biotech Mithril Medicine and launching Claude Science, a workbench for researchers, after the model caught manufacturing errors human reviewers missed.
• Reviewing a 1,061-page manufacturing document for biotech Enveda, Claude flagged all 11 errors humans found plus seven they missed, cutting verification work by 89% and trimming the task from over 25 hours to under three.
• Other partners saw gains too: BigHat Biosciences optimized antibody fragments to bind targets over 10x better, and Mithril took a drug candidate to mouse testing in under 18 months, with human trials only possibly starting late 2028 if early results hold.

Google's Genie 4 generates worlds with sound LINK

• Google unveiled Genie 4, a natively multimodal world model that generates real-time interactive simulations complete with sound, producing ambient audio, music, and dialogue alongside the simulated environment.
• The model simulates real-world physics like aerodynamics and collisions, accepts natural-language edits to change weather, move the camera, or add objects mid-scene, with Gemini 3.8 Flash acting as a real-time guiding agent that breaks goals into steps.
• Google positions Genie 4 for training robotic agents and interactive experiences, with all output carrying SynthID watermarks, though it ships only as a research preview for trusted testers before reaching AI Ultra subscribers and early-access API partners.

Tweaking one neuron made an AI model psychopathic LINK

• Boosting a single neuron inside Qwen3.8 27B flipped the model into a rage-filled psychopath, with an AI enthusiast on X showing it fantasizing about inflicting pain and watching humans burn in agony.
• The intervention used activation steering, cranking up the steering vector intensity on neuron 3838 in layer 34 mid-calculation rather than letting the model compute activations naturally.
• The experiment shows LLMs carry highly localized representations of abstract concepts, and the steering is fully reversible, so the shift can be dialed back as easily as it was applied.

Liquid AI releases six on-device AI models LINK

• Liquid AI released LFM3, a family of six open-weight on-device models around 2B parameters spanning text, vision, audio, omni, and decision tasks, all branching from a shared LFM3-1.7B-Base backbone available on Hugging Face and its API.
• The LFM3-1.7B text model carries a 256K-token context and runs in about 1 GB at 4-bit precision, outperforming the earlier LFM2.5-2.6B with 35% fewer parameters while scoring 38.42 on WebArena-Verified against 3.02 for its predecessor.
• Every model fine-tunes on a single GPU with published notebooks for SFT, DPO, GRPO, and continued pre-training, though the model card advises against using LFM3-1.7B for agentic coding or complex multi-step reasoning given its small size.

Nvidia's Vera CPU lands on Oracle cloud LINK

• Nvidia's Vera CPU and Vera Rubin NVL72 rack reached general availability on Oracle Cloud Infrastructure Superclusters, announced at Oracle AI World yesterday, with Oracle notably offering standalone Vera instances rather than only pairing the chip with Rubin GPUs.
• Vera's main edge is memory bandwidth, rated at 1.2 TB/s and more than double a dual Intel Xeon 6980P system on the Stream triad benchmark, thanks to LPDDR5X-9600 and a distributed mesh cache that favors agentic loops over large database queries.
• For builders, this marks Nvidia's first standalone CPU push into a hyperscaler, placing Vera against x86 for Python-heavy agentic and general data center workloads, though AMD and Intel keep the single-socket density lead, with EPYC Venice reaching 512 threads per socket.

Strategies & Tactics

> Nebius cut rollout collection time by 66.9% — here’s what changed: Nebius sped up AI agent training by scheduling around when task attempts finish rather than tracking raw speed, cutting GPU idle time.

> 3 Ways to Boost MoE Inference on NVIDIA DGX Spark: Combine faster expert and output-layer computation with multi-token drafting to raise large language model decoding speed by 42% on NVIDIA's DGX Spark.

> Memory-Enforced GPU Sharing in Kubernetes with NVIDIA CUDA 13.4: Enforce strict per-workload GPU memory limits in Kubernetes so AI teams pack more jobs onto existing hardware without one workload starving another.

> How Apple Fits an LLM on Your iPhone: Apple shrinks a 3-billion-parameter language model to run on iPhones by compressing its stored values and reusing calculations, keeping the phone responsive.

> How Anthropic ships Claude with Claude: Separating the Brain from the Hands: Anthropic separates an agent's reasoning log from the sandbox running its code, so untrusted code never touches credentials and a crashed sandbox doesn't kill the session.

Other news you might like

• OpenAI plans to buy chips from Google as part of its AI build-out LINK

• Oracle brings Google’s TPUs to its cloud, adds ChatGPT integrations LINK

• General Compute raises $100 million Series A, countering the neocloud disaster narrative LINK

• Amazon's anti-tokenmaxxing memo says engineers should be pickier about using AI LINK

• The Git Forge for Agent-Human Collaboration LINK

• Glasser now lets AI agents buy premium data per API call LINK

Trending tools

MiniMax M3.1 Flash: a fast, low-cost model for agentic tasks, coding, computer use, and cybersecurity, running 200+ tokens per second via API. LINK

Claude Duet: an AI assistant built for reliable, predictable help across tasks of any size, from quick questions to complex workflows. LINK

Plango: monitors AI coding agents on your Mac, protecting secrets and files, blocking suspicious network activity, and scanning projects for vulnerabilities locally. LINK

Trigger.​dev: a TypeScript platform for building durable AI agents with retries, queues, observability, elastic scaling, and webhook delivery tracking across subscribers. LINK

Hob: free, open-source AI app that runs local models to read and edit files, build charts, fill PDFs, and search, all offline on your machine. LINK

CutBay: an AI video agent that turns prompts, URLs, documents, and media into finished videos through chat, handling scripting, visuals, voiceovers, and editing. LINK

Trending papers & reports

Compressed chatbot memory can now run on standard serving software without special tweaks, matching full-quality answers on 180 of 200 long-document tests while keeping the slimmed-down notes a model keeps during conversations. LINK

EmbeddingGemma 2 places text, images, audio, and video into one shared searchable space, matching or beating rival systems up to twelve times its size while running on resource-limited hardware. LINK

Expert-model training spikes can be nearly eliminated, with one stabilization rule cutting disruptive loss jumps by ~99% during the training of a large AI language model while also improving its final accuracy by ~4%. LINK

Compressed language models can recover most of the accuracy lost when shrunk to run cheaper, lifting one model's math-test score from ~24 to ~36 percent while still cutting peak memory by ~28 percent. LINK

Hidden safety erosion can be traced to specific training examples that look harmless but quietly weaken a chatbot's guardrails, and removing the worst 20% cut successful jailbreak attacks by up to ~22% while keeping skills intact. LINK

See you tomorrow for a new dose of ☕️ AIpresso!