|
BIG TECHAWAYS
Claude designs proteins that bind scorpion toxins
• Anthropic put Claude on a drug-discovery problem that receives little commercial attention: designing proteins that can bind scorpion neurotoxins, with the longer-term goal of developing an antivenom. Over 64 hours, the model used 12 tools to generate 318,484 designs and screen 553,052 design-toxin pairs.
• The team selected 180 designs, added 31 more, and sent all 211 to Adaptyv Bio and Twist Bioscience for testing. Seventeen met the criteria against one of four neurotoxins, with Cn2-B1 and Cn2-B2 performing best against the Cn2 toxin. No design targeting multiple toxins met the criteria against two or more toxins.
• In this experiment, AI generated designs, selected targets and binding sites, coordinated tools, filtered candidates, and moved the results out of the model for testing. The proteins have been confirmed to bind toxins created for testing; the next steps are tests with purified venom toxins, checks on whether the proteins block the toxin’s effect on sodium channels, and mouse studies. Anthropic has published all 211 designs and the binding data.
OpenAI buys Orchestra Bio for drug research
• OpenAI acquired Orchestra Bio and will bring its team and technology into Rosalind, its set of models and tools for life-science research. Orchestra helps drug-development teams connect scientific discoveries with research, decisions, and next steps. In the system, agents track research and campaigns while people guide the work, make decisions, and handle problems.
• One example involves a manufacturing batch failure that could affect a clinical-trial schedule. An agent can identify the activities affected, collect information, and compare response options. OpenAI says developing a new drug in the U.S. typically takes about 10 to 15 years from target identification to approval, and the announcement focuses on using agents to manage specialized processes alongside Rosalind’s tools.
Samsung and SK Hynix plan NAND upgrades
• Reuters reports that Samsung Electronics and SK Hynix are planning to upgrade their China plants to produce more advanced memory chips, citing four people familiar with the matter. Samsung is reportedly targeting 286-layer V9 NAND at Xi’an, while SK Hynix is targeting NAND with more than 240 layers at Dalian. The plans come as AI data centers increase demand for memory and YMTC remains a competitor.
• Xi’an currently produces chips with around 200 layers and accounts for about 40% of Samsung’s global NAND output. SK Hynix operates the Dalian NAND plant after acquiring Intel’s NAND business. The upgrade plans therefore sit between demand for AI data-center memory and restrictions on manufacturing equipment in China: both companies need annual U.S. licenses, while the BIS has said it does not intend to authorize technology upgrades or capacity expansion at these plants. Samsung says its plan complies with U.S. rules and does not expand capacity or build a new facility.
|
|
|
AI RESEARCH
AI judges need more than a default score
• Decision models can score AI automatically, but they should not be used with a default threshold straight away. Hamel Husain tested six models on 100 traces, meaning logs from an apartment-leasing agent’s sessions, with tasks such as finding failed bookings or missed handoffs to a human. After removing two QA samples, the set contained 98 production traces: 61 labeled failures and 37 traces labeled without a failure. Each model returned a score from 0 to 1 for the likelihood that a trace contained an error.
• A 0.50 threshold is only a default. A model could score every failing trace at 0.40 and every passing trace at 0.10: a 0.30 threshold would catch everything, while 0.50 would catch nothing. Husain recommends tuning the threshold on labeled data with five stratified splits, choosing it on the training portion and measuring it on the held-out portion, while requiring at least 80% precision and recall for alerting. If no threshold qualifies, the sample prompt asks the model to report that result instead of relaxing the condition.
• Performance has to be weighed against cost: Perplexity reached an F1 score of 0.79 at 0.82 cents per 100 evaluations, while OpenAI Decisions reached 0.87 at 3.72 cents. When a score represents confidence, calibration also needs to be checked: one Clef Flash score group at 0.08 still contained errors in 60% of traces, while OpenAI’s calibration was slightly worse in this experiment. Husain recommends building a separate judge for each failure mode and measuring again after every prompt or model change. The experiment used this set of 100 traces.
|
|
|
AROUND THE AI WORLD
• Figma adds 20 MCP partners: At Config in India, Figma announced 20 new MCP partners, including Amp, ClickUp, Dia, JetBrains’ Junie, Linear, Meta Muse, Mistral, Notion, OpenCode, and Sarvam Agents. MCP connects AI agents to external data and tools, while CEO Dylan Field says the standard still needs improvement and that exposing more data can increase abuse risk. He also raised the problem of asymmetric context sharing across an ecosystem.
• Uber uses an agent to clean up stale feature flags: Uber’s AutoClean finds feature flags that have finished rolling out, rewrites the code, and creates a diff for the owning team to review; the company says it has processed more than 16,000 flags and over 1 million lines of code. The workflow combines an agent with PolyglotPiranha, CI, static checks, and CodeCourt reviewer agents to check for errors, preserve behavior, and avoid removing the wrong flag before human review.
• Gemini for government workflows: Google Public Sector announced Gemini for government services in preview for U.S. federal, state, and local agencies, with specialized skills, data connectors, and partner agents. The workflows include policy analysis, FOIA processing, and procurement compliance, while the system can connect to Workspace, SharePoint, Outlook, Salesforce, ServiceNow, and federal regulatory sources. Google says the product inherits existing access controls and links answers to source files or regulations.
• Anthropic’s AI safety bill: Anthropic’s draft describes cases in which a model release could be blocked when an imminent risk could kill more than 50 people or cause more than $1 billion in property damage, involving biological weapons, a model evading developer controls, or a cyberattack. It proposes independent evaluators registered with the U.S. Commerce Department; if they find a catastrophic risk, they would have 72 hours to report it so the department could ask a Washington, D.C. judge to halt the release.
• Muse reaches 10.2 million U.S. downloads: Sensor Tower data cited by TechCrunch puts Muse at 10.9 million global downloads in its first 40 days, including 10.2 million in the U.S., with 30% seven-day retention and 23% 14-day retention. Meta allocated 39% of its total advertising spend in October to Muse, while daily active users reached 2.5 million on October 16, up 15% week over week.
|
|
|
GOODREADS
Research across multiple model sessions
• The paper presents Épi, a way to organize research around persistent history, hierarchical memory, branching, independent review, and fixed verification steps. It is for people designing research agents that need to continue work across multiple sessions instead of starting from an empty context. The paper reports performance results in the settings it tested.
Three scaling axes in agentic RL
• The paper separates three sources of scaling in agentic reinforcement learning: tasks, environments, and trajectories. The distinction helps agent builders see whether they are increasing the number of problems, the diversity of environments, or the number of attempts on the same task. The paper reports a 1.7x efficiency result across the models, domains, and RL settings it tested.
Check Point’s Agentic Readiness Ten
• The checklist covers visibility, identity, least privilege, traces, approvals, prompt injection, and pause controls. It can support an initial review for teams putting agents into production, with questions about who can observe, limit, and stop an agent.
A web built for agents could be worse for users
• James O’Donnell of MIT Technology Review treats agents as a third group of visitors alongside users and crawlers. The article covers websites blocking agents, separate checkout lanes for agents, and the possibility of interfaces, prices, and advertising being optimized for machine traffic. It examines how an agentic web could change the rules and the relationship between agents, websites, and users.
Context engineering for agents
• DeepLearning.AI’s course, updated on October 20, organizes agent reliability around agentic search, context management, memory, traces, and evaluations. Its discussion of context rot explains why a longer context can still contain repeated, conflicting, or outdated information. The course presents a framework for structuring work with context.
|
|
|
