- NATURAL 20
- Posts
- AI's Next Tests Are in Labs, Power Infrastructure and Security
AI's Next Tests Are in Labs, Power Infrastructure and Security
PLUS: Meta upgrades Muse memory, Antigravity gets voice mode, and Meta's strategy chief joins Fireworks.
Wes Roth
November 09, 2026
In partnership with
You know the problem. Now build the fix.
You've spent years close to a broken industry or process. You know exactly what would fix it. But you’ve never had the time or resources to make it happen.
With Lovable, your expertise is all you need. Describe the solution to your problem — a website, a dashboard, an internal tool — and get professional-grade software back. You make every call on how it works, because you're the one who knows what it needs to do.
Start today and take $5 off your first month.
Discount of $5 applies to first month of any paid plan. You need to sign up via this webpage in order to claim the discount. The discount applies to new and returning users without an active paid plan. Terms apply.
Today:
OpenAI tests AI agents on 100 raw wet-lab biology tasks
Anthropic reportedly secures $3.4 billion in gas generators for data centers
Nvidia opens an independent review of reported foreign intelligence targeting of its CEO
Meta improves Muse and Meta AI with specialized memory and task models
Google Antigravity brings live voice conversation to coding agents
AI's Biggest Questions Now Reach Beyond the Chat Window
OpenAI has released a new test of whether advanced AI can interpret messy biological experiments. Meanwhile, Anthropic is reportedly securing power-generation equipment on a massive scale, and Nvidia's board has opened an investigation after reports raised national security questions involving its chief executive.
These stories reflect the growing reach of the AI industry. Better reasoning models need to understand real evidence, expanding data centers need dependable electricity, and the companies supplying critical chips face increasing scrutiny over how sensitive technology is protected.
OpenAI introduced BioBasicsBench, a benchmark developed with biotechnology company Ligo to measure how well AI agents understand the evidence scientists handle in everyday wet laboratories. Its 100 tasks use raw experiment materials rather than only the curated charts and images commonly published in scientific papers.
The tasks cover protein methods, microscopy and flow cytometry, quantification and assays, nucleic acids, and molecular cloning. They include DNA gels, colony plates, microscope images, sequencing data, instrument exports and laboratory notes. Models must interpret measurements, notice flawed controls and sometimes explain why an experiment cannot support the conclusion being requested.
OpenAI evaluated seven models at 37 combinations of model and reasoning effort. Across 18,500 scored responses, GPT-6 Astra performed best in the reported xhigh comparison, earning an average of 66.7% of available credit. GPT-6.1 Sol earned 62.0%, while GPT-5.6 Sol scored 43.8%.
The strongest model handled some experimental ambiguities better than earlier systems. In one example, it recognized that overlapping microscope fluorescence channels could not establish whether cells were dead. In another, it spotted an incorrectly labeled antibiotic plate that made a cloning control unreliable.
The report also documents failures. Astra chose the wrong colonies after misinterpreting the bands in a PCR gel, overlooked observations contradicting a preferred explanation, and overcounted colonies on one plate by about 12%. Those mistakes show why better benchmark scores do not yet make AI a dependable substitute for experienced lab scientists.
BioBasicsBench measures interpretation of laboratory evidence, not whether an agent can independently perform physical experiments. The researchers hope it will make scientific assistance more measurable and show where stronger controls and human review remain essential.
Anthropic has reportedly agreed to secure approximately $3.4 billion worth of gas-fired generators for AI data centers. The deal covers commitments for about 1.7 gigawatts of heavy-duty generation equipment sourced from OnePWR Solutions, a supplier based in San Antonio, Texas.
The commitments were reportedly purchased in late September and include engineering and construction capabilities needed to install the generators. Anthropic did not acquire OnePWR, and the supplier's employees were not transferred to the AI developer.
The scale is significant. A 1.7-gigawatt equipment commitment shows how much electricity frontier AI infrastructure may require as companies train models and serve growing numbers of users. Securing equipment could help Anthropic prepare power capacity as data-center projects compete for grid connections and other resources.
The arrangement forms part of a broader push to expand computing capacity. Reuters previously reported that Anthropic expected to spend more than $500 billion over the following decade on AI infrastructure, although that figure was not a detailed schedule of completed projects.
The precise locations, construction timelines and operating plans for the reported generators were not specified in the accounts reviewed. Reuters said it could not independently confirm the transaction, and Anthropic and OnePWR had not responded to its requests for comment.
Nvidia's board announced an independent investigation on November 8 following reports that CEO Jensen Huang may have been the target of a foreign intelligence effort. A committee of independent directors, supported by outside counsel, will oversee the review.
The inquiry follows reporting about Huang's relationship with Alice Huang, a former executive at Singapore-based Nvidia customer Megaspeed. U.S. officials reportedly identified her as working with Chinese intelligence, while the company has faced separate scrutiny related to access to advanced Nvidia chips and possible export-control concerns.
Jensen Huang acknowledged that his relationship with the executive was personal and apologized publicly to his family. The central business and security question is whether any confidential corporate information or national security interests were affected.
Nvidia said its board had no reason to believe either company secrets or national security information had been compromised. The company also said Huang supports the inquiry and will cooperate fully, and that he will continue serving as CEO during the review.
🧠RESEARCH
From Autonomous Agents to Cognitive Symbionts argues that AI should enhance human thinking, not only automate tasks. It proposes persistent companions that remember context, notice overlooked details, and assist at moments. The paper is conceptual rather than an experimental demonstration and examines privacy, dependence, fairness, and governance risks of assistance.
VFlash accelerates visual planning for robots by predicting points, boxes, and motion trajectories together instead of generating coordinates token by token. Researchers report roughly 25 times faster preliminary 3D motion forecasting and 70.2% success on four MolmoSpaces task suites, compared with 62.6% for a baseline world action model in tests.
Muir proposes a more efficient linear-attention design that allocates gating capacity where state updates are most sensitive. Researchers report lower language-model perplexity than Transformer and Gated DeltaNet-2 baselines, improved long-context recall, and 2.7 times faster training than Transformer at a 32K context. These are results from evaluated small research models.
📲SOCIAL MEDIA
🗞️MORE NEWS
Meta released specialized models trained with capabilities from Muse Wonder for its Muse and Meta AI assistants. The company says the updated Muse model scored 47% on its Personal Superintelligence Index, up from 38%, and improved its ability to retain critical context, with 90% success on an internal memory evaluation versus 76% previously. Meta also reports a 67% improvement in completing everyday tasks when unexpected obstacles arise.
Google launched voice mode for the Antigravity app and CLI, allowing developers to give coding tasks, steer running agents and respond to permission requests through spoken conversation. Powered by Gemini Live, it can read back results while displaying the agent's work on screen, with nine voice choices and adjustable playback speed. Destructive actions still require on-screen approval, and CLI voice mode is supported in version 1.5.0 or later.
Anthropic updated Claude Code cloud sessions to make project setup more automatic for Pro, Max and Ultra users. Claude can determine a repository's required dependencies, network permissions and API keys, then save the prepared environment for future sessions, while keeping keys in protected storage rather than showing them to the model. The change is intended to reduce setup work before agents begin coding and testing.
Dave Wehner announced that he is leaving Meta, where he served as chief strategy officer after previously holding the chief financial officer role, to become CFO of AI infrastructure company Fireworks. He said his experience at Meta convinced him that companies will want specialized AI models built around their own data and expertise. His move brings a longtime senior Meta executive to an AI inference and model-serving business.
What'd you think of today's edition? |
Reply