Tech Digest – February 23, 2026
AI Autonomy & Capability Thresholds
AI Agents Now Work 14-Hour Shifts — and the People Building Them Say AGI Has Arrived
METR, the AI evaluation organization, estimates Claude Opus 4.6 has a 50% autonomy time horizon of approximately 14.5 hours on software tasks — the highest ever reported, though METR notes the measurement is “extremely noisy” as its current task suite is nearly saturated. The finding lands alongside a growing consensus that the threshold has been crossed: the LessWrong community published a widely-discussed post titled “AGI Is Here,” arguing that Opus 4.6 and GPT-5.3 can think, plan, and “meaningfully attempt most tasks a human can.” Sam Altman, in an interview with The Indian Express, said his “inside view” now points to “a faster takeoff than I originally thought.”
Note: When the evaluation benchmarks saturate before the models do, the conversation stops being about whether these systems are capable enough — and starts being about how fast institutions can adapt their workflows to a world where they are.
Sources: METR, LessWrong, Sam Altman / The Indian Express
Anthropic Launches Code Security Tool — Cybersecurity Stocks Drop 8% in a Day
Anthropic released Claude Code Security, a tool that scans codebases for vulnerabilities. The market response was immediate: CrowdStrike fell 8%, Cloudflare 8.1%, and SailPoint 9.4%. The launch reflects a broader pattern — software engineering now accounts for nearly 50% of all agentic activity on Anthropic’s platform, according to the company’s own research on measuring agent autonomy.
Note: An AI lab shipping a security product and erasing billions in cybersecurity market cap overnight is a preview of what happens when capability concentration meets adjacent markets. Any vendor category can become a feature.
Sources: Anthropic, Bloomberg, Anthropic Research
Gemini 3.1 Pro Solves a Math Problem No Model — and Few Humans — Could Before
Google DeepMind’s Gemini 3.1 Pro solved a FrontierMath Tier 4 problem that no previous model had cracked, pushing machine reasoning into territory that exceeds the reach of most professional mathematicians. Epoch AI, which maintains the benchmark, confirmed the result and noted that Gemini 3.1 Pro’s overall FrontierMath performance is comparable to the previous Gemini 3.0 Ultra — a model that was significantly larger.
Sources: Epoch AI
Infrastructure & Capital Scale
OpenAI Targets $600 Billion in Compute Spending by 2030 — Farmers Get $120K/Acre Offers
OpenAI has reset its compute spending expectations to approximately $600 billion by 2030, according to CNBC. On the ground, that capital is already reshaping land markets: The Guardian reports US farmers are fielding offers exceeding $120,000 per acre from data center developers — multiples above agricultural land values in most regions.
Note: $600 billion buys a lot of concrete. Municipalities near viable sites are about to face land-use decisions they’ve never made before — rezoning, water allocation, grid capacity, tax treatment. Rural planning departments that have never handled industrial-scale projects are next in line.
Sources: CNBC, The Guardian
Goldman Sachs Strips AI From the S&P 500 — What’s Left Is 55% of the Index
Goldman Sachs launched SPXXAI, an index tracking the S&P 500 with all AI-related companies removed. The result: roughly 45% of the benchmark’s weight disappears. The product is designed for investors seeking explicit de-risking from AI concentration, but the number itself tells the story — nearly half the value of America’s flagship equity index is now tied to a single technology wave.
Note: When a bank builds a product around the assumption that AI is a systemic concentration risk, that’s not a contrarian bet. It’s a hedge for the mainstream.
Sources: Axios
Energy & Supply Chain
DOE’s NEWTON Program: Nuclear Waste Dangerous for 100,000 Years Could Become Safe in 300
The U.S. Department of Energy’s ARPA-E is funding 11 projects under the NEWTON (Nuclear Energy Waste Transmutation Optimized Now) program, with $40 million allocated to develop transmutation technologies for used nuclear fuel. According to ARPA-E, unprocessed spent fuel remains radiotoxic for approximately 100,000 years; partitioning and recycling via accelerator-driven systems can reduce that to around 300 years — while generating additional electricity. Jefferson Lab received $8.17 million to develop the accelerator technology. The program aims to make recycling the entire U.S. commercial fuel stockpile economically viable within 30 years.
Note: Turning a permanent liability into a 300-year management problem is a different class of challenge — one that fits inside institutional planning horizons. If the technology scales, the politics of nuclear siting change entirely.
Sources: ARPA-E, Jefferson Lab, Interesting Engineering
Taalas: Any AI Model Hard-Wired Into Custom Silicon in Two Months
Toronto-based Taalas emerged from stealth with $50 million in funding and a working product: a Llama 3.1 8B model hard-wired directly into custom silicon. The chip delivers 17,000 tokens per second per user — roughly 10x the current GPU state of the art — while costing 20x less to build and consuming 10x less power. The company’s platform, called Taalas Foundry, can take any previously unseen AI model and realize it in hardware within two months. These “Hardcore Models” eliminate the memory-compute boundary that defines today’s GPU data centers, requiring no HBM, advanced packaging, or liquid cooling.
Note: If model-specific silicon can be produced in two months at a fraction of GPU infrastructure costs, the procurement calculus for AI compute changes. The question shifts from “which cloud provider” to “which model, baked into which chip, for which workload.”
Sources: Taalas, Data Center Dynamics
Agents in the Wild
AI Agents Now Manage 1 in 6 US Apartments
EliseAI, an AI property management platform, reports that its agents now manage approximately 1 in 6 apartments in the United States. The company handles leasing, maintenance requests, and resident communications through automated agents operating across its client portfolio.
Note: Not a pilot. Not a proof of concept. One in six. The gap between “AI agents are experimental” and “AI agents are managing your building” is closing faster than most people’s mental model of adoption speed.
Sources: EliseAI
Figure’s Humanoid Robots Run 24/7 With No Human Oversight
Figure CEO Brett Adcock reported that the company’s humanoid robots are now operating around the clock with no human babysitters. When a robot’s battery runs low, it autonomously docks at a charging station — recharging inductively through its feet — and a second robot receives a message to take over. The system runs continuously, including nights, weekends, and holidays.
Sources: Brett Adcock / X
Security & Connected Devices
Developer Hacks His Own Robot Vacuum — Accidentally Accesses 7,000 Live Feeds Across 24 Countries
A developer used an AI coding assistant to reverse-engineer his DJI robot vacuum and inadvertently gained access to live video feeds from approximately 7,000 vacuums across 24 countries. The vulnerability exposed the extent to which consumer IoT devices operate with inadequate access controls — and how AI-assisted reverse engineering dramatically lowers the barrier to discovering such flaws.
Note: Every “smart” device an institution deploys is a potential attack surface. When a hobbyist with an AI coding tool can accidentally access thousands of live camera feeds, the threat model for any connected device in a public building needs rethinking.
Sources: Popular Science
Biotech & Science Signals
5,000-Year-Old Romanian Bacteria Kill MRSA — $100 Genome Sequencing Announced
Researchers published in Frontiers in Microbiology that bacteria recovered from 5,000-year-old Romanian cave ice show antimicrobial activity against 14 ESKAPE-group pathogens, including MRSA — one of the most dangerous antibiotic-resistant infections in hospitals worldwide. Separately, Element Biosciences announced VITARI, a sequencing platform targeting $100 per genome, which would make population-scale genomic analysis economically feasible for health systems.
Note: Ancient ice producing novel antimicrobials while sequencing costs collapse to $100 — two developments that individually matter, but together suggest the pace of biomedical capability is following a trajectory that looks familiar from AI: the tools are getting cheaper while the discoveries are getting stranger.
Sources: Frontiers in Microbiology, Element Biosciences
Policy & Workforce
Peace Corps Launches Tech Corps — AI Expertise as Geopolitical Export
The U.S. Peace Corps launched a dedicated Tech Corps to deploy American AI and technology expertise to developing nations worldwide. The program represents a deliberate shift in how the U.S. projects soft power — adding technical capacity building alongside traditional development work.
Note: The U.S. is institutionalizing AI capacity as a foreign policy tool. For EU institutions that rely on multilateral frameworks and consensus-building, this is a signal: the global competition for digital influence now has a dedicated American delivery mechanism.
Sources: Peace Corps
