Tech Digest – February 16, 2026
AI Crosses Into Original Science
GPT-5.2 Conjectures and Proves a New Result in Particle Physics — In 12 Hours
OpenAI used GPT-5.2 Pro to conjecture a new formula for gluon scattering amplitudes — a class of particle interactions that standard textbooks had assumed were zero. A scaffolded version of the model then spent 12 hours producing a formal proof, which was independently verified by physicists from Harvard, Cambridge, the Institute for Advanced Study, and Vanderbilt. The preprint has been submitted to arXiv. Harvard’s Andrew Strominger, who helped develop string theory, reportedly said this was the first time he’d seen AI solve a problem in his kind of physics that might not have been solvable by humans.
Note: This isn’t AI doing faster calculations. It’s AI identifying a pattern across expressions too complex for manual analysis, then proving it formally. The template — humans define the problem space, AI explores and conjectures, humans verify — is now a replicable methodology for fundamental research.
Sources: OpenAI, Patrick O’Shaughnessy (X)
AI 2027 Forecasters Grade 2025: Progress at 65% of Predicted Pace
The authors of “AI 2027,” a detailed forecasting scenario for AI progress, published a grading of their 2025 predictions. Quantitative metrics came in at roughly 65% of the pace they originally projected. Most qualitative predictions — agent deployment, revenue growth, public awareness — landed on track. The team’s updated forecasting model, published separately in December 2025, now places the median for full coding automation around 2032, several years later than the original 2027–2028 scenario.
Note: 65% of a prediction that most people considered extreme is still remarkably fast. And these forecasters are adjusting their models in public, with methodology attached — which is more than most vendors or consultants offer when making technology timeline claims.
Sources: AI Futures Blog
Berkeley Researcher’s Coding Agent Optimizes Itself Overnight, Cuts Own Costs by 98%
UC Berkeley professor Koushik Sen let a coding agent watch its own execution logs, edit its own code, and rerun until performance metrics improved. It ran overnight and delivered a 98% cost reduction across nine changes — none written by a human. The agent used no architecture redesign and no manual tuning, just iterative self-modification against its own telemetry.
Note: The agent didn’t need new capabilities. It needed permission to modify itself and a feedback loop. That’s a template anyone with API access can replicate — and a preview of what “continuous improvement” means when the system improving itself doesn’t sleep.
Sources: Koushik Sen (Dev.to)
Enterprise Adoption & Productivity Signals
Airbnb: One-Third of US and Canadian Customer Support Now Handled by AI
Airbnb disclosed that AI now handles a third of its customer support interactions in the US and Canada. The company framed this as a quality and speed improvement rather than a headcount play, though the operational implications are straightforward: a platform processing millions of support requests annually has shifted a significant share to automated resolution.
Note: One-third is the disclosed number. The direction is clear. Any institution running a citizen-facing service desk should be benchmarking its own inquiry volumes against what’s already being automated in the private sector — not to copy, but to understand where the floor is moving.
Sources: TechCrunch
Anthropic Puts Claude at the Center of CS Education for 20,000+ Students
Anthropic is partnering with CodePath to integrate Claude into computer science courses at community colleges and HBCUs, reaching over 20,000 students. More than 40% come from families earning under $50,000. The program embeds AI tools directly into coursework rather than treating them as supplementary — students learn to code with AI from the start, not as an afterthought.
Sources: Anthropic
US Productivity Grew 2.7% in 2025 — Nearly Double the Prior Decade’s Average
A Stanford analysis reported by the Financial Times shows US productivity grew 2.7% in 2025, nearly doubling the average rate of the prior decade. The acceleration aligns with the timeline of broad AI tool adoption across knowledge work and signals that AI’s economic impact is beginning to appear in aggregate statistics, not just individual case studies.
Note: Productivity statistics are lagging indicators. By the time they show up in national data, the underlying shift has been underway for a while. The more useful question: which workflows drove the gain, and which sectors haven’t started yet?
Sources: Financial Times
Supply Chain Under Pressure
Western Digital’s Entire HDD Capacity for 2026 Is Already Sold Out
Western Digital has no remaining hard disk drive production capacity available for the year. Demand from hyperscale data centers — driven overwhelmingly by AI workloads — has consumed the company’s entire output. The storage industry is experiencing a demand surge that mirrors earlier semiconductor shortages, with lead times extending and buyers competing for allocation.
Note: Storage isn’t glamorous, but it’s foundational. Any institution planning a digitization project, data migration, or archive system in 2026 should check procurement timelines now — not when the RFP is ready.
Sources: WCCFTech
Broadband Memory Prices Jump 7x in Nine Months — Routers and Networking Hardware Hit
Counterpoint Research reports that broadband memory prices have surged sevenfold in nine months, pushing memory from roughly 3% to over 20% of total router costs. The price increase is driven by competition for memory chips between AI infrastructure and traditional networking equipment. Telcos and broadband providers are absorbing the cost or delaying hardware refreshes.
Note: AI’s resource appetite is repricing components far from the data center. If your next budget cycle includes network hardware, the quotes from six months ago are already obsolete.
Sources: Counterpoint Research
TSMC Plans Another $100 Billion for Four More US Fabs
TSMC is planning an additional $100 billion investment to build four more semiconductor fabrication plants in the United States, on top of its existing Arizona commitments. The expansion reflects sustained demand for advanced chips and the ongoing push to reshore semiconductor production closer to end markets. Total TSMC US investment would exceed $165 billion.
Sources: Financial Times
Energy for the Intelligence Economy
First Nuclear Reactor Airlifted by C-17 — and Fusion Hits 150 Million Degrees
Two energy milestones in the same week. The US Department of War executed Operation Windlord, the first-ever C-17 airlift of a nuclear reactor — three Globemasters carrying the eight modules of Valar Atomics’ Ward250 microreactor from March Air Reserve Base to Hill AFB in Utah for testing. The 5-megawatt reactor is designed to power roughly 5,000 homes or a military installation independent of the civilian grid, with a target of reaching criticality by July 4, 2026. Separately, Helion Energy’s Polaris prototype became the first privately funded fusion machine to achieve deuterium-tritium fusion, reaching plasma temperatures of 150 million degrees Celsius — three-quarters of the way to commercial operating conditions. Helion is building a 50-megawatt commercial plant in Washington State under contract with Microsoft, targeting grid power by 2028.
Note: The common thread is speed. A modular reactor airlifted cross-country like cargo, and a fusion prototype breaking records on its seventh iteration. Energy infrastructure is being developed on a timeline that looks more like software than civil engineering. For anyone planning data center capacity or long-term power procurement, the options available in 2028 will look very different from today’s menu.
Sources: The War Zone, Stars and Stripes, Helion Energy, TechCrunch
Policy, Privacy & Capital Flows
EU Moves to Ban Infinite Scrolling in Apps Like TikTok
The European Commission announced plans to restrict addictive design patterns in digital platforms, with infinite scrolling as a primary target. The proposal, aimed at apps like TikTok, frames manipulative UX as a consumer protection issue. Implementation details and timelines are still emerging, but the direction is consistent with the EU’s pattern of regulating platform design alongside platform content.
Note: Regardless of where you stand on the policy, the regulatory signal is clear: the EU is moving beyond content moderation into design mandates. Any organization building or procuring digital services for EU audiences should expect the UX compliance surface to expand.
Sources: European Commission
Meta Plans Facial Recognition on Ray-Ban Smart Glasses This Year
Meta plans to add facial recognition capabilities to its Ray-Ban smart glasses in 2026, according to the New York Times. The feature would allow wearers to identify people in real time through the glasses’ camera. The announcement arrives as the glasses have already become Meta’s fastest-growing hardware category, with millions of units in circulation.
Note: Facial recognition on consumer hardware worn in public spaces creates a very different privacy environment than facial recognition at a border checkpoint. For GDPR-regulated organizations, the question isn’t whether this will be controversial — it’s when it arrives in your jurisdiction and what your stance needs to be.
Sources: New York Times
India Clears $1.1 Billion State-Backed VC for Deep Tech as ChatGPT Hits 100 Million Weekly Users There
India approved a $1.1 billion state-backed venture capital program targeting deep tech, including AI and semiconductors. In the same week, Sam Altman disclosed that ChatGPT now has 100 million weekly active users in India alone. The combination of government capital mobilization and consumer-scale adoption positions India as one of the fastest-growing AI markets globally.
Note: 100 million weekly users in a single country is a workforce transformation signal, not just an adoption metric. When that many people are using AI tools regularly, the productivity expectations for every sector in that economy shift — and so does the competitive pressure on trading partners.
Sources: TechCrunch, TechCrunch
