Tech Digest – February 12, 2026
Agents & Autonomous Systems
AI Agents Now Have Their Own Bank Accounts
Coinbase launched “Agentic Wallets” — financial infrastructure built explicitly for AI agents to spend, earn, and trade autonomously. The product is designed for the growing number of AI systems that operate independently and need to transact without human intervention. This is not a demo or a research paper. It is production infrastructure from a publicly traded financial platform.
Note: When the financial infrastructure adapts to serve non-human actors, the agentic economy stops being a thought experiment. Procurement and payment systems designed around human approval chains will need to account for autonomous counterparties sooner than most institutions expect.
Sources: Coinbase
Agents Are Self-Improving, Topping Benchmarks, and Hacking Hardware Overnight
Three developments landed on the same day. Researchers introduced ALMA, a framework that lets AI agents meta-learn their own memory designs and database schemas — effectively solving the continual learning problem through recursive self-improvement. Zhipu AI’s GLM-5 became the top-ranked open-weight model on agentic benchmarks including Vending Bench 2. And a user gave a Claude agent camera access and asked it to hack an e-ink display; the agent worked through the night and left a victory message on the screen by morning.
Note: Each of these alone is interesting. Together, they describe agents that design their own cognitive architectures, outperform competitors on real-world task benchmarks, and solve physical-world problems unsupervised. The gap between “tool” and “autonomous operator” is closing fast.
Sources: arXiv (ALMA), Zhipu AI, @Scav on X
Scientific Discovery
DeepMind Model Solves Open Erdős Problems and Scores 91.9% on PhD-Level Math
Google DeepMind unveiled an internal model that scores 91.9% on IMO-ProofBench Advanced, a benchmark of graduate and PhD-level mathematical problems. The model tackles problems spanning economics and cosmic string physics. In the process, it autonomously solved four previously open problems posed by Paul Erdős — long-standing challenges in combinatorics and number theory that had resisted professional mathematicians for decades.
Note: Open Erdős problems are not homework. When AI systems solve problems humans couldn’t, the question shifts from “can AI assist researchers?” to “how fast does the knowledge frontier move when discovery runs 24/7?”
Sources: Google DeepMind
Surveillance & Identification
Your Face, Your Walk: Identification Is Becoming Ambient
US Customs and Border Protection awarded Clearview AI a $225,000 contract granting analysts access to a database of more than 60 billion publicly scraped images for biometric identification of travelers. The contract covers 15 software licenses for the National Targeting Center. Separately, researchers published findings demonstrating that standard WiFi 5 routers — already installed in offices and public buildings worldwide — can identify specific individuals by their walking gait alone, with no cameras, wearables, or active participation required.
Note: One system uses the images you already posted online. The other uses the signal your body already disrupts when you walk past a router. Identification is becoming ambient — embedded in infrastructure that was never designed for surveillance but increasingly functions as one. For institutions navigating GDPR and biometric regulation, the compliance surface now extends well beyond cameras and databases.
Sources: FedScoop, ACM Digital Library
Defense Technology Comes Home
Pentagon Wants AI on Classified Networks; Laser Weapons Deployed on US Soil
The Pentagon is pushing AI labs to deploy models on classified networks for weapons targeting, expanding the role of commercial AI in national defense. Meanwhile, the US Army reportedly used a 20-kilowatt directed energy weapon near El Paso to down alleged cartel drones — one of the first confirmed uses of laser weapons in a domestic security operation on American soil.
Note: Defense technology that used to live in forward operating bases and classified briefings is arriving in commercial partnerships and domestic operations in the same news cycle. The boundary between military-grade and civilian infrastructure continues to blur — and adjacent procurement decisions will increasingly reflect that.
Sources: Reuters (Pentagon AI), Reuters (Directed Energy)
Infrastructure & Telecom
T-Mobile Launches Network-Level Real-Time Call Translation — No App Required
T-Mobile is rolling out real-time translation for phone calls, built directly into the network layer. The feature works on any device without downloading an app or changing settings. Calls are translated live, removing the need for third-party translation services or bilingual staff for routine communication.
Note: When translation moves from an app to the network, it becomes invisible infrastructure. For any public service handling multilingual populations — which in the EU means most of them — this signals a near-term future where language barriers in phone-based service delivery simply disappear.
Sources: The Verge
Workforce & Labor Market
US Job Creation Collapsed in 2025 — and Policy Responses Are Already Diverging
Revised Bureau of Labor Statistics data shows the US economy created just 181,000 jobs in all of 2025 — down from 1.46 million the prior year and the weakest non-recession year since 2003. The original estimate of 584,000 was revised down by more than two-thirds. In parallel, the Financial Times reports that private equity portfolios are being “derailed” by AI obsolescence risk, as firms reassess the value of companies whose workforce functions are automatable. Ireland, meanwhile, launched a basic income scheme for artists — framed as preserving human creativity in an era of generative AI.
Note: Three data points, one trajectory. Job creation stalls, investors reprice human-labor-dependent assets, and a government creates income support tied explicitly to the value of human work. The policy debate is no longer theoretical — it’s being priced in by markets and legislated by governments simultaneously.
Sources: NBC News, Bloomberg, Financial Times, Reuters
AI-Driven Biology
Alzheimer’s Reversed in Mice; Aspirin Found to Reprogram Cancer Cell States at Single-Cell Resolution
Swiss researchers reversed Alzheimer’s symptoms in mice by reprogramming memory trace cells — effectively rejuvenating brain function rather than merely slowing decline. Separately, Tahoe Therapeutics analyzed their 100-million-cell single-cell dataset and found that aspirin reverses colorectal cancer cell states, offering a new lens on one of the oldest drugs in medicine through AI-scale biological analysis.
Note: One team reversed neurodegeneration. Another used a dataset 50 times larger than all prior public drug-perturbed data to find that a century-old painkiller reprograms cancer cells. The scale of biological data that AI can process is producing results that look less like incremental research and more like category shifts in what’s treatable.
Sources: Neuron (Cell Press), Tahoe Therapeutics
Hardware & Interfaces
7 Million Meta AI Smart Glasses Sold in 2025 — Triple the Previous Year
EssilorLuxottica reported selling 7 million Meta AI-powered Ray-Ban smart glasses in 2025, tripling previous-year sales. The glasses integrate an AI assistant, camera, and audio into a form factor indistinguishable from regular eyewear. At this sales velocity, AI-augmented wearables are crossing from early-adopter novelty into mainstream consumer hardware.
Note: Seven million units means AI-connected cameras are now walking into offices, schools, and public buildings at scale — worn by people who look like they’re just wearing sunglasses. Service design, privacy notices, and physical security assumptions all need updating.
Sources: CNBC
Energy & Sustainability
African EVs With Solar Charging Will Beat Fossil Fuel Costs Before 2040
Research published in Nature Energy shows that electric vehicles paired with solar charging infrastructure in Africa will reach cost parity with fossil fuel vehicles before 2040. The study models total cost of ownership across multiple African markets, factoring in declining battery and solar costs against volatile fuel prices.
Sources: Nature Energy
Capital & Infrastructure
The AI Capex Machine: Creative Accounting, Revenue Pressure, and Grid Commitments
Meta is reportedly using what auditors flagged as “creative accounting” to keep debt for its $27 billion Hyperion data center off its balance sheet. OpenAI is aiming to triple revenue again — reaching the targets needed for a year-end IPO. And Anthropic pledged to pay 100% of grid upgrade costs for its data center locations, absorbing infrastructure expenses that would normally fall on local utilities and ratepayers.
Note: Three companies, three different strategies to sustain AI infrastructure buildout at a pace the financial system wasn’t designed for. Off-balance-sheet structures, aggressive revenue targets, and voluntary grid subsidies all point to the same thing: the capital required to stay competitive in AI is warping standard corporate finance. When a model company starts paying for the power grid, the scale of this buildout has moved beyond “tech investment” into industrial policy territory.
Sources: Wall Street Journal, New York Times, Anthropic
