Tech Digest – March 16, 2026
AI Capabilities & Self-Direction
Anthropic: 70–90% of Future Claude Code Is Already Written by Claude Itself
According to a Time profile, 70 to 90 percent of the code behind future Anthropic models is now written by Claude, with chief science officer Jared Kaplan stating that fully automated AI research is less than a year away. Alignment lead Evan Hubinger was direct: “Recursive self-improvement is not a future phenomenon. It is a present phenomenon.” In a separate demonstration, startup Percepta hard-coded a WebAssembly interpreter into transformer weights — executing arbitrary C code as tokens for millions of steps — providing empirical grounding for the claim that neural networks are now practical, general-purpose computers.
Note: The practical consequence is compression of the development cycle: if AI is writing and evaluating its own research, the 18-month planning horizon that institutions use to anticipate capability changes may already be obsolete.
Sam Altman at Stanford: Today’s Models Are Smart Enough to Find the Architecture That Replaces Them
Speaking at Stanford’s TreeHacks in February 2026, OpenAI CEO Sam Altman stated: “I bet there is another new architecture to find that is gonna be as big of a gain as transformers were over LSTMs — and I think you finally have models that are smart enough to help do that kind of research.” The statement, which has circulated widely without an official transcript, was confirmed across multiple independent sources. Altman stopped short of claiming OpenAI had found such an architecture; the point was that the tools to search for it are now available.
Note: Every major institutional AI procurement decision today is implicitly a bet on transformer-based architecture. A post-transformer shift — even if years away — would reset capability assumptions, cost models, and integration roadmaps across the board.
Sources: Gary Marcus / Substack, Verified quote via X
Institutional Adoption & Tools
Anthropic Releases 1-Million-Token Context Window for Opus 4.6 and Sonnet 4.6 — Generally Available
Anthropic has moved the 1-million-token context window to general availability for Claude Opus 4.6 and Sonnet 4.6. A context window of this size can hold roughly 750,000 words in a single session — enough to process an entire legal archive, a multi-year procurement record, or a full institutional policy library in one pass.
Note: Document-heavy institutions — administrations, legal departments, procurement offices — gain a qualitatively different capability here. The constraint was never whether AI could answer questions; it was whether it could hold enough context to answer them accurately. That barrier is gone.
Sources: Anthropic
AMA: 81% of Physicians Now Use AI — Double the 2023 Rate
A new American Medical Association survey found that 81 percent of physicians now use AI in their practice — up from roughly 38 percent in 2023. Confidence in the technology grew alongside adoption. Separately, a Sydney entrepreneur used ChatGPT to design a personalized mRNA cancer vaccine for his rescue dog; the tumor shrank considerably. The anecdote is one data point, but it illustrates the gap between what institutional guidelines permit and what individuals are already doing with publicly available tools.
Note: A profession that spent decades resisting protocol changes doubled its AI adoption rate in two years. Any institution still treating AI adoption as a leading-edge behavior is describing 2023, not today.
Sources: American Medical Association, Times of India
US Senate Formally Permits AI Tools for Official Staff Work
New US Senate guidelines now allow aides to use Gemini, ChatGPT, and Microsoft Copilot for official work, according to the New York Times. The move formalizes what had likely been informal practice, and positions the US legislature as the most visible public institution to have moved from AI restriction to AI permission at the staff level.
Note: The guardrails question in public institutions has shifted from “should staff use AI?” to “under what conditions?” EU public bodies watching this development will face the same question on a shorter timeline than most expect.
Sources: New York Times
Infrastructure & Supply Chain
Qatar Helium Outage Puts Global Chip Supply Chain on a Two-Week Clock
Iranian drone strikes on QatarEnergy’s Ras Laffan facility knocked roughly 30 percent of global helium supply offline. Nine days after the attack, production had not restarted, according to Tom’s Hardware. Semiconductor fabrication depends on ultrapure helium for cooling and leak detection; a prolonged outage would affect chip production globally. The downstream pressure is already visible: RAM kits are reportedly now shipping with one real stick and one dummy unit, described by one retailer as offering “desperate psychological relief” during the shortage.
Note: Hardware procurement timelines for any institution planning data center expansion or device refreshes in Q2–Q3 2026 are now directly exposed to this disruption. Single-source dependencies in critical materials remain one of the more predictable — and consistently underestimated — risks in digital infrastructure planning.
Sources: Tom’s Hardware, Tom’s Hardware (RAM shortage)
On the energy side, a contrasting signal: America’s first large-scale offshore wind farm, Vineyard Wind, has completed construction off the Massachusetts coast — a reminder that the energy infrastructure required to power AI-scale compute is being built, if unevenly and not without setbacks along the way.
Sources: WBUR
Workforce Signals
Meta Plans to Cut 20% or More of Workforce as AI Infrastructure Costs Mount
Reuters reports that Meta is planning sweeping layoffs of 20 percent or more across the company, driven by the mounting cost of AI infrastructure investment. The cuts come as Meta simultaneously accelerates AI model development and builds out its data center footprint. The pattern mirrors a broader dynamic: the companies most aggressively investing in AI are also the ones reducing their knowledge worker headcount fastest.
Note: The message is structural, not cyclical. AI infrastructure is expensive enough that even trillion-dollar companies are trading headcount for compute. For institutions planning workforce and skills strategies, this is a directional signal about where value is concentrating.
Sources: Reuters
Scientific Automation
Physical Superintelligence Releases the First Open-Source Agentic AI Physicist
Physical Superintelligence PBC (PSI) has released Get Physics Done (GPD) — an open-source agentic system built by physicists for physics research. GPD scopes problems, plans research phases, runs derivations and numerical checks, and verifies results against physical constraints. It covers 18 physics subfields, locks notation conventions across a project, and can operate in autopilot mode for well-scoped problems. It runs inside Claude Code, Gemini CLI, and other AI runtimes; the code is Apache 2.0 on GitHub. One early user described inference cost projections from GPD’s own analysis as showing a potential 100x to 1000x drop by 2035. Former White House science advisor Tom Kalil called for using GPD to map the “civilizational tech tree” at granular resolution.
Note: Physics is the upstream bottleneck for materials science, energy, semiconductors, and much of applied engineering. Tooling that compresses the cycle time from question to verified result — from weeks to hours — would accelerate every field downstream. The institutional implication is not immediate, but it resets the timeline on when AI-driven R&D moves from lab curiosity to procurement-relevant capability.
Sources: Physical Superintelligence PBC, GitHub (psi-oss/get-physics-done)
Autonomous Systems & Defense
Airbus Targets German Air Force Drone Wingman Deployment by 2029
Airbus announced it is preparing two uncrewed combat aircraft from US manufacturer Kratos for first flight, as part of a program targeting integration as a drone wingman for the German Air Force by 2029. The announcement positions Europe as an active participant — not just observer — in the accelerating deployment of autonomous combat systems.
Note: A three-year timeline to operational drone wingman deployment in a NATO member state is not a long-range signal — it is a near-term procurement and training reality. Defense institutions and adjacent public bodies managing related infrastructure, airspace, or dual-use technology frameworks should treat 2029 as a near-horizon date, not a future planning exercise.
Sources: Airbus
