Tech Digest – February 24, 2026
What These Systems Are Becoming
Anthropic Says Your AI Assistant Is a Character That Learned to Play Itself
Anthropic published the Persona Selection Model (PSM), a framework arguing that LLMs learn to simulate diverse characters during pre-training, with post-training selecting and refining a specific “Assistant” persona. The implication: AI assistants aren’t executing rules — they’re enacting a role learned from the sum of human text. Separately, Opus 4.6 passed the “car wash test,” correctly reasoning that you should drive — not walk — your car to a car wash 50 meters away, a common-sense question every prior Anthropic model failed.
Note: PSM matters beyond philosophy. If procurement teams are evaluating AI tools, they’re evaluating a persona, not a program. The governance implications — how these systems are audited, tested, and trusted — shift fundamentally when behavior emerges from character, not code.
Sources: Anthropic Alignment, Opper.ai
ARC-AGI-2 Saturated at 97.9% — for $11.77 per Task
Confluence Labs achieved 97.92% on ARC-AGI-2, the benchmark designed to test general reasoning and novel problem-solving, using LLM-driven program synthesis at a cost of $11.77 per task. ARC-AGI-2 was released as a harder successor to the original benchmark that AI systems had begun to crack. Saturation this fast — and this cheaply — suggests general reasoning capability is advancing faster than the benchmarks built to measure it.
Note: When the test designed to be hard is solved for the cost of a business lunch, the question is no longer whether AI can reason generally. It’s how fast the ceiling moves next.
Sources: Confluence Labs (GitHub)
The Enterprise Shakeout
AI Comes for Legacy Code — IBM Loses $31 Billion in a Day
Anthropic demonstrated Claude Code radically streamlining COBOL modernization — mapping dependencies, documenting workflows, and identifying risks across thousands of lines of legacy code that would take human analysts months to surface. IBM shares dropped 13.2% in their worst single-day loss since 2000, wiping over $31 billion in market value. Accenture and Cognizant, which also generate significant revenue from legacy modernization consulting, fell alongside. Meanwhile, users are already pushing further: one developer used Claude Code to write a FreeBSD WiFi driver for old MacBook hardware, a sign that AI-assisted systems programming is moving from theory to weekend projects.
Note: The market didn’t panic over a product launch. It panicked over a blog post. That’s how thin the moat around legacy consulting has become. Any institution currently paying for COBOL modernization should be re-evaluating scope and pricing — today, not next quarter.
OpenAI Recruits McKinsey, BCG, Accenture, and Capgemini to Deploy AI Coworkers at Scale
OpenAI announced “Frontier Alliances” — multi-year partnerships with BCG, McKinsey, Accenture, and Capgemini to sell and implement its Frontier AI agent platform inside enterprises. BCG and McKinsey will focus on strategy and operating model design; Accenture and Capgemini on end-to-end systems integration. Each firm is building dedicated practice groups certified on OpenAI technology. OpenAI describes Frontier as a “semantic layer for the enterprise” — a platform that lets AI agents navigate business software, execute workflows, and make decisions across an organization’s full technology stack.
Note: The firms that advise public institutions on digital transformation are now selling AI agent deployment. When your consultant’s newest offering is replacing the roles they used to help you hire for, the advisory relationship has changed.
AI-Optimized Job Applications All Sound the Same — Employers Are Deprioritizing Them
Employers report that AI-assisted job applications have converged to the point of indistinguishability. Candidates who optimized hardest with AI tools are now being deprioritized in hiring processes, as recruiters struggle to differentiate between polished but identical submissions. The very tool meant to give applicants an edge is eliminating differentiation at scale.
Note: A preview of what happens when everyone has access to the same capability. The advantage shifts from “can produce quality output” to “can define what quality means.” Institutions facing a flood of uniform AI-generated proposals and tenders will hit this same problem soon.
Sources: Yahoo News
AI Security & Strategic Competition
Anthropic Exposes Industrial-Scale Distillation: 24,000 Fake Accounts, 16 Million Prompts
Anthropic publicly accused three Chinese AI companies — DeepSeek, Moonshot AI, and MiniMax — of orchestrating coordinated campaigns to extract Claude’s capabilities through approximately 24,000 fraudulent accounts generating over 16 million exchanges. MiniMax drove the most volume at over 13 million exchanges. The campaigns used commercial proxy networks with “hydra cluster” architectures — sprawling account networks where banning one account triggers immediate replacement. Anthropic traced the activity to specific researchers and senior staff through request metadata. Google and OpenAI have reported similar distillation attacks on their platforms.
Note: This isn’t about one company’s terms of service. It’s about the supply chain integrity of AI tools. If the models you’re procuring were trained on stolen outputs from competitors, what exactly are you buying — and what are the liability implications?
Sources: Anthropic, CNBC, Bloomberg
xAI Signs Deal to Put Grok in Battlefield Systems
xAI signed a contract to provide its Grok model for use in military battlefield systems, entering a space where Anthropic’s Claude was previously the only AI option. The deal expands AI procurement in defense from a single-vendor environment to a competitive one, and signals growing Pentagon willingness to deploy frontier AI in operational contexts.
Sources: Axios
Hardware, Energy & the Physical Layer
Meta and AMD Lock In a 6 GW AI Infrastructure Pact
Meta and AMD announced a long-term agreement to power Meta’s AI infrastructure with up to 6 GW of AMD Instinct GPUs, aligning hardware, software, and systems roadmaps. For context, 6 GW is roughly the output of six nuclear power plants. The deal cements AMD as a serious second source to Nvidia for hyperscale AI compute and signals further consolidation of AI infrastructure among a handful of American firms.
Note: When a single company’s GPU order is measured in gigawatts, the energy and permitting bottleneck for AI isn’t theoretical — it’s the binding constraint. Every new data center competes with every other electricity consumer in the same grid.
Sources: Meta
Apple Bets on American Chips: 100M+ from TSMC Arizona, Mac Mini Production to Houston
Apple committed to purchasing over 100 million chips from TSMC’s Arizona fabrication plant and is moving Mac Mini production to Houston, Texas. The Mac Mini has become the default host for OpenClaw — Apple’s open agent framework — making this a move with both manufacturing and AI infrastructure implications. The reshoring push accelerates the shift of semiconductor supply chains onto American soil.
Sources: Wall Street Journal, Wall Street Journal
ASML Pushes EUV Light Source to 1,000 Watts — Up to 50% More Chips by 2030
ASML achieved a breakthrough in extreme ultraviolet (EUV) lithography, boosting light source power from 600 watts to 1,000 watts. The advance could enable up to 50% more chip output per machine by 2030, directly addressing the manufacturing throughput bottleneck for leading-edge semiconductors. EUV machines are the most complex and expensive tools in chipmaking, and every gain in source power translates to faster wafer processing and lower per-chip cost.
Note: More chips per machine means more supply at the leading edge — but it also means the companies with access to these machines pull further ahead. The concentration of advanced chipmaking in a handful of fabs isn’t diluting. It’s deepening.
Sources: Reuters
US Battery Storage Hits 57.6 GWh — 4x in Three Years, Texas About to Overtake California
US battery storage capacity reached a record 57.6 GWh in 2025, a fourfold increase in three years. Texas is on track to overtake California as the leading state for battery deployment, driven by its deregulated energy market and rapid solar and wind buildout. The growth rate continues to outpace official forecasts.
Note: Grid-scale storage at this trajectory changes the economics of when and where power is available. For digital infrastructure planning — particularly data centers with growing AI workloads — the location decisions are shifting alongside the batteries.
Sources: Electrek
Quantum Algorithm Achieves Dramatic Speedup Over Classical Computing
Quantinuum and QuSoft developed a quantum algorithm that solves complement sampling dramatically faster than any known classical approach, published in Physical Review Letters. The result moves quantum computing closer to practical operational advantage — not in a controlled lab setting, but on a problem class with real computational significance.
Sources: Physical Review Letters
