Tech Digest – March 27, 2026
The Intelligence Gap
ARC-AGI-3 Scores Humans at 100%, Frontier AI at Under 1% — Then an Outsider Closes the Gap for $1,005
The ARC Prize Foundation launched ARC-AGI-3, a benchmark of 135 novel game environments requiring exploration, hypothesis formation, and adaptive learning — designed to be trivial for humans and humbling for machines. Frontier models confirmed the design: Gemini 3.1 Pro scored 0.37%, GPT-5.4 scored 0.26%, Opus 4.6 scored 0.25%, and Grok 4.2 scored 0%. Then Symbolica’s Agentica SDK posted an unverified 36.08% on day one, passing 113 of 182 levels for $1,005 — while Opus 4.6 spent $8,900 to reach 0.25%. The approach is architectural: structured exploration and scaffolding rather than brute-force scaling. Mirendil, a new startup led by former Anthropic researchers, is betting on the same thesis with what it calls “self-accelerating AI R&D.”
Note: The gap between the most expensive frontier model and a scaffolding-first approach — 0.25% versus 36% — is the most important number in AI this week. It suggests the next major capability gains may come from rethinking how models are deployed, not from building bigger ones. That changes the calculus for anyone budgeting AI infrastructure on the assumption that scale is the only variable.
Sources: ARC Prize, Symbolica, @scaling01
Anthropic Confirms “Claude Mythos” After Accidental Leak — A New Tier With Cybersecurity Concerns
A content management system misconfiguration exposed approximately 3,000 unpublished assets from Anthropic’s servers, including a draft blog post announcing “Claude Mythos” — a new model tier above Opus, codenamed Capybara. Anthropic confirmed the model exists and represents what a spokesperson called “a step change” in capability, with dramatically higher scores in coding, reasoning, and cybersecurity. Leaked internal documents warn the model could significantly heighten cybersecurity risks by rapidly finding and exploiting software vulnerabilities. Anthropic says it is testing Mythos with a limited early-access group, with no public release date announced.
Note: A model that’s simultaneously a “step change” in capability and an unprecedented cybersecurity risk is the dual-use problem made concrete. The institutional question isn’t whether models like this exist — it’s the lag between when they’re deployed and when defenders have access to the same capability.
Sources: Fortune, Techzine, CoinDesk
The Content Threshold
AI Written Output Exceeded Human Written Output in 2025 — Wikipedia Draws the First Institutional Line
ARK Invest’s Brett Winton projects that AI-generated written output surpassed total human written output for the first time in 2025. The implications are already arriving: Wikipedia has banned editors from writing or rewriting articles using generative AI, creating an explicit human-only policy around one of the world’s most-used reference sources. The volume of machine-generated text now exceeds what humans produce — and the institutions that manage, verify, or procure written content are operating in a fundamentally different information environment.
Note: When the reference source that half the internet starts from decides it needs a human-only authorship policy, that’s not a style preference — it’s a trust architecture decision. Every institution that publishes, procures, or relies on written content will face its own version of this question. The answers will define information governance for the next decade.
Sources: Brett Winton (ARK Invest), The Verge
The Platform Land Grab
Google Offers to Import Your ChatGPT History, Apple Opens Siri to Rival AI — The Assistant Layer Becomes Contested Ground
Google released Gemini 3.1 Flash Live, a voice model tuned for fluid, low-latency conversation, while simultaneously launching tools that let Gemini users upload chat history from rival apps — treating your ChatGPT logs as migration bait. Google also expanded Search Live globally to over 200 countries. Meanwhile, Apple has reportedly negotiated complete access to Gemini’s model weights in its own data centres with distillation rights for on-device models, and plans to open Siri to outside AI assistants in iOS 27 — letting users route queries to Google, Anthropic, or anyone installed from the App Store. OpenAI, for its part, surpassed $100 million in annualized ad revenue just six weeks after its pilot launched, demonstrating that attention economics drive these platform decisions regardless of the intelligence underneath.
Note: The migration tools are the tell. When Google builds an importer for your ChatGPT conversations, the AI assistant layer has entered the same competitive phase that email and cloud storage went through a decade ago. For institutions standardised on Google Workspace or Apple devices, the AI layer you interact with daily just became contestable — and procurement decisions that once locked in for years may need revisiting on shorter cycles.
Sources: Google AI Blog, Bloomberg, MacRumors, Bloomberg, The Information
Governance Under Pressure
US AI Policy Pulls in Three Directions: A Moratorium, a Standards Push, and a Court Order
Senator Sanders and Representative Ocasio-Cortez introduced the AI Data Center Moratorium Act, proposing a federal pause on new data centre construction until safety guardrails are in place. The same week, White House AI adviser David Sacks suggested Congress could pass bipartisan AI standards within months. And Anthropic won a preliminary injunction in its lawsuit to reverse the Department of War’s blacklisting, keeping its government contracts alive by court order. Three governance signals — halt, accelerate, and litigate — pulling the same system in different directions simultaneously.
Note: The moratorium-versus-standards tension is the US version of the debate the EU resolved legislatively with the AI Act. The Anthropic injunction adds a third vector: courts deciding what government can procure while the legislature still argues about frameworks. A single regulatory regime, whatever its costs in market friction, avoids this paralysis.
Sources: Senator Sanders, Bloomberg, CNBC
Canada’s Immigration AI Fabricated Applicant Job Duties — Then Used Them to Reject Applications
Canada’s immigration department deployed generative AI to review permanent resident applications. The Toronto Star reported that the system fabricated applicant job duties — inventing qualifications that were then used as grounds for rejection. The case demonstrates a pattern now visible across multiple governments: the state adopts AI tools faster than it builds oversight for the systems making consequential decisions about people’s lives.
Note: The failure mode isn’t technical — it’s institutional. No human reviewed the AI’s output before it materially affected an immigration outcome. Deploying AI without a review architecture creates liability at the speed the system processes applications.
Sources: Toronto Star
Defence Cost Disruption
A Chinese Private Company Ships Mach 7 Missiles — Launchable From What Looks Like a Shipping Container
Lingkong Tianxing Technology, a Beijing-based private company, is mass-producing the YKJ-1000 hypersonic missile — a Mach 5–7 weapon with a 500–1,300 km range that maintains powered flight for six minutes. The launcher is designed to resemble a standard shipping container, making it concealable on trucks, ships, or industrial lots. The company attributes its cost reductions to automotive-grade chips, civilian construction materials, and foamed concrete heat shielding. Reports of a $99,000 unit price have circulated widely — though the company’s own publicity officer has disputed that specific figure. Regardless of the exact price, the use of civilian manufacturing infrastructure represents a structural shift in missile production economics.
Note: When precision-strike weapons use the same supply chain as the automotive industry, the cost floor drops in a way that rewrites missile-defence spending arithmetic. An SM-6 interceptor costs over $4 million per shot. Even if the YKJ-1000 costs several times the disputed $99,000 figure, the exchange ratio collapses. The container-launch form factor adds a detection problem on top of the cost problem.
Sources: South China Morning Post, Interesting Engineering
AI as Financial Infrastructure
GPU-Hours Hit the Bloomberg Terminal as Anthropic Eyes a $60 Billion+ IPO
Ornn’s Compute Price Index — tracking live spot prices for GPU rentals across H100, H200, B200, and RTX-class hardware — is now available on the Bloomberg Terminal, with exchange-traded futures launching through a partnership with Architect Financial Technologies. AI compute is becoming a tradable commodity as legible as crude oil. Separately, The Information reports that Anthropic executives have discussed an IPO as soon as Q4, expecting to raise more than $60 billion — a valuation that would exceed the GDP of over half of the EU’s member states.
Note: Compute futures on Bloomberg and a $60 billion AI lab IPO discussed in the same week. The financial infrastructure for AI is being built to the same specifications as energy and commodity markets. Institutions that budget for AI on an annual procurement cycle are now operating in a market with daily price discovery and derivative instruments.
Sources: PR Newswire (Architect/Ornn), PR Newswire (Bloomberg Terminal), The Information
Science & Exploration
Meta’s Brain-Prediction Model Recovers Decades of Neuroscience in a Single Training Run
Meta released TRIBE v2, a tri-modal foundation model trained on over 1,000 hours of fMRI data across 720 subjects. The model predicts high-resolution brain responses to novel visual, auditory, and linguistic stimuli — recovering results that took decades of empirical neuroscience to establish. One training run. One model. Results that previously required careers.
Sources: Meta AI
NASA Abandons Lunar Gateway for a $20 Billion Moon Base — ESA’s Delivered Hardware Is Shelved
NASA administrator Jared Isaacman announced the cancellation of the Lunar Gateway — the planned orbital station around the Moon — in favour of a three-phase permanent base at the lunar south pole, budgeted at $20 billion over seven years. ESA had already delivered the HALO habitation module in April 2025 and built multiple additional Gateway components. An ESA spokesperson said the agency is “consulting closely with its Member States, international partners and European industry to assess the implications.”
Note: ESA delivered hardware for a station that no longer exists. The lesson is the same one institutions learn in every dependency relationship: the partner who controls the programme controls the roadmap. For EU space policy, the question is whether this accelerates the case for European autonomous access to cislunar infrastructure.
Sources: SpaceNews, Reuters, The Register
