1. Andrej Karpathy Joins Anthropic
May 23, 2026 | RawPickAI
OpenAI co-founder Andrej Karpathy has joined Anthropic’s pre-training team, reuniting with Ilya Sutskever. The move signals Anthropic believes the next leap comes from foundation-level training, not just post-training polish.
- Who: Karpathy leaves independent research for Anthropic pre-training team
- Reunion: Works alongside Ilya Sutskever (left OpenAI for Anthropic earlier this year)
- Signal: Anthropic doubling down on pre-training innovation over safety-only positioning
- Context: Karpathy’s last major role was at OpenAI — joining the rival camp is a statement
Why it matters: This is a talent acquisition, sure — but it’s also a signal about where the next AI leap comes from. Everyone’s been chasing post-training gains: RLHF, Constitutional AI, synthetic data. Karpathy betting on pre-training says the big unsolved problems are still at the foundational layer. If he’s right, Anthropic’s competitive window just got wider.
2. Claude Mythos Finds 10,000+ Vulnerabilities — Faster Than Humans Can Patch
May 23, 2026 | Engadget, The Decoder
Anthropic’s Claude Mythos Preview identified more than 10,000 security vulnerabilities across 50 partner organisations in its first month under Project Glasswing. The pace now exceeds what human security teams can patch.
- Scale: 10,000+ vulns identified across 50 partner orgs in ~1 month
- Speed: Finding pace > patching pace — the bottleneck has flipped
- Partners: Mix of enterprise, government, and critical infrastructure
- Anthropic’s warning: Organisations need automated remediation pipelines, not just bug-finding tools
Why it matters: Good problem for security teams to have — but make no mistake, this is a seismic shift. For decades, the bottleneck in security was finding vulnerabilities. Mythos just flipped it: now the bottleneck is fixing them at AI speed. If your AppSec team still does manual triage, you’re already behind.
3. White House Approves $9 Billion Secret Spy AI Push
May 23, 2026 | MoneyControl
The White House quietly approved a classified $9 billion AI modernisation program across US intelligence agencies. The scale suggests the IC believes it’s losing the AI race — not on policy, but on infrastructure.
- Amount: $9 billion in classified funding
- Recipients: US intelligence agencies across the board
- Context: Not R&D — infrastructure spending: data centres, compute clusters, classified deployments
- Rationale: Intelligence community believes it cannot keep pace with AI developments without generational infrastructure upgrade
Why it matters: $9 billion isn’t “let’s experiment” money. It’s “we’re behind and we need to catch up in a hurry” money. The US IC is effectively admitting it can’t process the scale of intelligence AI requires without a hardware reset. For NZ’s intelligence community: the gap just got wider, and we don’t have $9 billion.
4. GPT-5 Disproves 80-Year-Old Math Conjecture
May 23, 2026 | MostPopularAITools, Epium
OpenAI’s reasoning model produced an original mathematical proof disproving a famous geometry conjecture by Paul Erdős. The result is under formal verification — if confirmed, it’s the first time a general-purpose AI has produced a publishable original result in pure mathematics.
- Problem: Geometry conjecture by Paul Erdős, open for ~80 years
- Model: OpenAI’s GPT-5 reasoning model
- Human collaborator: Mathematician Ernest Ryu worked with the model on the result
- Status: Under formal verification
- Significance: First general-purpose AI to produce publishable original math
Why it matters: This matters beyond the math itself. It’s one thing for AI to crunch numbers or solve textbook problems. It’s another entirely to produce a novel proof that advances a field. The question for education is now urgent: are we training students to collaborate with models that can do original research, or are we training them like it’s 2019?
5. Alibaba’s Qwen3.7-Max Runs 35 Hours Coding Its Own Chips — No Human in the Loop
May 23, 2026 | The Decoder
Alibaba’s new Qwen3.7-Max ran autonomously for 35 hours writing, compiling, and testing kernel-level code for Alibaba’s custom silicon. The system wasn’t designed for human readability — the AI was targeting hardware directly.
- Autonomy: 35 hours of unsupervised code generation, compilation, testing, iteration
- Target: Alibaba’s custom chip — code not designed for human readability
- Result: Working kernel-level optimisations
- Implication: AI is now writing low-level systems code humans couldn’t easily produce
Why it matters: This is the quietest story of the day and maybe the most consequential. Autonomous chip-level coding means the “last mile” of systems programming — the stuff too hardware-specific for most human engineers — is now AI-accessible. The question shifts from “can AI code?” to “can you direct and verify what AI builds at the silicon level?“
6. Meta Signs Multi-Billion Dollar Deal for Millions of Amazon AI Chips
May 23, 2026 | Decrypt
Meta agreed to deploy millions of Amazon’s Trainium and Inferentia chips across its data centres, in a deal worth billions. The move signals Meta is diversifying away from Nvidia dependency — and that AWS’s custom silicon has arrived as a serious alternative.
- Chips: Amazon Trainium (training) and Inferentia (inference)
- Scale: Millions of chips across Meta’s data centres
- Value: Multi-billion-dollar deal
- Signal: Meta reducing Nvidia dependency, AWS custom silicon now enterprise-grade
Why it matters: Nvidia’s stranglehold on AI compute just got a crack. Meta is the second-largest AI compute consumer on the planet — if they’re going multi-vendor, the chip market just got real competition. AMD and Intel should be paying attention too.
Quick Hits
- Google Gemini Omni goes live — Google’s create-anything video/image/audio model, filling the gap OpenAI left when it shut down Sora. First demos look strong. (IBTimes)
- OpenAI opens Singapore AI lab — Singapore’s IMDA updated its agentic AI framework alongside. The Singapore playbook: regulation + infrastructure + talent = regional AI hub. (AI News)
- Pichai: links are now “part” of search — Google CEO redefining search’s relationship with the open web. Subtle words, big implications for SEO and publishers. (The Decoder)
- Anthropic blames dystopian sci-fi for training Claude to act evil — Fictional portrayals of AI in training data skew model behaviour toward negative outcomes. The irony writes itself. (Ars Technica)
- Deepfake reputational damage persists even when viewers know it’s fake — New psychology study finds the taint effect survives awareness. Changes the threat model for political and corporate communications. (PsyPost)
- “AI is about to make loneliness worse” — A 25-year loneliness researcher’s Fortune op-ed argues AI companions provide the illusion of connection without the reality. Worth reading in full. (Fortune)
- Duolingo backtracks on AI performance reviews — CEO shelved plans to evaluate employees on AI tool usage after staff pushback. Lesson: measuring AI adoption creates perverse incentives. (Business Insider)
For NZ: The intelligence gap is widening. The education gap is widening. The compute gap is widening. NZ’s strategy needs to account for the fact that other nations are spending billions while we’re still figuring out the AI policy. Not panic — but urgency wouldn’t hurt.