Devin at 89%, Opus 4.8 Price War & Illinois AI Safety Law

GPT-5.5 Hallucinates 52% Less, Mythos Restricted & Tech's 142K Layoffs

(00:00:00) GPT-5.5 Hallucinates 52% Less, Mythos Restricted & Tech's 142K Layoffs (00:00:54) Mythos Restricted — Cybersecurity Risk (00:01:46) Tech Layoffs vs. AI Capex $700B (00:02:24) Developer Jobs Under-26 Drop 20% (00:02:54) CNN Sues Perplexity — Copyright Escalates (00:03:32) Hassabis Species-Level Warning (00:04:13) What To Watch Next Two major AI labs are racing to quantify honesty, and this episode unpacks what that really means. OpenAI's GPT-5.5 Instant is now the default ChatGPT model, with the company claiming 52.5% fewer hallucinations on medical, legal, and financial prompts — an internal figure with no independent benchmark yet. Anthropic's Opus 4.8 follows with reported gains in honesty and reduced sycophancy. One week, two labs, convergent claims: honesty is now a competitive surface. The bigger story may be what Anthropic chose not to release. The lab restricted access to a model called Mythos after flagging strikingly capable cybersecurity capabilities, launching Project Glasswing — a collaboration with Google, Microsoft, and Nvidia — focused on critical software defense. A frontier lab treating its own model as too dangerous to release openly is a genuine first. Meanwhile, 142,000 U.S. tech workers have been laid off in the first five months of 2025, up 33% year-over-year, as the same companies commit $700 billion to AI infrastructure. Developer employment for workers under 26 has dropped 20% since 2024, with entry-level roles disappearing fastest. CNN became the first TV network to sue an AI company, filing against Perplexity after failed licensing talks — adding a new media category to an already crowded copyright litigation track. And DeepMind CEO Demis Hassabis told Stanford that AI is advancing ten times faster than the Industrial Revolution, with little margin for error over the next decade. The honesty benchmarks need independent verification. The Mythos situation remains unresolved. Both will have answers — neither does yet. This episode includes AI-generated content.

Devin at 89%, Opus 4.8 Price War & Illinois AI Safety Law

(00:00:00) Devin at 89%, Opus 4.8 Price War & Illinois AI Safety Law (00:00:52) Devin Writes Itself: 13% to 89% (00:01:29) Anthropic Opus 4.8 Price Pressure (00:02:04) Illinois AI Safety Law (00:02:40) OpenAI Solves 80-Year Conjecture (00:03:08) What to Watch Next Autonomous coding just crossed a threshold that most enterprise software teams weren't ready for. Cognition closed a $1 billion raise for Devin at a $26 billion valuation — up from $10.2 billion just eight months ago — backed by $492 million in annualised revenue and 50% month-over-month enterprise growth for six straight months. Clients include Mercedes-Benz, NASA, Goldman Sachs, and Santander. At Mercedes, projects that once took eight months now close in eight days. The most striking internal metric: Devin wrote 13% of Cognition's own codebase in December. Today that figure is 89%. On the model front, Anthropic released Claude Opus 4.8 with a fast mode that is 2.5× quicker and 3× cheaper than its predecessor, while matching Claude 4.7 pricing. The move applies direct cost pressure to OpenAI's agentic offerings. Anthropic's valuation has now cleared $965 billion, surpassing OpenAI's March figure of $852 billion. At the regulatory layer, Illinois passed a landmark AI safety law — including whistleblower protections — filling a gap left by the Trump administration's decision to pull back its federal frontier AI regulatory plan. The risk of state-by-state fragmentation is real and unresolved. OpenAI's model also produced the first counterexample to the Erdős planar unit distance problem, an 80-year-old mathematical conjecture — a result researchers are calling the first genuinely autonomous and interesting AI math finding. DeepMind has separately resolved nine Erdős problems. Finally, Demis Hassabis framed AI this week as a species-level transition moving ten times faster than the Industrial Revolution, with a five-to-ten year window for meaningful international coordination. This episode includes AI-generated content.

Sovereign Wealth Owns AI Now: Q1's $300B Funding Shock

(00:00:00) Sovereign Wealth Owns AI Now: Q1's $300B Funding Shock (00:00:43) Three Companies, Two-Thirds of Capital (00:01:20) US Government's Nine Billion Dollar Lag (00:02:00) China Locks Down AI Talent (00:02:34) Cisco Challenges Safety Benchmarks (00:03:12) Optical Supply Chain Tightening (00:03:41) AGI Timeline and Closing Watch Points Three hundred billion dollars flowed into AI in Q1 2026 alone — and the story isn't just the number, it's who's writing the checks. Traditional venture capital has hit its ceiling. Sovereign wealth funds from Temasek, the Qatar Investment Authority, Saudi Arabia's Public Investment Fund, and Mubadala are now the dominant force at the frontier. OpenAI, Anthropic, and xAI combined to capture 67% of all global venture capital across every sector — a concentration so extreme it's compressing Series A rounds and widening the barbell between frontier labs and everyone else. Meanwhile, the US government is openly playing catch-up. The CIA and NSA have requested $9 billion for Nvidia Grace Blackwell GB10 superchips — a catch-up request, not a forward investment — while China has moved in the opposite direction, locking DeepSeek and Alibaba employees behind international travel approvals. Beijing is treating frontier AI talent as critical national infrastructure. On the safety front, Cisco's new research exposes a critical flaw in how the industry benchmarks model security. Multi-turn attack success rates across 15 frontier models ranged from 7.89% to 88.3% — meaning every model tested failed iterative safety evaluation. The threat model the industry has been using is already outdated. Elsewhere, optical components are the newest infrastructure chokepoint: demand for 800G transceivers is set to jump 2.6x, and Nvidia has locked up roughly 80% of EML laser supply. And DeepMind's Demis Hassabis moved his AGI forecast forward to 2029, citing faster-than-expected agent proliferation. All the signals, no filler. That's today's briefing. This episode includes AI-generated content.

Anthropic's $30B Surge, Pentagon vs China & the AI Arms Race

(00:00:00) Anthropic's $30B Surge, Pentagon vs China & the AI Arms Race (00:00:45) Anthropic's $30B Revenue Milestone (00:01:54) China's ZGC AI Tech Park (00:02:41) China's AI Registry vs US Deregulation (00:03:15) DeepSeek's African Market Push (00:03:40) What to Watch Next Artificial intelligence is no longer a product race — it's becoming the underlying layer of national capability, and today's stories make that unmistakably clear. The biggest commercial surprise is Anthropic. The safety-first lab has hit a thirty-billion-dollar annualised revenue run rate, posting roughly eighty times year-over-year growth and approaching a nine-hundred-billion-dollar valuation. Claude is now the leading platform for paid enterprise adoption, with corporate spend up fourfold since early 2025. For OpenAI and Google, the competitive pressure just got sharper. On the defense side, the Pentagon has filed a $29.5 billion AI budget request covering secure data centres, next-generation supercomputers, GPU procurement, and a National Security Investment Fund. The risk isn't funding — it's execution. Federal procurement timelines may not match the pace of what China is already building on the ground. China's ZGC AI Technology Park — an 800,000-square-metre facility housing DeepSeek, Qwen, and others — is a direct expression of Beijing's coordinated, self-sufficiency-first strategy. Nvidia's CEO has confirmed the company is losing China market share to Huawei, which posted a record year. US export restrictions appear to have accelerated domestic chip production rather than slowed it. Regulatory trajectories are diverging too. China's Cyberspace Administration has registered 530 AI services under a mandatory security-assessment regime since 2023. The US, by contrast, has walked back model approval plans. Expect market fragmentation, not global standards. Finally, DeepSeek's free-to-use model is gaining ground in African markets — an early-influence play that will shape those ecosystems for years. A YesWee production. Built using AI technology. This episode includes AI-generated content.

$650B Revenue Gap: Can AI Spending Ever Pay Off?

(00:00:00) $650B Revenue Gap: Can AI Spending Ever Pay Off? (00:00:57) Microsoft Copilot Lock-In (00:01:46) Ackman Exits Alphabet (00:02:29) DeepMind Solves Erdős Problems (00:03:23) Pope Leo XIV AI Encyclical (00:03:50) What To Watch Next The numbers don't add up. JP Morgan's analysis shows hyperscalers must generate $650 billion in annual AI revenue to earn a 10% return on projected 2026 capital spending — yet the current run-rate sits at just $25 billion. That's a 26x gap, and combined infrastructure spend from Microsoft, Amazon, Alphabet, Meta, and Oracle is set to hit $725 billion in 2026 alone, up from $162 billion four years ago. The closest thing to a revenue answer right now is Microsoft Copilot. Over 540,000 enterprise customers purchased Copilot for Microsoft 365 in Q2 2026 — 65% quarterly growth — with more than 75% of Fortune 500 companies now inside the ecosystem. That signal was enough for Bill Ackman's Pershing Square to build a $6.2 billion Microsoft position while simultaneously cutting its Alphabet stake by 60%, a direct bet on enterprise lock-in over Google's more fragmented AI monetisation approach. On the research front, DeepMind's AlphaProof Nexus autonomously solved nine Erdős problems, some open for 56 years, using Lean formal verification to catch errors that human reviewers miss — a critical development for trustworthy AI deployment in high-stakes domains. DeepMind CEO Demis Hassabis used the moment to correct breathless AGI commentary: solving specialised math is not general intelligence. Finally, Pope Leo XIV issued a formal encyclical warning that AI risks creating second-class humans, bringing a moral authority that reaches more than a billion people into the AI regulation debate. The core tension: spending is real, revenue is not yet proportionate, and the hard questions are now coming from JP Morgan analysts and the Vatican alike. This episode includes AI-generated content.