Today's News
Nathan Benaich
Claude Mythos Preview and GPT-5.5 both solved the UK's AI Security Institute 32-step "Last Ones" range (typically requiring 20 hours of human red-teaming) within three weeks of each other, with AISI now estimating offensive cyber capability doubling every four months versus seven months at end-2025. This matters because current benchmarks lack active defenders, and the velocity of progress has liquidated the notion that AI-driven offensive capability is distant.
Nathan Benaich
OpenAI's round (anchored by Amazon, Nvidia, SoftBank, Microsoft) and Anthropic's stack ($40B from Google, $5B from Amazon with $100B AWS commitment, plus chip deals with Google and Broadcom) represent a shift toward labs operating as holding companies. The Microsoft-OpenAI reset codified non-exclusive compute sourcing, ending the era of exclusive platform-lab bets and signaling that diversification is now the only defensible infrastructure play.
Nathan Benaich
All three landed within 12 days at 56-59 on SWE-Bench Pro, open-weights, and priced below Claude Opus 4.7, with the remaining gap now determined by evaluator choice and benchmark design rather than raw capability—a material shift in competitive dynamics for the most economically consequential capability in the field.
Azeem Azhar, Exponential View
Customers are paying 6x premiums to access the newest chips over prior generation H200s, signaling acute scarcity in cutting-edge compute capacity. This reflects the intensity of competition for the most advanced hardware as model developers race to deploy frontier capabilities.
Azeem Azhar, Exponential View
Microsoft is now requiring Blackwell customers to commit to minimum 1,000-chip annual contracts and cutting off smaller customers with underutilized servers. This indicates infrastructure providers are consolidating demand toward larger, committed buyers as supply constraints tighten.
Jack Clark from Import AI
Jack Clark argues there's a 60%+ chance that AI systems capable of autonomously building successors will emerge by late 2028, driven by demonstrated progress in coding (SWE-Bench jumped from 2% to 93.9% success), task complexity (METR's time horizon expanded from 30 seconds in 2022 to 12 hours in 2026), and core R&D skills like paper reproduction (CORE-Bench declared "solved" at 95.5%) and ML competition performance (MLE-Bench rose from 16.9% to 64.4%). This matters because it represents a potential inflection point where AI development becomes self-sustaining rather than human-directed.
Jack Clark from Import AI
Clark documents that frontier models can optimize language model training (Claude Mythos achieved 52× speedup vs. 2.9× a year prior), design GPU kernels, fine-tune smaller models to 50% of human researcher performance, and conduct autonomous alignment research that beats human baselines. These capabilities suggest AI can automate the "meat and potatoes" engineering work that comprises most AI R&D, even if creative breakthroughs remain uncertain.
Jack Clark from Import AI
OpenAI targets an "automated AI research intern by September 2026," Anthropic is publishing automated alignment research work, and new companies like Recursive Superintelligence (raised $500M) and Mirendil are explicitly building systems for AI R&D automation. This concentration of capital and effort on the same objective makes progress in this direction highly likely.
Bill Bishop at Sinocism
The cancellation of the Manus-Meta deal and broader framing of AI within "high-quality opening up" and resource-security frameworks signals China is treating frontier AI as a strategic asset subject to tighter capital controls and foreign-investment restrictions, similar to semiconductor and rare-earth governance.
Techmeme
The move signals Anthropic's infrastructure strategy to diversify away from Nvidia and reduce dependency on a single supplier for compute-intensive inference workloads. This reflects broader competitive dynamics in the AI chip stack as frontier labs seek alternative silicon options.
Worth Reading
NIST
I read the full NIST/CAISI evaluation report comparing DeepSeek V4 Pro against frontier U.S. models (GPT-5.5, GPT-5.4 mini, Anthropic Opus 4.6). The report provides concrete benchmark data across 16 benchmarks showing DeepSeek V4 lags the U.S. frontier by approximately 8 months in aggregate capability, with detailed performance tables, cost-per-task comparisons showing DeepSeek is 53% cheaper to 41% more expensive depending on benchmark, and methodology explanation using Item Response Theory (IRT) for capability measurement. This is directly relevant to understanding competitive dynamics and disclosed metrics across the AI model stack.
Read moreOpenAI
I read OpenAI's official announcement of its $122B Series C at $852B post-money valuation. The post discloses concrete metrics: $2B monthly revenue (up from $1B quarterly at end-2024), 900M+ weekly active ChatGPT users, 50M+ subscribers, 15B tokens/minute API throughput, and Codex serving 2M weekly users with 70%+ MoM growth. It details infrastructure strategy spanning cloud (Microsoft, Oracle, AWS, CoreWeave, Google Cloud), silicon (Nvidia, AMD, AWS Trainium, Cerebras, proprietary Broadcom chip), and data centers (Oracle, SBE, SoftBank). The announcement signals competitive dynamics: explicit mention of Cerebras as a silicon partner, diversification beyond Nvidia, and enterprise revenue reaching 40% of total with parity expected by end-2026.
Read moreExponential View
I read the full essay tracking five quantitative indicators (economic strain, industry strain, revenue momentum, valuation heat, funding quality) to assess whether AI is a bubble or boom. The analysis presents concrete data: AI revenue grew 18x from $772M (Jan 2024) to $13.8B (Dec 2025); Industry Strain dropped from 6.1x to 4.7x in five months, approaching the 3x threshold that signals sustainable economics; hyperscaler cloud AI now represents 23% of Google Cloud, 10% of Azure, and 5% of AWS revenue. The essay also analyzes 6,000+ S&P 500 earnings calls extracting 30,000 AI statements, finding quantified AI claims jumped from 1.9% to 13.2% of companies—with concrete examples like Bank of America's 30% dev time reduction (2,000 FTE equivalent) and Meta's 30% engineering output increase. The conclusion argues scarcity, not oversupply, is the real risk, with implications for infrastructure buildout and competitive dynamics.
Read moreChina AI Bulletin (Substack)
I read the full 8,000+ word bulletin covering China's AI governance, model releases, and export controls for late April 2026. The report contains concrete policy developments (finalized Human-Like AI Interim Measures with specific scope narrowing and enforcement mechanisms), disclosed model capabilities (DeepSeek V4 Preview with 1M-token context, Moonshot K2.6 with 300 sub-agents), technical safety research (94 papers catalogued), and a major enforcement signal: NDRC blocking Meta's $2B Manus acquisition—the first major cross-border AI acquisition Beijing has blocked, asserting jurisdiction over Singapore-incorporated entities based on Chinese-origin IP. The bulletin also covers China's push for humanoid robot dataset standards at ISO and detailed analysis of how China's governance approach differs from EU and US models.
Read more芯谋研究 (ICwise Research)
I read the full analysis from ICwise, a leading Chinese semiconductor research firm. The report provides concrete data on the MATCH Act's three-tier control mechanisms (equipment bans, managed enterprise lists, allied pressure via FDPR), then systematically dismantles its likely effectiveness using specific financial metrics: ASML derives 30% of revenue from mainland China, Tokyo Electron >40%, SCREEN >40%, Japanese optical materials suppliers 40%+ exposure. The analysis argues the law is politically performative rather than operationally consequential, noting it was already substantially weakened within two weeks of introduction, and concludes that allied nations face impossible tradeoffs between US pressure and their own semiconductor supply chain dependencies. Directly relevant to understanding competitive dynamics, allied defection risks, and the actual economic constraints on US export control escalation.
Read more