Anthropic reveals Claude now handles 26% of its AI R&D work

Inside Gemini: How Google's Model PM Balances Customer Requests Against Unpublished Research

Tulsee Doshi, who leads product for Gemini models across Google's apps, API, and internal products, describes a structural tension unique to frontier model teams: protecting researcher momentum on capabilities customers can't yet see while consolidating requests from three customer types (internal products, other Google teams, external developers). The model's customers—Gemini app, Search, YouTube, Gmail, and API developers—have disparate requirements and different tolerance for instability; internal surfaces can iterate live while API developers need stable versions. Rather than simply stack-ranking customer requests, Doshi's team weighs them against where the research is heading, a direction customers "don't know actually what we're investing in or where we think the research is going." The scarce skill on model teams is not improving models but specifying what "good" means precisely enough for researchers to optimize toward. Doshi starts with 20 concrete prompts where wins and losses are obvious, then turns those into measurable evals—a practice that takes days or weeks to build well. She hires for thoroughness over computer science credentials, recruiting from finance, consulting, and operations as readily as engineering, looking for people who dig "2 levels lower than 95%" rather than taking metrics at face value. Top model teams like NotebookLM deliberately prototype features 60% working today, 20% mediocre, and 20% reach—keeping capabilities the current model can't handle, then shipping immediately when new checkpoints arrive. Model releases ship on conviction about key signals (user behavior changes, conversation length, task difficulty) rather than traditional launch checklists, a departure from Google's standard product machine.

Huawei Accelerates Ascend 960DT Chip Launch to Q1 2027, Three Quarters Early, With 2x Compute and Memory Specs

Huawei pulled forward its Ascend 960DT AI chip from Q4 2027 to Q1 2027, a three-quarter acceleration. The chip doubles compute, memory bandwidth, memory capacity, and interconnect ports versus the 950. However, the announced 960 SuperPoD scales to 4,096 chips rather than the previously stated 15,488, suggesting Huawei may have restructured how it divides SuperPods versus SuperClusters. Chinese AI labs are targeting 10-40 trillion parameter models over the next two years, and Huawei continues discussing eventual connections of up to 1 million processors, signaling aggressive infrastructure buildout despite US export controls.

Zuckerberg, Huang, and Musk Blocked Trump's AI Regulatory Plan From Demis Hassabis

Mark Zuckerberg, Jensen Huang, and Elon Musk spoke with President Trump and successfully blocked a regulatory proposal, according to sources cited by the Wall Street Journal. White House officials also raised concerns about potential conflicts of interest involving AI company entanglements in government roles. The episode signals major tech CEOs' direct influence over AI policy direction at the highest levels.

SEC Greenlights Tokenized Stock Trading on Permissioned Onchain Venues With 75-Symbol Cap and Five-Year Sunset

The SEC's Division of Trading and Markets issued conditional relief allowing tokenized National Market System stocks to trade through permissioned liquidity pools, capped at 75 Tier 1 symbols trading up to 0.25% of prior-month average daily volume, with issuers retaining veto rights over third-party tokenizations. The framework operates under the principle of "innovation without arbitrage," preventing market participants from using blockchain-based listings to escape obligations that apply to traditionally traded instruments. SEC Trading Division Director Jamie Selway framed the tokenized securities framework as part of a parallel effort with the CFTC to harmonize rules on swap data reporting, portfolio margining, and product definitions. Paul Atkins's characterization of the relief as "interim" signals that durable rulemaking will follow, creating the first regulatory pathway for equities onchain in the U.S., though the framework's constraints suggest a measured approach to scaling.

AI Researchers and Economists Locked in $425,000 Public Bet on Whether AI Will Drive 15% GDP Growth Within Seven Years

The "fast growth" side, mostly AI researchers, forecasts explosive economic expansion from AI, while the "slow growth" side, mostly economists including Alex Imas, point to adoption friction, diffusion delays, J-curve effects, and physical-world automation bottlenecks as limiting factors. The forecast divergence reflects a fundamental disagreement about AI's near-term economic impact across the industry.

TypeSafe Launches Jev, a Decision-Focused AI Model Claiming 20-200× Faster Responses and 40-400× Lower Costs Than Frontier Models

TypeSafe AI, founded by Diogo Almeida (ChatGPT/RLHF co-inventor), released Jev, a "System One Model" built for fast, structured decisions rather than text generation. Jev returns structured decisions and probabilities in parallel rather than generating token-by-token, targeting use cases like ticket routing, fraud detection, and churn prediction. The model uses a new training method called Reinforcement Learning for Calibrated Decisions (RLCD) and achieves similar intelligence on System One tasks compared to existing LLMs while being two orders of magnitude faster and more efficient. Jev guarantees type-safe outputs (mathematically impossible to hallucinate or produce type errors), costs $0.042 per million input tokens with free output tokens, and demonstrates the broader shift toward optimizing for specific task economics rather than general-purpose capability.

High-Growth AI SaaS Companies Trade at 17.5x NTM Revenue Versus 3.5x for Low-Growth Peers

The median SaaS company trades at 4.2x NTM revenue with 13% projected growth, while the top 5 companies command 36.4x multiples. High-growth companies (>22% NTM growth) trade at 17.5x versus mid-growth at 6.5x, with median gross margins at 76% and FCF margins at 21% across the sector, illustrating the valuation premium commanded by AI-driven growth in public SaaS markets.

Measuring the Pace of AI Development

Anthropic published a report on measuring three critical dimensions of AI development pace: AI-led R&D automation, oversight of AI agents, and compute allocation to safety. Main takeaways:

• Claude leads 26% of Anthropic's AI R&D work (up from ~0% at start of 2026), with 90%+ of work involving AI collaboration at AL3+ automation levels
• Approximately 30,000 agents operate simultaneously on Anthropic's research platform with 100% online monitoring coverage; 0.002% of agent decisions blocked, ~1-2 per 1,000 flagged for review
• 6% of AI R&D compute allocated to safety work; 12% of compute in AI-driven R&D allocated to safety; metrics designed for cross-lab verification and third-party validation
• Anthropic proposes embedding independent third-party evaluators with access comparable to internal risk teams to verify safety practices and monitor key metrics

Will AI Soon Lead to Double-Digit GDP Growth?

Alex Imas and Benjamin Moll in Alex Imas's Substack published an essay titled "Will AI Soon Lead to Double-Digit GDP Growth?". Main takeaways:

• Standard growth theory models can produce explosive growth (12-20%+ annually) when automation reaches full economy-wide coverage, but this requires all five key assumptions to hold simultaneously in practice
• The authors identify five critical assumptions unlikely to be satisfied: fast economy-wide automation, continued high spending on automated goods, sufficient demand/investment, no AI-driven cyber incidents, and explosive R&D automation—each represents a potential failure point
• They argue 4-5% annual growth is a more reasonable baseline for the next 15 years, which would still represent a doubling of per-capita wealth and be historically massive
• The authors have put $460,000 against $115,000 in a bet that U.S. real GDP per capita will not grow 15%+ in any single year through 2033, implying a 20% probability they assign to the fast-growth scenario
• Measurement issues mean GDP may undercount welfare gains from AI, but this doesn't resolve the fundamental economic constraints on near-term growth acceleration

Introducing System One Models & Jev

TypeSafe AI released Jev, a new class of frontier model optimized for structured decision-making rather than text generation. Main takeaways:

• Jev achieves frontier-level intelligence on decision tasks while being 193.6x faster and 444.6x cheaper than GPT-5.6 Terra and Fable 5.1 on workflow evaluations, with output tokens priced at effectively free vs. ~5x input cost for LLMs.
• Model guarantees zero hallucination and type-safety through parallel probabilistic sampling instead of autoregressive token generation, making it suitable for production automation where latency and reliability are critical.
• Trained with novel Reinforcement Learning for Calibrated Decisions (RLCD) method that optimizes for epistemically honest confidence scores rather than human preference, enabling consistent decision-making with calibrated uncertainty estimates.