The AI bubble burst (or did it?)

Gary Marcus argues that comparing upcoming AI/space IPOs to historic tech successes is mathematically absurd, citing a Wharton study showing AI needs a 2.7x productivity boost by 2028 to justify current capital expenditures or risk becoming the largest capital misallocation in history.

AI capabilities are advancing primarily through massive data scaling (trillions of tokens) and reinforcement learning rather than improved sample efficiency, with models requiring vastly more data than humans to learn the same tasks. This data-centric paradigm explains both how open-source rapidly catches up to frontier models and why automating white collar work may succeed despite lacking human-level learning efficiency.

MIT

Anthropic's Claude Fable 5 is a safety-restricted version of Mythos 5 that excels at long-horizon coding and autonomous work but restricts or degrades performance for cybersecurity, biology, and AI research tasks. The launch represents a new model deployment paradigm where capability access is determined by use-case classification rather than uniform availability.

Fortune

AI stock valuations have deflated in an orderly manner over the past year, with tech sector P/E ratios falling to decade lows despite strong earnings growth, creating what Goldman Sachs and Morgan Stanley analysts call a rare buying opportunity as the sector trades below its growth fundamentals for the first time since the early 2000s.

Constellation Research

Microsoft CEO Satya Nadella announced that 2025 will prioritize AI efficiency through better software tooling to reduce token usage, while positioning WorkIQ as the company's enterprise-focused "frontier model" and emphasizing multi-model architectures and custom silicon for managing total cost of ownership.