Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
OpenAI launched GPT-5.5 on April 23, its most capable model yet, with significantly improved coding, computer use, and deep research capabilities. The model matches GPT-5.4 per-token latency while using fewer tokens to complete tasks, and ships with GPT-5.5 Thinking and Pro variants for paid subscribers. Available via ChatGPT, Codex, and API — free-tier users are excluded.
Chinese AI lab DeepSeek released V4 in preview on April 24, featuring a 1M-token context window (up from 128K), built on Huawei Ascend chips. The V4-Pro-Max variant outperforms GPT-5.2 and Gemini 3.0 Pro on some reasoning benchmarks while dramatically undercutting them on price — input tokens start at $0.14 per million. Fully open source.
Google committed $10B now at a $350B valuation, with up to $30B more contingent on performance milestones, making this the largest single AI investment to date. This follows Amazon's parallel deal of $5B now plus up to $20B more. Anthropic's ARR has surged past $30B, up from roughly $1B at the start of 2025.
At Cloud Next '26, Google announced TPU 8t (training) and TPU 8i (inference) — its first generation to split into purpose-built architectures. A single TPU 8t superpod scales to 9,600 chips delivering 121 ExaFLOPS. TPU 8i delivers 80% better price-per-dollar than the prior gen with 3x more on-chip SRAM. Both generally available later this year.
Google co-founder Sergey Brin assembled a dedicated team within DeepMind, led by pre-training engineer Sebastian Borgeaud, focused on catching Anthropic's lead in agentic coding. Brin's internal memo stated Google "must urgently bridge the gap in agentic execution." Anthropic reportedly uses AI for nearly all of its own coding; Google is at roughly 50%.
China's NDRC officially blocked and ordered the unwinding of Meta's acquisition of Manus (Butterfly Effect), a Singapore-based agentic AI startup. Meta had largely completed the deal in December 2025, with staff already relocated. Beijing is treating agentic AI as strategically critical IP and escalating efforts to block U.S. acquisition of Chinese-origin AI talent.
Meta is laying off 10% of staff (~8,000 people, effective May 20) while projecting $115–135B in AI capex for 2026. Microsoft offered voluntary separation to 7% of its U.S. workforce (~8,500). Nearly 50% of the 92,000 tech layoffs in 2026 so far are attributed to AI automation replacing human roles — a threshold that now looks structural, not cyclical.
Analysis by Pangram Labs found that 21% of ICLR 2026 peer reviews (~15,900) were fully AI-generated, with over half of all reviews showing AI involvement. This was compounded by an OpenReview API bug that exposed author and reviewer identities for ~10,000 papers. The conference assigned reviewers five papers each in a two-week window — an unsustainable load that incentivized automation.
Former staff from OpenAI, DeepMind, Anthropic, and xAI are raising nine-figure rounds for months-old ventures including Periodic Labs, Ricursive Intelligence, and Humans&. VCs have funneled $18.8B into AI startups founded since early 2025, on pace to surpass last year's $27.9B. The core driver: commercial pressure at frontier labs is squeezing out exploratory research.
ICLR 2026 concludes today after accepting 3,462 papers from 11,617 submissions (29.8% acceptance rate). Outstanding Paper went to "Transformers are Inherently Succinct," which offers a formal lens on why Transformers outperform RNNs by proving they encode concepts more compactly. The conference also awarded Test of Time honors to papers from ICLR 2016.
