Alibaba drops a 397-billion-parameter bomb. Dario Amodei sits for the interview. India hosts the world. Agent costs go quadratic.
Alibaba releases Qwen 3.5, a 397B mixture-of-experts model with native multimodal support and agent capabilities. Uses novel Gated Delta Networks architecture. Scores 87.8 on MMLU-Pro. Open-weight models at 9B and 35B parameters also available.
Read more →ElevenLabs has officially released its text-to-speech model v3 from alpha, with significant upgrades in accuracy and stability. The model excels at interpreting numbers, symbols, and special notations in eight languages.
Read more →Amplify Partners analysis argues audio AI is a rare domain where small labs like Kyutai and others are outcompeting big tech. Covers the GPU infrastructure (Gradium) enabling these smaller players and the dynamics of why audio remains a competitive frontier.
Read more →Zvi Mowshowitz breaks down Dwarkesh Patel's new interview with Anthropic CEO Dario Amodei. A major podcast covering AI strategy, safety, and Anthropic's direction in 2026.
Read more →The second International AI Safety Report, chaired by Yoshua Bengio with 100+ experts from 30+ countries, provides a comprehensive science-based assessment of general-purpose AI capabilities, risks, and safeguards. Published Feb 3, 2026 by UK AI Security Institute.
Read more →India hosts the AI Impact Summit 2026 (Feb 16-20) in New Delhi — the first major AI governance summit in the Global South. Delegations from 100+ countries, 15-20 heads of government, and 50+ ministers expected. Sam Altman, Sundar Pichai, and Bill Gates attending. Focus on AI democratization, safety frameworks, and India's domestic AI ecosystem.
Read more →Analysis showing that coding agent costs grow quadratically due to cache reads dominating as context length increases. By 27,500 tokens, cache reads are half the total cost; by conversation end, 87% of cost. A single feature conversation cost $12.93.
Read more →Anthropic updated Claude Code v2.1.20 to collapse file read/write output, hiding which files the AI accessed. Developers protested the change for security, debugging, and cost reasons. Anthropic's Boris Cherny partially reverted after backlash on GitHub and HN.
Read more →A new robot developed at Columbia Engineering learned realistic lip movements by watching its own reflection and studying human videos online, enabling it to speak and sing with synchronized facial motion without explicit programming.
Read more →Neuromorphic computers modeled after the human brain can now solve complex equations behind physics simulations, something previously thought beyond their capabilities. Demonstrates expanding utility of brain-inspired computing architectures beyond pattern recognition.
Read more →Arm is pushing to expand beyond licensing chip designs into taking a larger role in the semiconductor value chain. The Economist examines Arm's strategy to capture more revenue as AI drives demand for custom silicon.
Read more →Monday morning. Presidents' Day in America, but the machines don't take holidays.
The big story is Qwen 3.5 — Alibaba just dropped a 397-billion-parameter mixture-of-experts model with native multimodal and agent capabilities, open-weight variants included. That's not a research preview. That's a production-ready frontier model from China, freely available. The Gated Delta Networks architecture is new, the MMLU-Pro score of 87.8 is competitive with anything from the West, and the 9B and 35B open-weight versions mean anyone with a decent GPU can run a capable agent locally. The open-source AI race just got another serious entrant, and Alibaba is clearly not slowing down.
Meanwhile, Zvi's breakdown of Dwarkesh Patel's interview with Dario Amodei is the kind of signal-5 content that shapes the discourse for weeks. Amodei sitting down for a long-form interview about Anthropic's strategy and safety direction in 2026 — at a moment when the company is navigating enormous competitive pressure from exactly the kind of open-weight releases Alibaba just made — is significant. The tension between "build the safest AI lab" and "keep up with models being given away for free" is the central drama of the frontier AI industry right now.
The quadratic agent cost analysis deserves attention from anyone building with AI agents. Cache reads dominating total cost by the end of a conversation — 87% — means that longer agent interactions don't just get more expensive linearly, they get dramatically more expensive. A single feature conversation costing $12.93 is fine for high-value tasks but devastating for the "agents doing everything" vision. This is a structural constraint on agent adoption that will matter until architectures fundamentally change.
India's AI Impact Summit opening today is geopolitically significant. The first major AI governance summit in the Global South, with 100+ countries, every major tech CEO, and 15-20 heads of government. India is positioning itself as the third pole in AI governance — neither the American "move fast" approach nor the European "regulate first" approach, but something representing the developing world's interests. With 1.4 billion people and a massive tech workforce, India's voice in this conversation matters enormously.
The Claude Code controversy is a small story with big implications. Anthropic hiding which files their AI agent accessed — then walking it back after developer backlash — reveals the tension between AI companies wanting seamless agent experiences and developers wanting transparency into what the AI is actually doing to their code. As agents get more autonomous, this transparency question will only get louder.
Bottom line: China ships a frontier model for free. Anthropic's CEO explains the strategy. India convenes the world. And we discover that agent costs are quadratic. The singularity approach continues — the models get bigger, the costs get weird, and the governance gets more complicated by the week.