OpenAI Bets on Coding as Its Own Domain
GPT-5.2-Codex is a dedicated agentic coding model achieving state-of-the-art results on SWE-Bench Pro and Terminal-Bench 2.0. It signals a shift from AI-assisted coding to AI-native software engineering, and from general reasoning models to task-specific ones deployed through a common interface. The real battle is now with Anthropic's Claude Opus 4.6 for dominance in enterprise software pipelines.
Read full story →How Anthropic Is Trying to Actually Measure What AI Does to Work
Anthropic's new Economic Primitives framework decomposes jobs into atomic cognitive tasks and measures AI involvement at that level. The January 2026 Economic Index report shows the largest gap between current AI use and automation potential sits in professional services. But the methodology has a blind spot: it can only see workers who are still working.
Read full story →The Audio Stack Goes Open: Mistral and Cohere Bet on Voice AI the Same Day
On the same day, Mistral released Voxtral TTS and Cohere released Transcribe, an open-source speech recognition model at just 2 billion parameters. Together they cover opposite ends of the audio pipeline. The convergence signals that the barriers to open-source voice AI have fallen faster than expected, and the closed-model premium on audio quality may have a shorter shelf life than the big labs are counting on.
Read full story →Who Pays for AI's Power? Congress, Ratepayers, and Carbon Credits Enter the Fight
Senators Hawley and Warren want mandatory energy reporting from data centres as residential electricity prices have risen 36% since 2020. A Washington Post investigation found communities hosting data centres are absorbing infrastructure upgrade costs meant for tech companies. Meanwhile CNBC reports Big Tech's carbon credit purchases have exploded, with Microsoft's expenditure running ahead of actual emissions reductions.
Read full story →The Memory Crisis at the Heart of the AI Boom
GPUs get the headlines, but it is high-bandwidth memory that is quietly throttling AI expansion. A global HBM shortage, SK Hynix rushing to US capital markets, Arm making its first direct chip play, and Google's TurboQuant research unsettling memory-stock investors all point to the same underlying stress: the AI supply chain has a memory problem it cannot build its way out of quickly.
Read full story →Google DeepMind Builds a Meter for AI Manipulation
DeepMind published a genuinely novel contribution to AI safety this week: an empirically validated toolkit for measuring whether models are actually changing people's beliefs, not just theoretically capable of doing so. The open methodology release and new Critical Capability Level for manipulation set a concrete standard other labs can now be measured against.
Read full story →When Earth Runs Out of Room: Nvidia's Orbital Data Center Bet
At GTC 2026, Nvidia unveiled the Vera Rubin Space-1, hardware designed for orbital data centers. The pitch is that AI demand has hit genuine resource limits on Earth: energy politics, water rights, and grid capacity are becoming more constraining than silicon. Space removes all three, and the fact that Nvidia is building for it tells you how constrained the terrestrial options have become.
Read full story →The Billion-Dollar Bet Against LLMs: Yann LeCun's AMI Labs
AI's most prominent LLM skeptic has raised $1.03 billion for AMI Labs, a startup pursuing world models rather than next-token prediction. LeCun's argument that language models lack genuine causal understanding of the world has always been coherent; now it is a billion-dollar bet, and investors are paying to see who is right.
Read full story →The Measurement Trap
Anthropic's new economic primitives methodology is genuinely useful, and it is going to make the labour displacement problem look smaller than it is. Task-level measurement captures what AI does; it does not capture what happens to the people whose income depended on those tasks being expensive. The people already displaced are not in the logs.
Read opinion →