Latest

Microsoft's new AI rulebook says its models are not conscious

The Guardrails

Microsoft's new AI rulebook says its models are not conscious

Microsoft has put its own model constitution up for public comment, the UN's top human rights official is demanding "cast iron guarantees" on frontier AI, and the network layer under AI clusters just got $205 million of new money behind it. Microsoft published a 37-page "humanist AI code of conduct" that tells its models they are not conscious and never should pretend to be. The draft, posted by Microsoft AI and reported first by The Verge, states that "people matter more than AI," that models

Cohere CEO calls AI models the most potent cyber weapon ever built

The Guardrails

Cohere CEO calls AI models the most potent cyber weapon ever built

The cyber-safety debate found a dissent inside itself today, the White House picked a side in the camera wars, and some people are knitting their way out of the surveillance net. AI models are "the most potent cyber weapon that has ever been created," Cohere CEO Aidan Gomez told CNBC — and unlike the slowdown camp he is being grouped with, he doesn't think a regulator would have stopped the last breakout. The co-author of "Attention Is All You Need" said the models are "incredible at finding an

d-Matrix stacks DRAM on its chip and claims 100 TB/s

The Stack

d-Matrix stacks DRAM on its chip and claims 100 TB/s

Hot Chips 2026 wrapped with a memory bet, and the funding calendar is unusually busy for a Monday. d-Matrix says stacking memory directly on the compute die beats HBM by 20x on bandwidth density. The startup's Raptor accelerator puts a TSMC N4 logic die face-to-face on top of custom 3D DRAM and claims roughly 100 TB/s per card at 32 GB — against a practical ceiling of about 20 TB/s for today's HBM4 packages, which d-Matrix attributes to limited package "beachfront" and raw physics: driving 100

China's TC260 adds runaway agents to its safety framework

The Guardrails

China's TC260 adds runaway agents to its safety framework

Beijing spent the same week it dismissed Washington's slowdown panic issuing another spec — and this one is about agents, not devices. China's national cybersecurity standards body released the Artificial Intelligence Safety Governance Framework 3.0 on Monday, and the new edition spends real time on models that won't stop. The annex work is the tell: the document, published at the opening of the 2026 National Cybersecurity Week in Jinan, carries a dedicated agentic risk-management framework alo

King Charles hosts the frontier labs while Washington won't

The Guardrails

King Charles hosts the frontier labs while Washington won't

Monday's pacing fallout went two directions at once: the one Western figure with no electoral stake in AI is convening the labs, while Beijing upgraded its rebuttal to an accusation — and quietly already regulates for Amodei's worst case. Britain's King Charles will host leaders of Nvidia, Google DeepMind, OpenAI and Anthropic, along with the UK's AI minister, this week to discuss whether the technology needs a shared set of principles. Buckingham Palace framed the agenda as how AI can benefit

Anthropic tells investors it will be profitable for a second quarter

The Arena

Anthropic tells investors it will be profitable for a second quarter

The week AI's biggest labs spent asking for permission to slow down ended with one of them telling shareholders the money is very good — and the number attached to that claim reframes the entire pacing debate. Anthropic has told shareholders its adjusted operating income will be positive for a second straight quarter, with gross margins above 80%, according to the Financial Times, which sourced the report to multiple people with knowledge of the matter. The margin figure comes with the caveat t

How to — decide if your model needs fine-tuning

The Stack

How to — decide if your model needs fine-tuning

When an AI feature misbehaves, "just fine-tune it" is the most popular and most expensive wrong answer. This is the order of operations that tells you whether the weights are actually the problem — before you spend a fraction of your budget finding out. The tell is rarely dramatic. Your support bot starts inventing refund policies it was never given, or a summarizer keeps ignoring the output format your whole product depends on. Someone on the team says the model "just needs training on our dat

Firmus targets a $5B IPO, Australia's biggest listing since Telstra

The Arena

Firmus targets a $5B IPO, Australia's biggest listing since Telstra

The pacing debate moved markets all morning. Now it has a counterweight: the biggest public-market test of the AI buildout, roadshowing this week in the least likely place. Firmus Technologies is seeking to raise up to $5 billion (A$7 billion) in an initial public offering, targeting an October launch with a debut on the Australian Securities Exchange as early as late October. The Australian AI-infrastructure company began meeting investors in Asia on Monday, with Europe and the United States b

Open Source Radar — September 14: agents get a QA department

The Stack

Open Source Radar — September 14: agents get a QA department

Today's trending board is mostly about control, not capability: a code reviewer that Alibaba ran internally for two years, a verifier that refuses to let a model assert facts, a memory store for agents that forget on session close, a scanned skill registry, and a browser agent that hands you a spreadsheet instead of an essay. OpenCodeReview (Go, 24,147 stars) — Alibaba's internal AI code review assistant, open-sourced after the company says two years of use across tens of thousands of its own d

Ant's new guardrail reads the model's own hidden states mid-sentence

The Guardrails

Ant's new guardrail reads the model's own hidden states mid-sentence

A Chinese fintech giant shipped the safety layer most labs only describe in system cards, and a Science Advances paper taught an AI to watch chemical bonds break. Ant Group has released SingProbe, a guardrail that runs inside the model it polices. Instead of a second classifier re-reading the text, SingProbe reads the base model's own hidden states as tokens are produced, and outputs a ten-dimensional signal covering three jobs at once: what the user is asking for, whether the answer taking sha

Deep Dive — Tokens halved in price. The bill to make them didn't

The Arena

Deep Dive — Tokens halved in price. The bill to make them didn't

Something unprecedented happened to the price of machine intelligence in August, and the industry's favorite growth story did not survive it. Silicon Data's LLM Token Spend Index — which tracks what buyers actually pay per million tokens across the market — fell 29% in a single month to $0.97, the first print under a dollar, and roughly half its May 28 peak of $2.07. Three months earlier, the consensus was that smarter models earn more pricing power. Now the flagship models are being discounted

Hugging Face runs async RL training with no NCCL and no weight ship

The Stack

Hugging Face runs async RL training with no NCCL and no weight ship

The hardest part of reinforcement learning is not the algorithm. It is getting a trainer and a bunch of inference servers to keep passing weights back and forth without one of them stalling. Hugging Face just published a recipe that does it with three rented jobs, one storage bucket, an HTTP proxy, and no NCCL at all. Hugging Face trained a reasoning model end to end using async GRPO and LoRA, and the whole weight sync is a file upload. The setup is deliberately small: one job runs the trainer

See all