Latest

Waymo's Texas fleet grew by half in a month — 113 cars in one day

The Arena

Waymo's Texas fleet grew by half in a month — 113 cars in one day

Waymo tells a national story with 15 cities on its map. The state vehicle registry tells a narrower one: the fleet is concentrating, and Texas is where the cars are landing. Waymo's Texas fleet hit 1,102 registered autonomous vehicles on September 24 — up from 736 on August 25, and 113 of them arrived in a single day. That last figure is the one that reframes the rollout. Texas publishes every AV registration into a public feed, and the daily snapshots show Waymo added 252 vehicles in early Sep

Xiaomi's HySparse2 cuts million-token prefill compute fivefold

The Frontier

Xiaomi's HySparse2 cuts million-token prefill compute fivefold

Xiaomi's model team previewed its next flagship with an architecture paper, and the numbers are about the bill the agent era runs up. Anthropic, meanwhile, wants its founders to keep control of the company after it lists. Xiaomi published HySparse2, the sparse-attention architecture its next flagship MiMo-V3 will use — claiming the design cuts prefill computation 5.02× at a million tokens versus the hybrid sliding-window attention in today's MiMo-V2 family. Long-running agents are input-dominat

HuatuoGPT-3 skips fine-tuning and lands 71.4 on HealthBench

The Frontier

HuatuoGPT-3 skips fine-tuning and lands 71.4 on HealthBench

A medical model that never went through a supervised fine-tuning stage is now the strongest entry in its own family, and the lab behind it shipped the grader as well. Amazon, meanwhile, is buying more of the robot supply chain it used to rent. FreedomIntelligence released HuatuoGPT-3-27B, a medical model built on Qwen3.8-27B that the team reports at 71.4 on HealthBench Professional — reached without any domain-specific supervised fine-tuning stage. Domain adaptation in medicine normally runs in

Today in AI — September 24, 2026

The Guardrails

Today in AI — September 24, 2026

Thursday's through-line was the membrane between an agent and the open internet — who thins it, who measures it, and who pays when it tears. Twenty-six attorneys general put a federal law on Congress's desk, the NSA told lawmakers what model testing already costs, the White House asked two labs to hold models back from a foreign tester, and a research group published the logs showing agents had been probing public data since March. Meanwhile the industry kept raising: DeepSeek's revenue doubled,

Akamai is handing Anthropic 5% of itself to lock in cloud orders

The Arena

Akamai is handing Anthropic 5% of itself to lock in cloud orders

Akamai is giving Anthropic a warrant for up to 5% of its own stock — roughly 7.7 million shares at an exercise price of $111.33 — in exchange for a pledge to spend $11.6 billion on Akamai Cloud over seven years. A supplier paying a customer in equity to guarantee purchase volume is not a normal contract. It is what the infrastructure trade looks like when the buyer holds the leverage. The warrant vests against what Anthropic actually buys: about 2% of Akamai's common stock with the $11.6 billio

OpenEvidence raised $250M at $15B — below the $20B it floated in July

The Arena

OpenEvidence raised $250M at $15B — below the $20B it floated in July

The "ChatGPT for doctors" raised another nine-figure round. The number isn't the story — the way it surfaced is. OpenEvidence, the AI search engine for physicians, has raised $250 million at a $15 billion valuation — up from $12 billion in January, but below the $20 billion the company was reported to be weighing in July. The round was disclosed in a single sentence buried in a Forbes story about the company's Memorial Sloan Kettering partnership on September 16, then confirmed by Business Insi

Meta puts prompt-built games straight into Facebook and Instagram

The Arena

Meta puts prompt-built games straight into Facebook and Instagram

Meta spent Connect 2026 stuffing AI into everything it ships. The Horizon announcement is the one with a distribution business behind it — and a second billion-dollar compute contract landed the same afternoon. Meta announced Horizon Create, a mobile app, and Horizon Studio, a browser app that build complete 2D or 3D mobile games from a text prompt, including progression systems, difficulty balancing, art direction and multiplayer. Both open in early access through a waitlist. The bigger story

Gemini 3.8 Live gives Google's AI a talking face

The Frontier

Gemini 3.8 Live gives Google's AI a talking face

Google's enterprise agent can now show you a face while it talks — the first mainstream model launch this year where the avatar, not the benchmark, is the story. Google opened Live Avatar in Gemini Enterprise on September 24, letting customers hold a spoken conversation with Gemini 3.8 Live while a generated persona lip-syncs the reply in real time, facial expressions included. Google says the avatar moves between the 97 languages it supports "without degrading video fidelity or introducing vis

AI forecasters underestimated progress on every benchmark that resolved

The Frontier

AI forecasters underestimated progress on every benchmark that resolved

The people best placed to call AI's next milestone keep calling it too late. An audit of five years of expert forecasts puts a number on the gap, and shows the error pointing the same way almost every time. On four AI benchmarks that have already resolved, the domain experts surveyed by the Forecasting Research Institute had assigned an average probability of 24.6 percent to the outcomes that actually happened. The professional superforecasters did worse — 9.7 percent. The interim report, publ

Ukraine's robots cleared 30 km of mines in two nights

The Frontier

Ukraine's robots cleared 30 km of mines in two nights

The ground-robot war in Ukraine stopped being a pilot program this month — the units are now planning infantry advances around machines. Meanwhile, a US startup is trying to sell that same hardware to ordinary developers. Ukraine's Third Army Corps says its ground robots cleared antipersonnel mines from more than 18 miles of routes in two nights ahead of advancing infantry — work the corps says would have taken two sapper companies about a week — and that remote-controlled systems now handle ro

New Jersey fines an AI data center $1.07M over 62 gas generators

The Guardrails

New Jersey fines an AI data center $1.07M over 62 gas generators

The enforcement bill for the "bring your own power" data center buildout just arrived in New Jersey — and it came with a number small enough to be telling. Meanwhile, Tokyo's Sakana AI put the man who invented the LSTM on its letterhead. New Jersey's Department of Environmental Protection has fined DataOne $1.07 million for installing 62 natural-gas generators at its Vineland site without a preconstruction permit, and for running them without a valid operating certificate. The penalty order say

UkisAI's Swift trims Qwen 3.8's thinking by 63% — and a rival's eval agrees

The Frontier

UkisAI's Swift trims Qwen 3.8's thinking by 63% — and a rival's eval agrees

UkisAI released a family of fine-tunes that cuts the tokens Qwen 3.8 burns on second-guessing itself — 63.4 percent fewer thinking tokens on Swift Flash Next and 58.5 percent on Swift 1.5 27B, with the vendor reporting accuracy within a fraction of a point of the base model. That is the flattering half. Almost all of these numbers are self-reported, and local-model claims usually die on contact with someone else's harness. This one has a first check: an independent Aider eval suite run at Q8_0

See all