Six Chinese AI Firms Got Caught Copying America's Best Models. The US Fix: Quietly Make Their Answers Worse.

Wednesday’s AI news split into three lanes that rarely share a headline: a security fight over who gets to use the best models, a scientific tool that could change how doctors read your DNA, and a hardware company betting its next phone on a chattier assistant. In India, the government approved its most ambitious rail expansion in years, the unglamorous kind of infrastructure spending that quietly decides how goods and people move for a generation.

Washington says six Chinese AI companies have been copying America’s best models. Its fix is to make their answers worse.

On September 8, the National Security Agency, the FBI and the Cybersecurity and Infrastructure Security Agency published a joint advisory accusing six Chinese AI companies, DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI, of running an industrial-scale distillation campaign against American AI companies since late 2024. The agencies say the six firms pulled billions of tokens across millions of queries from Anthropic’s Claude, OpenAI’s GPT and Google’s Gemini, then used that output to train their own models, a shortcut that skips much of the expensive trial and error that produced the originals.

The advisory gets specific. It says DeepSeek drew on multiple versions of Claude, GPT and Gemini to build its R1 and V3 models, and argues the company’s widely cited $5.6 million training cost is misleading because it leaves out the cost of that distilled data. It says Moonshot AI pulled output from Anthropic’s Claude Fable and OpenAI’s GPT-4o to build its Kimi K2 and K3 systems. The companies allegedly used fraudulent accounts, bulk premium subscriptions and proxy routing services, sometimes called transfer stations, to get around usage limits and regional blocks.

The advisory’s most unusual recommendation is not to simply block the accounts it flags. Instead, it tells American AI companies to quietly serve worse answers to accounts identified with high confidence as running malicious distillation, on the theory that an outright ban just tips off the operator to switch accounts, while a silent quality drop poisons their training data without them knowing why. The agencies also want US firms to build a shared intelligence network to track these actors across the industry. A spokesman for the Chinese embassy in Washington, Liu Chang, dismissed the advisory as a deliberate attack on China’s AI progress. None of the six companies named responded to requests for comment.

Google DeepMind mapped what every possible mutation to human DNA would do

Also on September 8, Google DeepMind released the AlphaGenome Atlas, a free public database of predicted effects for all roughly 9 billion possible single-letter changes to the human genome, plus more than 100 million short insertions and deletions seen in real people’s DNA. It is a 1-petabyte dataset, more than 30 times the size of DeepMind’s AlphaFold protein database, and it is free for non-commercial research immediately, with commercial licensing on Google Cloud to follow.

Alongside the Atlas, DeepMind released a single number called the AlphaGenome Variant Impact score, which combines predictions from AlphaGenome with an earlier DeepMind tool called AlphaMissense to estimate how damaging a given mutation is likely to be. The idea is to save researchers from running their own predictions for a mutation that might matter, since the answer now sits in a lookup table. DeepMind says a team at the Broad Institute of MIT and Harvard used the Atlas and its scores to help flag a non-coding variant, the kind of mutation that sits outside genes and is notoriously hard to interpret, as a likely cause of a case of severe epilepsy. Genome sequencing has gotten cheap. Knowing what a given mutation actually does has not, and that is the gap this tool is aimed at closing.

Apple built AI into the iPhone launch instead of around it

Apple’s September 9 event, titled “Surprise and Shine,” was nominally about new iPhones, but the company built its pitch around AI more than it has at any launch since it first announced Apple Intelligence. Apple introduced its third generation of Foundation Models, five named models built for different jobs, including a compact on-device model with roughly 20 billion parameters and a larger Cloud Pro model that runs on Apple’s own servers. The new iPhone 18 Pro ships with the A20 Pro chip and pairs the new models with a rebuilt Siri that can hold a conversation, take actions across apps, and reason about what is on screen instead of answering one question at a time.

Siri’s new AI capabilities are launching in beta in English only, with French, Japanese, Korean, Portuguese and Spanish support due in October. The feature will not be available in the European Union or China at launch, two markets where regulatory requirements, and, in China’s case, licensing rules for foreign AI systems, have repeatedly slowed Apple’s rollout plans. The iPhone 18 Pro starts at $1,199, with preorders opening September 12 and general availability on September 18.

India approved its biggest rail expansion in years

On September 9, the Cabinet Committee on Economic Affairs, chaired by Prime Minister Narendra Modi, approved eight railway multitracking projects worth a combined 20,804 crore rupees, roughly $2.5 billion, spread across 31 districts in nine states. Multitracking means adding extra parallel lines alongside existing routes. It is less dramatic than laying a brand new railway, but it is often the more useful fix, since it lets more trains run on a corridor that already exists instead of waiting years for fresh land acquisition and route surveys.

Five of the projects, costing 10,021 crore rupees, cover 17 districts across Tamil Nadu, Andhra Pradesh, Karnataka and Telangana and will add about 540 kilometers of track, including new lines and doubling work on the Arakkonam-Renigunta, Whitefield-Bangarapet, Hosur-Omalur and Salem-Karur-Dindigul routes, plus multitracking between Secunderabad and Kazipet. The other three projects, costing 10,783 crore rupees, cover 14 districts across West Bengal, Jharkhand, Odisha, Madhya Pradesh and Chhattisgarh, adding about 656 kilometers, including a fourth line between Kharagpur and Jharsuguda and a fourth line between Katni and Pendra Road. That second package alone is expected to add 27 million tonnes a year of freight capacity on routes that mostly move coal, cement, iron and steel.

Together the eight projects are expected to add roughly 1,196 kilometers of track and connect an estimated 6,911 villages with a combined population above one crore, or 10.8 million people. One crore rupees works out to a little over $1.2 million at current exchange rates, which puts the overall price tag in perspective: this is a government spending real money on the parts of a railway network nobody photographs. The government has targeted 2029-30 for completion.

Every story here is really about the same question: who controls the flow of something valuable, and what happens when someone else starts taking more of it than they are supposed to. The US wants to control the flow of frontier AI knowledge, and its chosen answer is quiet deception rather than a wall. DeepMind wants to make the flow of genetic information free, betting that opening it up helps more people than it costs. Apple wants to control the flow of your attention through a phone, now with a chattier assistant built in. India, meanwhile, is just trying to keep more trains moving on tracks that already exist, which turns out to be one of the more reliable ways a country gets richer.