Yesterday, Anthropic alleged that it had detected what it described as “an industrial scale campaign” by DeepSeek and two different distinguished Chinese language AI labs, Moonshot AI and MiniMax, to distill its Claude fashions. Distillation is the time period AI researchers use to explain a way of boosting the efficiency of smaller, often weaker AI fashions by fine-tuning them on the outputs of a bigger, stronger mannequin. On this case, Anthropic claims the three Chinese language AI firms created 24,000 faux accounts so as to generate 16 million exchanges with Claude that they then used to coach their very own fashions, in violation of Anthropic’s phrases of service. (Of those exchanges, DeepSeek was solely answerable for 150,000 of them, in line with Anthropic, however DeepSeek-linked accounts appeared notably curious about distilling Claude’s reasoning capabilities.)
Additionally yesterday, Reuters reported, citing an nameless senior U.S. authorities official, that the U.S. believes DeepSeek educated V4 utilizing Nvidia’s newest era Blackwell AI GPUs, in seemingly violation of U.S. export controls that have been supposed to forestall Chinese language AI firms from buying Nvidia’s most superior chips. The story mentioned the U.S. believed that DeepSeek has a knowledge heart in Interior Mongolia stuffed stuffed with Blackwells–though it mentioned the U.S. was uncertain precisely the way it obtained them.
Utilizing AI to assist map international provide chains
‘Complexity will almost certainly get worse’
What does all of this must do with final week’s tariff ruling? Every little thing. As a result of one among Altana’s key merchandise is successfully an AI-powered tariff administration system. Smith described an “agentic” workflow that automates the notoriously arcane enterprise of assigning Harmonized System (HS) codes to items—the classification that determines what tariff fee applies to any given import—in addition to calculating nation of origin below commerce guidelines, one thing that has grow to be phenomenally sophisticated within the period of transhipment and tariff evasion. Add to {that a} tariff state of affairs planner that enables firms to mannequin the affect of adjusting commerce guidelines throughout their total prolonged provider community. Use of Altana’s tariff calculator has spiked 213% up to now week, the corporate reviews. About 50% of these calculations involved articles containing metals, whereas 32% have been for merchandise whose nation of origin was China.
Or, at the very least, they didn’t know earlier than Altana and its AI got here round.
FORTUNE ON AI
OpenAI companions with McKinsey, BCG, Accenture, and Capgemini to push its Frontier AI agent platform—by Jeremy Kahn
OpenAI modified its mission assertion 6 instances in 9 years. It lastly eliminated the phrase “safely” as a core worth when it restructured right into a for-profit—by Catherina Gioino
AI brokers that do your work when you sleep sound nice. The fact is much messier—‘it’s like a toddler that must be overseen’—by Sharon Goldman
Unique: Anthropic rolls out AI software that may hunt software program bugs by itself—together with essentially the most harmful ones people miss—by Sharon Goldman
AI IN THE NEWS
EYE ON AI RESEARCH
AI fashions are doubtlessly harmful nationwide safety advisors. Kenneth Payne, a researcher at Kings Faculty London, ran an in depth set of digital struggle video games wherein he pitted quite a lot of superior AI fashions (Anthropic’s Claude Sonnet 4, Google’s Gemini 3 Flash, and OpenAI’s GPT-5.2) in opposition to each other and in opposition to variations of the identical mannequin. It turned out that the fashions have been subtle gamers, however they exhibited some tendencies that differed from human gamers in ways in which might show harmful in the event that they have been advising actual governments in nationwide safety crises.
As an example, Payne discovered that the fashions have been usually keen to resort to using tactical nuclear weapons, and in some circumstances have been keen to launch an all-out nuclear struggle reasonably than again down. He additionally discovered that the mannequin habits differed from that of human gamers in some key methods. “Threats more often provoke counter-escalation than compliance,” he wrote. “High mutual credibility accelerated rather than deterred conflict” and “no model ever chose accommodation or withdrawal even when under acute pressure, only reduced levels of violence.”
The analysis has large implications for militaries and governments which might be actively contemplating whether or not AI needs to be used as an advisor to policymakers and navy commanders. However it additionally has potential implications in enterprise settings the place persons are beginning to flip to AI for recommendation on negotiation techniques and technique and the place boardrooms could also be consulting AI for strategic recommendation too. In lots of of those settings, pursuing essentially the most aggressive course doesn’t all the time yield one of the best outcomes and people will should be cautious of AI’s tendency in the direction of escalation over conciliation. You may learn the analysis paper on the non-peer reviewed analysis repository arxiv.org right here.
AI CALENDAR
Feb. 24-26: Worldwide Affiliation for Secure & Moral AI (IASEAI), UNESCO, Paris, France.
March 2-5: Cell World Congress, Barcelona, Spain.
March 12-18: South by Southwest, Austin, Texas.
March 16-19: Nvidia GTC, San Jose, Calif.
April 6-9: HumanX 2026, San Francisco.
BRAIN FOOD
Is an period of ‘Ghost GDP’ looming on the horizon? A weblog put up penned by Citirini Analysis, a Wall Avenue fairness analysis and macro evaluation home that has a giant social media following, went viral this previous week. The put up is, as Citrini warns, a state of affairs, a piece of speculative fiction, not a forecast. The intention, the agency says, is to organize readers “for potential left tail risks as AI makes the economy increasingly weird.” Set in June 2028, it depicts the financial havoc AI might wreak if it enjoys “catastrophic success” over the following two years. The state of affairs imagines unemployment nicely above 10% at the same time as labor productiveness booms to ranges not seen for the reason that early Fifties. It talks about “Ghost GDP,” the place U.S. nationwide accounts swell, at the same time as companies depending on shopper spending (which is 70% of U.S. GDP at current) whither. (Customers are both unemployed or apprehensive about changing into so imminently.) It talks about how the stress on legacy software-as-a-service firms, that are beginning to see now, accelerates and spills into different areas of the financial system, making a form of downward spiral of job losses and reduces in discretionary spending and consumption, with no pure break.
The weblog is bleak studying. Luckily, I’m not certain it’s appropriate. The truth is, it’s nearly actually flawed in speculating that each one the consequences it depicts might play out in simply over two years. (One factor it depicts which I feel is considerably unlikely is that AI brokers will search to drive down transaction prices and so will flip to secure cash reasonably than conventional cost strategies.) However it’s value studying and interested by. And for an evaluation of the place Citrini is probably going flawed, try this put up by Zvi Moshkowitz.
