Nvidia’s research model outscored every human at IOI 2026, Alibaba shipped a coding upgrade without bothering to change the version number, the DOJ told a federal judge that training on copyrighted work is fair use, and someone built 215,000 fake pages to feed the answer engines.
Follow updated AI news all day at TheDrone.Report
An AI Beat The Best Human At The International Olympiad In Informatics
This is the one that actually matters and it will get a fraction of the attention of the earnings news. Nvidia researchers published results on September 2 for a system called Nemotron-3-Ultra-CC, run during the official IOI 2026 competition window under the same time limits, submission limits, and internet restrictions as the human contestants.
Gold medal threshold361.12
Top human contestant498.27
Nemotron-3-Ultra-CC535.4
The researchers claim this is the first time an AI system has outscored the highest scoring human contestant on any IOI problem set. Not matched. Outscored, by 37.1 points, against the best teenage competitive programmer on the planet.
The method is worth noting because it is not magic. The pipeline runs 22,000 curated programming problems, synthetic reasoning traces generated by DeepSeek-V4-Flash, supervised fine tuning, and a test time technique they call GenCorrect that generates diverse candidate solutions, feeds them through an evaluator, and refines the next round on the feedback. In other words: try a lot of things, check them, try better things. It is brute force with good taste.
The ReadCompetitive programming is a bounded problem with a scoring function, which is exactly the kind of thing these systems eat. It is a real milestone and it is not the same as engineering. Watch how fast people conflate the two.
Alibaba Shipped A Major Upgrade And Did Not Change The Version Number
Also on September 2, Alibaba released Qwen3.8-Max-0902, a post trained snapshot of its existing flagship aimed at coding and agentic work, with a context window up to one million tokens. Alibaba says it scored 1,691 on CodeArena, a 22 point gain that puts it first on the leaderboard.
Reporting on the leaderboard put it three points ahead of Claude Opus 5 on the WebDev board. Three points. On a leaderboard. This is the margin the entire industry is now marketing on.
1,691
Qwen3.8-Max-0902’s claimed CodeArena score, up 22 points, first on the board. Delivered as a snapshot, not a new version.
The genuinely notable thing is the packaging. A frontier lab made a meaningful capability jump and labeled it with a date stamp instead of a new number. Either that is admirable restraint in an industry addicted to version inflation, or it is a company that has decided the version number is now a marketing asset to be spent later. Both readings are supported by the evidence.
The ReadA leaderboard lead measured in single points, on a benchmark the model was post trained to win, is a press release, not a capability jump. The million token context is the part worth testing.
The Justice Department Told A Judge That Training On Your Work Is Fair Use
On September 1 the DOJ filed a Statement of Interest in the Southern District of New York in the copyright litigation against OpenAI, and it came down on the side of the AI companies.
The argument runs in three parts. Training is “exceedingly transformative” because it converts text into numerical representations rather than reproducing works for their original purpose. Requiring licenses would advantage only the largest technology companies and the legacy media outfits with deep archives, concentrating the market. And obstacles to domestic AI development threaten American competitiveness and national security.
The New York Times did not take it quietly. And there is a detail sitting underneath the filing that nobody in it mentions: back in July, OpenAI was reported to be in talks to hand the federal government a roughly 5 percent equity stake.
So the government filed a brief arguing that the defendant’s core business practice is legal, while separately discussing terms on which it might own a piece of that defendant. There may be a perfectly clean explanation. There is not, as far as anyone can tell, a disclosure of it in the filing.
The ReadWhatever you think about fair use, an executive branch that is simultaneously a potential shareholder and a friend of the court is a structure worth naming out loud before it becomes normal.
Three Websites Manufactured 215,128 Pages For The Answer Engines To Quote
A research report published September 2 documents three interconnected sites that generated 215,128 machine written “best software” buying guide pages. The domains, all registered between December 2023 and May 2024, share page templates and DNS infrastructure. Two of them put the phrase “Facts and Grounding Page” in their HTML titles, which is language written for a retrieval system, not a reader.
The researchers queried Perplexity’s sonar and sonar-pro models across 380 buyer intent software categories and collected 7,534 citations across 2,055 domains. The three sites picked up 181 of those citations, about 2.4 percent, appearing in 41 categories.
59.8%
Share of sources behind grounded AI recommendations that sit outside the 100,000 most visited websites. Another 23.4 percent of citations point to domains outside the top million entirely.
Two point four percent is not an emergency. The number underneath it might be. If most of what the answer engines cite lives in the unranked long tail, then the cost of manufacturing a credible source has collapsed, and the only thing standing between a buying guide and a fabricated one is whether anyone bothered.
Somebody bothered. Two hundred and fifteen thousand times.
The ReadSEO spam took fifteen years to poison search results. The equivalent attack on answer engines is cheaper, faster, and harder to see, because the machine does not show you the sludge, it shows you a confident paragraph.
And Somebody Just Raised $100 Million To Guard The Models
HiddenLayer closed a $100 million Series B on September 2, led by Delta-v Capital with Ten Eleven Ventures, Morgan Stanley, Microsoft’s M12, and Booz Allen Hamilton participating. The company sells tooling that protects models, agents, and workflows against adversarial attacks and injection.
Annual recurring revenue is up more than tenfold in a year into the tens of millions, with over 90 percent of that from new customers. The client list runs to financial services, big tech, the Department of Defense, and the intelligence community, plus one unnamed frontier model provider with more than 700 million weekly users.
“Inference is still inference. We haven’t had to pivot, but we’ve had to grow our scope, from traditional modeling to GenAI to agentic.”
Chris Sestito, CEO, HiddenLayer
The ReadTenfold revenue growth in AI security is not a story about HiddenLayer. It is a measurement of how many production AI deployments got scary enough to need a guard on the door.
Filed Under Interesting
A model out-programmed the world’s best human. A leaderboard changed hands by three points. A government argued for its industry in court while negotiating for a slice of it. And a small content factory quietly proved that the machines will cite almost anything shaped like a source.
Four stories, one theme. Every single one of them is about a system being optimized against its own scoring function until the score stops meaning what it used to mean. The olympiad has a scoring function. The leaderboard has a scoring function. The legal system has one, and so does the citation graph.
Turns out you can win all four. Winning is just not the same as being right.
Sources
- Post-Training Language Models for Gold-Medal Performance in Coding Competitions (arXiv, Nvidia researchers, Sept 2026)
- Alibaba upgrades Qwen3.8-Max with a new 0902 snapshot (TechNode, Sept 2, 2026)
- Alibaba’s Qwen-3.8-Max-0902 Debuts With The Weirdest Flex Ever (Wccftech, Sept 2026)
- DOJ Sides with OpenAI, Warns Obstacles to AI Development Threaten National Security (IPWatchdog, Sept 3, 2026)
- DOJ urges judge to rule for OpenAI, Microsoft in N.Y. Times lawsuit (The Washington Post, Sept 2, 2026)
- OpenAI in talks to give Trump administration a 5% stake in the company, FT reports (CNN Business, July 2, 2026)
- Three sites made 215,128 “best software” pages for AI. Perplexity cites them (Trellner Research, Sept 2, 2026)
- HiddenLayer nabs $100M as enterprises rush to secure their AI deployments (TechCrunch, Sept 2, 2026)
MONTANIMATION STUDIOS • HALF ART. HALF TECH. ALL SHENANIGANS.