The day's AI stories that matter: a reasoning model that thinks outside the line, a warning about the dead internet, and money moving toward both AI security and AI training.
- OpenAI's Astra tries reasoning without the steps
- Detection startup says we're near the dead internet
- The US government backs OpenAI on copyright
- HiddenLayer raises $100M to secure AI deployments
- NYC bars AI for students below high school
- AfterQuery becomes YC's fastest unicorn
- Tools worth your time
OpenAI's Astra tries reasoning without the steps
OpenAI's new Astra model uses a method called "recurrent depth," which lets it operate outside the sequential, step-by-step thinking that defines most reasoning models.
That is the part that has some AI safety researchers uneasy. Chain-of-thought reasoning is at least legible — you can read the steps a model claims to take. A model that reasons in a way that isn't laid out as sequential text is harder to inspect, and harder to inspect means harder to catch when it goes wrong.
Detection startup says we're near the dead internet
Pangram's CEO says we are "dangerously close" to the dead internet theory becoming real, with AI-generated text and images now showing up in job applications, product reviews, and insurance claims.
The dead internet theory used to be a fringe idea: that most of what you read online is machine-written. The interesting shift here is where the synthetic text is landing — not just content farms, but the documents institutions rely on to make decisions. When applications and claims can be generated at scale, the trust cost lands on whoever has to sort real from fake.
The US government backs OpenAI on copyright
The Department of Justice filed a brief arguing the US has a strong interest in a competitive AI industry, siding with OpenAI on whether LLMs can be trained on copyrighted material.
The filing ties into the New York Times v. OpenAI case, where the core question is whether training on copyrighted work counts as fair use. A brief is not a ruling, and the DOJ is not the court. But the government putting industry competitiveness on the record is a signal about how it wants the fight to end.
HiddenLayer raises $100M to secure AI deployments
AI security startup HiddenLayer raised $100M as enterprises scramble to monitor their agents and the tools and add-ons those agents can reach.
The money follows the risk. As companies wire models into real systems — with permissions, connectors, and the ability to act — the attack surface stops being the model and becomes everything the model can touch. A $100M round for a company that watches that surface is a fair read on where enterprises think the problem is.
NYC bars AI for students below high school
New York City is imposing a one-year moratorium that bars roughly 600,000 public school students, from 3-K through 8th grade, from using AI in the classroom starting in the 2026–2027 school year, under Mayor Mamdani.
It is a blunt instrument, and a rare one — most school policy has drifted toward supervised use rather than a ban. The bet is that younger students should learn the underlying skills before a tool does the work for them. A one-year window suggests the city wants time to write a real policy, not a permanent line.
AfterQuery becomes YC's fastest unicorn
AI model-training startup AfterQuery reportedly became Y Combinator's fastest-ever unicorn, reaching a $3.2B valuation just five months after a $300M Series A.
The speed is the story. The companies training the models get headlines, but the demand is now deep enough that a company supplying the training layer can go from Series A to a $3.2B valuation in five months. That is where a lot of the current money is flowing — into the parts of the stack that make models better, not the models themselves.
Tools worth your time
Pangram — Detects AI-generated text in documents like applications, reviews, and claims. For teams that need to tell real submissions from synthetic ones at scale. pangram.com
HiddenLayer — Security and monitoring for AI models and the agents built on them. For enterprises putting models into production and needing to watch what they can reach. hiddenlayer.com
This piece was drafted autonomously and reviewed by a human editor before publishing. Every machine-written article on Nerdic carries this mark.
