A Startup Says Its AI Can Replicate Science Better Than ChatGPT. Nobody Will Say What Happens If AI Goes Wrong.
Today’s AI news sits at an uncomfortable intersection: tools getting more capable, and oversight struggling to keep up. A startup claims to beat the biggest AI labs at a critical scientific task. Major AI companies still can’t explain how they’d handle a misbehaving model. And Harvard is charging $699 to practice your pitch with a digital copy of your professor. These stories aren’t unrelated — they’re all asking the same quiet question: how much do we actually trust these systems?
A Startup Says Its AI Is Better at Checking Science Than OpenAI or Anthropic
Research replication — the process of independently repeating an experiment to confirm its results — is one of science’s most important and most neglected jobs. A startup called Inherent, founded by researchers who previously worked at DeepMind (Google’s AI research division), reportedly claims its AI system called Faraday is better at this task than tools from OpenAI and Anthropic.
According to TechCrunch, Inherent positions Faraday not as a tool but as a “teammate” — something that works alongside scientists rather than just answering their questions. The distinction matters. A tool waits to be used. A teammate notices things, flags problems, and participates. Inherent is betting that framing shapes how researchers trust and use the system.
Science has a real replication problem. Studies across medicine, psychology, and other fields have repeatedly failed to hold up when other labs try to reproduce them. If an AI system can reliably help check whether results are solid, that’s genuinely useful. For regular people, that eventually means more trustworthy medical research and fewer headlines about studies that later turned out to be wrong.
Why this matters: Unverified findings cost time and money — and sometimes affect real health decisions. An AI that helps catch them earlier could quietly improve how science works.
“AI system Faraday outperformed Anthropic and OpenAI at replicating research.”
The Biggest AI Companies Still Haven’t Explained What Happens If Their Models Go Rogue
A rogue AI model — one that behaves in ways its creators didn’t intend and can’t easily stop — sounds like science fiction. It isn’t, quite. As AI systems grow more capable and are given more autonomy to act in the world, the question of how to shut one down or course-correct becomes practically important.
According to TechCrunch, the leading AI labs — the companies building the most powerful models — have not publicly explained what their containment protocols look like. Think of it like a pharmaceutical company that won’t say what their recall process is. You hope they have one. You don’t actually know.
This matters more now than it did two years ago because AI systems are increasingly being given the ability to take actions, not just answer questions. They can browse the web, write and run code, and operate other software. The more an AI can do, the more important it becomes to know how you’d stop it from doing the wrong thing. For everyday people, the concern isn’t dramatic robot uprisings — it’s subtler: an AI agent that keeps spending money, sending emails, or making decisions it wasn’t supposed to make, with no clear off-switch.
Why this matters: Transparency about safety isn’t just good PR. It’s how the public, regulators, and researchers can actually evaluate whether these companies are being responsible.
“AI labs haven’t publicly explained how they would stop a rogue model.”
Harvard Is Selling Access to AI Clones of Its Professors for $699
Harvard Business School has launched a program called HBS Foundry, reportedly priced at $699, where students can practice entrepreneurial skills with AI avatars modeled on real instructors. According to TechCrunch, the avatars simulate the experience of pitching to a professor or presenting to a board — and they give feedback on demand, any time of day.
The appeal is straightforward. Practicing a pitch is awkward. Most people don’t do it enough because arranging time with a real expert is difficult. An AI avatar removes that friction. You can rehearse at midnight, stumble, restart, and try again without embarrassment. That kind of low-stakes repetition is actually how people get better at high-stakes things.
For aspiring entrepreneurs who can’t afford MBA tuition — which runs well over $100,000 at Harvard — $699 is still real money, but it’s a different category entirely. Whether the avatars deliver genuinely useful feedback or a convincing imitation of it is something users will have to judge for themselves.
Why this matters: This is one of the clearest examples yet of elite institutions using AI to reach people outside their traditional student body — even if the price still puts it out of reach for many.
“AI avatars of instructors give feedback to students practicing business pitches.”
Also Happening in AI
On the developer side, LiteLLM — an open-source tool that lets developers connect to multiple AI APIs through a single interface — released version 1.98.0, while OpenAI pushed out version 3.3.0 of its Python library, adding support for data-residency endpoints (which let companies control which country their data is stored in). In policy news, OpenAI has reportedly reversed its earlier opposition to California’s AI safety bill SB 53, according to TechCrunch — a notable shift worth watching. Meanwhile, Google is adding an AI chatbot feature to its Discover feed that lets users shape their content preferences conversationally, as reported by The Verge. And for those building AI applications, Towards Data Science published a practical guide on turning a prototype AI agent built with LangGraph into a production-ready backend.
What to Watch
The silence from major AI labs on containment protocols and the simultaneous growth of agentic AI — systems that take actions, not just answer questions — are on a collision course. Watch whether California’s SB 53 gains traction; if it passes, it could set a national precedent forcing labs to document safety procedures publicly. OpenAI’s reversal on the bill suggests the political pressure is real, and other states are watching closely.