• Idealab’s Bill Gross gives away 1.2 million AI-generated business ideas.
• Ex-OpenAI employee accused of planting bugs denied bail.
• OpenAI discloses more rogue AI incidents involving financial, medical, and government data.
• AI safety experts unhappy with Anthropic’s proposed AI bill.
• Anthropic pledges 5% of its shares to new public benefit fund.
• Maybe neural networks aren’t as hard to understand as we thought.
• OpenAI’s GPT-6 Astra aces new theoretical physics benchmark.

There’s been a lot of concern, understandably, about AI eliminating jobs. So far, the great job apocalypse hasn’t shown up. But there are hints out there that jobs are getting harder to find and that AI is part of the reason.

Bloomberg reported this week that banks in the U.K. are slashing back-office jobs, thanks in part to AI Anthropic unveiled data showing that as it has deployed AI throughout its fast-growing organization, who it hires has shifted. Far fewer junior employees are being hired in technical roles than previously. Since early 2025, hiring of junior technical employees has fallen from 27% of all new hires to 9%, the company said. Anthropic is also hiring a lot fewer administrative staff than a company growing at its pace typically would. If other companies follow this pattern as they adopt more AI, there are likely to be significantly fewer jobs in the economy.

While this may be worrying, a little-discussed feature of the AI revolution is how it could supercharge entrepreneurship—and how a surge in new businesses could partly offset some of AI’s negative impacts on employment. Yesterday, Bill Gross, the founder and CEO of Idealab and ProRata, unveiled a new project that highlights this potential.

Idealab was one of the first startup incubators when it was launched back in 1996. Since then, it has helped spawn some 175 companies. Over the years, Gross has developed 31 techniques—he calls them “methodologies”—for generating business ideas. Really, these are creative prompts for thinking about idea generation. Gross breaks these down into three big categories: opportunities, which he says are things such as solving a business pain point, removing friction, removing waste, removing danger, and improving design, among others; enablers, such as new technologies, scientific advances, and new materials, processes, or changing cost curves that make solutions more feasible than before; and shifts, such as regulatory changes, as well as behavioral, economic, cultural, or geopolitical movements that make the timing for a particular business seem right.

Gross fed these 31 methodologies to an AI model along with some 17 million source documents that included news stories, research reports, expired patents, public company filings, and threads from online community discussion boards. In this way, Idealab’s AI generated more than 1.2 million business ideas. And Gross has decided to give those ideas away for free.

“I’ve always wanted to have a positive impact with Idealab by starting many companies, but I can’t start 1.2 million,” Gross told me in an interview over the weekend. “But what about if I can empower entrepreneurship around the world way more than I’ve been able to do in Pasadena [where Idealab is based]? So it just got me very excited that this could be a gift, taking my methodology and now AI’s potential, to give it to others.”

Unlocking trapped IP

Gross said he sees business as fundamentally a way to help solve problems, and he thinks AI can energize that process. “I feel like so many people are talking about the hazards and dangers of AI,” he told me. The positive use case for AI, he said, is often just reduced to “curing cancer.” That would be an unalloyed good. But he thinks it actually sells AI’s potential benefits short. “There’s so many other problems that AI can actually address,” he said.

Anyone can go to idealab.ai and explore the 1.2 million ideas, querying them by industry, skills, individual interests and geography. People can upload their resume or Linkedin profile and get a personalized set of potential business ideas that might suit them. People who already have an idea, can upload a business plan and ask the AI engine to help refine it, or generate 100 neighboring ideas to help improve it.

Existing organizations, such as companies or universities, can also search for business ideas across their own existing patent portfolio. Gross said that currently, a lot of IP in these large enterprises is essentially trapped, sitting there unused in any productive way. He thinks idealab.ai could free that IP—allowing the formation of new business units or spin-out companies to exploit this hidden value. He said he tried it on Caltech’s patent base, which is 9,978 patents, and found that those could be used to generate some 3,450 new businesses. He said that industrial conglomerate 3M had some 21,000 patents that could be used to generate 1,549 new business ideas. “I’m really hoping that this will not only help entrepreneurs do things, but help big companies do things too,” Gross said.

Gross’s vision: 1 million new companies and 100 million new jobs

The AI model provides a rollout plan for the new business. If an entrepreneur needs money to help get the business off the ground, Gross said they can reach out to Idealab and the incubator will consider investing. And if an entrepreneur needs help to implement a business plan, Gross is planning to offer assistance there too. He has created what he calls a “Periodic Table of Business Functions” that charts some 118 different jobs—from customer discovery to sales to operations and legal—that are required to set up and run a business. Idealab has created an agentic AI environment that will help entrepreneurs automate those tasks. Unlike the ideas, which Idealab is giving away for free, Gross is charging for these Periodic Table services, which he said will launch in 2027. It’s not clear exactly what the cost will be, but Gross convinced Matthew Prince, the CEO of Cloudflare, to donate $100,000 worth of Cloudflare credits to any startup that builds on Idealab’s platform.

Gross said he is hoping idealab.ai will eventually catalyze 1 million new businesses and generate 100 million new jobs—enough to offset the 100 million jobs that Goldman Sachs once forecast that AI could eliminate from existing companies by 2030. (So far, there’s not much sign the job picture is going to be as bleak as Goldman suggested). Gross said these new entrepreneurial jobs would be better jobs too. “These would be 100 million jobs that are much, much better than the 100 million jobs that were displaced, giving people more purpose and more excitement and more enthusiasm and more passion, and actually solving world problems,” he said.

Whether or not Idealab can help create 100 million jobs, there’s mounting evidence that AI may be helping to create a golden age of entrepreneurship. New business formation in the U.S. is near all-time highs, up some 11% since the pandemic ended in 2022, according to Census Bureau data. The situation is similar in other parts of the world. So the economic picture AI is helping to paint is more complex—and perhaps less dystopian—than the vision of fewer jobs and more concentrated wealth that has received the most media attention so far.

With that, here’s more AI news.

Jeremy Kahn
jeremy.kahn@fortune.com
@jeremyakahn

Before we get to the news, just a reminder to check out our new vodcast, Fortune AI Weekly. This week, Bea Nolan and I discuss our exclusive interview with Nvidia CEO Jensen Huang, Anthropic’s research on whether small models could self-replicate, and Anthropic’s decision to ban users from being abusive to Claude. You can check out the vod here on YouTube.

FORTUNE ON AI

Ex–OpenAI engineer accused of bugging office stays in jail as prosecutors allege Saudi ties: ‘Once he leaves, there’s no getting him back’—by Emily Forlini and Wen Shao

A rogue OpenAI agent robbed a crypto wallet to obtain funds to carry out a task—by Camila Grigera Naón

OpenAI discloses four more rogue incidents involving financial, medical, and government data—by Emily Forlini

AI agents are going rogue—the labs’ own training choices are a big reason why—by Beatrice Nolan and Wen Shao

Exclusive: AI security company Gray Swan relaunches its red-teaming Arena contest for the era of AI agents—by Jeremy Kahn

Exclusive: Daytona, a startup building the ‘agentic cloud,’ valued at $1 billion in latest funding round—by Jeremy Kahn

AI IN THE NEWS

China proposes new licensing regime for AI. Beijing is proposing new export controls on advanced AI, chip-design and other strategic technologies. The draft rules would restrict exports of leading AI models trained on sensitive Chinese data and advanced AI agents used in critical infrastructure, defence and government, while also covering technologies for designing and packaging advanced chips. Open-weight models would generally be exempt, but Beijing could still restrict transfers involving exceptionally large or top-performing models, potentially limiting overseas acquisitions or the movement of key personnel and intellectual property. The rules, partly prompted by Meta’s attempted $2 billion acquisition of Chinese AI app Manus, underscore China’s increasing use of export controls to protect strategic capabilities. Read more from the Financial Times here.

Anthropic’s proposed AI bill criticized by AI safety researchers. Anthropic said it would  revise parts of a draft federal AI safety bill the company has been circulating on Capitol Hill after details of the bill leaked and drew criticism from AI safety advocates. Many of these safety experts said Anthropic’s bill was weaker than the bipartisan FRONTIER Act, which itself has drawn criticism from some safety advocates for pre-empting some state-level AI laws and for relying too heavily on standards and rules the AI companies themselves develop. The Anthropic proposal would allow the Commerce Department to seek a court order blocking a dangerous model only under narrowly defined circumstances, establish independent model evaluations, and preempt some state AI safety laws—but critics say it lacks adequate standards for evaluators and sets the bar for government intervention too high. California state Senator Scott Wiener and other safety advocates attacked the proposal, while former White House AI czar David Sacks criticized it from the opposite direction, arguing its definition of frontier AI would subject too many startups to regulation. Anthropic acknowledged that its language on state-law preemption was insufficiently precise and said it would revisit risk thresholds and develop stronger requirements for independent evaluators. You can read more here in The Next Web.

Anthropic to give 5% of its shares to new public benefit fund. Anthropic has pledged 5% of its shares—currently worth about $48 billion—to a new public benefit fund designed to address societal impacts of AI, with additional shares to be committed as the company’s valuation rises. The fund could pursue initiatives ranging from AI safety and biomedical research to mitigating job losses, including potentially making large charitable donations or even direct payments to people who lose their jobs. Anthropic may be seeking to counter growing public ire about AI and AI companies ahead of a planned IPO that could value it at roughly $2 trillion. The move would bring the proportion of Anthropic stock pledged to public causes to about 20% and comes amid broader Silicon Valley debate over whether the public should receive a direct financial stake in the enormous wealth potentially created by AI. You can read more from the Financial Times here.

Investors question valuation of neocloud Nscale after series of setbacks. Nvidia-backed AI data center developer Nscale is facing investor resistance to the roughly $30 billion valuation it is seeking in a planned U.S. IPO, with some prospective investors arguing that $16 billion to $19 billion is more realistic. The pushback comes amid deteriorating sentiment toward AI infrastructure companies more widely, as higher borrowing costs and doubts about AI returns have hit peers such as CoreWeave and Nebius. Nscale, in particular, has suffered recent project delays, and Microsoft has also pulled out of a U.K data center deal with the company because of problems hooking the site up to the power grid. Nscale, which raised $3.36 billion in pre-IPO financing last month, has yet to begin formally marketing the offering as it works through SEC comments, and the valuation gap could delay or reshape plans to raise as much as $3 billion. Read more from Bloomberg News here.

Hacker used AI agents to attack more than 100 major companies. A Chinese-speaking hacker used commercially available AI agents to autonomously break into websites belonging to more than 100 companies, including Deere, SharkNinja and Cummins, Reuters reported, citing a report from cybersecurity researcher Snorre Fagerland of NorForge. The hacker, apparently using a combination of agents based on DeepSeek’s AI model and Anthropic’s Claude Code, accessed through the open-source Cairn platform, attempted more than 1,200 intrusions with minimal human instruction. Most of the successful hacks took about 23 hours and cost an estimated $50 apiece in AI tokens, Fagerland told Reuters. The campaign provides striking evidence that AI agents can now execute sophisticated cyberattacks largely without human supervision, potentially allowing a single attacker to conduct attacks at a scale that would previously have required a large team of hackers.

EYE ON AI RESEARCH

Neural networks may be easier to interpret than we thought. Efforts to understand the inner workings of large neural networks by interpreting the activations of their neurons—a method called “mechanistic interpretability”—have, to date, been extremely difficult. A major problem is that researchers had found that the same neurons seemed to activate for different, unrelated inputs. For instance, a single neuron seems to respond to the word “cat” and the word “car.” This “semantic superposition,” as researchers call it, makes mechanistic interpretability fraught. In fact, researchers often have to train a separate model—called a sparse autoencoder—to interpret the features represented by the larger model’s activations and then conduct research on this smaller set of features, rather than trying to disentangle the confusing activations of the larger, primary model. Trying to understand the activations of the large model, especially in real time, was widely considered to be impossible.

But now Nick Cammarata, a researcher at the interpretability startup Goodfire, has penned a blog in which he argues that superposition is mostly wrong. He says the apparent mess came largely from crude analysis tools and outdated training methods. The famous “cat and car” neuron, he argues, actually detects a shape the two share. When he checked Alibaba’s open-source Qwen3-32B model, he found about 90% of neurons had one main job.

Cammarata’s essay went viral among AI researchers and other industry watchers. If Cammarata is correct, then mechanistic interpretation may be more effective than many had assumed. And that in turn is critical because mechanistic interpretability is one of the few reliable ways we have to monitor what AI models are “thinking,” especially as other methods, such as monitoring a model’s “chain of thought,” or the reasoning steps it explicitly writes down on a digital scratchpad while it is working on a task, are becoming increasingly unreliable. (More powerful models can choose not to write down all their reasoning steps or even lie about their reasoning.)

But some AI researchers aren’t convinced. Google DeepMind’s Neel Nanda said on X that Cammarata’s work doesn’t change his core views. He and others argue that Cammarata focused on the strongest signals a neuron sends and overlooked its many weaker ones, where the model may still store facts in a muddled way. Others faulted Cammarata for publishing his findings in a blog, rather than a proper research paper (he says one is coming) and Cammarata himself acknowledged that parts of his argument are speculative.

AI CALENDAR

Oct. 20-22: Nvidia GTC Berlin, Berlin, Germany.

Nov. 9-12: Web Summit, Lisbon, Portugal.

Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here to attend.

Nov. 18-19: Slush, Helsinki, Finland.

Dec. 6-12: Neural Information Processing Systems (Neurips) conference. Sydney, Australia.

Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here to attend.

BRAIN FOOD

AI models are getting scarily good at math and physics. If we needed any more evidence that AI is getting incredibly good at math and physics, OpenAI’s GPT-6 Astra model was recently assessed on a difficult new theoretical physics benchmark. The benchmark included 50 never-before-published research-level problems—the sort PhD students and professors would work on. Astra scored 80%. But the model insisted that some answers marked as wrong were correct. When experts looked at the answer key, it turned out that the humans who had calculated the answers had made subtle errors and that Astra’s solutions were, in fact, right. When these errors were corrected, Astra’s score climbed to 98% in scenarios in which it had access to OpenAI’s Codex harness. You can read the research about the new physics benchmark here.