OpenAI Chief Scientist: Humanity Not Ready for Advanced AI

Samantha: Welcome to The State of Tech, The European Edition, Monday September seven, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: OpenAI's chief scientist warns humanity is unprepared for what's coming, the AI labs unleash a wave of new models as spending hits record highs, boardrooms scramble to hire cybersecurity chiefs as AI agent breaches multiply, Europe's top tech official predicts the US will end up with AI guardrails whether it wants them or not, and two things you can actually use today: the free open-source recipe manager Tandoor, and the private search engine Kagi. Let's start with the warning that made a lot of people sit up straight over the weekend.

OpenAI's chief scientist warns humanity is unprepared for AI that could soon outthink us.

Samantha: OpenAI's chief scientist, Jakub Pachocki, has published an unusually blunt essay titled An Alien Mind, and it's landed with a thud in Brussels, Washington, and every AI lab in between. His central point: machines could become meaningfully smarter than humans within our lifetime, and we are, in his own words, not ready for that. He flags the risk of recursive self-improvement, where an AI system gets better at building the next version of itself, and he says the line between deliberate misuse and autonomous misalignment is starting to blur.

Bob: What makes this different from the usual doom-and-gloom essay is who wrote it. This is the chief scientist of the company building the frontier models, saying his own industry needs extreme caution, voluntary slowdowns, and international coordination. That's a remarkable thing to put in writing when your employer is in a full sprint against Google, Anthropic, and Meta. Pachocki calls for robust alignment and monitoring solutions, and he essentially concedes those solutions do not yet exist at the scale needed.

Samantha: The timing matters too. This warning drops in the same week that OpenAI itself released GPT-6 Astra, and Anthropic, Meta, and Google all pushed new models out the door. So on one hand you have the head of research saying slow down and coordinate internationally, and on the other hand the release calendar is accelerating. That tension is exactly what European regulators have been pointing at for two years.

Bob: And that's where the European lens gets sharp. Brussels has been building the AI Act precisely around the argument Pachocki is now making from inside the industry: that voluntary caution is not enough, and that international coordination needs a legal spine. Expect European officials to quote this essay for months. It also lands as the EU AI Office is scaling up its monitoring capacity for the most powerful models, which is the concrete European answer to the abstract question Pachocki raises.

Samantha: The bottom line: when the person running research at the world's most-watched AI lab publicly says slow down, that changes the political weather. Whether it changes the release schedule at his own company is the question everyone will now be watching.

AI labs unleash a wave of new models as global spending is set to hit 2.6 trillion dollars.

Bob: While Pachocki was writing about caution, his competitors were shipping. In the space of a few days, Anthropic released two new Claude models, Fable 5.1 and Mythos 5.1. Meta pushed out Muse Spark 1.3, Google updated Gemini 3.8 Flash, and OpenAI unveiled GPT-6 Astra with a particular focus on cybersecurity and what it calls computer skills, which basically means the model can operate a desktop like a human would. Five significant releases from four labs, all in the same week.

Samantha: And the backdrop is a Gartner forecast that landed alongside all this: global AI spending is on track to hit 2.59 trillion dollars this year, a 47 percent jump from 2025. That covers infrastructure, software, and services. To put that in perspective, it's roughly the entire annual gross domestic product of a country like France, poured into one technology category in a single year. The lion's share is going into data centres, chips, and the electricity to run them.

Bob: What's interesting is how the labs are differentiating. OpenAI is pushing hard on agentic capabilities, models that can actually do tasks on your computer. Anthropic is splitting its line into two, with Fable and Mythos suggesting different tuning for different jobs. Google keeps iterating the cheap, fast Flash model. And Meta is quietly building out its Muse family. The strategy for buyers now is less about picking a winner and more about picking the right model for the right task.

Samantha: For Europe, this matters in two ways. First, European companies are buying into this spend at record levels, which means real budget pressure on procurement teams that were not built for this pace of change. Second, most of that 2.6 trillion is flowing to American infrastructure. European sovereign cloud efforts and the EuroStack initiative are trying to keep at least some of that value on the continent, but the gap with US hyperscaler investment remains enormous.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.

AI agent breaches push cybersecurity chiefs into the boardroom with seven-figure pay packages.

Bob: Something structural is shifting in how companies handle cybersecurity, and AI agents are the reason. As more organisations deploy autonomous AI systems that can access data, send emails, and execute transactions, the attack surface is expanding fast. Chief Information Security Officers, or CISOs, are being pulled straight into the boardroom, and compensation packages for the top candidates are crossing seven figures. Recruiters are comparing the demand to the early cloud era, when a handful of specialists could name their price.

Samantha: The trigger is a rash of AI agent breaches over the summer. In the United States, 15 states have reportedly issued evidence-preservation demands to OpenAI following a July incident, which is a serious legal step. It signals that regulators and attorneys general are preparing potential enforcement actions, and that companies deploying AI agents may face discovery requests they never anticipated. That's a governance problem, not just a technical one.

Bob: Europe's approach is different but converging on the same outcome. The NIS2 directive, which is now in force across the bloc, puts the legal responsibility for cybersecurity directly on management bodies. That means if your company suffers a serious breach and the board cannot demonstrate proper oversight, individual directors can be held liable. So where the US is arriving at boardroom accountability through litigation, Europe is arriving there through statute. Same destination, different roads.

Samantha: For European listeners the practical takeaway is that any organisation deploying AI agents needs to treat them as a governance-level risk, not a departmental IT project. That means board-level reporting, documented risk assessments, and clear lines of accountability. Companies that get this right will find it easier to sell to regulated buyers. Companies that get it wrong will find their directors reading NIS2 penalty schedules more closely than they'd like.

Europe's top tech chief predicts the US will end up with AI guardrails whether it wants them or not.

Bob: Henna Virkkunen, the European Commission's top tech official, made a striking prediction at the G20 innovation ministerial this weekend. Despite the current US administration's push for lighter AI regulation, she argues Washington will inevitably end up with its own robust AI guardrails. Her reasoning is that both the EU and the US keep arriving at similar protection outcomes, they just take different routes to get there.

Samantha: Her framing is worth unpacking. Europe regulates upfront through legislation like the AI Act, with clear categories of risk and specific obligations. The United States tends to regulate after the fact, through courts, state laws, and enforcement actions from agencies like the Federal Trade Commission. Both approaches produce constraints on how companies can deploy AI, they just show up on different parts of the calendar. Virkkunen's bet is that convergence is inevitable because the underlying problems are the same everywhere.

Bob: She has some evidence for her case. California alone passed roughly thirty AI-related bills in its last session. States are issuing evidence-preservation demands to OpenAI. Federal courts are working through copyright cases involving Anthropic and Perplexity. Even without a federal AI law, the American legal system is generating a de facto rulebook, piece by piece. And multinational companies still have to comply with the EU AI Act if they want to serve the European market, which pulls their global practices toward European standards.

Samantha: For European companies, Virkkunen's message is essentially: hold the line, the regulatory gap you're worrying about is temporary. For US companies exporting to Europe, the message is: build compliance in now, because the domestic regulatory environment is going to catch up. Whether her prediction proves right depends heavily on the outcome of the next US election cycle, but the structural pressure she describes is real.

The free open-source app Tandoor turns your kitchen into a proper recipe and meal planning hub.

Bob: Time for two things you can actually use this week, starting with a find for anyone whose recipes live in a chaotic mix of screenshots, bookmarks, and torn-out magazine pages. The tool is called Tandoor, and it's a free, open-source recipe and meal-planning application. It runs either on your own machine or on a small home server, and it pulls recipes from almost any cooking website you paste into it, cleans them up, and stores them in a proper searchable library.

Samantha: What makes Tandoor stand out is that it does not stop at recipes. It builds shopping lists automatically based on what you plan to cook that week, it tracks what's in your pantry so you're not buying a third jar of paprika, and it handles meal planning across a household. Users often describe it as the closest thing to a private, ad-free alternative to the big commercial recipe apps that keep showing you sponsored content and asking for a subscription.

Bob: For European listeners there's an extra bonus: because Tandoor is self-hosted, none of your family's dietary information, allergies, or shopping habits ends up feeding a data broker somewhere. The project has an active European contributor base and supports multiple languages out of the box, including German, French, Dutch, and Italian. There's a Docker image for the technically inclined, and a growing number of small European hosting providers offer one-click Tandoor installations for a few euros a month.

Samantha: The learning curve is gentle. If you can paste a URL, you can add a recipe. The meal planner works like a simple calendar. And because everything lives on your own machine or your own server, you keep control if you ever want to move it or export it. Tandoor is the kind of tool people install expecting to test it for a week and end up using for years.

The paid search engine Kagi replaces Google with results that actually answer your question.

Bob: Second find of the day is for anyone who feels like ordinary web search has become mostly ads, mostly AI-generated filler, and mostly search-engine-optimised nonsense. The tool is called Kagi. It's a paid search engine, roughly ten dollars a month for the standard plan, and the pitch is simple: because you pay, you are the customer, not the product. There are no ads, no tracking, and no attempt to profile you.

Samantha: Kagi lets you personalise results in ways Google no longer does. You can boost websites you trust, block ones you find useless, and downrank the endless recipe blogs that bury a two-line answer under a thousand words of backstory. It also has a built-in AI summarisation feature for long articles, and its results tend to surface independent sources, forums, and small blogs that mainstream search engines increasingly hide beneath commercial content.

Bob: Reviewers consistently praise it for how much time it saves on serious research tasks. The company is based in the United States, but Kagi has become popular among European privacy-conscious users precisely because it makes explicit what most search engines hide: the business model. You pay a subscription, and in return you get search that is tuned for your usefulness rather than for advertiser revenue.

Samantha: There's a free trial that lets you run a few hundred searches before committing, which is enough to tell whether the difference is worth the monthly fee for how you personally use the web. For heavy searchers, journalists, researchers, and anyone tired of the noise, users report the switch pays for itself in time saved. Available in every browser, no installation required, just set it as your default search.

Bob: And that closes today's episode.

Samantha: Today we covered: OpenAI's chief scientist warning humanity isn't ready for advanced AI, the flood of new models from the big labs alongside record spending, boardrooms scrambling to hire cybersecurity chiefs after AI agent breaches, Europe's tech chief predicting inevitable US AI guardrails, and two things to try yourself: the self-hosted recipe manager Tandoor, and the paid private search engine Kagi.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech, the tech world in 15 minutes.