Google had a chatbot before ChatGPT existed. Demis Hassabis now says keeping it in the lab was “obviously in retrospect probably the wrong decision.”
He said it in a conversation with Sebastian Mallaby, author of the Hassabis biography The Infinity Machine, that Intelligence Squared published on October 15. Six days later he told WSJ Magazine, which named him its 2026 Innovator in Tech, that he now expects AGI within 2 to 3 years.
The company that sat on LaMDA is now the one where Sundar Pichai says Gemini 4 Argon left Google capacity constrained. This interview explains the miss behind that comeback, and the method Hassabis describes for hard problems works for a 10-person startup as well as it works for DeepMind.
I watched the full 14 minutes so you don’t have to. Here are the 10 lessons that matter.
1. A Bot That Is Wrong 20% of the Time Is “Doubly Tricky” for Google (the Trap of Being Too Close to Your Own Tech)
Hassabis sees the logic of waiting, then names the blind spot behind it.
“I think that’s an interesting thing of where you can be sometimes too close to your own technology.”
Hassabis says he had no part in the decision. Back then DeepMind ran mostly as an independent research group, and LaMDA belonged to the Google Brain side. He can still see the logic. Google Search exists to organize the world’s information, so a bot that gives the wrong answer 20% of the time is a far bigger risk for Google than for a startup. A startup, he says, “especially someone with someone like Sam leading it,” can “YOLO it.”
The phrase I keep coming back to is “too close.” The team that builds a model sees every hallucination up close, while users judge the product on the jobs it gets right. Anthropic faced its own version of this call when it limited the model that can hack anything to a small group of partners. If your brand rests on accuracy, fix the error rate in the system design first, as our agent reliability playbook lays out, and then let users tell you where the bar sits.
Try this: before you delay a launch, ask 5 users outside the team which error rate they would accept for each job.
Sponsored by HubSpot for Startups
Hassabis is talking about the launch bar, and every AI founder hits the next one right after it: turning early AI users into customers who stay.
IGNITION 2026 by HubSpot for Startups brings 200 AI-native founders and GTM leaders to San Francisco on Nov 10 for founder and investor talks plus 1:1 GTM clinics. It’s free for approved applicants.
2. ChatGPT’s Clunky Name Shows Nobody Knew How Big It Would Get
Hassabis reads the product name as OpenAI’s own forecast.
“ChatGPT is a programmer’s name. They would have named it something catchier if they knew how big it was gonna get.”
To be fair to his own side, Hassabis says, nobody knew how big ChatGPT would get, OpenAI included, and he adds that Sam Altman has admitted as much. The name gives it away. He says OpenAI thought of the launch as a research experiment, and as a startup with no existing products that relied on reliability, it could live with the mistakes.
A working title tells you how big the team thought the idea was. For a look at how OpenAI pitched its long-term ambition years before the chatbot, the 64 slides of its 2018 AGI plan are worth an evening.
Try this: name and staff your next experiment as if it might be the one that takes off.
3. Flawed Models Found Their First Jobs Where Errors Were Cheap
Users found the forgiving use cases before the labs did.
“It didn’t matter if it made some errors that weren’t to do with information seeking.”
What caught everyone by surprise, OpenAI included, was how quickly people found uses for models that still hallucinated a lot. Once millions of people had access, Hassabis says, they found low-hanging fruit like summaries and homework help, jobs where a wrong detail cost little. Demand for those early use cases turned out to be huge, and he says that is what showed the technology was mature enough to leave the lab.
That is the oldest pattern in AI products. The wedge is the job with the highest tolerance for error. Only 4.5% of Americans pay for AI, according to our breakdown of consumer AI spending, so the forgiving job that gets people coming back is where you earn the right to charge. For the build side, read the AI product-market fit playbook from OpenAI’s product lead next to PostHog’s 38 lessons from shipping 10 AI products.
Try this: rank your use cases by what a wrong answer costs the user and launch on the cheapest one.
4. Go Has More Positions Than the Universe Has Atoms, So Brute Force Loses
Hassabis picks 2 problems where checking every option is impossible.
“In Go, there are more possible board positions than there are atoms in the universe. So that means you can’t brute force and enumerate every option.”
Proteins are harder still. An average-sized protein has 10 to the 300 possible shapes, Hassabis says, so you can’t find the right fold by looking at every one.
DeepMind’s answer to both problems was the same. Train a neural network on what the domain already knows, from Go games to known protein structures, until it holds a model of how that domain works. His 2019 MIT deck on self-learning systems shows how long he has been betting on this approach.
The founder version is simple to run. Before you budget more compute for a problem, estimate how many options it has. When the count is astronomical, extra GPUs buy a slightly bigger slice of an impossible search, and our token cost playbook shows where that kind of spend leaks out of an AI bill.
Try this: write down the size of your search space before you choose the method.
“You’re going to have to do something much smarter.”
5. Let the Model Pick What to Search, Then Write the Objective Function
A model shrinks the haystack, and a clear goal tells the search where to climb.
“The final thing you need is an objective function. You need to have a clear goal you can hill climb against.”
With a model in hand, the search gets smart. A Go position has around 100 possible moves, Hassabis says, and the model tells you which few are worth exploring, so the compute budget goes to the fruitful ones. He calls it a very generic solution for any problem you can frame as a needle in a huge haystack.
The goal is the last piece. Games make it easy, since you win or maximize the score, which is why DeepMind trained on Atari before Go. Science has clear goals too, like minimizing the energy in a system, and he says a surprisingly large share of problems can be framed this way.
The best agent setups today follow the same recipe. Give an agent a metric and you can let it optimize overnight, the idea at the heart of Karpathy’s autoresearch method.
Try this: state your product’s objective function in 1 sentence, and get the team to agree on it before you add compute.
Setting a goal your system can climb? Start with these:
Google’s Agent Broke a 56-Year Math Record. Yours Forgets Yesterday
The Harness Is the Product: 6 Decisions That Turn an AI Agent Into a Worker You Can Leave Alone
6. Hassabis Ranks Curing Disease First and Clean Energy Second
Asked to weigh benefits against risks, he gives a ranked answer.
“I think the number one thing we can apply AI to is curing terrible diseases.”
The question comes after a run of dark topics, from AI in warfare to cybersecurity fears. Hassabis answers that AI is the ultimate tool for science and medicine. Second on his list is helping the environment and finding new renewable energy sources, and he points out that this is where he has spent his own time, through AlphaFold and Isomorphic Labs.
A ranked list forces the trade-offs a vision statement lets you skip. Put your use cases in order and fund the top one first. If you invest in hard tech, our Physical AI Investment OS maps where the money sits across 12 layers.
Try this: write your top use case and your runner-up in 1 line each, and share both with the team.
7. Hassabis Says AI Could Bring a Second Renaissance Within 10 Years
He backs a 10-year forecast with a science group that’s nearly 10 years old.
“There’s almost no area of science that I think AI can’t help with. And I think if we do that right, in the next 10 years we’ll have almost a second Renaissance.”
DeepMind has had an AI for science group for nearly 10 years, Hassabis says, and that is where AlphaFold came from. Some rivals are only now setting up science teams. His group works on fusion, weather prediction, climate models and material science.
Sam Altman reached for the same word last month when he described the renaissance he wants AI to build. Big forecasts are easy to make, and a lab that shipped AlphaFold has earned a hearing for this one.
Try this: when you hear a bold forecast, check what the person making it has already shipped.
8. Even the Best Scientists Hit a Limit Without AI
Hassabis read the biographies of his heroes and saw a ceiling.
“There’s a limit to how much we can comprehend, I think, with just unaided, with our human minds, even with scientific method and other things.”
He grew up reading the biographies of scientists from Feynman to Turing and calls them all really smart. He also says he was born with an “insatiable curiosity” about the deep mysteries of the universe, like the nature of time and consciousness. From the start, AI was a 2-factor idea for him: an amazing tool for science, and an artifact that could help us understand our own minds.
If the greatest scientists hit a ceiling working alone, your best engineers will too. Pair them with AI on the parts no single person can hold in their head, and keep the judgment calls with people. Companies are already hiring for judgment as AI takes over execution, and The Agent Workforce Kit gives you Anthropic’s 8 metrics for running an agent workforce.
Try this: pick 1 problem too big for one person’s head and put an AI system on it this week.
9. DeepMind Worried About Safety From Its Start in 2010
He dates the concern to the founding year and explains why.
“We were worried about safety from the very beginning of DeepMind in 2010, and we still are now, because it’s general. So it can be used for anything.”
The technology is dual purpose, Hassabis says, and Google DeepMind knew it from day one. He says he would have done this work no matter what, because to him it is the most fascinating question and technology out there. The job for the industry, in his view, is to push the areas that are great for society, like medicine and energy, more quickly.
A date tells you more than a values page. Dario Amodei built a whole company around the same worry, and our read on his long game of safe AI covers how. The same dual-use logic applies to the AI features you ship, so run the AI app security checklist before launch.
Try this: ask every AI vendor you buy from when its safety work started and who owns it today.
10. Hassabis Loves the AGI Race and Still Wants International Cooperation
He enjoys the competition, then says something bigger sits behind it.
“I’ve got this gnawing feeling that there’s something much more important, much bigger than the commercial race, which is overall getting AGI safely over the line for humanity.”
Hassabis says the competitive part of him loves the “ferocious commercial battle,” leans into it and finds it helps him lead. He also says DeepMind is one actor among 5 or 6 leaders, with the Chinese labs in the mix, and that the story of the next few years is still to be written. He wants more cooperation on safety, while admitting that is very hard with today’s geopolitics.
When someone leading the race says the race leaves something out, it pays to listen. For your own planning, use our 4-scenario framework for AI policy so your roadmap holds up whichever way governments go.
Try this: add 1 slide to your next board deck on what your product does if AI rules tighten.
“there needs to be more cooperation and coordination at an international level.”
What This Means for You
The Hassabis method fits on an index card: a model of the domain plus a goal it can climb, launched first where users forgive mistakes.
Founders: Write your objective function in 1 sentence before you pick a model, and launch on the job where errors cost the least. When you raise on that story, The AI-Era Fundraising Toolkit has the investor workflows ready.
Investors: Ask each team how big its search space is and what error rate its users will tolerate. Test the answers against the 9 bottlenecks deciding who wins in AI and the payoff data in State of AI 2026.
Operators: Price every search run against your agent’s ROI, and count the options before you add compute.
Everyone else: Give an AI tool 1 task this week where a mistake costs you nothing, and check which AI plan is worth paying for before you upgrade.
The 5 Principles to Steal
Test Launch Bars Outside the Team. The people closest to the flaws are the worst judges of what users will forgive.
Pick Forgiving Jobs First. Let users show you where errors are cheap, then widen the scope.
Size the Search Space. Estimate how many options exist before you choose a tool or a budget.
Write the Goal Down. Every search needs 1 clear objective to hill climb against.
Date Your Safety Work. Ask each lab and vendor when it started and what it still worries about.
Watch the full conversation to hear how Hassabis connects Go and protein folding.
If this breakdown saved you the full watch, send it to one founder or investor who needs it.
Keep reading
Muse Just Beat ChatGPT’s Launch 3-to-1. Here’s the Operating Manual Nobody Wrote
OpenAI Declared the AGI Era. Then Greg Brockman Moved 25% of Its Production Engineers to Defense
Cover photo of Demis Hassabis by Arthur Petron, CC BY-SA 4.0.