Thinking sand
Hi everyone, if you support England and/or France, I hope you’ve recovered from this intense World Cup week… I’d like to thank Box once again for supporting Hypertext. Feel free to share this email with a friend, it really helps 💌 Sign up here.
Thinking sand
It’s summertime, which means it’s slow news season. There are simply less things happening in the business world over the summer. Maybe that’s why Demis Hassabis and his team thought it was a good time to float ideas around global AI regulation.
The CEO of Google DeepMind wrote an article on X to push a few ideas forward. Like Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman, he thinks it’s time to create a proper framework to regulate the release of new AI models.
Instead of appearing reactive, he positions himself as proactive. The real reason why they all want to define a set of rules around AI is because things have become too unpredictable with the Trump administration (reactive).
Claude Fable 5 was released then banned, then re-released. GPT-5.6 launch was delayed then released. You know the drill. DeepMind, the AI lab behind Google’s Gemini models, is probably worried about its next model release, too.
And yet, Hassabis says it’s time to change AI regulation because artificial general intelligence (AGI) is right around the corner (proactive).
“AGI cannot be compared to standard technological breakthroughs, not even ones as consequential as the internet or mobile - it is much more akin to the discovery of electricity or fire,” he wrote.
He then offered a poetic observation that has lingered in my mind ever since: “If you stop to think about it, we’ve essentially found a way to make sand think.”
Coming back to his regulatory framework, Hassabis is calling for the creation of a new Standards Body modelled after the Financial Industry Regulatory Authority (FINRA). AI labs would have to share new models with the new regulator up to 30 days before any public release. This group of technical experts and open-source representatives would conduct large-scale testing during that time.
Like Sam Altman, Hassabis believes the US should lead the charge here. “The US is well positioned, given its economic and technical standing, to take the first step in developing such a framework,” he wrote.
I still don’t understand why it should be a purely US-led initiative given that confidence in the US to do the right thing regarding world affairs is now lower than confidence in China across dozens of countries, according to a new Pew research. It would make more sense to involve at least the US, China and the European Union.
Hassabis is walking a fine line of “cautious optimism,” saying AI labs should be free to innovate and push the frontier of intelligence “while also incentivising responsibility and security.”
What’s remarkable is that Demis Hassabis, Sam Altman and Dario Amodei are all advocating for roughly the same approach. According to Axios, Mark Zuckerberg is working on his own memo. He probably wants a seat at the cool kids table.
But there’s a much bigger problem with this entire proposal: open-source models. Moonshot AI, one of China’s leading frontier AI labs, just released Kimi K3, a model that rivals OpenAI’s GPT-5.5 and Anthropic’s Opus 4.8 in benchmarks. Those models were released in April and May, respectively.
Side note: as always, you should treat benchmarks with a grain of salt. “Moonshot has always been pretty scammy with benchmark scores,” a source told me.
Nevertheless, when companies drop open-source models, any company that has access to a large cluster of GPUs can download and run these models without asking for permission.
Once frontier models are released as open weights, regulation shifts from controlling access to controlling compute. At that point, there’s no company left to stop.
Box’s State of AI report is here. Grounded in a global survey of 1,640 IT decision-makers, the report reveals that security and compliance remain paramount, with 49% of organisations having already experienced an AI-related data exposure incident. Read the full report to discover how organisations are establishing a governed, secure content layer to achieve trusted automation without increasing risk.
Bicycle for the mind
Coming back to the “thinking sand” metaphor, it made me think about the famous Steve Jobs quote about computers. According to Jobs’ storytelling, he once read in Scientific American that humans performed poorly against many animals when it comes to energy efficiency, except when they were riding a bicycle.
“What a computer is to me is the most remarkable tool that we have ever come up with. It’s the equivalent of a bicycle for our minds,” he said.
Let’s make sur AI doesn’t break this metaphor. A bicycle helps you go farther using your own effort. The promise of AI is that it can amplify our thinking (an electric bike). But AI slop is mindless content. It’s what happens when you stop pedalling.
Have a good day ☀️
Romain