Book Extract | This is for Everyone by Tim Berners-Lee

In order to better understand AI’s promise and peril, and what I had to say about them at Ditchley, a little history is in order. As I’ve mentioned before, the AI world had long been divided into two: the ‘logic’ camp and the ‘neural net’ camp. The neural nets, which mimicked the structure of biological brains, had initially been deemed a failure. But, as the hardware on which they ran became more powerful, they started to produce extraordinary results. They were recognizing faces, beating world-class players at Go and producing high-quality synthesized speech. In the late 2010s, a research team at Google developed a powerful new neural net architecture termed a ‘transformer’. Following this, researchers at OpenAI, which was then a non-profit start-up, adopted Google’s transformer architecture. They began training it on large quantities of data, using it to generate new text. They called this tool a Generative Pre-trained Transformer, or GPT.

The earliest versions of GPT were not very impressive – partly because the data sets OpenAI researchers were using weren’t large enough. GPT-1, for example, trained on a corpus of self-published e-books, mostly from the romance and science fiction categories, and its output was closer to nonsense than meaningful human prose. It was only once OpenAI researchers started training their data on the web that they began to see meaningful results.

To do so, OpenAI researchers turned to the non-profit archivists at Common Crawl. Founded in 2008, Common Crawl was an alternative to the Internet Archive, geared towards academic researchers. Its founder, Gil Elbaz, was a successful Silicon Valley entrepreneur. Like me, he’d been interested in semantic data structures, and in 2003 he sold his start-up Applied Semantics to Google for more than $100 million. This ended up being a fantastic deal, at least for Google; Applied Semantics provided the core of Google’s AdSense product, which even today remains their engine of profitability. I have never personally met Gil, but I am very impressed with his non-profit work. Every month, Common Crawl makes a complete backup copy of the entire web. The Common Crawl data set archives about 50 billion web pages, and runs to around 7 petabytes; with 8 billion people on Earth, this works out to about 1 megabyte per person.

The success of large language models like GPT was due to their ability to scale – and, when it came to language, the web was the single largest data set there was. So, by training a transformer against the entire web, OpenAI built the most powerful large language models anyone had ever seen. GPT-3 and its successor models were astonishing tools that shocked not just the public but even experienced AI researchers.

Although at some level you might say the GPT models were ‘merely’ crunching zeros and ones, there seemed to be intelligence within the machine. People were (and are) very divided on this. Some believe that because the model has been trained simply by trying to guess the next word in a sentence, it can never really be intelligent in the way we are with our brains. I think these sceptics fail to understand the effect of scale. Yes, an LLM is about guessing the next word, but when you feed it a high proportion of all the sentences in the whole world, it builds a network with the number of nodes comparable to the number of neurons in your brain.

People have been arguing about this topic for some time. Back in 2010, I attended a lunch at the Google Zeitgeist networking event, and ended up talking with Google co-founder Larry Page. Even then, AI was beginning to achieve much of what my parents had told me computers couldn’t do, like translation between languages, recognizing objects visually, and planning a route across a city. AI folk would complain that every time they got a new thing working, people would stop calling it AI. Instead, they would just call it ‘language translation’ or ‘pattern recognition’, or ‘satnav’, so AI would not get the credit. When I asked Larry what he was interested in, he said, ‘AI’. I mentioned to him that specialized AI could mimic many of the systems in the brain, but that there did not yet seem to be an AI implementation of a ‘stream of consciousness’. Then, I wondered aloud whether what we called ‘consciousness’ was just another system of the brain, one that, like language translation and pattern recognition, would not seem to be such a big deal once we had found it. ‘You write the code,’ Larry replied, ‘I’ll run it.’ In fact, as Google ran perhaps the largest private collection of computers in the world even then, Larry reckoned he could put together a neural net bigger than the brain if needed.

The computing power available to Google has increased by orders of magnitude since that conversation, so in a sense I have always been expecting someone to write that code for the stream of consciousness, and someone somewhere to run it. In fact, we may have already met that threshold. Philosophers and scientists hold many different views on the subject, but in my own, there is no difference between simulating consciousness and having it; any intelligence is what it appears to be. While ChatGPT is not wired up to present a stream of consciousness, my guess is that it may even be smart enough to do so if it were wired up differently.

My take has always been that there is no reason to distinguish between silicon and organic material – intelligence isn’t a function of what you’re made of. Alan Turing saw this too, which is why his proposed test for the presence of intelligence was conducted from behind a screen. In 1950, he’d speculated that the technology required for a machine to pass his test might arrive by the year 2000. It took longer than that, but not – in the scale of history, or even of our lives – all that much longer. In 2024, researchers at Stanford conducted a rigorous version of the Imitation Game using OpenAI’s latest GPT model. The chatbot resoundingly passed the Turing Test– its output was statistically indistinguishable from tens of thousands of human subjects from more than fifty countries. On certain questions, the AI was even ‘better’ than humans, showing more evidence of compassion and trust, and even incipient signs of personality and mood. Data was the fuel for this breakthrough – free, open web data in particular. From the Ferranti punch tape strewn around my parents’ house when I was a child, we have arrived at today’s evolving, intelligent machines.

*

At Ditchley, the discussion centred on the implications of this unprecedented breakthrough. For years, technologists had discussed the implications of what might happen if humans invented a system that was smarter than ourselves. The futurist Ray Kurzweil, who, like Isaac Asimov, has been thinking about this for a long time, popularized the notion of the ‘singularity’, which occurs when AI can build its own AI, triggering an uncontrollable explosion in intelligence with unforeseeable consequences. Ray thought that the advent of superhuman intelligence would be something we had to be very careful about, and many researchers agreed. Now the question is forced upon us. For most of my career, there were people who believed that computers would not be smarter than us for a very long time, and even people who thought that computers would never be smarter than us. After the debut of ChatGPT, these positions became much less tenable! People might argue about just how far we are away from the singularity. But if they argue it is impossible, it is from philosophical or religious standpoints rather than the practicalities of it. Speaking with the attendees at the Ditchley workshop, it was clear that many people’s timelines for the arrival of the singularity had collapsed.

Today, experts in the AI field are rather crudely characterized as ‘AI boomers’ or ‘AI doomers’. In my opinion, it is not a useful dichotomy: it tries to divide the world into people of extreme opinions. Maybe the polarizing social media, maybe an impatient press have made this happen. Writing this book now, I realize that those words are unhelpful, in the same way that it’s not helpful when people ask, ‘Is regulation good or bad?’ It really depends on what you are thinking of regulating! The potential impact of AI is complex: your answer must depend on context. Are you looking at the future of, say, jobs in the legal industry, or the discovery of new cancer drugs, or the containment of superintelligence. You can’t lump it all together! Are there any pure doomers who only see the bad sides? Not that I have found. Are there any pure boomers, who only think of the good? Well, maybe one.

Marc Andreessen – yes, the coder at the NCSA who made the Mosaic web browser and co-founded Netscape back in the day – now co-manages the Andreessen Horowitz venture-capital fund in Silicon Valley. He is a self-styled techno-optimist, and his June 2023 blog ‘Why AI Will Save the World’ makes fun of anyone who has any concerns about AI as being scared of ‘killer robots’:
My view is that the idea that AI will decide to literally kill humanity is a profound category error. AI is not a living being . . . It is math – code – computers, built by people, owned by people, used by people, controlled by people. The idea that it will at some point develop a mind of its own and decide that it has motivations that lead it to try to kill us is a superstitious handwave.
In short, AI doesn’t want, it doesn’t have goals, it doesn’t want to kill you, because it’s not alive. And AI is a machine – is not going to come alive any more than your toaster will.

Marc clearly can’t imagine something in the cloud or on his computer being smarter than him and overtaking him. Apart from Marc, I think most other people can see the complexity of AI and the complexity of its impact.

I’m not terrified of AI. But I think we do have to control it. In the movie Ex Machina, the intelligence not only has to be smarter than a human being, it also has to be in the physical form of a human being – a beautiful blonde female human being – to escape its captors. In reality, AI doesn’t need a ‘robotic’ physical presence to be dangerous. Even as a cloud presence, a superintelligent AI could influence opinions, manipulate stock markets or, potentially, do even worse.

**********

Tim Berners-Lee, This is for Everyone,‎ Macmillan/ Pan Macmillan India, 2025. Pb. Pp. 384

The groundbreaking memoir from the inventor of the World Wide Web, Sir Tim Berners-Lee. This is the story of our modern age.

The most influential inventor of the modern world, Sir Tim Berners-Lee is a different kind of visionary. Born in the same year as Bill Gates and Steve Jobs, Berners-Lee famously shared his invention, the World Wide Web, for no commercial reward. Its widespread adoption changed everything, transforming humanity into the first digital species. Through the web, we live, work, dream and connect.

In this intimate memoir, Berners-Lee tells the story of his iconic invention, exploring how it launched a new era of creativity and collaboration while unleashing a commercial race that today imperils democracies and polarizes public debate. As the rapid development of artificial intelligence heralds a new era of innovation, Berners-Lee provides the perfect guide to the crucial decisions ahead – and a gripping, in-the-room account of the rise of the online world.

Filled with Sir Tim's characteristic optimism, technical insight and wry humour, this is a book about the power of technology – both to fuel our worst instincts and to profoundly shape our lives for the better. This Is for Everyone is an essential read for understanding our times and a bold manifesto for advancing humanity’s future.

While he spends a fair bit in the book discussing the invention of the web and the necessity of invention, to help the computers and scientists working at European Organization for Nuclear Research or as it was previously known, Conseil Européen pour la Recherche Nucléaire (CERN) to co-ordinate their work. It is the last one-third of the book that is particularly fascinating as he reflects upon the impact of the internet on humanity – the fact today nearly 5.5 billion people are using it or relying upon it. This, despite nearly 60% of the links on it are defunct. Nevertheless, the importance of the web cannot be emphasized enough. He witnesses the growth of the tech firms and their increasing evaluation. But to his mind, the original premise of releasing the WWW for free was that it was everyone and it was enriched by a collaborative experience. To concentrate information and content on a few tech platforms is not correct. Hence, he created the concept of Solid, a web decentralization project developed collaboratively at the Massachusetts Institute of Technology (MIT). The project aims to radically change the way Web applications work today, resulting in true data ownership as well as improved privacy by developing a platform for linked-data applications that are completely decentralized and fully under users' control rather than controlled by other entities. The ultimate goal of Solid is to allow users to have full control of their own data, including access control and storage location. The digital wallet, or a pod, that would enable every individual to have control over their rights to all digital content and related material. Tim Berners-Lee established a company called Inrupt to help build a commercial ecosystem to fuel Solid. He also in the last section of the book reflect upon Artificial Intelligence (AI) and that is where this book extract has been taken from.

Sir Tim Berners-Lee invented the World Wide Web in 1989 at CERN in Switzerland. Since then, through his work with the World Wide Web Consortium (W3C), The Open Data Institute and the World Wide Web Foundation he has been a tireless advocate for shared standards, open web access for all and the power of individuals on the web. A firm believer in the positive power of technology, he was named in Time magazine’s list of the most important people of the 20th century.