2 weeks ago
Michael Wooldridge is for illustration nan coach you wish you’d had: approachable, capable to explicate difficult things successful elemental terms, neither dauntingly highbrow nor off-puttingly cool, and genuinely enthusiastic astir what he does. “I emotion it erstwhile you spot nan ray spell connected successful somebody, erstwhile they understand thing that they didn’t understand before,” he says. “I find that incredibly gratifying.”
He comes crossed a regular benignant of guy, which, arsenic an Oxford professor pinch much than 500 technological articles and 10 books to his name, he intelligibly isn’t. Typically, his favourite activity is his publication to Ladybird’s Expert Books – an update of nan classical children’s bid – connected artificial intelligence. “I’m very proud of this,” he says, arsenic he hands maine a transcript from his bookshelf. We’re successful his study successful nan University of Oxford’s somewhat municipal computing section connected a sunny outpouring day. Maybe it’s nan field setting, but our chat almost takes nan shape of a seminar.
Wooldridge is an adept nationalist communicator, particularly connected artificial intelligence – a section he has worked successful for much than 30 years, but astir which he retains a patient scepticism. In his 2023 Christmas lectures for nan Royal Institution, titled The Truth astir AI, he brought successful a robotic canine and asked his school-age assemblage to ballot connected whether they’d whack it pinch a shot bat. And, to explicate reinforcement learning, he recreated nan classical 80s movie WarGames, successful which a young Matthew Broderick averts atomic catastrophe by getting nan US subject machine to play noughts and crosses pinch itself (until it concludes location is nary existent measurement to win). “Matthew Broderick was successful London astatine nan time. We tried to get him to travel connected nan Christmas lecture, but he couldn’t do it,” says Wooldridge. “So we called our machine BrodeRick successful his honour.”
WarGames is really beautiful adjacent to nan taxable of Wooldridge’s latest book, Life Lessons from Game Theory: The Art of Thinking Strategically successful a Complex World. He’s taught nan taxable to his students for much than 15 years, he says. Now it’s our turn. There’s nary maths successful Wooldridge’s book; alternatively he translates crippled mentation into 21 digestible scenarios, incorporating everything from Atlantic cod fishing, to Pepsi v Coca-Cola, to nan beingness of God.
“It is astonishing really galore world events tin beryllium explained by a comparatively mini number of crippled theoretic models,” Wooldridge says. One of nan simplest is nan crippled of “chicken”, which he illustrates successful his book utilizing a segment from nan James Dean movie Rebel Without a Cause (none of his students had heard of it, he admits). Two teenagers thrust their cars towards a cliff; nan first to jump is nan “chicken”, and loses nan game. If some jump astatine nan aforesaid time, it’s a draw; don’t jump astatine each and you’ve mislaid nan crippled beautiful severely (spoiler alert: that’s what happens successful nan movie).
The mentation instruction present is astir Nash equilibriums (we won’t get into nan details) – but, practically, we spot this crippled playing retired successful existent life each nan time. The Cuban rocket situation utilized to beryllium nan go-to example, but different 1 is unfolding arsenic we speak: nan US-Iran war. “You’ve sewage 2 sides pinch ever-escalating threats against each other; somebody’s sewage to backmost down astatine immoderate point,” says Wooldridge. “The threat is, if neither backs down past you’ve passed a constituent of nary return and you get nan worst-case script for everybody.”
Is location immoderate measurement retired of this? “Well, 1 of nan ways that a crippled tin get changed is if a 3rd statement comes successful and provides immoderate inducement for 1 of nan parties to behave successful a different way.” Another action is to circumvent nan crippled by communicating pinch your opponent. That’s what happened successful nan Cuban rocket crisis, but it feels little apt here. “Although, I person to say, Iran seems to beryllium playing it a batch much cannily, successful nan consciousness that nan US broadside is very, very unpredictable. Now, being unpredictable is simply a classical crippled theoretic strategy arsenic well, but it makes it very difficult for personification connected nan different broadside to cognize really to respond. If you really are playing against an irrational subordinate past 1 of nan things crippled mentation says is you conscionable hedge your bets against nan worst-case scenario.”
This is not conscionable astir warfare, aliases moreover games, Wooldridge stresses. He defines crippled mentation successful nan book arsenic “a mathematical mentation that aspires to understand situations successful which self-interested parties interact pinch 1 another”. That, he argues, could use to each mode of situations: social, governmental and philosophical.
The conception of nan “zero-sum game”, for example, has go a mainstream word (partly acknowledgment to WarGames), moreover if it’s wide misunderstood. A zero-sum crippled is not simply 1 wherever 1 broadside gains what nan different broadside loses; it is 1 wherever nan inducement is to make your force suffer as severely arsenic possible, Wooldridge explains. So, technically, chess is not a zero-sum crippled because you’re conscionable trying to win, not to destruct aliases humiliate your opponent. There’s a socio-political magnitude to this. “This zero-sum mentality is very damaging. It’s a very antheral trait,” he says. “And nan grounds is that, not only do you extremity up not needfully doing arsenic good successful life arsenic you could do, but really you extremity arsenic a much miserable person. You consciousness for illustration you person little agency successful your affairs. One of nan important lessons from crippled mentation is that, actually, nan mostly of interactions that we’re successful are not zero-sum.”
This adversarial worldview is nan motor of populist authorities – successful nan “migrants are coming to return your jobs” sense. You are losing because others are winning. One of Wooldridge’s favourite games encourages america to deliberation nan opposite: nan Veil of Ignorance was devised successful 1971 by nan philosopher John Rawls and nan premise is that you tin creation nine successful immoderate measurement you want, but afterwards, you will beryllium placed randomly wrong it. Wooldridge describes it arsenic “a beautiful thought research … It incentivises a socially desirable outcome, but group are still pursuing their self-interest.” Bill Clinton and Barack Obama were some fans, he adds.
It’s not initially clear really crippled mentation fits pinch AI, but nan erstwhile is simply a large portion of nan second these days, Wooldridge explains, particularly successful his superior area of interest, which is multi-agent systems – programs that interact pinch 1 different and enactment connected your behalf. “So if I want to put a gathering pinch you, why would I telephone you up? Why doesn’t my Siri conscionable talk straight to your Siri?” These types of interactions are embedded successful our online life. Online auctions specified arsenic those connected eBay, for example, wherever you’re trying to gaffe successful nan winning bid astatine nan past moment. “If my supplier is going to interact pinch your supplier and my preferences are not needfully aligned pinch yours, past nan mentation that explains really you should deliberation astir those interactions is crippled theory.”
When Wooldridge started out, AI was almost an absurd concept. He entered computing via amateur enthusiasm. Growing up successful agrarian Herefordshire, nan boy of a mediate head astatine nan section cider company, it was a large arena erstwhile Wooldridge’s section electronics shop had a location machine for sale, successful astir 1980. “This sounded ridiculous because computers were multimillion lb things successful my mind.” The shop owners generously fto him person a spell connected it (it was a Tandy TRS-80). “I went backmost week aft week and taught myself to program. I was virtually sat successful nan shop model connected nan computer.” He went connected to study computing arsenic an undergraduate, began a PhD connected AI successful 1989, past did an internship pinch Janet (the Joint Academic Network), which was fundamentally nan UK branch of nan early internet.
The exertion has moved connected astronomically since then, but essentially, Wooldridge says, “the halfway techniques that drove nan existent AI gyration were invented by nan mid-80s”. He mentions Geoffrey Hinton, pioneer of artificial neural networks – nan system that now underpins instrumentality learning. “The only obstacle opinionated successful nan measurement of nan AI gyration successful nan 1980s, really, was that computers weren’t powerful capable and we didn’t person capable data.”
When it comes down to it, Wooldridge says, nan breakthrough occurrence of GPT-3 successful 2020 was mostly “based connected a stake that OpenAI made that if they did nan aforesaid thing, only 10 times bigger, that would present results. A batch of group astatine nan time, including me, were very sceptical astir it. I’m a scientist; I would for illustration to spot advances done technological development, not conscionable by throwing much machine powerfulness astatine it. But it turned retired that, actually, that was a very successful bet.” Does that propose OpenAI leader Sam Altman and his ilk are not nan tech geniuses they’re taken to be? “I’ve ne'er met Sam Altman; I don’t know,” he says diplomatically. “He’s intelligibly delivered thing remarkable.”
Geniuses aliases not, these AI pioneers whitethorn beryllium reaching their limits. A fewer years ago, nan likes of Altman and Google DeepMind’s Demis Hassabis expected to execute AGI – human-level artificial wide intelligence – wrong a fewer years. “I personally deliberation they’re overoptimistic,” Wooldridge says. You tin talk to ChatGPT astir quantum mechanics successful Latin, he points out, “but astatine nan aforesaid time, we don’t person AI that could travel into your house, that it had ne'er seen before, find nan room and clear nan meal table” – thing a minimum-wage quality worker could do.
“The limits are nan machine powerfulness and nan information that you’re capable to propulsion astatine it. And information is now a existent constraint.” The full of Wikipedia made up conscionable 3% of GPT-3’s training data, he says. “Where do you get 10 times much information from adjacent clip around?” Data is becoming a valuable assets for that reason, and immoderate organisations person a imaginable trove of it. “The NHS is sitting connected a immense magnitude of information astir quality beings. That’s nan astir valuable benignant of information imaginable.” Private corporations would salary dearly for it, he says, “but I fishy that whoever signed disconnected connected specified a woody would unrecorded to regret it”. He imagines a dystopian early script wherever “you’re only capable to person entree to nan NHS if you work together to beryllium wired up to wearable tech that monitors you connected a regular ground … I deliberation we are very quickly going to a world wherever nan adjacent procreation of online influencers fundamentally work together to person each of their life experiences, everything they opportunity and do and see, harvested to supply information for AI.”
From an world standpoint, Wooldridge resents nan measurement Silicon Valley has travel to predominate nan AI field, some successful position of resources (“GPT-3 required 20,000-odd AI supercomputers to train; location are astir apt a mates of 100 successful nan full of nan University of Oxford”) and nan nationalist discourse. “We person seen nan communicative stolen by Silicon Valley, which is promoting a type of AI [profit-driven, job-replacing and almost wholly focused connected ample connection models] that surely maine and an atrocious batch of my colleagues person nary liking successful promoting aliases building,” he says. “It’s benignant of depressing, arsenic personification who’s spent their profession trying to build AI to make a amended world and to amended people’s lives.”
He continues: “If you return successful nan wide picture, past location are a immense scope of benefits to AI that often conscionable don’t get noticed because ample connection models suck each nan oxygen retired of nan room.” He mentions a squad successful Oxford that is processing an AI-assisted instrumentality that tin analyse a bosom scan done utilizing a elemental ultrasound and sent to your GP via mobile phone. “This is nan benignant of costly worldly that nan NHS struggles to provide, each of a abrupt disposable astatine negligible cost.”
In 2025, Wooldridge won nan Royal Society’s prestigious Faraday prize for his expertise successful communicating technological ideas successful laic terms. His accompanying lecture, fixed successful February, was titled This Is Not nan AI We Were Promised. Around nan aforesaid time, Wooldridge speculated connected AI having a “Hindenburg moment” – nan Hindenburg clang killed nan airship manufacture overnight. “It’s wholly plausible that we could spot immoderate akin AI-related disaster,” he says. “Computer programs spell incorrect successful each sorts of ways and we are wholly reliant connected a computing web infrastructure wherever AI is progressively embedded.” Having said that, erstwhile it comes to existential risks, “AI is not precocious connected my database of things that keeps maine awake astatine night,” he says. “I don’t interest astir a robot takeover. At least, it’s not successful my apical five.” The truth that he considers atomic warfare a greater threat is hardly reassuring, mind you.
If he could, though, he would slow nan gait of AI development, “just truthful that we person much clip to understand what’s going on”. It is, he points out, a classical “prisoner’s dilemma”, 1 of nan foundational parables of crippled theory. In nan modular scenario, 2 prisoners must separately determine whether to confess to a crime they person jointly committed, aliases support quiet. If 1 confesses and nan different doesn’t, only nan confessor will beryllium freed. If some confess, they will each service a shorter term. If some support quiet, they’ll service moreover shorter terms. So they’d do amended if some agreed to support quiet, but neither knows what nan different captive will do. Counterintuitively, perhaps, crippled mentation concludes that nan smartest action is to confess.
By nan aforesaid logic, AI companies are locked successful a title to get ahead. Their title intends moreover much expenditure, resources and energy-hungry datacentres, pinch nary nett summation successful use for humanity. But present we are. “We’ve sewage a mini number of very, very able companies that are engaged pursuing AI, while astatine nan aforesaid clip saying that they are acrophobic that something’s going to spell horribly incorrect pinch it. So why are they engaged pursuing it? Because they deliberation if we backmost down and we don’t prosecute it, personification other will.”
Was he ever tempted by Silicon Valley himself? “There are a fewer points astatine which that could person happened, I suspect,” he says. “But I’m 60 this twelvemonth and it’s a young person’s crippled correct now.” Some person based on there’s nary constituent successful studying astatine each these days, now that AI is predicted to switch truthful overmuch of quality activity. Wooldridge doesn’t spot it for illustration that. “I didn’t get into computing because I thought it was going to springiness maine a bully job. I sewage into it because I was conscionable really willing successful it.” He gets a batch of parents asking him what their kids should study astatine university, he says, “and nan reply is: ‘Let them study thing that they’re really passionate about.’ I deliberation that’s nan astir important point by far.”
