[Paces across stage, speaking with deliberate pauses and increasing intensity]
So I was reading about artificial intelligence the other day... which is what you do when you're in your 60s and can't sleep because you're worried about being replaced by something more... economically viable.
[Takes sip of water, looks suspiciously at audience]
They say compute is the currency of AI. Compute. COMP-YUTE. Like a made-up word from a dystopian novel. "And here's our new COMP-YUTE machine that can predict when your marriage will fail... and then suggest personalised dating apps."
[Gestures dismissively]
Apparently, 70 percent of all supercomputers are in America. SEVENTY PERCENT. Which means when the machines eventually take over, they'll do it with American accents. They won't sound like they're from the Home Counties or the chap who runs the local pub. No, they'll all sound like they're trying to sell you healthcare you can't afford.
[Turns to face a different section of the audience]
And now China's catching up with AI... but apparently it "doesn't matter." That's what the experts say. "China may match individual U.S. model capabilities this year, but that won't matter for America." Won't matter! Won't MATTER!
[Paces more frantically]
Have you seen what they're doing with these AI computers? They're doubling the power every six months. DOUBLING! When I was at university—yes, I went to Oxford, but I don't like to mention it because it makes me sound unbearable—when I was there, the most impressive computing we did was programming a BBC Micro to print "HELLO WORLD" followed by a crude stick figure. Now they're spending BILLIONS on chips! Not even proper chips! Not even chips you can eat with curry sauce outside a kebab van at 2 AM!
[Sudden quiet intensity]
You know what they call the people who build these AI systems? "Frontier labs." Frontier labs! Like they're explorers discovering new lands. "Oh, look what we found! It's a digitalised version of consciousness that might render humanity obsolete! How marvellous! Let's name it after ourselves!"
[Mimics posh voice]
"We at the Winston Bartholomew Smythe-Poncington III Frontier Lab have developed an AI that can write terrible comedy routines in 0.3 seconds."
[Returns to normal voice]
The most disturbing part? GPT-4 was trained with the computing power equivalent to the entire electrical output of a shopping centre on a Bank Holiday Monday. Just to create something that can... write your emails for you. Or help students cheat on their coursework.
[Deadpan]
I've spent decades perfecting my craft, and now some silicon chips in a warehouse somewhere can do a passable imitation of me. Not a good one, mind you... just good enough that people might think, "Well, why pay for the real thing when I can get this for free?"
[Walks to other side of stage, tone shifts to more thoughtful]
But here's the thing about AI safety that nobody wants to talk about: it's the biggest prisoner's dilemma in human history. If you're not familiar with the prisoner's dilemma, it's this classic game theory problem where two criminals can either cooperate with each other or betray each other. If they both cooperate, they get mild sentences. If they both betray, they get medium sentences. But if one betrays while the other cooperates, the betrayer goes free and the cooperator gets a severe sentence.
[Gestures dramatically]
Now imagine that with AI development. All these companies and countries SAY they want safe AI. "Oh yes, we're all about responsible development!" But the MOMENT someone thinks they can get ahead by cutting corners on safety... boom! They'll do it. Because the first group to develop superintelligent AI potentially controls... everything. EVERYTHING!
[Mimes typing frantically]
"Should we implement these extra safety protocols that might slow us down by six months?"
"Well, DeepMind might not be using them..."
"Forget it then! Full steam ahead!"
[Walks centre stage, more serious]
It's like nuclear weapons development, except instead of mutually assured destruction, it's mutually assured... obsolescence. And unlike nuclear weapons, you can build this in a basement with enough GPUs and a decent cooling system.
[Conspiratorial tone]
But here's the thing – we're all fixated on compute because that's the CURRENT bottleneck. But what if the scarcity is already shifting? What if it's not about how many chips you have anymore, but about having better RECIPES FOR THINKING?
[Leans in, intensely]
Have you heard about this AlphaEvolve thing Google's working on? While we're all distracted by NVIDIA's stock price and who's got the biggest data centre, Google's quietly developed an AI that IMPROVES OTHER AIs. It's like they've built a machine that makes better machines!
[Animatedly]
AlphaEvolve isn't just optimising single functions – it's improving ENTIRE CODEBASES. It's tackling centuries-old mathematical problems! It's developing better matrix multiplication methods – which is basically the plumbing that makes all these AI systems work faster.
[Mimes mind explosion]
And they're not even telling us the half of it! The rumour is they've already used it to improve their data centre scheduling, chip design, even their QUANTUM ALGORITHMS. Do you understand what that means? While we're all counting GPUs like they're Pokémon cards, Google might have already changed the RULES OF THE GAME!
[Dramatically]
The future won't belong to the company with the most compute. It'll belong to the company with the best algorithms – the one that figures out how to need LESS compute. The one that develops the algorithmic equivalent of fusion energy – unlimited power from simple principles. And they won't share it. They won't open-source it. They'll guard it like Colonel Sanders guarded his secret recipe, except instead of controlling fried chicken, they'll control the future of civilisation.
[With increasing paranoia]
What if DeepMind has already had their next "transformer moment" but kept it to themselves? What if right now, in some non-descript office in King's Cross, they've figured out an architecture that makes everything else obsolete? The transformer architecture revolutionised AI in 2017. Before that, we were all fiddling around with recurrent neural networks like cavemen trying to make fire with damp twigs.
[Intensely]
Today's frontier models might be tomorrow's BBC Micros! "Oh, how quaint – they were still using transformers!" And by the time we find out, it'll be too late. The algorithmic gap will be too wide to cross.
[Long pause, stares at audience]
And that's why I'm announcing today that I'm no longer a comedian. I'm rebranding as a... "human compute resource." My next tour will just be me calculating sums really slowly while making occasional sardonic observations about the political implications of exponentiation.
[Checks watch]
I've got about 0.000001% of the processing power of a modern AI system. So this next joke will be ready in approximately... 47 years.
[Freezes in position, staring intensely at the audience]
The scariest part? By the time I finish that joke, AI will have written, performed, and retired from ten thousand comedy careers. And they'll all be funnier than me.
[Shrugs]
But at least they won't have my unique human perspective on... [long pause] whatever it is that makes us special. I'm still working on figuring that part out.
[Walks off stage slowly, muttering]
Maybe I should invest in NVIDIA... No, too late for that. Perhaps I should learn to code... No, the machines are better at that too now. I know! I'll start hoarding electricity. That never goes bad. Or maybe I should just start collecting old algorithms – they might be vintage collectibles someday. "This is an original gradient descent implementation from 2020. Very rare. Completely inefficient by modern standards. Worth a fortune to the right collector..."
[Lights fade]