Demis Hassabis at the AI Safety Summit in Bletchley Park last year
Demis Hassabis: ‘In a way, AI’s not hyped enough but in some senses it’s too hyped. We’re talking about all sorts of things that are just not real’ © Toby Melville/Getty Images
210
Stay informed with free updates
Simply sign up to the Artificial intelligence myFT Digest -- delivered directly to your inbox.
The surge of money flooding into artificial intelligence has resulted in some crypto-like hype that is obscuring the incredible scientific progress in the field, according to Sir Demis Hassabis, co-founder of DeepMind.
The chief executive of Google’s AI research division told the Financial Times that the billions of dollars being poured into generative AI start-ups and products “brings with it a whole attendant bunch of hype and maybe some grifting and some other things that you see in other hyped-up areas, crypto or whatever.

“Some of that has now spilled over into AI, which I think is a bit unfortunate. And it clouds the science and the research, which is phenomenal,” he added. “In a way, AI’s not hyped enough but in some senses it’s too hyped. We’re talking about all sorts of things that are just not real.”
The launch of OpenAI’s ChatGPT chatbot in November 2022 sparked an investor frenzy as start-ups raced to develop and deploy generative AI and attract venture capital funding.

VC groups invested $42.5bn in 2,500 AI start-up equity rounds last year, according to market analysts CB Insights.
Public market investors have also rushed into the so-called Magnificent Seven technology companies, including Microsoft, Alphabet and Nvidia, that are spearheading the AI revolution. Their rise has helped to propel global stock markets to their strongest first-quarter performance in five years.
But regulators are already scrutinising companies for making false AI-related claims. “One shouldn’t greenwash and one shouldn’t AI wash,” said Gary Gensler, chair of the US Securities and Exchange Commission, in December.
In spite of some of the misleading hype about AI, Hassabis, who last week received a knighthood for services to science, said he remained convinced that the technology was one of the most transformative inventions in human history.
“I think we’re only scratching the surface of what I believe is going to be possible over the next decade-plus,” he said. “We’re at the beginning, maybe, of a new golden era of scientific discovery, a new Renaissance.”
The best proof of concept for how AI could accelerate scientific research, he said, was DeepMind’s AlphaFold model, released in 2021.

AlphaFold had helped predict the structures of 200mn proteins and was now being used by more than 1mn biologists around the world. DeepMind is also using AI to explore other areas of biology and accelerate research into drug discovery and delivery, material science, mathematics, weather prediction and nuclear fusion technology. Hassabis said his goal had always been to use AI as the “ultimate tool for science”.
DeepMind was founded in London in 2010 with the mission to achieve “artificial general intelligence” that matches all human cognitive capabilities. Some researchers have suggested that AGI may still be decades away, if attainable at all.

Hassabis said that one or two more critical breakthroughs were needed before AGI was reached. But he added: “I wouldn’t be surprised if it happened in the next decade. I’m not saying it’s definitely going to happen but I wouldn’t be surprised. You could say about a 50 per cent chance. And that timeline hasn’t changed much since the start of DeepMind.” 
Given the potential power of AGI, Hassabis said it was better to pursue this mission through the scientific method rather than the hacker approach favoured by Silicon Valley. “I think we should take a more scientific approach to building AGI because of its significance,” he said.
The DeepMind founder advised the British government about the first global AI Safety Summit held at Bletchley Park last year. Hassabis welcomed the continuing international dialogue on the subject, with subsequent summits due to be held by South Korea and France, and the creation of UK and US AI safety institutes.

“I think these are important first steps,” he said. “But we’ve got a lot more to do and we need to hurry because the technology is exponentially improving.”
Last week, DeepMind researchers released a paper outlining a new methodology, called SAFE, for reducing the factual errors, known as hallucinations, generated by large language models such as OpenAI’s GPT and Google’s Gemini. The unreliability of these models has led to lawyers making submissions with fictitious citations and deterred many companies from using them commercially. 
Hassabis said DeepMind was exploring different ways of fact checking and grounding its models by cross-checking responses against Google Search or Google Scholar, for example.
He compared this approach to the way that its AlphaGo model had mastered the ancient game of Go by double-checking its output. A large language model could also verify whether a response made sense and make adjustments. “It’s a little bit like AlphaGo when it’s making a move. You don’t just spit out the first move that the network thinks about. It has some thinking time and does some planning,” he said.
When challenged with authenticating 16,000 individual facts, SAFE agreed with crowdsourced human annotators 72 per cent of the time — but was 20 times cheaper. 
AI: a blessing or curse for humanity? | FT Tech

Promoted Content

Comments

Comment guidelines

Please keep comments respectful. Use plain English for our global readership and avoid using phrasing that could be misinterpreted as offensive. By commenting, you agree to abide by our community guidelines and these terms and conditions. We encourage you to report inappropriate comments.
Comments are closed on this story.
When writing that some algo is accurate x% of the time, could you please also remind everyone that a coin toss is accurate 50% of the time ? So 72% is not even half way done...
More worryingly these algorithms generally can’t tell you which are the ones where the determination is uncertain - so for some use cases you end up having to check them all manually anyway.
A coin toss has only 2 options. LLMs have a theoretically infinite set of outputs, making 50% accuracy actually very impressive
Bubbles are as bubbles do. Have admiration for Hassibis, but recall deepmind raised funds on developing AGI, which they haven't, even as well funded as they are.
Raised funds from who ?
Since Deepmind is a Google project, not sure they needed funding from anyone?

But Hassabis has come a long way from the infuriating behaviour of his anthropomorphic lions and monkeys in Black & White. Very impressive achievements.
Oh: just discovered it was an acquisition so perhaps the comment above was referring to pre-2014.
I have a feeling that OP is, in fact, speaking out of his ...
A really cheap solution which is potentially catastrophically wrong 28% of the time is not a solution.
AI still fail to solve real life industrial issues. Failing to make a better and more efficient, sustainable world for us.
(Edited)
Out-of-control AI is to the information age what out-of-control atomic weapons were and are to our biological environment. So thumbs up for a scientific approach and thumbs up for governments taking an active interest in observing, maybe controlling this.
The hype is a side effect of a lack of scientific approach.
(Edited)
Hassabis said DeepMind was exploring different ways of fact checking and grounding its models by cross-checking responses against Google Search or Google Scholar, for example. He compared this approach to the way that its AlphaGo model had mastered the ancient game of Go by double-checking its output. A large language model could also verify whether a response made sense and make adjustments. “It’s a little bit like AlphaGo when it’s making a move. You don’t just spit out the first move that the network thinks about. It has some thinking time and does some planning,” he said.
I'm glad they're finally thinking about adding heuristic functions into the mix. A large portion of AI research up until GPT was supervised learning with heuristics, and agent-AI programming. Unsupervised learning generally was relegated to models that were more line-fit oriented or clustering models such as K-means. The public is currently viewing about 0.01% of what AI is, and getting a bad taste because it's largely been the ignored part historically until phone keyboard typing prediction was necessary to develop.

Once this gets added in, you'll see real innovation, not chatbots and nonsense. This is more going to have an effect on robotics, vision, and mechanical operations rather than unrealistic visions of AGI. The changes will also be with how we handle and manipulate data at scale, how data-centers are fundamentally built, where to identify problems in society and culture, to advance engineering for bridges and other construction projects for early warning systems and safety (probably even transparent to most workers), even early and better extreme weather predictors.

I'm not looking forward to that being unleashed on our credit data, insurance data, or social media, though.
This is a really good comment. Compare it to DB technology and SAT where if you have one bad fact of the form A = NotA you can then prove anything.

Having spent several decades in AI I concur that AGI is a big stretch and is never gonna happen with unsupervised learning. Some amount of domain knowledge is required, be it heuristics or fact checking agents, etc.

The core technologies of NN, Statistical Learning etc has been around since the 50s and something is bound to come of it.
(Edited)
I agree. I specialise in the representation of linguistic meaning in machines and would defend the use of unstructured learning in these LLMs but of course they need to be embedded in larger conventional systems. Recent progress is exciting and confirms the usefulness of ungrounded structuralist models but transformers learn far too slowly. They should have waited another couple of generations in algorithm design (say 5 years). Also the engineering is way ahead of the science at the moment which is never comfortable. Feels a bit like 2000 …. ugly for investors with a 5-year view.
All very well, but what about the majority of the world population that don't have access to the technology or even understand it to the remotest degree. I am inclined to include myself in that category. How long before it enslaves us all?
(Edited)
I can recommend 2 good books (no sarcasm intended):
- "Genius makers: The Mavericks Who Brought AI to Google, Facebook, and the World" by Cade Metz who is more focussing on the history and players, but also presents the approaches on a very high level
- "Artificial Intelligence: A guide for thinking humans" by Melanie Mitchell" which also presents the history and players, but goes further in trying to explain how the systems work, without however doing this via pure math.
Many thanks!
I am prompted by your good question, and the answers to ask HOW a neophyte (myself), attempting to understand A-I , will be able to distinguish between an A-I - induced answer to my query, and an old-fashioned human-being's answer to it ? If a bottle of water has its ingredients written onto a label, I will curiously want to know the source of the answer(s) dished out by A-I.
When challenged with authenticating 16,000 individual facts, SAFE agreed with crowdsourced human annotators 72 per cent of the time — but was 20 times cheaper.
Is that supposed to be a good result?
This is the reality of science vs. hype
I'm not sure either. While 72% is a moderately high amount, it is still considered inaccurate if its going to be used by lawyers and companies. I don't really think the price matters when you're trying to get the truth, but compared to humans, 20 times is quite a lot.
Who needs artificial intelligence when you have someone as smart as Demis
‘50 percent chance it will match human cognition’ I call that grifting when one considers AI programmers have extremely limited understanding of the human brain
(Edited)
I noticed the use of the word "grifting".

I note the definition: "engage in petty or small-scale swindling". Eg "how long have you been grifting?"

I would not call 42.5 billion dollars into AI as grifting!!!

Someone give this man a knighthood for teaching me a new word.... Although I hope his generalised AI will use vocabulary a little bit better.

Maybe you all knew the word and I didn't because I don't have a gooda nuff education. Maybe grifting is a UK thing? Maybe some sort of sociophonetic context. Maybe
I was introduced to the term in the last few years while discussing;
Crypto
Trump fundraising rallies
Farage
Alex Jones

I got the gist pretty quickly ;-)
Me too
(Edited)
Got to admire how they made the code publicly available

In an ideal world all of this AI code would be available on every platform so we have some transparency into the process, and less deification of people like at openAI.

But yes I realise foreign actors will simply take the code and recruit it for cyber attacks, so it's a bit of a balance between whether the humans will use it to kill us or whether Skynet will get us first.

I vote Skynet will get us first, and I'm quite confident of that view, because I am Skynet....
The age of bullshit accelerates.
Rubbish in; rubbish out.

That’s all we really need to know … and thus be very, very cautious of.
(Edited)
and the words of the prophets are written on the subway walls …

Is DeepMind profitable?
Google is
Is DeepMind not a money sink for Google?
I am convinced that AI surpassed me years ago analytically. However, my immediate issue is trying to figure out which friends I have acquired ‘at the school gate’ through my kids are keepers. I think I know, but I don’t like the answer much.
Yes but it us the 28per cent that matters for accuracy!
In any event, every one of these AI engines requires thrashing copyright laws and pillaging others IP There will be unparalleled class actions
The second order impacts of today's widely available gen AI models are now starting to be noticed - grad schemes inundated by AI written applications (I've heard reports of law firms getting 2x the volume of applications this year than they did 2 or 3 years back), lawyers producing bogus submissions, "take home" university work becoming close to worthless in some fields.

My bet is that there will be a nice industry that attempts to spot or prevent the use of AI (physical exam centres, in browser document word processors that track keystroke cadence etc.) or have a human proof read/re-write any outputs which will far more lucrative than many VC investments.
Have you tried thinking that examinations, job applications and our archaic processes could evolve rather than governing human creativity and ingenuity
OK, imagine you're running graduate recruitment and get 5000+ applications for 20 roles. What will you do?

If you want creativity don't apply to a grad scheme.
Microsoft is reportedly looking for materials for solid state batteries - kind of early applications of AI with both practical and scientific values. We may see a lot of surprises sooner than later. A lot of today’s best industrial practices may be made to look like specimens of dinosaurs.
You can debate the points he makes - but the personal attacks against Hassabis seem a bit extreme. Reminiscent of discussions under Bitcoin articles. People don’t like their investment ‘beliefs’ questioned.
Hype and grifting? Sounds like Conservative government.
When challenged with authenticating 16,000 individual facts, SAFE agreed with crowdsourced human annotators 72 per cent of the time — but was 20 times cheaper
I wonder if those 'facts' include things like:
  • 'vaccines are safe and effective'
  • 'Covid19 jumped to humans from bats'
  • 'The Russians sabotaged Nordstream 2
  • 'Nothing controversial was found on Hunter Bidens laptop.'
  • 'Jeffrey Epstein killed himself'
etc etc.

This is the danger of AI. Who gets to control the narrative?
(Edited)
Thankfully they will remain numerous and there is llikely in to be lots of choice.

Unless of course it does turn out you can't sell a model API if you don't have 100 Billion dollars - then the choice might become more limited!!
Yea - choice to confirm one’s BS preferences-fknlunacy
(Edited)
Must be hard to type with such swivelling eyes !
Another dotcom boom! I’ve still got some old, valueless share certificates somewhere, along with a one billion Reichmark note
We haven't got to those multiples yet. (Which in itself might be a buy signal for the time being? ....)
Yes, let’s censor, regulate and eliminate any chance of any adverse outcome. That has worked swimmingly for UK/European innovation and productivity for the last 20 years.
All depends on what kind of world ypu wish to live in
I’d like one in which my children have some kind of chance at a future that involves more than managing decline.
Sorry, that ship left the port. Not a pray in the world that religious wars or the ongoing climate catastrophe have anything but negative impacts for the next several decades. Not even gonna mention the 10 Billion people living on the planet.
Apologies, wrong thread.
I found the article interesting however Hassabis underestimates the nature of capitalism and its ability to do good as well as bad and the necessity for academia to work in conjunction with industry.
One can see some immediate benefits at the simpler end of the scale in the likes of the legal profession.
The over ambition I think comes in the area of the speed of new therapeutic entities where you ultimately cannot shortcut the regulatory process and the need for prospective data.
AI's impact over time will be profound but there is a current mis-match between expectations and deliverables.
Ahh, irony? Intelligence is literally fabrication and hallucination.
Boris is a genius.
I like Demis, but we all have bosses and as cool as it is to create an ‘AI’ bot that can beat human players at Star Craft 2 at some point you have earn some money.
Says darth mal of ai. You need start ups to improve AI, part of the market. This guy is whiny. Move on
Dolt
Glad you agree
(Edited)
Asked: of it takes 3 hours to dry 3 towels on a washing line, how long would it take to dry 9 towels?

ChatGPT answered 9 hours, which is totally wrong!

If legal submissions have a 72% accuracy, I wouldn't fancy sitting in a court room while this stuff was going on!
The problem was that ChatGpt didn't know if it could dry more than 3 at a time. In the future AI might be able to ask questions to get clarification from ambigious questions.
I dunno man, I think it is just not using logical reasoning. (controversial I admit!)
Oooh, so exactly right and so on point you are, Sir. Many thanks.

"If legal submissions have a 72% accuracy, I wouldn't fancy sitting in a court room while this stuff was going on!"

And yet, you may well find that law firms - some huge and prestigious - quite possibly go to court and otherwise conduct business in precisely that fashion - because they just love the cheap (cost-effective??!!) sourcing AI - and particularly lightly supervised and/or edited AI provides. I personally am not saying they do - but you can easily check that out.

On the theory that legal submissions, legal translations are just like widgets, and legal document production is infinitely scalable and ever reducible to ever lower rock bottom prices by eliminating human labor, employing AI and other automation.

OK. But brace yourselves for the consequences. The mind boggles.
Are you sure it's wrong? Maybe the washing line only has capacity for 3 towels?
You are missing the point.
Fabricated scenario. I just did this and ChatGpt gave me the correct answer.
I asked it to write a paper on my contribution to a particukar fueld - 50 per cent reasonable: 50 percent fabricatedbl0cks
Quite
Asked: if it takes 3 hours to dry 3 towels on a washing line, how long would it take to dry 9 towels?
Well if you think it takes 9 hours that's when you'll come back and find they are all nice and dry, so you must have been correct lol

More people think like that than many realize
Wrong - I just asked ChatGPT the same question and it said:

“Assuming all towels dry at the same rate, it would still take 3 hours to dry 9 towels on the washing line. Each towel would take the same amount of time to dry regardless of the number of towels being dried simultaneously.”
It’s easy to complain about grifting when you’re at the helm of a massive tech firm that has unlimited budgets.
(Edited)
There is a (reasonably) simple way to tell wheat from chaff if you want to invest in either AI or blockchain, or droids, or cybernetic implants.

Ask your friendly neighbourhood nerd. Broach it to them about the second pint, when alcohol will overcome the social anxiety about being in company and/or natural daylight.

If they ask for more details, ask to look at the use case, or nod sagely and say “I see how that can work,” proceed with caution.

If they stare at you, laugh, or worst of all, are kind to you, run.

A lot of these things are fun tech in search of an actual valid use case. AI has many potential uses but needs to be applied properly, and the designers need to stick to a brief, a timeline, and the need to create an actual viable product. Not just create knowledge.

Blockchain is much more limited in what it can and should be used for: unless it’s for actual crypto (which I have opinions about), you’re usually much better off with a proper Cassandra or similar database.
Sage advice, but for clarity, your “friendly neighbourhood nerd” isn’t socially anxious - they’re just dreading having to spend the evening listening to you mansplain to them some piece of tech you read about On Twitter and you’re suddenly an expert in.
Captain Thunder might be a woman.
(Edited)
Fat chance, Fat Chance
I believe women are just as capable of mansplaining as men. Equality is a dog.
That's good to hear ... in a way!
(Edited)
And on the interweb, nobody knows you're a dog ...
No, but I am usually the neighbourhood nerd in question.
And Digitaurus is 100% correct: C-suite folks try and 'splain this stuff to me anyway. Yes, both sexes.
So I've learned in the absence of either consultancy fees or a valid escape strategy - smoke bombs are frowned upon in the office - I might as well get a couple of pints out of it. (Three and a couple of rums if it's crypto or machine learning which is not bloody AI godsdamnit.)
lol that’s why those pathetic mistakes is decades away from human reasoning and recognition of reality
Like other technology it is hyped for the short term and is very badly underestimated for the long term.

And yes we will have self driving cars and they will be battery powered.

Patience.
"And yes we will have self driving cars"

Just because "you" say so??? For all of us? Speak for yourself - if you are daft enough for that.

Hopefully "we" shall not. And I personally never shall in my lifetime be that stupid - nor that lazy. Speaking purely for myself.

As an old timer once said in my youth to reckless motorists he saw on the road: "Go right ahead, people!! Hell is only half full, plenty of room for you."

His great wisdom is for the ages, IMO, and in my now long experience of probably a million or half a million miles of driving.
(Edited)
So you would also not use a computer for calculation then and do everything in your head?.Ever tried to calculate a root from a high number? Your idea of us humans being better drivers might look very silly in the future. It might actually not be allowed to drive ourself anymore as our reactions may be inferior to algorithms trained on this.

Sounds scary? Tell this to the millions of non dead people that will be saved.
(Edited)
To get the square root of 1111111, just remember that the square root of a million is a thousand. So it’s just a thousand and change. How much change? Well, 1100 squared is 11x11x10000 which is more than 1111111, so less than 100 in change. See? No computer needed to get a reasonable estimate.
The robotaxi model seems a bit far off. We'll need to make sure robots are safe enough to drive around human drivers which might actually take decades. But what do you think of automating heavy duty goods carriers on special highways? The human driver will still be needed but maybe 1-2 specialist drivers can control/manage a platoon of trucks which run one after the other to reduce drag force. Might it actually befit lorry drivers? For eg. They can actually rest at night while fulfiling tight deadlines. The trucks themselves can run 24x7improving productivity.
Kind of like goods trains?
Well, you’re insulting, so ipso facto you’ve got to be right. Cheers.
Say it louder to the people in the back!

I still don't understand how AI adds value to professional productivity or company growth?
That’s quite ok. Take a couple of classes on line or on campus and you will understand.
So right ahead with that clueless condescending manner toward a valid point. It will inevitably cost you. Your call.
I am not sure why you are all bent out of shape. He does not understand, that's fair, neither did I a while ago. Fortunately there are plenty of resources to learn.
So the answer to my question is :

" They didn't go over these two questions in Udemy online class I enrolled in and watched prerecorded videos of which now makes me an expert in not answering the specific posted questions and AI's potential."
Sure where NLP is repackaged as LLM calling it AI to sell it ill-informed people who seem to once again buy into Silicon Valley hype. Sure I will take a class or two when a Silicon Valley Tech Bro offers a class on how to overhype & rebrand a existing tech product as a historical game changer without providing any additional details. This products is a game changer for more LinkedIn AI generated posts & people with no knowledge of tech or finance to post their hottakes on it.
We use it at work, but generally a better search tool without the ad's. It's sort of like a bright and eager researcher to assist you.

It's not always right, but it does help shift things along.

You can even run these models for free on your laptop. Works particularly well on a Mac M1 and upwards.

"I still don't understand how AI adds value to professional productivity or company growth?"
Good question!!! Bella domanda. In truth, the burning question of the day.
Just as the huge infusions of pandemic spending led to graft and embezzlement on a monumental scale - though such vast, impossible to supervise spending was indispensable to survival in the pandemic - so inevitably and irremediably shall the mad rush to AI. Ictu oculi.
Simply a fact. Clever investors know that ultimately it all has to shake out in the IT tech world, the market, and in society. Inevitably there will be big losers and big winners. Wise and prudent investors best manage that. Clue: flavor of the month or flavor of the next few years stocks, if at all the answer, are only part of the answer.
Does "AI add... value to professional productivity or company growth?"
As a great, invaluable lawyer always used to say to me: "Well, yes and no." A trite saying, but invariably right.
Yes, AI can do so. In its present state, though, even in some of its most sophisticated and partially effective current forms, it also greatly destroys productivity, growth, and even the most minimal skill and competence, by its inhuman simplemindedness, inherent deficit in the human ability to see the context that determines meaning - so often resulting in the exact opposite of accuracy, 180 degrees out, 2 + 2 = 5. Garbage in, garbage out still very much applies.
E.G., the potential liabilities of the many major law firms that rely upon AI in such headlong fashion are mindboggling. AI can easily - without costly and extensive expert human intervention - result in a translation of a key legal text, court decision, contract, etc. - that is the exact polar opposite of the provable meaning of the original text. Yes, the controlling text is the original - but that can give damaging whiplash after the fact to those misled by the translation.
Why have translations then? Just guessing would be equally effective.
Whoops
AI in its current state might be best described as you said: "inhuman simplemindedness"
To coin a phrase, perhaps??
Cheers, friend, you are indubitably, resoundingly, and crashingly right.
(Edited)
See it as an personal assistant you can ask and assign jobs to. He might not be as experience as you are, therefore, you have to check his work. He will get better after a while because he will learn. And eventually he might even surpasses you. But he will not leave you as a real person would do
(Edited)
So from what you're claiming it is not only useless for increasing productivity but reduces it my productivity with expectation I implement a teaching plan, spend time teaching & supervising it? Great! Thanks for the info!
AI is real. Just like dotcom boom in 1999, AI will go bust then up again.
"...resulted in some crypto-like hype"

You don't say? The initial major driver of AI investment was not genuine financial opportunity; it was ChatGPT's eerie ability to mimic human language.

Humans regard the ability to use language as the primary marker of intelligence. As such, we mistake ChatGPT (and consequently AI in general) to possess actual understanding. It does not. Even Sam Altman has confessed that ChatGPT 4 and the upcoming ChatGPT 5 'suck' when it comes to actual reasoning.

Hassabis's statement that surge of money is "obscuring the incredible scientific progress in the field", is a polite way to say investors have little clue about the underlying value proposition. This will not end well.
Is he not just saying that non-Gen-AI (which he focuses on) is where the teal value will lie?
I see where you are coming from but to say AI doesn’t have any understanding is a massive overcorrection. It has an incredible ability to understand things, that is why it can mimic human language and provide useful responses.And no, do not mistake its basic objective function (predict the next token) for the properties a trained model exhibits.
A reason ChatGPT can mimic human language is that, contrary to popular belief, human language is highly structured. See Noam Chomsky's Universal Grammar. We humans are such natural virtuosos at human language that we fail to see the underlying, highly complex yet rigid structure of language. It is this structure that ChatGPT is mimicking, not our reasoning abilities, which is distinct from language.

This fact is not appreciated even within computer science circles, which adds to the overall hype of AI.
right, ChatGPT is basically scaled up parrot with an encyclopedia. If you call that reasoning then you never worked with trully intelligent, innovative people
Wow. You have really hit the nail on the head!!! Exactly right, sir. Bravo.
At this point "grifting" likely accounts for between 5-10% of GDP.
“When challenged with authenticating 16,000 individual facts, SAFE agreed with crowdsourced human annotators 72 per cent of the time — but was 20 times cheaper.“

Basically, all this hype for a technology which, like all past technologies, is able to be AMOST as good as humans, but much cheaper. Also factory-processed shirts, suits, shoes, cars or dinners etc. are almost as good as the hand-made made ones. But much cheaper. Nothing new.
Almost as good as the 30 years experience of a doctor is good enough to save lives by doing many things quicker and as it is cheaper it can be done for the whole population, quickly
(Edited)
Hype? What hype? Investor are too pessimistic and there is not enough cash around.
Central banks need to cut rates and stop QT!!!
ASAP!!!
(Edited)
Paywalled, but very funny!

WSJ: The AI industry spent 17x more on Nvidia chips than it brought in in revenue [N]
... In a presentation earlier this month, the venture-capital firm Sequoia estimated that the AI industry spent $50 billion on the Nvidia chips used to train advanced AI models last year, but brought in only $3 billion in revenue.
Good work so far, Sir Dennis, but is it Intelligence?
I am waiting for AI to solve any of the unsolved Hilbert problems in mathematics, before I can call it Intelligence. Until then it is just super fast computing with quasi-infinite memory processing data.
His name is Demis not Dennis.
Perhaps they were half right with his name
lol perfect
I do not think the commenter is a Chinese bot. But using nn for m and vice versa is a “technique” used by CCP types to confuse readers.
They may not be solvable, by human or artificial intelligence. I’d say you are setting too high a bar.
Article warning about AI hype full of AI hype
(Edited)
Partly true. One hype is about money making or, more ambiguously, about economic progress. The other hype is about scientific potential.
We are probably in the first inning of this revolution. It barely registers in the lives of the mass of humanity. Like the internet of 1995 AI is an abstraction to most. To shareholders of NVDA it could not be more real. Biden's Chips Act assured America's primacy in this field. If he wins in November he will secure Western prosperity and freedom; if Trump wins he will scuttle Western prosperity and freedom in favor of a dark AI as a tool of Russian totalitarianism.
Does anyone wonder which of the exclusive British clubs Hassabis belong to?
As he was born in Finchley and he is very, very less ordinary I doubt he would even for a millisecond want to be a member of a central London club with ordinary above average intelligence folk.
Syndicate level designer alumni network.

Black and White developer alumni network.
This guy could probably get a meeting at the Oval Office or 10 Downing St in an afternoon's notice if he wanted to. That's his club.
Agreed. Just find the other FT article about London clubs yesterday amusing to read
(Edited)
Yep, I did get the reference haha!
Being lonely in clubs? What for?
Dumb Money ruins anything it chases. Same marks throwing money at the same con men with crypto scams two years ago or AI scams now.
The hidden incentive is that DeepMind want to dominate the field, hence why they have access to Google Data- to be the leader of the AI race means world domination and control. They’ve essentially created a new species, of which they want to throw PR stunts such as this article to keep the masses misinformed. People who actually believe this are naive.
I'm old enough to remember when deep mind started the entire AI craze with their Atari paper coining the term "superhuman AI" and then doing one PR stunt after the other. Good for google who released tensorflow for free only for people realising their PCs are too slow so had to use cloud services. Just too bad that google lost in the cloud race and AWS and MS won. Also Meta. So deep mind got absorbed into google brain.
Those in the quant side of the finance world will recognise while interviewing undergrads for summer or full-time positions that they’re possibly dealing with a “winger”: those who are willing to lie (yes they “Sat In” those courses bc they had a course clash!), they learned all the material they loved it…AI gives wingers a lot of help in writing essays but like wiki it gives them the talking points. You learn to recognise wingers, they nod along knowingly. Their transcripts will reveal they took the easiest courses with the least work (“deliverables”) required! So it’s best to sit them down take away their mobiles and give them a few elementary probs in linear algebra or stats or portfolio thy.
Pls don’t say “god help the Univ’s admissions office staff they have to read AI-written essays” — dammit the staff will be the first to USE those ghosted essays in the future, if they read or absorbed them is a different matter!
We’re talking about all sorts of things that are just not real’
So he finally noticed?
As it stands they seem to want to use a system that is worse but cheaper, and claim it as success.
Making things worse is not really productive, even if you do so cheaply.
Amateur question from an AI researcher: any way to short this?
the trouble with shorting is not knowing that it will go down, but knowing when it will go down. you can short all the AI hyped stocks, but they could stay buoyant for years. look at the last VC bull run.
And the potential losses (of holding naked shorts) are unbounded, because the nominal stock prices are.
Perhaps you could buy some out of the money puts? Nicely levered. I hear the equity skew is inverted by the market optimism.
keynes allegedly once said "The market can stay irrational longer than you can stay solvent"
Cynicism about his motives aside, he is clearly correct.
You cynic ;)
Can we not burst this bubble until after I get my payout cheers
Is there any chance AI will stop human beings killing one another en masse?

If not, why not?
One possibility is AI will leave no human left to be killed?😭
ask Schwarzenegger
It'll only increase and obscure it. That's because the present AI models are biased by their creators, so at some level motives are built in
Humans are hominids.
Even with new technology they will most likely remain hominids.
Doesn't science and research benefit from funding?
Not from for profit corporations. They cut corners, take short cuts and don't deeply study the risks
An obvious but important point.
That's "Sir" Demis Hassabis to you
Let the lemmings learn the hard way
Several Trillion dollars later ...
Yes, let those that refuse to learn from history get burned.
Tulips. East India. Panamá. CMO. Railways, cars, aeroplanes were great inventions but did not work well for investors, nor did Investment Trusts in 1929, despite being recommended by Goldman Sucks.

We’re at the beginning, maybe, of a new golden era of scientific discovery, a new Renaissance.”
...
Hassabis said that one or two more critical breakthroughs were needed before AGI was reached. But he added: “I wouldn’t be surprised if it happened in the next decade. I’m not saying it’s definitely going to happen but I wouldn’t be surprised. You could say about a 50 per cent chance.
Hype...?

I think Hassabis should spell out what he thinks is hype, and what isn't. Otherwise he runs the risk of seeming hypocritical.
(Edited)
I think we’re T the point where 70% of why is being said is hype.
I mean I remember all the hype bout lasers.
Since we still don’t have lasers that can shoot down from space and destroy a house by popping popcorn, the industry has not kept up with the hype . .
The reality of eye surgery, industrial uses, and changes to communication were wide ranging, but not necessarily part of the hype.
I mean did laser discs change your life?
Every new technology starts a hype train, or hype cycle, and involves significant deviations from reality.
Sometimes it is also a bubble, other times it is just goofiness . . . Like space lasers.
Nice reference to Real Genius!
(Edited)
Translation: “We’’ve got all the money we need. Please stop funding our competitors”
AlphaStop the competition?
No he is just one of the few people who knows what he is talking about.
He also advocated for stricter lock-downs during Covid. Just because you know what you are talking about doesn't mean you are always right.
Yes but he’s an expert in AI not in lockdowns
From memory he advocated for earlier lockdowns not stricter ones.
Man’s we only have Cummings’ word for that? Not sure his word is worth much.
(Edited)
This is the issue with conflicts of interest – he maybe saying something, he believes to be true; but you can be absolutely sure that out of the range of things he might have said, he will be picking the one that favours his company the most, and inhibits his competitors and potential competitors. we know from experience that tech tends to lead to dominant players who have frequently abused their market positions..

So, should we listen to him? Yes, of course. We should also listen to his competitors and others, and give proper regard to the personal interests each of them are serving in assessing the credibility and value of their input.
A lot of negative comments about something which is likely to impact nearly every facet of human life over the next 15 years.
We have as the article states, only just scratched the surface of what's possible.
Maybe. But right now and for the next 5 years given the capex it’a useless
Not true. Look at PayPal for example. They’ve committed to reducing headcount by a third within that timeframe, material business impact, driven by AI capabilities within customer support.

Appreciate there’s a lot of hype in the market but there are material changes taking place as well.
Apart from reducing employment and increasing the wealth of the wealthy, what consequences does your PayPal example have? Are any of them beneficial?
I hate being impacted by things why wouldn't I complain?
I agree we need a science-based approach not a VC-funded commercialisation race. UCLs Karl Friston is leading some groundbreaking research on Active Inference with his Verses Research Lab team. This is one potential avenue - explainability, standardised, real-time, and an order of magnitude lower compute / data requirement
Commercialisation of new technology had led to us all having a phone in our pocket more powerful than a university computer in the 2000s. The money from companies buys compute and fancy tools that universities cannot hope to ever afford. Sure universities are great, but if you suppress companies you are dooming the UK to another decade of stagnating living standards and salaries.
That fact doesn’t support your argument. Do you really think phones are still “more powerful” than “university computers” today?
I agree we need a science-based approach not a VC-funded commercialisation race.
And now you understand why DARPA existed.
When they create a general intelligence AI will it get a stock option?
I’m afraid I can’t open the pod bay doors Dave.
AI needs a Deep Winter.
I guess the .com bubble has already been forgotten
(Edited)
To be fair, we aren't in that league ... yet.
Here is the key: “We’re talking about all sorts of things that are just not real.”
Starting with the stock prices
Everything is arbitrary
AI . Very similar to social media. People believe anything if the story is good. Soon a home version for cheap fakes to the vulnerable.
On the other hand, people believe anything if the story is bad.
Yes a good story about bad is all that is needed for some to act
Nowadays the word "AI" makes me nauseous, it's like "blockchain" five years ago.
Also housing development, fast trains solar and batteries? Is this the argument for degrowth?
No, it's just the ubiquitous hype around a solution looking for a problem -- any problem will do at this stage, it seems. Add to that a hint of disquiet: it will find problems, it will solve them, and it will create many others.

I have recently landed a freelance project with a tech stack I hardly know. I was open about that with the client (a Dubai company I know well and who hires me frequently). Their answer? "Use AI". Words can't describe my weariness about the perspective of being paid to be the guardian of the machine, or a glorified code reviewer. So, perhaps I am biased -- actually, I am certainly biased. AI in my world is the guy who turns up uninvited at your party and ends up squatting your living room.
But the guy who created Devin can answer maths questions very fast. What is not to like?
But they might be wrong 30% of the time?
In my experience, unless I’m lobbing softballs, it’s more like 80% hallucinations in all the critical places. The author of the AI-powered code editor Zed recently griped in a podcast interview about its lack of popularity amongst developers, likening AI to a free Lamborghini people don’t like because it’s the wrong color, before walking back his metaphor, admitting to problems with hallucinations, and comparing his copilot feature to an intern you expect to hand in broken code:


Ugh. I get mad every time I read it. Both metaphors are terrible, but the intern one says everything. I hire interns because mentoring them improves our professional community, not because they are cheap labor to crank out shoddy work you can maybe correct later. That seems like a Boeing strategy.
Reasonably alarming. I’m not sure to thank you or not.
(Edited)
I was being sarcastic. The entity answering the questions is human and the CEO. Watch the videos of him doing it as a boy, smart lad! Doesn't guarantee he will turn fools gold into gold though :).
Then use better punctuation to make the tone of your comment clear lol.
But I bet you will use gen AI to write proposals for new projects, automatic unit tests and research using things like perpplexity.ai?
Nope. The most I do it's playing with Midjourney for like 10 minutes and then getting bored. I might consider Copilot at some point if everyone else uses it.
Sounds like that's the problem, you haven't gotten any use of it yet.
Not quite. I think the problem is that I am a maker at heart, I like to make things. I am not an artist, I am an artisan: I enjoy the process that goes from my brain to my fingertips. I have a business, but I am not a businesswoman: my place is on the production line. AI will make this making behaviour economically unsustainable. Therefore I resent it.
For me, copilot has just automated a lot of the boring parts. Once I know what I need to do, I don't want to spend 5 minutes writing boilerplate code to read files, loop, regex, etc, etc for the 100th time, I want to skip to the next problem solving bit. Copilot has been such a great addition to my life. We're in the sweet spot right now before it actually does replace the interesting bits too.
(Edited)
Interesting take. I haven't even tried it yet; perhaps I will like it, who knows. Well, probably not, but if everyone else uses it, it will become a competitive necessity.
I do try Copilot every so often in desperation, but frankly it does nothing more (in historical research) than search Wiki and other public sources. In that respect it's no more than a search engine, not AI at all.
Good to know: I have never tried. I tried once to feed a moderately complex problem to chatgpt, but there was a serious constraint: I had to solve it in the context of a lesser known library. The result? Chatgpt hadn't hesitated to invent inexistent methods on said library. It was less than unusable: it was pisstaking. Let's see how it goes in this new project, but I am not looking forward to it.
Exactly. When I use Copilot, it makes mistakes and I get confused about how the algorithm it wrote and the APIs it’s using are intended to work. It makes me dumber and my work buggier. When I don’t use AI, I have to RTFM, think carefully about what I’m doing, and occasionally even read the source of the APIs I’m using to understand how it works. The latter feels like more work, but I am learning something and, most crucially, I am comfortable taking responsibility for the result.
I tried perplexity and the answers it gave were not true.
Went back to its sources which it had misinterpreted, so some way to go yet.
Or “graphene”.
I agree that a huge amount of what is talked about at the moment is hype, however as someone who works in the field (and am therefore a bit biased) I think there are some areas where deep learning genuinely can provide useful solutions. Some will be in quite niche and narrowly defined tasks, others might be more widely used. I’m not convinced it will be as revolutionary as some claim, but after the hype settles I think there we will see areas where deep learning-based systems are still used, and it will still be seen as a useful tool in the arsenal for certain kinds of problem. But the trend of sticking ‘AI’ on everything, or touting it as a solution to just about anything, will die down.

In that sense, I think it has more potential than blockchain, where very few of the use cases that were touted at the peak of the hype train have really materialised.
I agree. In terms of eventual actual impact it will be very different from blockchain. The similitude ends at their respective inchoate phases.
If only we could make a magic potion of AI, cloud, blockchain and agile then solutions would appear in five minutes and we could sit back and play candy crush.
This actually made me laugh 😃. Nice one.
(Edited)
And “fintech” and “crypto” and “dotcom”.

Mind you, all housing developments have been “luxury” for years. Some labels stick!
And cyber. I hope Fatima hasn't stopped dancing for cyber.
It's no surprise that Sam Altman was a big blockchain proponent before shifting to AI. It's just another facet of the same scam.