Latest / AI in the UK / UK Startup Nears $1bn Seed Round; 3 Startups Bet Big on "Physical World" AI; UK Data Scientists Get Funding For Synthetic Data
Transcript
- 0:00All right, your luck. Listen up.
- 0:15Hello and welcome to AI in the UK.
- 0:18I'm Neil's footman. And I'm Matt Wilkinson.
- 0:21Professionally, we help organisations, leaders and
- 0:24individuals learn about and implement this big scary
- 0:27behemoth called AI. There's some amazing stuff going
- 0:31on with AI in the UK that doesn't always get the attention
- 0:35it deserves. With AI in the UK, we want to
- 0:38change that. Hello and welcome to this week's
- 0:56episode of AI in the UK. This week's stories, Significant
- 1:02funding for three UK start-ups, suggests a future where physical
- 1:06world AI solves problems that LLMS can't.
- 1:10But have we been here before? One of the UK's most brilliant
- 1:14AI minds nears a 1 billion funding round for his super
- 1:17intelligence start up. Could this catapult the UK into
- 1:21the ranks of frontier AI nations?
- 1:24And drawing on experience forged during the dark days of COVID 2,
- 1:29UK data scientists are plotting a path for the future of
- 1:33synthetic data. But how much real insight can
- 1:36this actually generate? So to kick us off with the first
- 1:40story this week, Matt, over to you.
- 1:43Thank you, Nails. And yeah, this is really quite
- 1:46exciting. We've got stories from three
- 1:48different British start-ups developing AI systems for the
- 1:52physical world and collectively they've raised £40 million
- 1:57across neuroscience inspired computing, defence and
- 2:00infrastructure applications. This February Stanhope AI, which
- 2:04builds AI based on how the human brain processes and adapts to
- 2:08the world, secured £6,000,000 on the 12th of February.
- 2:13Their their founder and CEO Professor Rosalind Moran said.
- 2:17We're moving from language based AI to intelligence that
- 2:21possesses the ability to act, to understand its world.
- 2:24A system with a fundamental agency.
- 2:28Our approach doesn't just process words, it understands
- 2:30context, uncertainty and physical reality. london-based
- 2:37defence startup Ockham raised 2.6 million on 11th of February.
- 2:43The company embeds autonomous software into existing mass
- 2:47manufactured drones, so there's no hardware modifications
- 2:51required. Having completed field testing
- 2:54in Ukraine through the Brave One Defence cluster in January, it
- 2:57has now been cleared for a proof of concept with Ukrainian drone
- 3:01manufacturers with broader frontline deployment to follow
- 3:06and field whose AI platform supports safety, quality and
- 3:10delivery risk management for field teams across energy,
- 3:13construction, water and wider infrastructure.
- 3:16Closed a £32,000,000 Series B led by New York based Energy
- 3:21Impact Partners, the three rounds reflect Britain's growing
- 3:25investment in AI that operates in and adapts to the physical
- 3:30world. I think this is really exciting
- 3:32Neil. You know, we're seeing, you
- 3:35know, AI systems that are being used to hopefully save lives,
- 3:39you know, on the front line in sort of, you know, with, with
- 3:43sort of Occam's, you know, testing of, of the brave one in,
- 3:48in Ukraine. We're seeing Stanhope's active
- 3:53inference architecture being used to, to really understand
- 3:58what's going on in the world around it.
- 4:01And then, and then finally, we've got field who's, who's out
- 4:05there, you know, basically using sensors and, you know, images
- 4:10and videos to be able to improve, you know, risks, you
- 4:15know, even to sort of project, you know, timelines to be able
- 4:20to actually then predict problems before they actually
- 4:23become significant problems. So I think it's really exciting
- 4:26sort of seeing how technology's moving from just being, you
- 4:29know, a glorified spell checker to actually being able to help
- 4:33us in the physical. World yeah, very far from being
- 4:37a glorified spell checker certainly.
- 4:39And I can only agree that these developments, they are exciting.
- 4:43Some of the things that are being developed here and tried
- 4:46out in the UK with the support of UK academic institutions,
- 4:52It's it's very exciting. Far be it from me to pour cold
- 4:56water and I wouldn't claim to be an expert in these particular
- 4:59fields, but I think it is worth pointing out that there are
- 5:02always, as with these things, grounds for scepticism.
- 5:07The idea that further developments will require
- 5:13advances in areas other than LLMS is is an increasingly
- 5:16common one and it's one will return to a couple of times in
- 5:19today's episode. But apparently inference, what
- 5:24was it called? Active inference, active
- 5:27inference architecture. So this is not apparently the
- 5:30first time that this has been employed by start-ups looking to
- 5:34solve problems of this nature. It has been used by a couple of
- 5:38other companies and as of yet has failed to reach any anything
- 5:44like the scale or the effectiveness that LLNS have
- 5:48achieved. It's great that this kind of
- 5:51research is going on, but I think it's worth pointing out
- 5:54that there is a very, very large way, very, very long way to go
- 5:57between successful funding rounds, successful proof of
- 6:02concept and something that really gains traction
- 6:07commercially at scale. A similar point might apply to
- 6:11the to the drone technology. It is absolutely very impressive
- 6:18this this has clearly gone beyond just an idea and it has
- 6:21achieved a level of success in the test so far, but it's still,
- 6:27I gather, a very long way to achieve something resembling
- 6:32frontline effectiveness in in real wartime scenarios.
- 6:38Yeah, absolutely. I, I mean, I think that the, you
- 6:41know, the first point you raised about active inference it, it's
- 6:44definitely not a new idea. The idea has been around for for
- 6:4725 years or more. But what's changed is the
- 6:51deployment context. Obviously we've had huge
- 6:54advances in how large language models work, but also in how,
- 6:58you know, and, and actually the computer available on these
- 7:00devices at the edge of the network.
- 7:03So sort of the actual devices that where the technology, you
- 7:06know, where the, where the, the AI is being used.
- 7:09And that's allowing these sort of these, you know, this active
- 7:12inference to, to create these adaptive models so that they can
- 7:16actually see how things are working, you know, as they do
- 7:20happen. So I think that's really
- 7:22exciting to sort of see that that we've got that sort of, you
- 7:25know, we've got progress there and whether we're quite there or
- 7:27not, you know, time will only tell.
- 7:28But I think it's really exciting to sort of see that, you know,
- 7:32that extension and and sort of to to be able to get away from
- 7:36some of the brittleness that large language models sometimes
- 7:39have in these physical systems. You're absolutely right about
- 7:43the sort of the commercial logic here.
- 7:45I mean, it's pretty terrifying to think that an AI can be used
- 7:50in a drone and be making decisions at the same time, if
- 7:54that can be used to help us save lives or actually to stop, you
- 7:59know, to stop drones coming into, you know, into into Europe
- 8:02as an example. That's that's only a good thing,
- 8:05particularly when you know, you've got other technologies
- 8:08that are currently causing problems such as GPS blocking,
- 8:13you know, jamming, you know, those sorts of, you know, those
- 8:16sorts of signal errors are the things that are currently
- 8:19causing people to, you know, to lose their lives, you know, a
- 8:23lot of the time in Ukraine. And so by being able to defend
- 8:27against those sorts of things and, and have systems that can
- 8:31can operate in absence of the human operator at the other end,
- 8:35that's, that's pretty exciting, I think.
- 8:37It it is, it is exciting, but it opens a whole kind of worms
- 8:40itself about the degree to which humans can and should be
- 8:43involved in ultimately making these decisions.
- 8:45If, if they lose any kind of consequence or immediate
- 8:49consequence, then the decisions about how to use them become
- 8:55potentially considerably easier and the damage that they could
- 8:57wreak. I think in, in this instance,
- 8:59sure, drones have been a central part of Ukraine's ability to
- 9:04defend itself. And I think in this particular
- 9:06instance, something that can assist Ukraine, you'd have to
- 9:10say is, is almost by definition a good thing given the huge odds
- 9:14they're facing. We could have a whole episode
- 9:18devoted to the moral arguments around increased use of AI in
- 9:23warfare. I I think that the broader point
- 9:25really is just just the extent to which it actually does work.
- 9:28And even for, for this and for, for the other, the active
- 9:32inference ideas, you do wonder as well that if they, if they
- 9:37looked like they were going to gain traction, looked like they
- 9:39were gaining success, would they, would these companies in
- 9:44the UK actually have the, the compute and the scale and the
- 9:48capability to use them massively in a way that the incumbents
- 9:53could, You think perhaps the incumbents could just perhaps
- 9:56move into this field quite easily and, and absorb this
- 9:59technology. And I was listening to an
- 10:01interview with Gary Marcus the other day, actually, whose name
- 10:04will come up again in a moment. A well, a well known sort of AI
- 10:08sceptic, or at least generative AI sceptic.
- 10:11And he was essentially making this point.
- 10:13He said the bigger AI companies that have developed these,
- 10:18they've got enormous resources and have developed these LLMS at
- 10:21enormous expense are in his view pushing in this direction
- 10:25anyway. And, and we'll, we'll
- 10:26increasingly incorporate bits and pieces of this technology
- 10:30and then perhaps look to claim it as something they've solved
- 10:33themselves. So how far, how far the these UK
- 10:37companies can take it? I am, of course, I'm of course
- 10:41very hopeful. It's great to see this kind of
- 10:43innovation taking place. The extent to which they can
- 10:46capitalise on it, I think is remains an open question as
- 10:48well. Now, absolutely.
- 10:51And I and I think that you're, you're right to sort of to, to,
- 10:54to call out that the, you know, the large hyperscalers are
- 10:57likely to be looking at this as well.
- 10:59And with the capabilities that they have and all the compute
- 11:01that they've got, they would be able to, to catch up or, you
- 11:05know, very, very quickly if they decided that it wasn't worth
- 11:09acquiring these, you know, and you just look, have to look at
- 11:12the sums that went for, for buying open claw.
- 11:16And, you know, they, they saw something that they wanted and
- 11:20they snapped it up just like that.
- 11:21So I think it's I think it's a very credible risk that these
- 11:25companies do do face that, that, that either they get replaced by
- 11:29the big company, you know, the big hyperscalers, or they get,
- 11:32you know, gobbled up by them. Which is a pretty neat Segway,
- 11:37as it happens into our next story.
- 11:40And in the front runner for lame headline of the week
- 11:44competition. This is high ho.
- 11:46Silver X Deep Mind scientist in talks to secure a $1 billion
- 11:51seed round. So a London AI startup is in
- 11:55talks to raise $1 billion in what would be the largest seed
- 11:58round in European history. The company, Ineffable
- 12:01Intelligence, is 3 months old and has no product.
- 12:05It's founder is David Silver, one of Britain's most decorated
- 12:09AI researchers who left Google DeepMind late last year after
- 12:12more than a decade at the lab. The round is being led by
- 12:16Sequoia Capital at a $4 billion valuation.
- 12:20NVIDIA, Google and Microsoft are also reported to be in talks to
- 12:24invest. So a bit more a bit of
- 12:27background. David Silver is a fascinating
- 12:29character. He's best known for leading the
- 12:31team that built Alpha Goal, the programme that in 2016 defeated
- 12:35world gold champion Lee Sedol in Seoul.
- 12:39Have you seen? I can't remember.
- 12:40I asked you, mate. You've seen The Thinking
- 12:41Machine. That that documentary.
- 12:43Yeah. About about DeepMind and about
- 12:48the founders. So in the second game, Alpha
- 12:51Goal made a move so unexpected that expert commentators called
- 12:55it a mistake. Live on air.
- 12:57It wasn't. It won the game.
- 13:00Silver's argument has always been that this is what AI looks
- 13:03like when it goes beyond human knowledge, and that training
- 13:07systems on human feedback would have prevented that move from
- 13:11ever being made. But the company itself This is
- 13:16not the first company to get these enormous seed funding
- 13:20rounds without actually having a product ready.
- 13:22Such is such is life in in the AI ecosystem at the moment.
- 13:28Ineffable intelligence is built on that premise.
- 13:31That is, Silver believes large language models are structurally
- 13:34limited because they're trained on human data.
- 13:38His company will instead develop systems that learn continuously
- 13:42from their own experience, generating knowledge no human
- 13:46has ever articulated. He describes the goal as
- 13:49building an endlessly learning super intelligence.
- 13:53Silver's also a professor at UCL and he has spent his entire
- 13:57career in London. So plenty going on here, not
- 14:02least when a phrase and endlessly learning super
- 14:05intelligence pops up and that should certainly start stir some
- 14:11conversations. So there's a bit of context
- 14:16around this and it does have some, it ties into some degree
- 14:19to what we were just talking about.
- 14:20This LLMS have a ceiling argument isn't fringe or isn't
- 14:26anymore. Gary Marcus is a well known
- 14:29sceptic about the the the possibilities of LLMS and the
- 14:33limits of scaling. So as you know, the idea was,
- 14:37yes, LLMS and generative AI make mistakes.
- 14:41If we scale them enough, they will get to a point where that
- 14:44is no longer an issue. Hence these absolutely
- 14:46gargantuan spends on CapEx and these data centres and ever
- 14:50bigger, more powerful models with ever more data.
- 14:53But this, this feeling that scaling may not be the answer is
- 14:59gaining increasing currency. And more recently, did, did you
- 15:02hear this interview with Ilya Suitscaver on the, I forget the
- 15:06name of the podcast, but it, it, it was about two hours long and,
- 15:09and this was essentially the point he was making.
- 15:12Did did, did you have that? I did.
- 15:14It was a pleasure by Yann Lakun, who recently left Meta as well.
- 15:20So there's a, there's quite a lot of scepticism out there
- 15:23about the, you know, about the, about how far LLMS can go.
- 15:29Yes. So this is no longer a fringe
- 15:33movement. I think it it, it was and and
- 15:35Gary Marcus is always very quick to point out that you know, he's
- 15:38he's been he has been making this particular point for quite
- 15:43some time and has been ridiculed for it and and now there seems
- 15:46to be a move towards his point of view.
- 15:51It's, it's a, it is potentially a really exciting development if
- 15:55this comes off. And yes, a billion, these are
- 15:58kind of crazy days we're living in when a billion is not
- 16:01actually considered that big an amount of money, but it, but it,
- 16:04it is a billion for a seed round for a company with no product,
- 16:10but with a clear enough idea of what they're trying to do.
- 16:13So this is with a 4 billion valuation for something that
- 16:15would facilitate the move to. So currently it's all about LLMS
- 16:22and it's about generative AI and it's about scaling.
- 16:25And the idea is that will reach a limit.
- 16:27It requires something like this to get us over the line of to
- 16:30AGI. Now, if this is a company that
- 16:33can do it, and by all accounts, David Silver is a deeply
- 16:36brilliant man and he was involved in Deep Mind, which has
- 16:39been hugely influential. This could elevate that.
- 16:44This could put UK among the ranks of countries with actual
- 16:48frontier AI models. Potentially a very exciting
- 16:51development and the the amount of excitement around it can be
- 16:56gauged by the kind of people he was attracting.
- 16:58Sequoia apparently flew to London specifically to court
- 17:02him, so they weren't trying to lure him over to Silicon Valley
- 17:06along with the other very big name investors and the the
- 17:10timing just seems right. DeepSeek and Open a Eyes O
- 17:15series Reinforcement learning is delivering breakthroughs, so
- 17:20Silver arrives with more experience in this approach than
- 17:23almost anyone at exactly the right moment, so I think there's
- 17:26a lot to be looking out for here.
- 17:30Yeah. And and I mean, there's a lot of
- 17:31excitement there. I mean, but that you know.
- 17:36It's it's. Hard to look at sort of things
- 17:40like how for go and, you know, alpha proof the, you know, alpha
- 17:46fold and and sort of be able to look at that and say, well, most
- 17:51of those had perfect rules and clear wind conditions and the
- 17:54real world doesn't. And I think that's really the
- 17:55issue with reinforcement learning that that sort of
- 17:59generalisation beyond closed environments.
- 18:01You know, it's been promised before and and it really hasn't,
- 18:07you know, difficulty of that lead really shouldn't be
- 18:09underestimated. But you're right, this is a huge
- 18:12deal and it shows a lot of faith, but, you know, being made
- 18:15by people willing to take bets, you know, and, and and, and
- 18:20pretty huge bets. You know, we know that London
- 18:22has talent and ideas, but it doesn't yet have that compute
- 18:25infrastructure. And, you know, the energy costs
- 18:30here are going to, you know, significantly challenge whether
- 18:33the, whether the compute is built here or not.
- 18:36But I, I think the biggest question for me about this is
- 18:39just that you've got somebody like silver genius as he is, but
- 18:43he's never to the best of my knowledge and run a company.
- 18:47And so while he's building world changing AI, is he going to be
- 18:52able to build both the AI and the sustainable organisation?
- 18:56You know, and we can see the challenges that are happening,
- 18:59you know, at the likes of open AI and, you know, and then
- 19:03Fropic and others where actually building sustainable companies
- 19:06around these great, these fantastic tools is, is, is
- 19:11really a challenge. So I, I really look to sort of,
- 19:14you know, I'm, I'm excited to see where this goes, but I'm,
- 19:18you know, I, I with a help with a, you know, with a big dollop
- 19:21of healthy scepticism. So my answer to that question,
- 19:25David Silver is I don't know. He's not someone I was very
- 19:28familiar with previously. I read up on him a bit.
- 19:30And yes, I gather he hasn't run a big company before either.
- 19:36Obviously experience is helpful, but not always.
- 19:41There is always an element of of just sheer good luck to to these
- 19:45things. I don't know if you so.
- 19:48So if we take the example of Sam Altman and Dario Amalde, so Sam
- 19:53Altman did and has had a, a stronger business background
- 19:58than Dario Amalde, who as far as I gather, he was definitely more
- 20:02on the research and technical side.
- 20:04Now there's certainly people arguing now that he has become a
- 20:07formidable CEO. Anthropic.
- 20:11Anthropic appear to have, you know, you go back maybe a year,
- 20:1418 months, you would think, yeah, Claude's good.
- 20:17It's it's, it's, it's a good product.
- 20:21And it's picking up a lot of uses in certain niche
- 20:23categories. Fast forward to know when you'd
- 20:26have to say Claude. Claude's momentum at the moment
- 20:29feels unstoppable. And you'd have to attribute that
- 20:32in large part to Amodei's astuteness as a business person.
- 20:36They haven't gone for everything.
- 20:37And I realise I'm digressing it to some extent, but perhaps
- 20:41having that really strong background in research mixed
- 20:44with, mixed with the business acumen that he gained actually
- 20:49proved to be a strength. They, they've, they've
- 20:51Anthropica focused on what they're really great at rather
- 20:55than open AI who are doing everything.
- 20:58And, and you know, it's, it's, it's putting them in quite a
- 21:01difficult position now. So that was a, a digression.
- 21:05But certainly there are examples of a technically brilliant
- 21:09people who have gone on to be exceptional CEO.
- 21:12So David Silva could, and I certainly hope will be the next
- 21:17one. Now I'd, I mean, I'd agree.
- 21:19I mean, I, I, yeah, like I say, you don't have to, you know,
- 21:22some of the best CE OS, you know, they start a found a
- 21:26business and then run it for, you know, however long.
- 21:29But it is just that open question, you know.
- 21:31But again, it's a big bet and I'm sure, you know, investors
- 21:34will be making sure that the organisation is, is well run.
- 21:39You know, you mentioned Anthropic there and it was
- 21:42interesting. Just last week I was or the week
- 21:46before last I was I was in it's in Boston as a trade show and
- 21:50speaking to a company who was one of open A is case studies on
- 21:55their website as a, you know, as an early adopter of the
- 21:57enterprise technology, only to find out actually from, you
- 22:01know, that they're actually switching from open AI over to
- 22:07Anthropic, which is a huge, huge, you know, loss for an
- 22:12organisation like open AI. Big, big win for Anthropic and
- 22:16actually a huge switch. When you look at the just the
- 22:18sheer volume of custom GPTS and sort of the that's realising
- 22:23that it's actually not an easy way to transition from these
- 22:27these platforms to an between 1:00 and, you know, to another
- 22:30yet. So it's really interesting sort
- 22:32of looking at that and going, OK, this is a really interesting
- 22:36development. So sort of seeing that kind of
- 22:39shift and sort of seeing how, you know, an organisation like
- 22:43Anthropic is, is really focused on, you know, really the
- 22:48enterprise and the enterprise usage is rather than kind of
- 22:50being more of a toy, which I think Open AI is trying to be
- 22:54all things to all people. You know, it's a great play.
- 22:57So I'm really excited to see where this kind of turns up.
- 22:59Yeah. And I'm, I'm kind of making that
- 23:02transition myself. So seems Anthropic are really on
- 23:05a roll. Absolutely.
- 23:07Speaking of roles, nice EU KS been on a bit of a roll with in
- 23:15terms of simulation economy as well and sort of when AI models
- 23:20replacing reality. You know this, this UK start up
- 23:25electric Electric Twin has raised £14.4 million to build
- 23:31Ala models that simulate reality.
- 23:35And what they're doing is they're creating synthetic
- 23:38audiences to model human behaviour.
- 23:42It was a funding round led by Atomico and the company's
- 23:45founders are really quite an interesting they they've have
- 23:50government backgrounds, having run EU KS pandemic data and mass
- 23:54testing infrastructure from within, you know, from within
- 23:57Downing Street. So this is sort of dear to my
- 24:02heart because I've been sort of creating, you know, essentially
- 24:05customers and, and groups of customers within our lands and
- 24:08kind of getting aggregate data and from across sort of persona
- 24:13groups and then building out, you know, shall we say focus
- 24:17groups, buying groups, whatever you want to call them within
- 24:20large language models themselves.
- 24:21What they're doing here is they're combined.
- 24:23Taking that approach even further.
- 24:26That's the platforms combining survey data, large language
- 24:30models, as well as social social science research to model how
- 24:33real population segments respond to messaging campaigns or
- 24:37strategic decisions, Replacing weeks of focus group research
- 24:41with minutes of synthetic query. This is the sort of thing that
- 24:44I've been trying to do, but nowhere near the level of
- 24:46complexity that these guys are doing.
- 24:48So if this technology really scales, the implications stretch
- 24:52well beyond just market research.
- 24:55You know, So a government could model how different communities
- 24:58respond to a new policy before it's ever announced.
- 25:02They could stress test the messaging across age, income and
- 25:06regional lines in hours rather than months.
- 25:09You can imagine that that advertisers could sort of
- 25:13simulate responses to adverts before launching any campaign,
- 25:16making sure that they that they weren't misstepping.
- 25:18You know, so many ads sort of have landed badly because of
- 25:23sort of fringe ideas about, you know, responses to, you know,
- 25:28to, to to wording or to an idea sort of gathering pace and sort
- 25:32of spreading across social media.
- 25:34So really interesting to to look at that modelling, being able to
- 25:38model that in, in real time. Now that's meaningfully
- 25:41different to kind of capability and one with consequences well
- 25:44beyond sort of replacing sort of the focus group.
- 25:49I'm a bit torn here. It feels to me like it's
- 25:52augmentation, but it does have maybe a potentially scary
- 25:56underside to this. But this is, yeah, I do think
- 25:59this is genuinely transformative here.
- 26:02This value isn't just replacing, you know, human research.
- 26:05It's, it's, it's about compressing the iteration loop
- 26:09between insight and decision. You know, being able to, you
- 26:12know, go out find information, you know, and build, build an
- 26:16accurate model of populations and then test against it, I
- 26:18think is genuinely exciting. What?
- 26:21What do you think, Nils? I think so.
- 26:25This issue of synthetic data is one that pops up again and
- 26:28again. Another one when these enormous
- 26:32training models start to consume everything, then what do they
- 26:35start to train on next? I realise this is a slightly
- 26:38different issue, so perhaps you can help me out here.
- 26:42If they are developing synthetic audiences entirely on SO, these
- 26:48are synthesis. Synthesis based on data from
- 26:52actual human beings, correct? Yeah, so.
- 26:57How is this achieved at scale? This is something that is when
- 27:03you're initially running this, you've got a lot of pre-existing
- 27:06human data to run it on. How?
- 27:09How is this scaled? Yeah, that's and that's a,
- 27:13that's a great question. But I think that the, the answer
- 27:17that I often give here is not just how is it scaled, but it's
- 27:21also what does it replace? And very often the answer is a
- 27:28guess. So I think here what we're what
- 27:32we're looking at. Yeah, absolutely.
- 27:33If we can go off and get survey data and we can start, you can
- 27:36just matching the amount of data that that already exists within
- 27:40data sets such as you Gov, you know, the, the, the, you know,
- 27:43the UK government's, you know, polling system.
- 27:46And so if we've got all of that data and we've got all the data
- 27:49that we can then go off and find in other ways as well, we can
- 27:53build some really powerful models about how populations
- 27:56react to things about what about, you know, about really
- 28:00understanding, you know, you know, subsets of the population
- 28:04as well. And so when you're then looking
- 28:08at that and then being able to infer how things would respond
- 28:12and, you know, obviously there needs to be a feedback loop, you
- 28:14need to test that, you know, being able to go out and
- 28:18understand, OK, if we do XY and Z, the models suggest these
- 28:22outputs, these outputs could happen.
- 28:25What happens in the real world? I think that that feedback loop
- 28:28will help to then, you know, improve the models.
- 28:31But you'd have to be doing this a huge amount to really be able
- 28:34to, to be able to test. So how it scales and how they go
- 28:41about building that, that's the real trick.
- 28:42But I think it could be really, really, you know, really
- 28:45exciting. You know, they, they're claiming
- 28:48at the moment that they get sort of 95% concordance, you know, or
- 28:52accuracy on sort of those predictions at 10,000 times the,
- 28:56the, the speed that they normally would do.
- 28:59You know, they're obviously marketing numbers, but if
- 29:01they're, you know, if they are independently validated, then
- 29:05that's pretty exciting stuff. It is, yes.
- 29:09I, I did see that 95% figure. I think it would, you would
- 29:13really want to dig into that, wouldn't you?
- 29:14And, and know exactly what it is they're claiming it's 95%
- 29:17accurate against. Obviously trying to get
- 29:23dependable and, and up to date data on the attitudes and and
- 29:28particular circumstances of groups within society is both
- 29:33increasingly difficult and increasingly valuable.
- 29:36I think you regularly hear from the ONS the struggles that that,
- 29:39that, you know, the National Statistics Office, the struggles
- 29:43they now have in getting people to respond to surveys,
- 29:46particularly when they're trying to do the census and so on.
- 29:49It's really, really difficult. So there is, I would say
- 29:53unquestionably a need for a service of this type.
- 29:57The question would become, I guess, our own scale and our own
- 29:59methods. If they're using historical data
- 30:03to synthesise and provide, even even if you're saying it's
- 30:06better than a guess, and I've no doubt it is, clearly it is
- 30:10guesses will be, are inherently unreliable and, and, and you
- 30:14know, shot through with, with bias of who, of whoever's making
- 30:17the guess. But if they're basing this model
- 30:22on historical data, then the, the danger becomes, as it always
- 30:26would with something like that, that you're synthesising the
- 30:29past. You're not creating.
- 30:31It's not able to create genuinely new insights or bring
- 30:34into account rapidly changing circumstances on the ground.
- 30:39I, I should add, I'm, I'm saying this as someone who's, who's
- 30:42read about this a little bit. And no doubt if we spoke to them
- 30:46themselves, they would be able to provide sort of some
- 30:49responses to those kind of issues.
- 30:50But that, that's my concern. If it's synthetic data isn't a
- 30:54risk, it's also looking at it. It's synthesising the past
- 30:57essentially. I think that's, I think that's
- 31:01very fair. And, and of course you're then
- 31:03looking at predictions based off of data, but that's, you know,
- 31:06that is how we, you know, that's how we forecast the weather,
- 31:10looking at what's happened in the past and trying to
- 31:12understand how systems will respond in the future to a
- 31:16stimulus. I, I think for me, there's a,
- 31:19there's a real concern that we haven't touched on yet and
- 31:22that's the ethical 1 here. So actually, how could this be
- 31:26used to manipulate populations for nefarious means?
- 31:34And so I think that for me is really probably where you know,
- 31:38if you can really predict accurately how do peak
- 31:40populations respond. And if you can, you know, you
- 31:44can build these models, then it's not just about building the
- 31:47models. It's then actually the, the,
- 31:49the, the real task of regulation is looking at how do we stop bad
- 31:54actors make, you know, using that sort of insight to then be
- 31:59able to manipulate populations at scale.
- 32:02So I think that there's, you know, there's some, there's
- 32:04some, that sort of that sort of that sort of nefarious use of
- 32:12this sort of idea is actually pretty scary in my mind.
- 32:16And there's a few people that some of the investors might be
- 32:23coming from some of those organisations where we might not
- 32:26necessarily feel that they that we'd want them to be doing that.
- 32:30So I think that there's, I think that there's definitely a, you
- 32:33know, a risk there for manipulating populations at
- 32:37scale. And I think that that's it.
- 32:39That's something that I would be very cautious of and concerned
- 32:42about. Yes, I can only agree with that.
- 32:48Well, it's always good to end on a nice positive note like that
- 32:52nefarious use against broader populations.
- 32:54So that looks about all we'll have time for this week.
- 32:59It it has has ever been a pleasure of chatting through
- 33:01these UK issues that you might. Likewise, Nils, it's been, it's
- 33:06been great. And you know, hopefully, that
- 33:08this is all good and none of that nefarious detail ever comes
- 33:12to light. Yes, on that we can absolutely
- 33:15agree. All right.
- 33:17Thanks and I'll see you next week.
- 33:19See you next week. And that's it for this week's
- 33:35episode of AI in the UK. If you found it useful, you can
- 33:40subscribe wherever you get your podcasts and we'd really
- 33:43appreciate a great rating or review.
- 33:46It genuinely helps others find the show.
- 33:49You can also follow Neil's Footman and Matt Wilkinson on
- 33:52LinkedIn for links, notes and occasional extra content.
- 33:58Thanks for listening and we'll be back next week.
- 34:00Helping you make sense of AI as it lands here in the UK.