UK's Isambard-AI super powers up as government goes AI crazy Britain's beefiest supercomputer, Isambard-AI, is set to become fully operational this summer, as the government steps up its strategy to push AI everywhere as a driver for economic recovery. Announced about 18 months ago, Isambard-AI has been touted as a huge leap forward for AI compute power in the UK. Based on the HPE Cray …
No, I worry that you're betting big on a totally unproven, extremely expensive and power hungry technology that's unlikely to create more than a handful of jobs*.
* And they'll mostly be estate agents trying to sell unused data centres that were built only the foundations of hopeless and uninformed optimism.
I’m sure even WestminsterGPT would have got right ….
- not cutting winter fuel payments for pensioners
- found the £12m gap in funding that nixed the £1/4bn Astra-Zeneca Flu Vaccine factory expansion in Liverpool
- not be spaffing billions in Carbon Capture and Storage pie in the sky
- denied Heathrow/Gatwick airports expansion and eased congestion at those airports by helping the local authority in reopening mothballed Doncaster Robin Hood airport by moving some flights north- and avoided millions of unnecessary journeys to LHR/LGW by people north of Watford.’local airport for local people’.
- not spaffed £500m subsidy for a JLR battery factory in Conservative Somerset sticks automotive manufacturing wasteland a hundred miles away from JLR’s factories - and would have built it on the vacant site in Coventry - over the road from JLR’s Global HQ and close to the UK’s Battery Industrialisation Innovation Lab
- HS2 etc …
- be properly funding Foreign Aid/BBC World Service instead of making weapons - kinda keeps furriners in foreign. ‘Soft power’ which UK good at.
- have long ago taken water back into public ownership and tough shit to the investors
- long ago have cut off weapons supply to Israel and imposed sanctions for ethnic cleaning (that’s anti-Zionism, not anti-semitism - ironic considering this year will be the 80th anniversary of the public exposure of the holocaust).
About time the government realised (again) that the World Service is not (actually) part of the BBC, but part of the FCDO.
BBC only runs it because FCDO has no news (or other) journalists. FCDO could then take input from BBC/ITN/Sky etc.
FCDO should pay all costs, not lumping them onto the license fee.
TBH keeping it part of the BBC helps with trust as it’s not part of the Government but at arms length run by a much loved organisation that has faults and isn’t perfect…. Much indirectly by fiscal squeeze and the constant attacks on them by the former Tory Government.
FCDO can still pay for it though.
Much in the same way I have far-far more faith in the xclwnt content from USA’s PBS/NPR than Voice of America etc.
An entirely new sector for the British government to fail in, having failed in all the others.
Caveat. I am not a Reform supporter. I despise them. Simply stating a fact. After over a decade of even worse governmental failure than is usual in the UK, you can be as bad as Reform and still stand a chance of getting elected if the voters are racist enough, xenophobic enough, gullible enough and managed to avoid learning anything at school.
Is it about time that "Labour rebels" (100? at last count) left Labour and founded a new worker-oriented party - a bit like when Labour was set up by the Trades Unions. Combined with the LibDems, they could then be HM's Opposition.
Good thinking Citizen Smith! But you'd need a name. I had a look at the parties who contested the last GE, and the Socialist Labour Party is already taken, Workers Party of Great Britain that's gone too. Socialist Worker's Party, yep, one already. Trade Union & Socialist Coalition, already here. Cross-Community Labour Alternative, owned, Socialist Equality Part, that's a thing. Socialist Party of Great Britain, them too.
I suggest you either call your new party the People's Front of Judea, or the Tooting Popular Front.
Of course, a splinter group of Labour dire-hards who hark back to the golden age of Harold Wilson will do about as well as any of those groups named, but it will keep its members busy.
TBH more doubling down on Labour Principles and effective technocratic Government than racist-lite.
For example improving the immigration/deportation system so that - with due process - illegal immigrants were visibly but lawfully (express) deported within 2-4 weeks maximum than languishing state supported for 12+ months etc would do more to deter than threatens Rwanda or smashing gangs. Spend the £bn’s there to sort it.
Not opening the door to over a million culturally misaligned non-EU Willy-Nilly after closing the door to a couple of hundred thousand largely white Catholic European Heritage Poles a year post Brexit also just dumb.
Yeah, and they're all named Isambard over there anyways, with Isambard (aka Isambard-1) as the first all-ARM supercomputer, based on Thunder-X2 chips. Isambard-2 upgraded that to A64FX. Isambard-3 has 2 partitions, one is Grace-Grace and the other Grace-Hopper (GH200).
This here Isambard-AI is (I believe) part of the GH200 partition, with a test rig named Isambard-AI phase 1 currently at #155 in Top500 with 7.4 PF/s Rmax.
The Isambard-AI in TFA is about half the size of the Swiss Cheese Alps 10,752 GH200 super (#7 in Top500 with 435 PF/s Rmax), also an HPE EX design, which is used for both HPC and AI. I would hope (expect, really) that the whole Isambard-3 will similarly be usable for both types of workloads, despite the AI moniker of one of its partitions (iiuc).
I don't know.
Naming it after Bloody Stupid Johnson† might one day, in hindsight, prove to be incredibly prophetic.
Either Bergholt Stuttley or Boris de Piffle; your choice; either way a dismal half dozen of one or the other.
Advanced AI suffers ‘complete accuracy collapse’ in face of complex problems, study finds
‘Pretty devastating’ Apple paper raises doubts about race to reach stage of AI at which it matches human intelligence
So we have just enough AI to take jobs away from working in creative industries and call centres but not enough AI to solve any of the great problems the human race obviously needs help with (energy, climate, food...).
An interesting paper indeed, by the same Apple team who previously evaluated LLMs' abilities at grade-school mathematical reasoning using their own independent benchmarks (GSM-Symbolic), identifying similar limitations to the AI software.
Here, Figure 4, shows Towers of Hanoi collapsing at 4 or 5 disks without CoT in Claude Sonnet 3.7 and DeepSeek V3, and at 8 disks with the "reasoning" version of the models. Checkers jumping collapses at 2 vs 3 checkers, blocks world collapses beyond 2 blocks for either type of model, and river crossing with 2 people (not more) is ok with CoT but fails without it. Fig. 6 shows o3-mini is better at checkers jumping and blocks world by a bit but all tested LLMs fail "beyond certain complexity threshold" that is rather low (imho).
Makes me wonder how much Graph of Thought (GoT) might help in this, if at all ...
[edited to add]: The El Reg Expert take on this study is now up (the real McCoy!)!
"It's one thing to make a dumb tool work for you, but something with sentience and intelligence?"
Now go off into the weeds and contemplate "differently intelligent".
Humans grow up in a society where they absorb all sorts of social instruction that's not directly taught, yet we know, "hey, doing this would be wrong". We know that chowing down on a Limburger sandwich on the Underground is not acceptable. Our ideas and goals get shaped by those learning's and an AI without them isn't going to make the same decisions a human would. It might shed new light on something, but more often it's outputs are not going to be biased towards the dreams and desires of humans. Will using them train us to be less human?
Chowing down on a Limburger sandwich on the Underground is not acceptable.
I am not in much better position than an AI agent to understand that. No idea what Limburger sandwich is - obvious originally from Limburg. "On " the "Underground" which would seem unaccountable to a LLM that had a model of prepositions of place and anyone who only knew Metro.
The idea that AI in its current incarnation could ever reproduce the barest nuances of human socialisation would flatter the word ludicrous into blushing embarrassment.
Looked up Limburger - apparently an extremely (mal)odiferous cheese seemingly composed of toe jam and rotting cadavers' socks. Durian on steroids.
"but not enough AI to solve any of the great problems the human race obviously needs help with (energy, climate, food...)."
I'd set aside grand dreams of solving ill-defined problems and question what the tool* can do better today.
*AI is just a tool and should be seen as one.
Skroeder: Maybe it's pissed off.
Newton Crosby: It's a machine, Schroeder. It doesn't get pissed off. It doesn't get happy, it doesn't get sad, it doesn't laugh at your jokes.
Newton Crosby, Ben Jabituya: [in unison] It just runs programs.
Howard Marner: It usually runs programs.
I am puzzled by what is meant by AI 'matching human intelligence'.
Assuming everyone is agreed upon what the word 'intelligence' means in this context, there remains the matter of what exactly has to be matched. Human intelligence -measured on interval scales or by simple observation of fellow mankind - varies immensely in terms of signifying potential accomplishments. 'Average' - mean, medium, or modal, as one wishes - is mundane; indeed people at that level would replicate the errors manifested by AI, that is if they could even understand the task.
So, presumably, matching somewhere in the higher reaches of intelligence is sought, but where? If the problem were inverted, then one easily would identify the lowest range of functional intelligence as equivalent to the requirement for membership of the Welsh Assembly Government. Similarly, membership of the Westminster Commons hovers just below the 'average'. That leaves the puzzle of working out which occupational group(s) best typify the abilities required from AI. Suggestions please?
5000 GPUs for £225m
If these Nvidia GPU's come on Nvidia's own 8 x GPU sled, like all the other vendors have, that comes to 625 servers. So with that assumption, each server costs ~£360k, hoollleeee shhheeeet. I saw something early this year stating that grace hopper single gpu machines were costing ~$41k and up so I would say HPE are making a nice profit here.
Decent number of 64 bit flops (AI's 8 bit flops pretty useless for much else I suspect) which should permit some decent computational fluid dynamics, or operations research problems that require linear algebra with gazillions of equations with zillions of variables.
As might be easily predicted the facility will be obsolete before these punters finally throw in their hand and make the resources available to those that can use the facility to do useful work.
"Should know how to run itself, shouldn't it?"
At that point you have The Schumann Computer and once its solved the answer to life, the universe and everything, it will shut itself off and not answer any more silly questions from meatsacks.
(References from Niven, Adams and MiB all in one go)
Job loss is a concern for many. I have yet to see senior managers, directors et al being exposed to the same threat and speculation. Can they not be replaced by AI, and if so, why not?
Various layers of employees are suddenly perceived as outdated. Never mind the cladding. You can call it this or that, the outcome is the same.
What about those senior leaders, though? Who needs shepherds if there are no sleep?
I would suspect the speaker of being an arrogant c@nt†.
"this debate has been had many times. We need to push past it."
Perhaps realistically and honestly addressing all the issues and contrary views to reach some kind of consensus might be a better look as "pushing past " is redolent of the bully and as the "crash through or crash " mindset was of a former and equally obtuse generation.
AI/LLM is not going to solve the world's problems (ever); it will solve some well defined problems for some people (eventually); mostly it will create a massive surplus of problems from its innumerable unforeseen consequences and be responsible for damage that in some cases will take decades to repair.
"Fortunately" the way things are heading our current crop of loons are just as likely to cause a deep global recession that smothers this nonsense by removing its financial oxygen before too much irreversible harm is done.
† of course he necessarily won't always be the prime minister ...
-Steel mills
-Nuclear Power Plants
-Natural Gas
-Coal
-Chemical Plants
-Plumbers
-Skilled Workshop men at their CNC machine
-blacksmiths
That is, if you believe WEF Muppets like Starmer.
In fact these imbeciles moved industry to China and made all of Europe much poorer. And colder in the winter.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
