Keir Starmer’s AI Plans Are Not Very Intelligent

2 godzin temu
Zdjęcie: keir-starmer’s-ai-plans-are-not-very-intelligent


Keir Starmer’s AI Plans Are Not Very Intelligent

Authored by Matthew Kirtley via TheCritic.co.uk,

Britain needs growth. More urgently, Labour needs to reassure the markets that growth is coming. With the dust settling in the wake of the 2024 Budget, Reeves has presided over a surge in gilt yields that has by far surpassed the much-decried heights achieved under Truss and Kwarteng.

The government needs to balance the books. And unless Reeves reneges on her promise to follow her fiscal rules — likely tanking Britain’s remaining credibility on global bond markets — then tax rises or spending cuts must come. Unless we discover growth.

To that end, Reeves and Starmer have launched a growth blitzkrieg (“growkrieg”). Reeves has yanked the chains of regulators to discover growth opportunities, pried from China a whopping £600m of investment (around 1.2 days of NHS spending), and launched an all-new Industrial Strategy Council. The growkrieg’s latest action came this week, with Starmer taking point to announce its latest wunderwaffe: the AI Opportunities Action Plan.

The AI action plan comes courtesy of Matt Clifford, a darling of the UK tech scene. A former McKinsey consultant, Clifford has successfully spun out the quantitative parts of his MIT political science masters into a position of tech guruship. After establishing the Entrepreneur First startup incubator in 2011, Clifford has leveraged his exposure to the technology ecosystem to become an influential voice on AI policy — albeit one that has never worked as an engineer or within a tech company.

That is, Clifford is incredibly impressive in the Starmerite sense. He’s clearly good at impressing the professional managerial class by throwing platitudes at them. And he also has racked up plenty of credentials and is incredibly well-network. He completes the full package by also being the co-founder of Code First Girls. In the eyes of many policymakers, this makes Clifford eminently qualified to be the lead adviser on what’s supposed to be a critical growth lever.

But what of the growth itself? Clifford claims AI can create £400bn in value for the UK by 2030 through greater innovation and productivity — a boost to the annual growth rate of 2.6 per cent.

This figure comes from a 2023 report by lobbying and strategy firm Public First, commissioned by Google. Their methodology is a black box, relying on AI to rank the likelihood of how much particular occupations would be automated away and the consequent productivity gains. And it’s also countered by just as many credible parties who are far more bearish on AI: for example, MIT economist Daron Acemoglu argued last year that AI will only grow US GDP by 0.9 per cent — 1.16 per cent cumulatively by 2034.

At this stage, guesses as to productivity gains from AI are extremely speculative. But let’s remain bullish and assume there is a potential £400bn value unlock from AI, as the government clearly hopes. How exactly does Starmer hope to capture it?

For the most part, Clifford’s recommendations mostly concern liberalising data sharing and datasets, expanding “skills” and training, encouraging regulators and government departments to use AI, and encouraging public-private partnerships around AI. Some concessions are made to infrastructure investments and theoretically dedicating money to UKRI’s AI Research Resource, and establishing “AI growth zones” to build out AI data centres.

Their overarching purpose is to encourage the development of the next generation of foundation models — that is, the huge AI models like GPT that use tens of billions of parameters to classify vast quantities of training data which can be used in a range of contexts. To that end, Clifford’s report calls for the creation of a UK Sovereign AI unit to encourage the creation of AI models that are trained and operated in Britain — with the idea being that we would benefit first and foremost from the innovations enabled by these best-in-breed models if domiciled at home.

It’s a £400bn gambit, in Starmer’s eyes. But there are major challenges.

First, we’re already seeing cracks appear in the idea foundation models are governed by “scaling laws” — or, very roughly, they may not get noticeably better as we pump more training data into them.

However, even assuming such scaling laws do hold, there’s an even more existential problem for Starmer. Computing power for AI demands energy. A lot. Even if we were to overcome planning and regulatory hurdles to building the data centres, those data centres would be at a fundamental economic disadvantage to those operating virtually anywhere else in the globe.

Let’s take GPT-4, which consumed over 50 gigawatt-hours of power in its training run. In the US, energy prices are at 12.9 cents per kilowatt-hour — so we can roughly assume that GPT-4’s training run cost $6.5Mn in electricity. In Britain, our industrial energy prices are 24.2 pence per kilowatt-hour, or around 29.5 cents. That means the same GPT-4 training run would have cost $14.8Mn in electricity.

This 2.3-fold greater cost to train foundation models in Britain may seem tolerable while they “only” mean a $10mn premium. After all, OpenAI claims GPT-4 cost over $100mn total to train. But those very scaling laws that the optimism of foundation models is predicated on means that they’re hungry for more power.

Commentators such as Scott Alexander have observed that every GPT generation needs roughly 30x the amount of compute and energy for a training run. So, a hypothetical GPT-5 would need 1,500 gigawatt-hours of electricity. For the US, that’s around $200Mn in electricity. For Britain, that same run would cost around $450Mn. At this point, a $200Mn price premium just makes it impossible for any sane private investor to justify a model’s training run and its inferences being based in Britain.

This all assumes that foundation model scaling laws will hold, which is dubious. If they don’t hold, a lot of the optimism for AI’s transformative productivity benefits will come to nought. But if they do hold, that means that Britain has an unassailable structural disadvantage in attracting any private capital in domestically based AI. And that’s just the up-front cost of a training run, never mind the ongoing energy cost of the actual ‘inference’ work of subjecting a model to prompts and queries.

Even if we ignore that many data centres have to wait 13 years for a new grid connection, or that our planning system already means that new data centre projects have to wait the best part of a decade to get built, the fundamental economics of British energy costs make the problem unassailable. Any foundation model will be considerably costlier to train and operate in Britain than just about anywhere else just because of our energy prices.

… it shouldn’t be surprising that they’re divorced from the material fundamentals of the technology they’re relying on as a crutch

This runs directly into Labour’s own commitments and aspirations. Industrial energy prices are not going to go down without significant new baseload generation capacity, which entails new nuclear or fossil fuel plants. And just last month, Ed Miliband’s own Department for Energy Security and Net Zero announced that it was still doubling down on its commitment to fully powering Britain with clean energy by 2030.

Even if we were to adopt the boosterism around AI and believe £400Bn in value stands to be unlocked if we seize the opportunity, it’s an opportunity the government is poised to miss. Our current energy system virtually guarantees that foundation models and their putative benefits — at least as we know them — aren’t going to stay domesticated in Britain.

Do Starmer and Reeves know this? Probably not. But it shouldn’t be surprising that they’re divorced from the material fundamentals of the technology they’re relying on as a crutch. After all, their goal isn’t to deliver growth — it’s to try to reassure us that growth is coming.

Tyler Durden
Thu, 01/16/2025 – 05:00

Idź do oryginalnego materiału