Of course I miss water

 Cartoon by Jonesy

How good is "AI" programming?

 It turns out, not very good at all

From Jan Wildeboer

#Oops. A benchmark that checks how good "AI" coding works on real code from real companies. The TL;DR: Nope, it doesn't work. Not even close.

"Does AI code match the bar of a real-world enterprise? Our results show us that we're far from that reality. [...] We've found that today's models are weaker at understanding company coding patterns and frequently miss requirements or don't verify their assumptions."

https://withspecific.com/benchmarks/re

This is supposed to be one of the main uses of "AI".  How soon before the bubble pops?  

AI might be good at drawing elves or creating cat videos, but how lucrative a use case is that?

This is how the AI bubble pops

 This is a long but fascinating video.  Cory Doctorow talks about what AI is, what it can and can't do, and why the AI bubble will pop and what it will look like when it does.  

He talks about "process knowledge" —  how firms depend on the unwritten knowledge that makes them work.  His example is how a machine breaks down, and how new workers are told that you contact the guy who used to fix that machine or its predecessor, and who has retired but for £50 will come in and fix it.  My example is Mrs McGinty in dispatch who knows whom to contact when things go wrong, who will be late, who can be relied on, and who needs to be chivvied.  When she retires or is fired (to save money!) dispatch stops working.  His point is that workers will be laid off because AI will be able to do their jobs, except it won't, so the workers will have to be rehired, only they will have moved on.  So critical process knowledge will be lost.

He and his interviewer compare how the railway boom popped, but the capital stock (the railways themselves, the rolling stock, the property infrastructure) remained.  All governments had to do was to pay off the creditors, and give the bankrupt companies enough money to get them restarted.  Once restarted, they could keep going by themselves.  But AI isn't like that.  There is no equivalent of railways tracks, and the rolling stock depreciates in a couple of years rather than over 30 years, because Nvidia continually improves the chips used by making them more and more specialised, but doesn't bother to retrofit its newer chips to the older ones, which means that every three years, data centres have to be levelled to the concrete and completely refitted.  There isn't enough money, private or public, which will be able to save these companies in the same way the railways were saved. 

He discusses how Nvidia sells chips, then lends the money to buy those chips to its customers ("circular finance"), but nevertheless still counts those "sales" as sales, and says it is making a profit on them, when its customers are unlikely ever to repay their debts.  The debt structures needed to fund data centres are vulnerable to a credit crunch, because most of the credit is provided by risk-averse investors who have written many escape clauses into the contracts.  These include connecting to the grid, diesel generators, sourcing water for cooling, completing construction on time, and if these targets aren't reached because too many data centres are being built or opposition is too strong, there will be a cascading collapse.

There are many more insights and examples, too many to try to summarise here.  But he is the first analyst I have seen who clearly really understands AI and can also communicate well with non-nerds like me.

My two key conclusions from watching this video are:

• AI is not going to take over the world.  It can't, because it's not actually intelligent.
• the AI bubble will pop, and that is likely soon, because credit is tightening, Central Banks are raising interest rates, and bond yields are soaring.

World PMI very sluggish

This is my calculation of world manufacturing PMI, compared with J.P. Morgan's calculation.  I only started keeping the J.P. Morgan data in 2011, which is why I needed to make my own calculation to understand previous cycles.  Where I don't have back data for individual countries, I have used manufacturing business confidence, and estimated what each country's PMI would have been if it had been calculated by IHS Markit (which used to publish PMI data before S&P Global took over.)  In some cases, I have smoothed the input series (some, such as ABSA's PMI for South Africa, or the AIG PMI for Australia or Canada's Ivey survey, are very "spiky", i.e., have large month-to-month random errors.)  In other cases, I have extreme-adjusted the series before I used them to calculate my estimate of world PMI.  This mostly, in effect, reduced the down spike from COVID, but had some small effects elsewhere.

Why this chart is interesting is because, hitherto, in all recoveries, from deep recessions or shallower slow-downs, the rebound has been sharp.  This cycle, it's been a slow, and not especially steady ascent.  Observe that it actually began a steep-ish recovery at the end of 2023, before it fizzled out.

Obviously, Trump's tariff tango has something to do with this, but I suspect there's more to it.  Inflation isn't falling like it should be when the economy is so sluggish, and part of the reason for that is the growth of monopoly and oligopoly is the US, and the West's determination to stop China exporting its deflation to the world, particularly in cars, solar panels and batteries, via tariffs and quotas.  Why was inflation lower before Covid, when manufacturing was just as concentrated as it is now?  Because everybody expected inflation to remain low.   But in the Covid rebound, firms found that they could indulge in a bit of "greedflation", and pushed up their margins, and expectations have accordingly shifted.  Monopolies and oligopolies now know they can shove up prices every year by more than they used to, and get away with it.  To use more technical terms, inflation over the last few years has been more cost-push than demand driven.

Sluggish growth may well continue, even though Europe is clearly (finally) recovering.  But higher inflation means that Central Banks will be reluctant (=slow) to cut interest rates.   And if the AI bubble pops, the US will go into recession.   If that happens, the US dollar will plunge, pushing other economies themselves into slow-downs or recession.  

Of course, happy days may be here again.  But I hae me doots.

Use of AI suddenly drops

 

From Futurism

After three years of unprecedented tech spending and nonstop hype, the demand for AI in the workplace seems to be drying up fast.

Referencing data from a recent US Census Bureau survey, The Economist estimated that the percentage of Americans using AI to “produce goods and services” at large companies rang in at a modest 11 percent in October, the latest available survey date. It’s not just that the figure is a bit soggy for the supposedly world-changing technology, but that it’s suddenly moving in the wrong direction: the financial publication notes that the percentage is actually down from 12 percent in the prior survey, conducted two weeks previously.

Looking at the big picture doesn’t make it any prettier. Back in March, the number of businesses with 100-249 employees that reported not using AI within the last two weeks stood at 74.1 percent. The survey results show a steady uptick in “no” results over the past few months, culminating in a dreadful 81.4 percent as of the latest poll.

For big corporations with over 250 employees, meanwhile, the “no” reports have crept up to 68.6 percent, up from the year’s low of 62.4 percent recorded in February.

The data is nothing if not a major red flag for an industry which is expected to spend $5 trillion on AI infrastructure between now and 2030. To do so will require a massive increase in revenue from both business and personal AI use — the latter of which has been lagging.

Unfortunately for the tech industry, enterprise AI customers aren’t picking up the slack. Though various non-government surveys cited by The Economist varied wildly in their numbers, they all seemed to spell out the same results: AI remains more of an experimental plaything in the workplace than a serious driver of productivity.

One economist at Stanford who tracks the use of generative AI at work found a major drop in usage month to month: though 46 percent of respondents reported using the tech in June, that number had fallen to 37 percent by September. Another estimate, by Fintech firm Ramp, found that AI use at American corporations went through the roof earlier in 2025 to around 40 percent, but has since plateaued.

The results follow a disappointing summer for AI advancements, with models like OpenAI’s GPT-5 falling short of expected performance gains. Still, the cracks in enterprise AI adoption had begun to show as early as December of 2024, when an EY pulse survey of 500 senior executives found over half felt they were “failing in their role” of supporting AI in their companies.

Instead, executives pointed to a creep of “AI fatigue” among the rank and file — which a year of AI horrors probably hasn’t helped.

With a $600 billion gulf between AI revenue and AI spending, an immense amount is riding on whether the tech can start bringing home the bacon.

My wild and woolly guess:  it's not going to bring home the bacon.    And that's bad news for the US economy, because most of the rise in spending recently has been on AI.

Is the US economy just one big AI bubble?

 An interesting and, frankly, scary video from TLDR.