What caught my eye this week.
Bad news! Not only are the machines now coming from our cushy brain-based desk jobs, but our best response will be to hug it out.
At least that’s one takeaway from a report in the Financial Times this week on what kinds of jobs have done well as workplaces have become ever more touchy-feely – and thus which will best survive any Artificial Intelligence takeover.
The FT article (no paywall) cites research showing that over the past 20 years:
…machines and global trade replaced rote tasks that could be coded and scripted, like punching holes in sheets of metal, routing telephone calls or transcribing doctor’s notes.
Work that was left catered to a narrow group of people with expertise and advanced training, such as doctors, software engineers or college professors, and armies of people who could do hands-on service work with little training, like manicurists, coffee baristas or bartenders.
This trend will continue as AI begins to climb the food chain. But the final outcome – as explored by the FT – remains an open question.
Will AI make our more mediocre workers more competent?
Or will it simply make more competent workers jobless?
Enter The Matrix
I’ve been including AI links in Weekend Reading for a couple of years now. Rarely to any comment from readers!
Yet I continue to feature them because – like the environmental issues – I think AI is sure to be pivotal in how our future prosperity plays out. For good or ill, and potentially overwhelming our personal financial plans.
The rapid advance of AI since 2016 had been a little side-interest for me, which I discussed elsewhere on the Web and with nerdy friends in real-life.
I’d been an optimist, albeit I used to tease my chums that it’d soon do them out of a coding job (whilst also simultaneously being far too optimistic about the imminent arrival of self-driving cars.)
But the arrival of ChatGPT was a step-change. AI risks now looked existential. Both at the highest level – the Terminator scenario – and at the more prosaic end, where it might just do us all out of gainful employment.
True, as the AI researchers have basically told us (see The Atlantic link below) there’s not much we can do about it anyway.
The Large Language Models driving today’s advances in AI may cap out soon due to energy constraints, or they may be the seeds of a super-intelligence. But nobody can stop progress.
What we must all appreciate though is that something is happening.
It’s not hype. Or at least for sure the spending isn’t.
Ex Machina
Anyone who was around in the 1990s will remember how business suddenly got religion at the end of that decade about the Internet.
This is now happening with AI:
Source: TKer
And it’s not only talk, there’s massive spending behind it:
Source: TKer
I’ve been playing with a theory that one reason the so-called ‘hyper-scalers’ – basically the FAANGs that don’t make cars, so Amazon, Google, Facebook et al – and other US tech giants are so profitable despite their size, continued growth, and 2022-2023 layoffs, is because they have been first to deploy AI in force.
If that’s true it could be an ominous sign for workers – but positive for productivity and profit margins.
Recent results from Facebook (aka Meta) put hole in this thesis, however. The spending and investment is there. But management couldn’t point to much in the way of a return. Except perhaps the renewed lethality of its ad-targeting algorithms, despite Apple and Google having crimped the use of cookies.
Blade stunner
For now the one company we can be sure is making unbelievable profits from AI is the chipmaker Nvidia:
Source: Axios
Which further begs the question of whether far from being overvalued, the US tech giants are still must-owns as AI rolls out across the corporate world.
If so, the silver lining to their dominance in the indices is most passive investors have a chunky exposure to them anyway. Global tracker ETFs are now about two-thirds in US stocks. And the US indices are heavily tech-orientated.
But should active investors try to up that allocation still further?
In thinking about this, it’s hard not to return to where I started: the Dotcom boom. Which of course ended in a bust.
John Reckenthaler of Morningstar had a similar thought. And so he went back to see what happened to a Dotcom enthusiast who went-all in on that tech boom in 1999.
Not surprisingly given the tech market meltdown that began scarcely 12 months later, the long-term results are not pretty. Bad, in fact, if you didn’t happen to buy and hold Amazon, as it was one of the few Dotcoms that ultimately delivered the goods.
Without Amazon you lagged the market, though you did beat inflation.
And yet the Internet has ended up all around us. It really did change our world.
Thematic investing is hard!
I wouldn’t want to be without exposure to tech stocks, given how everything is up in the air. Better I own the robots than someone else if they’re really coming for my job.
But beware being too human in your over-enthusiasm when it comes to your portfolio.
The game has barely begun and we don’t yet know who will win or lose. The Dotcom crash taught us that, at least.
Have a great weekend!
From Monevator
Does gold improve portfolio returns? – Monevator [Members]
How a mortgage hedges against inflation – Monevator
From the archive-ator: How gold is taxed – Monevator
News
Note: Some links are Google search results – in PC/desktop view click through to read the article. Try privacy/incognito mode to avoid cookies. Consider subscribing to sites you visit a lot.
UK inflation rate falls to lowest level in almost three years – BBC
Energy price cap will drop by 7% from July [to £1,568] – Ofgem
House prices are modestly rising, driven by 17% annual spike in new build values – T.I.M.
Hargreaves Lansdown rejects £4.7bn takeover approach – This Is Money
Judge: Craig Wright forged documents on ‘grand scale’ to support Bitcoin lie – Ars Technica
FCA boss threatens private equity with regulator clampdown – CityAM
Sunak says it’s 4th July, in the rain, against a subversive soundtrack [Iconic]– YouTube
Sir Jim Ratcliffe scolds Tories over handling of economy and immigration after Brexit – Sky
No, it’s not all the Tories’ fault… but Sunak and Hunt were too little, too late – Bloomberg
Products and services
Pay attention to catches as well as carrots when switching bank accounts – Guardian
Which energy firm offers the cheapest way to get a heat pump? – T.I.M.
How to get the most from second-hand charity shops – Which
Get £200 cashback with an Interactive Investor SIPP. New customers only. Minimum £15,000 account size. Terms apply – Interactive Investor
Nine out of ten savings accounts now beat inflation – This Is Money
Problems when transferring a cash ISA – Be Clever With Your Cash
Nationwide launches a trio of member deals worth up to £300 – Which
Transfer your ISA to InvestEngine by 31 May and you could get up to £2,500 as a cashback bonus (T&Cs apply. Capital at risk) – InvestEngine
Seven sneaky clauses in estate agent contracts that can cost you dear – This Is Money
Halifax Reward multiple account hack: worth up to £360 a year – Be Clever With Your Cash
Hidden homes in England and Wales for sale, in pictures – Guardian
Comment and opinion
No, the stock market is not rigged against the little guy – A.W.O.C.S.
The life hedge… – We’re Gonna Get Those Bastards
…is easier said than implemented [US, nerdy] – Random Roger
Checking out a fake Ray Dalio Instagram investing scam – Sherwood
An open letter to Vanguard’s new CEO – Echo Beach
If you look past the headlines, London is charging ahead – CityAM
Most of us have too much in bonds [Search result] – FT
Why we still believe in gold – Unherd
Are ‘fallen angel’ high-yield bonds the last free lunch in investing? – Morningstar
For love or money – Humble Dollar
Naughty corner: Active antics
Fund manager warns putting £20k in the US now will [possibly!] lose you almost £8k – Trustnet
A deep dive into US inflation, interest rates, and the US economy – Calafia Beach Pundit
A tool for testing investor confidence – Behavioural Investment
When to use covered call options – Fortunes & Frictions
Valuing Close Brothers after the dividend suspension – UK Dividend Stocks
Meme stock mania has entered its postmodern phase [I’m editorialising!] – Sherwood
Kindle book bargains
Bust?: Saving the Economy, Democracy, and Our Sanity by Robert Peston – £0.99 on Kindle
Number Go Up by Zeke Faux – £0.99 on Kindle
How to Own the World by Andrew Craig – £0.99 on Kindle
The Great Post Office Scandal by Nick Wallis – £0.99 on Kindle
Environmental factors
Taking the temperature of your green portfolio [Search result] – FT
The Himalayan village forced to relocate – BBC
‘Never-ending’ UK rain made 10 times more likely by climate crisis, study says – Guardian
So long triploids, hello creamy oysters – Hakai
Robot overlord roundup
We’ll need a universal basic income: AI ‘godfather’ – BBC
Google’s AI search results are already getting ads – The Verge
AI engineer pay hits $300,000 in the US – Sherwood
With the ScarJo rift, OpenAI just gave the entire game away – The Atlantic [h/t Abnormal Returns]
Perspective mini-special
How much is a memory worth? – Mike Troxell
We are all surrounded by immense wealth – Raptitude
How to blow up your portfolio in six minutes – A Teachable Moment
My death odyssey – Humble Dollar
Off our beat
The ultimate life coach – Mr Money Mustache
How to cultivate taste in the age of algorithms – Behavioural Scientist
Trump scams the people who trust him – Slow Boring
Buying London is grotesque TV, but it reflects the capital’s property market – Guardian
The algorithmic radicalisation of Taylor Swift – The Atlantic via MSN
And finally…
“Three simple rules – pay less, diversify more and be contrarian – will serve almost everyone well.”
– John Kay, The Long and the Short of It
Like these links? Subscribe to get them every Friday. Note this article includes affiliate links, such as from Amazon and Interactive Investor.
RE: Google’s market cap loss… in the aftermarket:
Amazon: “Hold my beer!”
It has indeed been a blast. Certainly I’m none the wiser. And it feels more like the stadium lights are just coming on and they’re hosing down the bleachers, rather than the first innings…
I’m not going to be posting regularly here again (for one thing, I just haven’t got the time to try and keep up with 20 plus ‘link paste worthy’ pieces every day which I was latterly IDying); but, a bit like @TI with his Propegator Substack, I’ll still do the odd post here now and again of something significant.
However, they’ll each have a bit more narrative wrapping around any links.
Given the AI half of the dual themed weekend reading links yesterday, and not wanting to clog up that thread with another comment on AI, I thought I’d put my comment here, and just link to it from the w/e reading thread.
Recursive self improvement (RSI) in AI, where automated AI systems accelerate their own development, especially through algorithmic improvements and system architecture unhobblings, potentially leading to (perhaps very) rapid capability gains, is already leading (by Dario Amodei’s estimates) to 400% (5x) annual algorithmic efficiency improvements (compared to 2x to 2.5x p.a. increases in computation per watt from new generations of GPUs/TPUs/NPUs); algorithmic improvements which could double or (potentially far) more under RSI.
Frontier labs (e.g., OpenAI) are automating R&D, aiming for hundreds of thousands of AI “interns” in months and full automation in just a couple of years, enabling massive scale without human limitations.
On the bearish view, RSI yields faster, but merely incremental, progress within current generative AI paradigms, like accelerating a car from 200 to 300 mph. Exponential. yet familiar.
For the bulls, it enables transformative breakthroughs, akin to inventing flight, with multiple, successive, 99% costs’ reductions and new, unenvisionable, emergent capabilities.
Whoever is right on that score, diminishing returns in scaling laws (e.g., compute yields smaller reliability gains, like 9% to 90% then 90% to 99%); difficulties automating hypothesis generation, compute allocation, and strategic decisions, reliability/quality issues in AI experiments; all suggest (to me at least) that the bears are less wrong (and closer to truth) on this one.
But…..AI already writes most lab code; and, if course, Blackwell/Rubin chips enable more experiments.
If the bulls prove to be right, then RSI will compress GPUs’ economic life from years to months as algorithmic gains (potentially 800–4,000% annually) outpace hardware cycles, making chips obsolete faster and breaking synchronised depreciation models, creating a financing risks mismatch between long term debt (e.g., 5–7 year GPU backed securities) and shrinking asset relevance; risking writedowns, covenant breaches, and fragility in vendor financing loops (e.g., Nvidia funding customers).
With no derivatives market for compute neoclouds, lessors, and investors are exposed to spot price volatility, residual value drops, and obsolescence. Meanwhile hyperscalers (e.g., Microsoft, Google) are more insulated via internal repurposing.
If RSI leads to an acceleration in the rate of algorithmic improvement then computation likely becomes another commodity, like oil. If this scenario eventuates then they’ll be an urgent need for forwards for price certainty, put options for obsolescence insurance, swaps for volatility/risk transfer, and basis swaps for generational spreads.
RSI turns semipredictable AI economics into structural fragility, risking bubbles like the 2000s telecom ‘dark fibre’ crash. Emerging chaos.
Nvidia’s 12–18 month releases (e.g., H100 to Blackwell to Rubin), CoreWeave’s GPU lending, predictions of automated researchers scaling efficiency: all create tremendous potential for near term volatility and uncertainty which will likely undermine confidence even as the fundamentals transform with a shift from high cost GPU scaling (via ever bigger and more expensive data centres) to TPU/GPU lite algorithmically efficiency paradigms.
RSI might be highly unclear in its impacts, with a wide dispersion of effects, but, regardless, it is/would be a catalyst for profound changes and (being technically transformative, yet economically risky) ones necessarily including financial innovation to manage the acceleration of progress.
I found this piece especially useful in thinking around the issues and implications.
https://open.substack.com/pub/hyperdimensional/p/on-recursive-self-improvement-part
And this is the bonus piece for @old_eyes (in the weekend reading thread today (#62)), on Musk’s mad data centre missive, and the reasons behind the xAI and SpaceX merger:
https://youtu.be/cBExaZo29Q0?si=MIEe_CUJykr2-YGS
@Delta Hedge #603
Thanks for the link. My objections to space-based AI data centres were based on physics and engineering. I had not appreciated how much of the engineering was financial!
Yes. My motto is long engineers short politicians and evangelical entrepreneurs.
Musk didn’t found Tesla. That was engineers. He just brought in. Likewise, SpaceX may have been founded by Musk in 2002 but much of the early work it built upon in relation to it’s original Martian raison d’être was actually already done by aerospace engineer Robert Zubrin, and Musk didn’t get his BA and BSC awarded by Wharton until 1997 (he was born in 1971).
I also, like you, object to the idea of data centres in space on physical hard limits grounds, namely the fourth power law in relation to radiative cooling under Stefan–Boltzmann’s law (see my earlier comments on the same on this thread).
Any effective cooling in space is going to be necessarily non-convective in the absence of an atmosphere, and the radiator used will, consequently, have to be absolutely massive and, however efficient it is, and whatever materials are used, it would still be enormously heavy.
Even before considering maintenance and radiation shielding requirements (both huge issues in themselves), it’s hard indeed to imagine how this could ever be more economical than terrestrial data centres.
This will be one of the occasional update posts here.
Not an update on ‘news’ but on framing.
So less people and events, more ideas and issues.
We all realise that bottlenecks rule, and that half or more of the $50-$60 tn p.a. (of collective salaries) white collar burden to business is a problem of ingrained culture and organisational misalignment of mid manager (empire building / retention) incentives with shareholder value.
https://open.substack.com/pub/davidoks/p/why-im-not-worried-about-ai-job-loss
And that’s despite the incredible pace of improvement that continues at the frontier:
https://open.substack.com/pub/aidisruption/p/major-upgrade-to-gemini-3-deep-think
https://open.substack.com/pub/noahpinion/p/you-are-no-longer-the-smartest-type
https://open.substack.com/pub/aifutures1/p/grading-ai-2027s-2025-predictions
And see also Zvi’s comments recently here: “The most important thing to know about the METR graph is that doubling times are getting faster, in ways people very much dismissed as science fiction very recently.
METR: We estimate that GPT-5.2 with `high` (not `xhigh`) reasoning effort has a 50%-time-horizon of around 6.6 hrs (95% CI of 3 hr 20 min to 17 hr 30 min) on our expanded suite of software tasks. This is the highest estimate for a time horizon measurement we have reported to date.
Kevin A. Bryan: Interesting AI benchmark fact: Leo A’s wild Situational Awareness 17 months ago makes a number of statements about benchmarks that some thought were sci-fi fast in their improvement. We have actually outrun the predictions so far”:
https://open.substack.com/pub/thezvi/p/ai-155-welcome-to-recursive-self
And whilst Marcus continues to rail against hollow promises and points out the shortcomings:
https://open.substack.com/pub/garymarcus/p/about-that-matt-shumer-post-that
https://open.substack.com/pub/garymarcus/p/promises-are-cheap
From my own vantage point it’s getting like that famous short from my childhood ‘powers of ten’ out there at the frontier:
https://open.substack.com/pub/exponentialview/p/the-hundred-million-token-day
As Keynes said when the evidence begins to change you should change your mind…
One can be correct in one’s criticism and still be directionally wrong here!
But, in turn, the human side which makes up that half of that part of the picture is still only half of the overall picture.
The full view of the ‘horizon’ on the ‘ML side’ itself presents us with a different set of distinct and, for the most part, and upto now, rather poorly framed issues.
This piece helps to remedy that:
https://open.substack.com/pub/derekthompson/p/why-americas-ai-discourse-feels-so
Without fully understanding that there’s actually four separate ‘things’ which are going on now underneath the surface lump description of ‘AI’ then you can’t see the other half of the combined ‘AI visage’ clearly, and it’s all just a muddy muddle.
But ‘it’ can be both a bubble and world transformational, which both boosts productivity but is still (in some sense) ‘bad’ but without threatening humanity (at the species level) and which, whilst increasingly useful, is fundamentally unthinking (intelligence without precedent in evolution, but also intelligence without consciousness or presence).
With AI/ML being four overlapping (but neither coextensive nor coterminous) separately derived sets of issues (in this framing) it can be any combinations of possibilities at once which changes over time (there are twenty four simultaneous combinations of four variables).
That’s the complexity of what we’re dealing with. Even if it is a slow disaster, or instead ‘hitting the jackpot’, what that means (in it’s Gibsonesque way) then operates across and differentially affects all four separate domains of uncertainty:
https://open.substack.com/pub/doomsdaymachines/p/hitting-the-jackpot
Even just on the last, being most relevant to Monevator as the investing tier of issues; what set of futures derive from each set of issues then drives how the investing cohort of opportunities and problems then develops, and how that set of issues presents to us now. And how those futures derive is about not just what those futures are compromised (what they look like) but the pathway to them.
The journey matters as much as the destination here, both the shape of the route and how long it takes us to traverse it:
https://open.substack.com/pub/davefriedman/p/ai-takeoff-speeds-rule-everything
On the one hand we have computing (quantum computing no less) with light:
https://youtu.be/rbxcd9gaims?si=KEZqofhU95Vzlvon
But on the other hand the job replacement situation is not panning out as hyped:
https://youtu.be/17KvQYyrBEQ?si=AGbCIrXKDYXcV6qN
https://www.telegraph.co.uk/business/2026/02/16/markets-fooled-believe-ai-magic/
And the jagged edge of adoption is serated by institutional inertia:
https://open.substack.com/pub/davefriedman/p/when-ai-speed-meets-institutional
And here’s the interview that set WSM off, as summarised by Zvi:
https://open.substack.com/pub/thezvi/p/on-dwarkesh-patels-2026-podcast-with
Per one commentator (AI Disruption Substack): “The overall software US ETF IGV has already dropped a cumulative 30%. Among Big Tech, Microsoft, Google, Amazon, and Meta have collectively lost more than $1 trillion in market cap largely because of excessive AI investment! Stocks related to OpenAI such as Nvidia, Microsoft, Oracle, and AMD are also having a tough time.”
And yet check out these data points:
https://open.substack.com/pub/exponentialview/p/data-to-start-your-week-26-02-16
And, away from Substack, from the Model safety crowd, this on AI automation improvement metrics and timelines:
https://metr.org/notes/2026-02-10-simpler-ai-timelines-model/
But even if the Labs have been mostly right (or even too conservative on many measures!) most of the time that may not help much with AGI:
https://open.substack.com/pub/davefriedman/p/have-the-ai-labs-actually-been-right
This hasn’t stopped Noahpinion worrying he’s underestimated xRisk from AGI or even just plain old ML:
https://open.substack.com/pub/noahpinion/p/updated-thoughts-on-ai-risk
Because even if your not an AI optimist tech has changed our world:
https://open.substack.com/pub/noahpinion/p/how-technology-has-already-changed
The sceptic (gold bug) perspective is to play the secondary and tertiary effect of commodity bottlenecks rather than picks and shovel plays like ASML, AMD, ARM, Nvidia and TSMC:
https://open.substack.com/pub/bonnerprivateresearch/p/supply-chain-sovereignty-and-ai-violence
Whilst even those with a more optimistic hue are more focused on energy constraints than GPUs and wafer yields:
https://open.substack.com/pub/techfund/p/power-semis-in-the-ai-data-center
Or even ‘just’ high quality specialised glass semiconductor substrate:
https://open.substack.com/pub/polymathinvestor/p/investing-in-the-antipodes-of-the
But bubble not a bubble comes down to equipment cycle versus adoption cycle:
https://open.substack.com/pub/qualitystocks/p/the-two-acts-of-ai-investing-in-the
And, no, just because something gets exponentially cheaper, and better, doesn’t stop it being an investment bubble (as la TMT in 1999-2002):
https://open.substack.com/pub/ruben/p/bubble
And, with reference to this past Friday’s Moguls’ piece on AI: for @Ian (comment #1 in that thread, and per my asterisk footnote at #5 over there), here it is:
https://open.substack.com/pub/disruptivehorizons/p/the-galaxy-is-probably-already-taken
😉
@Delta Hedge #608.
Thanks for this ! I’m following up a lot of your other posts and references on AI developments and what to make of it all too. It seems you only have to take your eye off the ball for five minutes and the research frontier is so far off there’s no hope of ever catching up with what’s going on again. Reminds me a bit of cosmological inflation.
Yes indeed. A never ending fractal, just like Eternal (chaotic) Inflation (a paradigm in search of a theory 😉 )