Weekend reading: First they came for the call centres

Reasoning models are dead. Long live reasoning models!:

https://open.substack.com/pub/artificialintelligencemadesimple/p/reasoning-models-are-a-dead-end-breakdowns

Shorter TL:DR to the TL:DR: it’s the wrong architecture, similar to Marcus’s disparaging of the pre-trained, backward propagation, weight adjustment, deep layered, token shape prediction paradigm and advocacy of hybridised neural nets and formal symbolic language overlay / syncretic (deterministic programme like and probabilistic input/output net nodes) approaches.

The longer TL:DR is from the actual TL:DR: “Reasoning models are a dead end because they try to compress a dynamic control process into static weights. Reasoning is not a pattern you can train; it is an algorithm you must run. When you train on reasoning traces, you only capture the final surviving path”.

My views, we need a better model reward function and ability to learn continuously to update a realistic world model.

Honoured by the link to this thread in the Monevator Weekend Reading Robot Overlord roundup 🙂 Thank you @TI.

The financing doom loop (and ‘dark fibre’ parallels: “the fiber optic buildout of 1999 where $500 billion of infrastructure investment produced 2.7 percent utilization and 12.8 percent default rates”):

https://open.substack.com/pub/shanakaanslemperera/p/the-stargate-deception

‘And yet it moves’: only today I got a LRM over half a dozen iterations to produce a 100% correct 6,000 word, original advice on a novel issue in less than a day including iteration, checking, polishing up. Realistically that would take 3 days from scratch without any assistance from ‘AI’ (or whatever it should be called). The hyperscalers probably won’t raise the $3 tn to $8 tn needed for the data centre buildout by 2028 to 2030 nor will OpenAI (or anyone get to 250 Gigawatts data centre electricity use) by 2032 (equal to India’s entire electricity consumption). But that doesn’t mean it wouldn’t be desirable to try go get there given what the technology has already demonstrated it can do. It might be merely mundane utility and in no way AGI itself or even just comparable to meaningful notions of AGI but I can easily see it replacing 50% or more of White Collar / Knowledge economy jobs eventually (maybe sooner rather than later) saving tens of trillions of dollars annually in payroll costs and feeding through directly to the bottom line in the P&L.

Crikey, this is a bit wild as a theory of how to get to AGI:

https://open.substack.com/pub/josecrespo/p/the-math-openai-doesnt-want-you-to

It might, or might not, be b*ts**t, but I do definitely agree with this bit:

“You cannot build AGI if you cannot see what your model is doing. You cannot deploy industrial AI if you cannot audit its reasoning. You cannot trust a system that cannot demonstrate coherence.”

Jevons’ Paradox on turbo boost.

Why ‘AI’ (and by extrapolation and extension AGI) will require more workers (and create more jobs than it destroys) even if (eventually) 5 white collar workers end up being able to do the work of 50 in the pre-LLM era (<2023):

https://open.substack.com/pub/ruben/p/replaced

Intriguing, and at least faintly plausible; but, to begin with, the job losses must surely first occur as the quickest way for shareholder value orientated businesses to boost the bottom line.

In the Industrial Revolution eventually more jobs were created and wealth ultimately cascade.

However, initially, huge numbers of people in crafts were put out of work (the Luddite movement was right in this respect), and people working in factories were paid far worse, and had even more appalling working and living conditions, than even the essentially peasant conditions which they or their parents had lived under before the onset of urbanisation.

The immediate historical precedent is not good.

All these IPOs do seem a bit 1999ish.

https://open.substack.com/pub/aidisruption/p/mega-ipos-flood-2024-investors-cash

Too much supply coming on for the demand for new AI and frontier tech (SpaceX etc) shares? Time as always will tell.

A look ahead for the semis and revised hyperscaler Capex in the years ahead:

https://open.substack.com/pub/techfund/p/ai-and-semis-outlook-2026

This mentions token usage / demand growth coming in much stronger in 2025 than expected beforehand; but the sheer scale of growth in use and decline in prices is not covered but deserves another mention here: one model provider (can’t remember now if it’s OpenAI or Google for ChatGPT or Gemini respectively) has seen monthly token usage go from 9.7 trillion to over 1.3 quadrillion (1,300 trillion) (every thirty days!!), with the price per million tokens down 99.7%, all in 18 months (IIRC mid April 2024 to mid October 2025, I’d need to check though, this is from my memory). One token = 0.7 to 0.8 words.

A engaging take on how AI will, over the next 50 years or so lever the top percentile of workers/ entrepreneurs and businesses:

https://open.substack.com/pub/generativevalue/p/2025-annual-letter

The more I read + learn the more convinced I am that the optimum portfolio is a modestly leveraged barbell with a lower risk underpinning of:
– some low volatility, high moat, low disruption risk surface and intersection of quality + value stocks with a degree of ‘inflation protection’ (especially consumer staples and infrastructure/ utilities);
– plenty of ‘risk off’ assets (gold and precious metals, long duration TIPS, global macro HF strategies);
– some cyclical broad commodity and deep value energy producer and junior miners (for dividends and optionality respectively) and opportunistic exposure to deeply discounted HY REITs and other CEFs/ITs;
– With a bleeding edge overlay, at the other end of the risk / opportunity barbell, of high growth, tech and tech disruption stocks from the mega caps right down the cap weight scale to (at the app/platform end) small/ micro/ nano cap (including fintech like operators);
– Plus some sort of juicing with a low single digit allocation of starting capital to a DCA leverage equity rotation strategy (which, if the worse happens, and it goes to zero, then still only burns through a few percent of the starting value of the portfolio), contributing less to the DCA when valuations are high and more in the recovery (e.g. back above 200 DSMA) after a crash (e.g. more than 20% drawdown on 52 week high).

Money left of the table, where’s the value left at now in the AI stack?:

https://open.substack.com/pub/randomwalkwithdata/p/assets-or-software-and-at-what-price

Kernel optimisation:

https://open.substack.com/pub/importai/p/import-ai-439-ai-kernels-decentralized

Intriguing: “The most important takeaway is that decentralized training is growing quickly relative to frontier AI training, with decentralized training runs growing their compute by 20X a year versus 5X a year for frontier training runs. But the other important takeaway is that the sizes of these things are completely different – today’s decentralized training runs are still about 1000X smaller than frontier ones.”

So true.

https://substack.com/@mjreard/note/c-185945397?r=2kxl2k

Dwarkesh is my Go To AI Guru.

Thanks for the link there @TI. Fantastic piece of research by JPM.

So we’re looking at the New Deal Public Works Administration, the Manhattan Project, the electrification of industry, the Interstate Highway System, the Apollo programme and the Broadband rollout all rolled into one (p.4)!

What could possibly go wrong?

I was struck by two points on p.9.

First the 1.5x-1.75x cost improvement and 1.25x-1.5x speed improvement of current frontier model assistance to existing human task experts.

Obviously, this seems cautious (and necessarily subjective, even though precisely quantified here) in itself, especially given that earlier models, like GPT-4o, are shown here as actually slowing down experts to half unassisted speed and doubling the task cost. That seems a very harsh assessment of earlier models’ IRL performance.

But what really strikes me here is the disconnect between the chart and reality.

The chart suggests to me organisations could shead 30% of their staff costs now for the same output, or, if the market they’re selling into has capacity, that they could increase output now by 50% for the same costs base.

Yet when I look left and right I see no sign of mass redundancy and no productivity boom.

It’s like what Alan Greenspan said about the Internet in the mid 1990s (1996 perhaps?), namely that the impact was showing up everywhere except in the productivity and growth data (and, of course, a lot of the corporate profits of that era turned out to be baloney, as we now know).

So what the heck is happening???

We’ve got a demonstrably very capable technology.

Aside from motors and electricity this is perhaps the most capable (and in many ways it is *the* most impressive) tech that I’ve ever come across.

And yet I see near zero sign of it (yet) changing businesses truly fundamentally, or in many cases even very much.

Are people just using this to work less to output the same with the same numbers and mix of staff and, therefore, with the same cost base?

Will business act like Tito’s Worker Cooperatives and basically let staff run the show for their benefit, or will shareholders at some point demand their pound of cost cutting flesh out of payroll?

Maybe it’ll all just take a long time.

Organisational and institutional culture bottlenecks are in their way as significant as energy, data and financing availability to the success of this endeavour.

The second point is on the question of data bottlenecks.

The footnotes on p. 9 reference a total of 4,750 tn tokens of data on the Internet, on video and in image libraries. That sounds loads but in the context of LLMs is it really?

Frontier models are (already) using up to 10exp26 (100 septillion) effective Floating Operations in training and seeing demand of over 1,300 tn tokens per month.

Accordingly, it seems quite blase for JPM to note that those 4,750 tn tokens available for training data (3,100 tn + 300 tn + 1,350 tn) will be enough for training until 2030.

Zvi’s on the case for that Philip Trammel and Dwarkesh Patel piece on ‘Capital in the 22nd century’ which I previously linked to and which was also separately linked to by @platformer in the most recent W/e reading MV links:

https://open.substack.com/pub/thezvi/p/dos-capita

Or perhaps hydrogen fluoride is the ‘real’ AI bottleneck (for making silicon tetrafluoride for chip wafers):

https://open.substack.com/pub/shanakaanslemperera/p/the-invisible-chokepoint

Scaling walls, energy connection backlogs, power generation shortfalls, data availability insufficiency, financing issues, revenue shortfalls and operating losses. It’s not an obviously happy picture.

The periodic table od AI (a very useful conceptualisation of how it fits together):

https://youtu.be/ESBMgZHzfG0?si=zXyG7D_nSFZb42Y7

Nvidia Rubin CPX “a chip that analysts estimate costs roughly 25% as much to manufacture as a standard Rubin R200 while delivering approximately 60% of the compute performance”:

https://open.substack.com/pub/shanakaanslemperera/p/the-architecture-of-dominance-nvidias

A nice link to the 14 top AI resources (in the authors opinion) from 2025:

https://open.substack.com/pub/theaiopportunity/p/the-2025-ai-breakthroughs

Trouble at mill with Meta and LLMs:

https://open.substack.com/pub/garymarcus/p/breaking-marcus-weighs-in-mostly

Reinforcement Learning as part of Intelligence As A Service (‘IaaS’):

https://open.substack.com/pub/semianalysis/p/rl-environments-and-rl-for-science

Vibe coding with Claude: 39 ‘free’ models to use, apparently…:

https://open.substack.com/pub/aidisruption/p/39-free-models-to-use-with-claude

Interesting spin off from Google Ventures:

https://open.substack.com/pub/appeconomyinsights/p/how-motive-makes-money

Erratum: “do” in my intro to #432 should be “so”. Apologies. Typing on move again.

Robots on the move:

https://open.substack.com/pub/robopub/p/new-atlas-robot-heads-to-hyundai

The buzz is shifting from GPUs to memory. SanDisk, Seagate. One share tipping site today: “In February 2025, Western Digital spun off Sandisk. Wall Street yawned. The market was brutal. Sandisk opened at $52.20—then promptly crashed 7% to close at $48.60. That $5.6B valuation at close represented a 65% haircut from the $16B Western Digital actually paid for the company in 2016. Wall Street thought flash memory was a dead commodity. They were wrong. 11 months later, Sandisk is a ~$50B juggernaut. The “easy” money in GPUs has been made. The real money is now in the second-order effects—the bottlenecks that hyperscalers can’t engineer around, memory storage is just one of them.”

xAI going from $50 bn 2024 valuation to $200 bn on a $20 bn raise.

https://open.substack.com/pub/aidisruption/p/xai-raises-20b-more

With OpenAI looking to increase its next up round from a $500 bn to a $830 bn valuation this year and a float at $1.5 tn in 2027, SpaceX looking to IPO this year or next for $1.5 tn, Anthropic eying a $350 bn valuation round were headed into uncharted waters for private into public. Saudi Aramco can’t be considered a real comparator here IMO.

AI at the science frontline and frontier:

https://open.substack.com/pub/sciencewtg/p/americas-genesis-mission-artificial

Very true. The advances may come in the boring industries and mundane if impressive tech has longer adoption curves than current expectations credit:

https://substack.com/@chocolatemilkcultleader/note/c-196658825?r=2kxl2k

AI tools worth paying for?:

https://open.substack.com/pub/artificialintelligencemadesimple/p/the-ai-tools-im-actually-paying-for

In the author of the piece’s opinion ChatGPT and Claude trump Gemini on the top tier ($/£200 pcm)

Why isn’t AI taking all the (white collar jobs (already)?:

https://open.substack.com/pub/randomwalkwithdata/p/if-ai-is-taking-jobs-then-where-are

Answer perhaps: aging in place. Maybe we need more RE in the FIRE?

SK Hynix, Micron and Samsung highlighted on this one as (mixing metaphors again) the ‘gatekeepers’ to solving the data centre storage ‘bottleneck’:

https://open.substack.com/pub/uncoveralpha/p/2026-ai-landscape-who-benefits-the

SK Hynix tipped by Woodford last year since when it’s gone from ~200,000 South Korean Won a share to ~750,000, on a 12x TTM P/E today, compared to just 6x a year ago. Say what you like about Neil (and no doubt WPCT/ SUPP/ INOV has to be one of the biggest % disasters in Investment Trust history) he still called this one right.

The Apple/TSMC partnership (a deep dive):

https://open.substack.com/pub/semianalysis/p/apple-tsmc-the-partnership-that-built

The TSMC fundamentals over the last several years are amazing.

It’s all about context window memory: storage stocks melt up:

https://open.substack.com/pub/amritaroy/p/memory-and-storage-stocks-are-melting

Astonishing claims on I/O speed up and power requirements from fractal, distributed computing.

Is this a steaming load of BS or something worth investigating further???:

https://open.substack.com/pub/fractalcomputing/p/if-data-is-the-new-oil-what-if-prices

My suspicion antenna are screaming red alert, but it beggars belief there’s not more info on this given the magnitude of the claims made. Why are OpenAI et al not all over this? They’re invested in survival and winning, not propping up a failed data centre paradigm of bigger and more expensive is better. I’m struggling to see how *everyone* but the fractal computing people would be incentivised to downplay or ignore this. Maybe my framework (sociological motivation for sidelining) for looking at this question is wrong.

AI = energy:

https://substack.com/@denisgorbunov/note/c-197401471?r=2kxl2k

But Claude 4.5 and Claude Code apparently not to shabby either:

https://open.substack.com/pub/thezvi/p/claude-codes

Masayoshi Son, SoftBank, Switch and financing the AI boom:

https://open.substack.com/pub/netinterest/p/financing-the-ai-boom-2

“How long can Nvidia’s “insane demand” for GPUs last? And what happens to the company when the peak cash cow of their only viable revenue source is over? Nobody has given a good explanation. The semiconductor industry is cyclical by nature, Nvidia’s market cap rise has lifted the entire Semiconductor sector with it, but it’s boosted mainly on myths of scaling”:

https://open.substack.com/pub/futuresin/p/nvidias-2026-updates

Ohio gets it right? 6 bn cubic ft per day of new natural gas. Cheap electricity. Attract data centers. New construction jobs:

https://open.substack.com/pub/doomberg/p/intelligent-design

“AI 2027” now AGI 2034?? Some detailed and wild extrapolation here:

https://open.substack.com/pub/aifutures1/p/ai-futures-model-dec-2025-update

It (the AI Futures Model) is just so nuts here that it’s actually worth quoting Seth Llyod’s widely cited 1999 MIT paper on “The Ultimate Physical Limits of Computation”:

https://arxiv.org/abs/quant-ph/9908043

“A kilogram of ordinary matter holds on the order of 10exp25 nuclei. If a substantial fraction of these nuclei can be made to register a bit, then one can get quite close to the ultimate physical limit of memory without having to resort to thermo nuclear explosions. If, in addition, one uses the natural electromagnetic interactions between nuclei and electrons in the matter to perform logical operations, one is limited to a rate of approximately 10exp15 operations per bit per second, yielding an overall information processing rate of ≈ 10exp40 operations per second in ordinary matter. Although less than the ≈ 10exp51 operations per second in the ultimate laptop, the maximum information processing rate in ‘ordinary matter’ is still quite respectable.”

“The ‘ultimate laptop’ is a computer with a mass of one kilogram and a volume of one liter, operating at the fundamental limits of speed and memory capacity fixed by physics. The ultimate laptop performs 2mc2/π¯h = 5.4258 × 10exp50 logical operations per second on ≈ 10exp31 bits. Although its computational machinery is in fact in a highly specified physical
state with zero entropy, while it performs a computation that uses all its resources of energy and memory space it appears to an outside observer to be in a thermal state at ≈ 10exp9 degrees Kelvin. The ultimate laptop looks like a small piece of the Big Bang.”

Somehow, even with AI assisted accelerating progress (which, with respect to AI Futures, is one heck of a big assumption to make here) I can’t see us getting anywhere even close to just the room temperature version of the ‘ultimate’ computer, at ~10exp40 FLOPS/CPS per kg / litre. If so, this then, in turn, surely rules out the AI Futures main projection on grounds of practical if not physical credibility (AFAICT).

Some ordinarily, and all too plausibly, bad AI futures (no apocalypse/ no economic collapse, just sucks in a regular, if still profoundly depressing, way; think idiotocracy):

https://open.substack.com/pub/bloodinthemachine/p/four-bad-ai-futures-arrived-this

There’s nothing new under the sun whether (as Jesse Livermore, writing as Edwin Lefèvre, observed in his autobographical ‘Reminiscences of a Stock Operator’) in the arena of investing or, for that matter, in computer science and AI.

Just as the Panic of 1907 and the Great Crash of 1929 share many features, will we look back upon the later 2020s as being a 1974 like (first) outbreak of AI winter moment?:

https://youtu.be/hYnadoy8aQE?si=IWMKQk8zRPH4YGad

The first explicitly articulated AI bubble ran from the coining of the term as a grant research marketing tool at the Dartmouth conference of 1956; but it took until long after the General Problem Solver of 1957 and the Perceptron of 1958 failed (both already obviously so by the time that Eliza debuted in 1964, in a misfired attempt to show the shallowness of the field) for the UK to finally pull the funding plug with the Lillil report in 1973 and then DAPRA to do likewise in the US the following year.

How long might it take this time around? If the past is a prelude (which it rarely is), and taking the first LLMs as 2017 and comparable, in relative importance, to say the Perceptron of 1958; then we’d be looking at 2032/33 for things to fall apart.

But the cycle of hype and the vortex of VC and Private Lending funding is much greater now than then. That suggests to me an acceleration of developments and cycles now relative to the field of ‘AI’ in the 1960s.

This is the US orientated bull case for LLMs *already* creating a productivity boom extracted from the start of one of the many (many) investing related emails which I get now each day (I can’t share a link as there’s no associated Substack etc to it to share):

“Productivity is eating inflation for lunch
“Experts” told us what would happen. All the tariffs, deportations, pressure for lower interest rates, pointed to a guaranteed crash. They were wrong…Trade Deficit: Dropped to $29.4B in October, down from $136B in March. That is the lowest level since 2009. GDP Growth: The Atlanta Fed projects 5.1% for Q4. The historical average since 1947 is only 3.2%. We are growing nearly twice as fast as “normal.” Inflation: It isn’t spiking. It is dropping. Down to 1.9% (Truflation). How does that happen? Productivity. It’s up nearly 5%. Companies are figuring it out. They’re using AI. They’re navigating deregulation. They are producing more profit with fewer people.”

I have to say, and against my sceptical (cynical?), miserly curmudgeonly instincts, this is (I reluctantly concede) actually quite plausible. I’m using enterprise Copilot at work. Copilot is not exactly highly regarded by premium paid tiers LLM ‘super users’. But I’m impressed nonetheless. Surprisingly impressed.

On appropriate tasks (some of which are very highly complex and novel, multi layered questions), and with my guiding, shaping, evaluating and amending (where needed) outputs; I reckon (even with all that human domain ‘expert’ input) that, overall, I’m still getting at least (a conservatively estimated) composite 2x to 3x speed up (i.e. 3 days’ of work compression into 1; although, sadly, this just leaves more time for bureaucracy to fill the vacuum productivity creates).

On “What Happens When Superhuman AIs Compete for Control?” 🙁

https://open.substack.com/pub/aifutures1/p/what-happens-when-superhuman-ais

It’s either 6,000 words of ‘sci fi’ ‘faction’ presented both as a warning scenario and as an aid to prudent risk management,… or we’ve had it….

I can’t see how 2027 is still on the table as a serious timeline for AGI (or even ‘just’ for the timeline for fully Automated Coders) given the evident slow down in the rates of improvement of GPT4 to 4.5, GPT4.5 to 5, GPT5 to 5.1, GPT5.1 to 5.2 and from Gemini 2.5 to 3; as compared to the rate and the overall magnitude of incremental improvements in utility from going from GPT3 to GPT4.

Building moats at the AI application layer:

https://open.substack.com/pub/artificialintelligencemadesimple/p/how-openai-builds-amazing-products

Despite incremental increasing utility, core failure modes (most notably non specificity of output to instructions, of which hallucinations were the most obvious early examples) still seem to be ‘baked in’ to the current diffusion and transformer architectures, even after nearly a decade:

https://youtu.be/bv19nXfb0bc?si=OjDUYSY5IdLaoZpC

Not promising for AGI.

Finally, saying the silent part out loud:

https://open.substack.com/pub/aisupremacy/p/generative-ai-might-be-hurting-the-labor-market-future-of-jobs

Yes, it is, I think, likely that LLM/LRMs, and wider neural net ML applications, are now starting to cause measurable, and indeed significant, permanent job losses.

Kinda obvious that it probably would, sooner or later; and so it seems likely that it is now actually showing up.

To paraphrase and mangle misquote Sherlock Holmes, when you’ve eliminated everything else what remains is truth.

If this is indeed ‘AI’ related job losses (especially given the strong GDP prints State side), then the issue becomes not if but how far, and for how long, does this go on for???

Are we about to be entering a world transformed (and not necessarily for the better) job wise???

Although the Chinese seem to be doing pretty damm well on the innovation front to me.

They don’t exactly feel like Kodak in 1975, developing, but not then going on and commercially pursuing, a digital camera:

https://open.substack.com/pub/robopub/p/world-no-1-chinese-firm-open-sources

The AI blizzard is coming in thick and fast to my investment related inbox. I’ll quick fire the next nine in below in the interests of economy.

Here’s the first, starting with Mr Musk’s very bold claims:

https://open.substack.com/pub/aidisruption/p/musks-3-hour-bombshell-interview

For powering data centres China’s a hundred years ahead of the US, apparently…

https://open.substack.com/pub/exponentialview/p/data-to-start-your-week-26-01-12

AI capex is not slowing down…

https://open.substack.com/pub/crackthemarket/p/the-crack-the-market-signal-1

We’re on motorcycle and either about to career off a cliff, or head up a jump ramp to leap over half a dozen monster trucks!

Software AI was just the warm up act:

https://open.substack.com/pub/theaiopportunity/p/robotics-will-be-the-next-decades

And, lastly, this from the aforesaid commentary (from Money Machine Newsletter) on the effect of AI on jobs and productivity:

“There’s an old habit in the investing world. It goes like this…If companies aren’t hiring, the economy must be crashing. It’s time to break that habit. Everyone spent the last year panicking about a “cooling labor market.” They looked at the slowing hiring rates and flattened work hours and predicted the worst. They were wrong. While mainstream media is obsessing over headcount, real GDP accelerated to a 4.3% annualized pace. Productivity surged at nearly 5%. This is what efficiency looks like. We are seeing an economy that generates more output with fewer workers. That isn’t a recession signal. That is the holy grail of business. It’s a productivity-driven expansion. For corporate America, this is the perfect setup: Unit labor costs go down. Inflation pressure eases. Profit margins expand. Companies are realizing they can grow earnings without aggressive hiring or raising prices. They are doing more with what they have. That’s why earnings and revenues are hitting records even while payroll gains moderate.”

And on the current viability of the picks, shovels and power producers as investment themes:

“AI’s big checks just cleared, expect more zeros in the market. Microsoft spent ~$9B on IREN—a stock we called out early in the year, when it was on no one’s radar.
Amazon tossed in ~$5B for CIFR. Yesterday, these were bitcoin miners. Today, they’re AI power plants. AI’s no longer about compute—it’s about capacity.
Each new data center sucks down enough juice for 100,000 homes. Satya Nadella (Microsoft CEO) nailed it: “We have the compute. We just don’t have the power to plug it into.” You can buy GPUs. You can’t buy electricity. At least not fast enough. So the leverage quietly shifted—from silicon to supply. How long will this last, who knows. Could this change? Absolutely. But not at this very moment”.

And on Apple choosing Google Gemini to power Siri:

https://open.substack.com/pub/fundaai/p/researchgoog-google-gemini-may-become

Another win for Alphabet with TPUs, Gemini 3.0, AI search summaries, Waymo tie in, Google cloud, DeepMind, AlphaFold, and all the rest. A good year for Google.

Evening @TI.

Direct knowledge is a dangerous thing given (as Morgan Hounsel reminds us) that we each sample just 0.000000001% (1 exp minus 11) of the lived experiences of the 117 bn odd who’ve ever lived, but….that said: even though it’s clearly not at all what any of us ever though AGI would be, and, on any reasonable and fair minded view, it isn’t (at least yet) anywhere even remotely near to what a truly generalisable AI should be capable of; the top paid tier frontier LLMs still deserve more credit for what they demonstrably *can* do than what they’re currently getting from most people most of the time.

I’m using mid tier (£24 pcm) Copilot to do my self assessment this month. It seems to be getting it right (have to make sure all the docs are OCR’d first of course).

Just on educational YouTube today, ChatGPT 5.1 scores 88% on 2nd undergrad year quantum physics paper in just 30 seconds for a 3 hour paper:

https://youtu.be/JcQPAZP7-sE?si=vTd02DjjoTpLWeIh

How is it possible that this tech’s not having a positive effect on productivity? It beggars belief.

I’m not saying (unlike Elon) that this is the singularity. It certainly ain’t that.

But it is something, and probably something really quite important.

I am sure many companies will be able to downsize payroll and boost the bottom line P&L very soon. Whether they’ve the gumption and the brass neck to do so might turn out to be a different matter.

Of course, the socioeconomic and public finance effects (less payroll tax) *if* we see mass ‘head shed’ could be devastating if it’s handled badly.

But that doesn’t mean that the rate of profit in GDP can’t go up a lot even as a result of only mundane utility models. The aggregate profit margin is probably no longer a mean reverting series.

Of course, none of this means LRMs/LLMs can find out anything truly new (though they might, see AlphaFold and DeepMind). And without genuine innovation it’s an open question whether you can get long term economic growth out of ML. But you probably can reduce the labour share of output, make the Gini coefficient worse, make a few people very rich indeed, and raise the IRR for shareholders.

And of course the first to ‘lose’ their job will be the young would be starters who don’t ever get taken on to begin with because the cheapest lay off is the one you don’t make because you freeze recruitment. I think this is demonstrably what we’re starting to see in graduate employment stats.

My cohort (the fifty plus brigade) will be in the firing line eventually I guess, but at least I’m expensive to make redundant (21 months’ pay).

Interesting pointa.

On your a)., if the task is outside the circle of competence/ comfort zone of the person concerned then they’d have to outsource to a professional at much greater cost than a paid up LRM (yet alone a free one).

If the task was something the person could do then the LRM can do it so much quicker even with checking.

Any 130+ IQ undergrad physics savant could get 88% on a 2nd year paper on quantum mechanics but (remembering that the paper used was one which had never been made public, and so couldn’t have been within the model’s pre training set).

But no human could answer the paper at such a solid 1st class (>70%) degree level in matter of seconds (versus the 3 hours allotted the exam).

There must (surely?) be a huge (actual / potential?) speed up for tasks here.

*If* workflows can really be joined up / managed effectively by agentic automation imminently, then I’m finding it hard to see how there’s not then going to be some possibly really quite big productivity boost.

That doesn’t necessarily mean any more revenue/ sales for firms, but it would be disruptive, and it might be important.

On b)., these models are, in practice, becoming a heck of a lot more reliable, accurate and useful over time (so far).

It’s true that they still screw up some tasks and for really compute intense things like image creation they still can mangle the instructions.

But I can’t help feeling like we’ve passed the iPhone moment.

We’re not I think anymore in the Palmtop Palm Pilot / PDA / Apple Newton phase of a ‘nice idea’ but it’s clunky, hard to use, not that useful (of the GPT3 era).

Whether things can continue to improve given scaling walls, energy constraints (especially getting new generation onto the network), financing woes (Oracle CDS spreads) etc is an open question; although, as a general steer to action, I tend not to bet too heavily against progress.

On c). the outsourcing question is fascinating.

It might make outsourcing even cheaper and more useful / reliable because the offshore centre is fully utilising LRMs etc.

On the other hand, fully agentic models (if they do arrive soon, as promised) make automated reshoring / ‘inhousing’ much more economically viable too.

I suppose a virtual task doesn’t really have a location.

Although the comparison is inexact (to say the least), I’m slightly reminded of 1989, in the aftermath of Tiananmen, when my history teacher provoked derision for saying (IIRC, it’s stuck with me) that within 40 years the world would be looking to copy China’s economic model (state capitalism / market socialism). It seemed ridiculous then, with China’s midget peasant based economy, and the communist world then teetering on thr brink of collapse. Less than 37 years on and China leads the world on robotics, manufacturing, exports, high speed rail, fastest urbanisation in history, greatest and swiftest lift of people out of poverty in history, largest PPP GDP in world etc.

Maybe in the early 2060s we will look back on the shaky start to the ML era in 2022/23 in a similar way.

All agreed.

It’s a heck of a lot of work / iterating and what not to fully use these tools (and perhaps most aren’t) but, by God, for the right task, when they do work well, then they’re (or at least can be) downright impressive.

And I don’t impress easily on tech. I stayed well clear of dot.com companies and though social media in the aughts to be a profitless scam. More fool me! 😉

I think that the biggest issue is that it’s not obvious (at least me) that it’s a case of the ‘West and then the Rest’ on AI.

China may be a lot closer to realising the benefits of this and interrelated automation tech than the US realises or understands.

I’ll post some links in a second from the last couple of days on this. I mention it here in case it gets auto filtered into moderation (as more than one hypertext per comment / post).