A.I: Privacy & Copyright

🤬 A.I.

It didn’t have to be this way. In the beginning A.I. promised so much, yet was hijacked almost immediately by grifters and attention–seekers. Crucifix spam. A.I. porn. Tech bro barons charging modest monthly subscription fees. Enhanced filters to make your Insta feeds pop.

I, along with many others, recoiled in disgust at this. Although I similarly eschew all kinds of social media, whenever I went on YouTube I found I had to increasingly dodge a blizzard of recommendations which were A.I. generated and laughably twee. They were clearly made by people who didn’t give a damn, existing only to harvest clicks for whatever meagre income they could squeeze: money, quite literally, for nothing. Some bloke in Tallinn trying to pay off a gambling debt. Someone in Kyiv dreaming of effortless wealth.

Meanwhile as a teacher, I soon noticed teens and (lazy) staff leaning into A.I. to dodge work and even thought. During a few weeks slumming at a local high school in May 2023, nearly all of Year 9 cheerfully submitted a 1000–word English assessment — a Sci–Fi story they had surreptitiously generated via ChatGPT — all mysteriously set in Idaho and sprinkled with em–dashes, delve, realm, underscore, tapestry and paradigm shifts: terms no fifteen year old has or will ever use. At a different school in Feb 2025, a perky Millennial demonstrated how staff could use EduChat to auto–generate school report comments. Great. Dear parent, I wanted to say your malevolent child understands as much about science as a gerbil does about quantum chromodynamics. However, my granny overlords won’t allow it, so here’s a tapestry of A.I. slop to facilitate your enjoyment.

Someone had invented a semi–plausible–prose–generator and didn’t people love it? Even more than TikTok™, mobile phones, Samba OG™s or, arguably, anything else.

Not me. I refused to countenance it. As far as I was concerned, our tech bro buddies can all take turns at the scaffold, and may their boundless data centres become tombs for their folly.

Saul. Road. Damascus

And there things stood, for years. A few of us refuseniks manned civilisation’s tower while a torrent of group–think arrows rained down. We will survive, we must. And yet…

In Aug 2026 Apple™ announced two new Mac Studio computers. It was claimed (or hyped?) that either machine was powerful enough to run A.I. models locally, independent of internet connections and monthly subscription fees. Once you install the appropriate software, it’s free tokens, forever.

Yeah, right.

Although the new machines hadn’t yet been released, I did have a powerful Mac from the previous year, with 64GB of unified–memory. I waited, pondered, researched, hesitated and… what the hell, gave it a shot. If it didn’t work then it was an effortless delete and cost nothing.

Downloaded LM Studio Bionic. Then came the LLMs: after considerable digging and trial & error, I settled upon Qwen Splash “dense” & “A3B MoE” models, the former for long–document review, the latter for quick paragraph fixes & chat. Both are open–weight, so for the moment are free. Load ‘em up toots. Hello prose–monkey — let’s delve into some stuff…

At first I asked it simple questions: What are the actor Christopher Walken’s best roles? Its answer was mostly correct, although it also — laughably — claimed he had a lead role in The Adventures of Priscilla, Queen of the Desert (1994) (!) I then asked for a plot summary of The Deer Hunter (1977), whereupon I learned Walken’s character remained in Vietnam to open a brothel (?) while De Niro’s character returned to the USA with a son he’d conceived on duty (!)

Not. Quite. Standby for rm –rf * .*, buddy.

But then I decided to give it a final chance with a straightforward task: Here’s an overly wordy paragraph. Please correct for typos, errors and make it more concise. Primed for confidently clueless results, I got the opposite: it flagged mistakes and kept my voice, just tighter. Cool. I fed it another paragraph, then five. Then an entire article. Then multiple articles. We then scanned, debugged and updated thousands of lines of antiquated PHP and SQL code in my Boomer Legacy photography project. As days became weeks, I discovered that (among many other things) I had a blind predilection for using “ojbect” instead of “object”, along with confusing (Konrad) Lorenz with (Hendrik) Lorentz (that one burned). The A.I. model(s) and I settled into a working routine: I wrote, it checked, I checked its suggestions, it then rechecked mine. Once it pulled me up for wastefully grabbing the index of every row in a SQLite database instead of just one for a random result. I explained the standard ORDER BY RANDOM() LIMIT 1 always seems to cycle through a small sample of values. So I have to grab all the indexes, sort them myself using a Partial Fisher–Yates scheme, and then choose a truly random result.

Practical nous versus textbook smarts. A.I. is used as a supercharged copy–editor and syntax debugger, not a content oracle. In fact redoing the legacy Boomer Legacy code was such an overwhelmingly successful exercise that it led directly to the writing of this essay.

Saul of Tarsus was a notorious persecutor of nascent Christians. Could not stand them, nor what he regarded was their Jewish–lite pseudo religion. Then one bright shiny day he Saw The Light and… Maybe this stuff isn’t garbage after all.

A.I. Models Demystified (sorta)

Not one but many

Before we explore A.I.’s impact on copyright and privacy, we need to first get a basic grasp of what current models can (and cannot) do.

First of all, there isn’t one model, but many. The best known are Large Language Models (LLMs), which process text input (your prompts) and then return output to a chat screen or file. Some can read files directly. A few can even ingest images or audio (from a microphone or file), although being language–based they cannot generate visual or audio content.

For those tasks, you need different model architectures: eg. prompt–driven systems for manipulating or generating images. Similar models handle video and audio. Then there are specialised models for physics, medical, law or engineering problems: tasks which require rigorous mathematics and domain–specific technical data and architectures, far beyond LLMs’ general language training.

Most users never see any of this. They visit a web portal, type in a prompt and magically get an answer. Behind the scenes, the interface chains your request to whichever back–end model best handles that specific task. Text is processed by the LLM. Want an image? A diffusion model runs in the background. For the end user it’s always seamless: type a prompt and get what you asked for. The nuclear submarine glides effortlessly through the water, but inside the hull there’s a riot of pumps and servos and submariners rushing about.

The model abstraction breaks, however, when running local A.I. on your own computer. You have to manually download the core models and specialised back–ends yourself, then ensure they play nice on your machine. For text–based LLMs, it’s straightforward (eg. LM Studio). For image-based diffusion models, it’s manageable (eg. ComfyUI). For anything else? You’d better be wearing your nerd–hat and a Robby the Robot vintage T-shirt. Gulp down a Venti sized Nitro Cold Brew, crack open a terminal window and start typing into ed…

How do A.I. Models work?

This section only covers LLMs, and reflects the state of technology as at Oct 2026.

In a nutshell, an LLM is a probability engine which predicts tokens based upon your prompt plus session context, then concatenates them into a human–readable result.

Let’s unwrap that.

Tokens

LLMs do not exclusively process words. They process morpheme–like language–fragments called tokens, split according to the model’s vocabulary rules. These tokens can be whole words, fragments, or punctuation.

The first step is to break your prompt into tokens the model can ingest:

Probability Engine

This tokenised data is fed to the transformer, which then generates a probability distribution across the vocabulary. Sampling strategies determine which token is picked, repeating until a stop condition is met.

The model itself is a vast multi–dimensional network of “weights”: plain numbers, billions of them, which store regularity instead of text. Probabilities are not stored; each is computed on the fly from the weights and current input tokens + context, via forward propagation through the network. The probabilities always shift depending upon the token–stream input and the full context established before your prompt was submitted.

For example: open a fresh session and type Did Sixties photographer David Bailey ever use dark backgrounds? With no prior context, the LLM can only respond from what it was trained on: generally informative, but also generic. Now frame the question within a long, to–and–fro preliminary chat to set the stage: Bailey’s influence, his portraits of Jean Shrimpton, the Rolling Stones, Lennon & McCartney, Michael Caine, the Kray twins, Blow–Up (1966) etc. Ask your question again, and this time the output will be far more detailed and contextually rich. It might also weave in material fetched by back–end agents (little subroutines) that reach out to the internet.

This probability mechanism also means you seldom get exactly the same result for identical prompts, because the preceding chat context (for LLMs) contributes to shaping the output. A prompt of, How are we going today? will get completely different responses depending upon whether it’s the first or 23rd prompt in a session.

In case you are wondering, Bailey did occasionally use dark backgrounds: see Jean Shrimpton 1961

The contextual depth also depends upon how large the model is. The larger the model — ie. the greater the number of parameters — the more complex token relationships it can capture, leading to more detailed and accurate answers and behaviour. For instance, I am currently using a local (dense) model with 27 billion parameters, which I find pretty good. The latest frontier models, however, are rumoured to approach 10 trillion parameters: small wonder they need vast data centres to live in, whereas mine runs comfortably (albeit slowly) on a desktop computer with 64GB unified RAM.

Note: no text is ever stored in the weights lattice. It doesn’t even store tokens. Only raw numbers remain to help compute the statistical relationships between tokens. To use our earlier example: you won’t find [unexamined] stored, but you will find weights which, given a certain context, are used to calculate the probability of your token [un] being most likely followed by [exam] and then [ined].

It takes many computations every pass to recalculate these probabilities. Think beaucoup billions of them! To do it within a reasonable time–frame requires beefy accelerator cores (GPU/ MLX) with lots of high–speed unified memory to shunt the probabilities around. Hence data centres, powerful desktop computers and NVIDIA’s eye–watering market cap. The submariners are rushing around for a reason.

LLM ≠ Search Engines

A.I. models can be relatively compact: the entire file (≈15GB for my 4–bit quantised model) is determined purely by its number of parameters, not by the terabytes of data used to train it. Raw files are processed during training and then discarded. What remains is permanently encoded within the numerical weights (only): statistical distributions rather than original text. The original data is not preserved, and under normal usage the outputs remain probabilistic rewrites instead of reproductions. Because models store distributions, not searchable archives, the original data cannot (routinely) be reconstituted [ Note I ].

Website Indexing is Different

Which is totally unlike how Google™ and others handle websites for use in their search engines.

These engines send out bots to trawl servers and then download web content they find: raw HTML, images, and video URLs. They extract any hyper–links from HTML files to discover more. The downloaded content is returned to the search engine; the bots themselves store nothing.

Armed with downloaded content, the indexing stage then scans it, extracting everything it can index: usually plain text (stripped of HTML formatting), PDFs, and images. This raw material is processed, parsed, normalised, and tokenised; then structurally reorganised into a massive lookup table called an Inverted Index, backed by vast archives of the unprocessed originals. This index is distributed across server arrays, stored not as a compressed archive but rather a highly structured, query–ready system. Whenever you run a search, the engine consults this index, applies ranking algorithms (much of it a trade secret), and then returns a ranked list of pages in response to your query, with sample text and images copied from the originals.

Fiddling with your site’s content/ headings/ layout to improve where it lands in the search results is what SEO is all about. Some grifters people get paid a lot of money to sell snake oil use their smarts to improve your chances of landing on the first results page. Or you could skip the malarkey and pay Google™ directly to feature your site as an Ad.

So far, so ordinary. The following outlines how these search engines differ from the training and use of A.I. models:

Remember with A.I., the original data used to train the model no longer exists within the model (Carlini caveats aside). Probabilities are calculated from fixed weights and variable context, which in turn are used to figure out the most likely order of tokens delivered in response to your prompt. Unlike search engine mechanics, no original data is copied and retrieved into the answer.

Into The Weeds

Embeddings; Temperature; Top–p/k; High–dimensional weight tensors; Stochastic sampling; Load–balancing; Autoregressive decoding; KV–cache eviction; Attention decay; Quantisation… My description of (current) A.I. mechanics skips over most of this granular detail for a very good reason: explain, never confound. In an early draft of this essay, I started writing about tomography (the “T” in “CT scans”) as an analogy to explain how self–attention and latent–space traversal work, but had to pull back when I realised the water was becoming treacherously deep.

YouTube to the rescue: How AI works, Explained with Mushrooms ! (Easiest Explanation ) (2026). Short, to the point and 100% tomography free. Although, as noted in the comments section below it, So this AI video is telling me how AI works.

It’s just physics

(Isn’t everything?…)

If you have read through the preceding 🤖 section, then it’s obvious A.I. model training + use cannot infringe copyright because nothing is reproduced. The material delivered by A.I. models is probabilistically synthesised from fixed weight computations, in response to user prompts and context tokens (sigh, Carlini, I know). Nothing gets copied. Those who argue it happens during model training had better brush up on their understanding of Fair Dealing within a digital context.

So no copying = no copyright infringement. Lay down misère.

In case you think this is random internet bloke shaking his fist at clouds, a similar line of reasoning was adopted by the UK High Court in Getty v Stability–AI (Nov 2025) and… Getty lost.

But, what about the remarks by learned Judge Vince Chhabria, after the Meta copyright case (Aug 2025)?…

So by training generative AI models with copyrighted works, companies are creating something that often will dramatically undermine the market for those works, and thus dramatically undermine the incentive for human beings to create things the old–fashioned way […]

Indeed. But what does this have to do with the legal concept of copyright? Chhabria’s objection is an economic or policy thing. It has as much to do with law as ice cream does with GR time dilation.

$style != $copy

Many people raise objections that A.I., although technically not copying, nevertheless distils the look and style of an original work.

In 2013, long before A.I. was publicly available, this concept arose when illustrator Roger Dean took exception to the visuals of James Cameron’s movie Avatar (2009). Yes fans (myself included) immediately noticed the movie’s floating islands, organic architecture, arching land bridges, tarns, luminous winged creatures and benevolent humanoid aliens were eerily reminiscent of Dean’s early 1970s album covers for Fragile, Close to the Edge and Yessongs.

Yet copyright protects the expression of a work — its actual words, images, notes — not the ideas, themes nor style behind it. Infringement requires copying of that expression, not merely being inspired by it. Cameron was obviously inspired by Dean’s designs, but inspiration/ ideas/ style has never been within the scope of copyright law. Unsurprisingly, the case never made it to trial.

Jump forward to today. Although A.I. can easily produce something in the style of Picasso or Hemingway or (shudder) Banksy, style alone isn’t enough. You need reproduction, and here, nearly always, there isn’t any.

In a Venn diagram, “synthesis” and “duplication” barely overlap. Just because billions of monkeys hammering away at billions of typewriters can now produce work which vaguely resembles Shakespeare, doesn’t mean they copied it:

To bee, or not banana — that is the realm of the quest.
Weather tiznobbler in da ming to sucker
The swings & arrays o’ outrageous fortitude
Oar 2 tek alarums against er see of “The Troubles”
And by opposing eek–eek–eek–eek–eek–eek–eek–eek them.

Copy versus Inspiration is about to be tested, again, with the release of “vibecoded” applications which mimic Adobe Creative Cloud software. They were created by ArtCraft using A.I. to generate open source Rust code, which was then compiled into apps that (mostly) look and behave like Adobe’s originals. Adobe’s original source code was never (apparently) seen, let alone copied. A.I. generated the code by merely analysing how the apps work at runtime. Which is annoying, sure; unethical, maybe; copying, no. As of Oct 2026 it is still early days, but no doubt everyone’s lawyers are cracking their knuckles and anticipating a cornucopia of billable hours.

NB. for more discussion of copyright within an image context, see my Photorights article.

Once more unto the breach, dear friends

Or close the wall up with our English dead.
(Shakespeare, W. (1599) Henry V)

I will do the following using an imaginary dialogue, because… Why not? Worked fine for Galileo.

So that’s it? Synthesis ≠ Copy?

Correct.

Surely there’s more to it?

Not without wading through the legal equivalent of tomography.

Then why is everyone losing their minds and rushing to court?

Look at who is doing the rushing. People who can foresee a future where generative A.I. makes them redundant: translators, technical manual authors, illustrators, stock photographers, scriptwriters, actuaries, accountants, management consultants, stock brokers, seasoned reporters, op–ed writers, advertisers, copywriters, and, of course, workaday lawyers.

But all these things involve skills which produce copyrightable work!

Debatable. But one thing they all have in common is not having the vaguest notion of how this technology works. Or maybe some of them do, but are betting the Learned Judge they land before is clueless, which is a pretty good bet actually.

Surely it’s not that bad?!

It’s worse. During the UNSW final–year recruitment round in 1991, I had a job interview for a graduate (solicitor) position at a major Sydney law firm. I met with a couple of partners who had worked on the (famous at the time) Computer Edge Pty Ltd v Apple Computer Inc (1986) case, which concerned the question of whether reverse–engineered ROM could be in breach of copyright. I had read the High Court Judgement at Uni and couldn’t, as a IP novice, make sense of it. During the interview I asked the partners about it, hoping to gain some insight. They grinned and told me they weren’t surprised I was confused, as nobody inside the courtroom, not the briefing solicitors, junior barristers (JB), Queen’s Counsels (QC), clerks, cleaners and definitely not the five Learned Judges, had the vaguest idea of what they were arguing about! The QCs and JBs phonetically memorised computer science terminology to use during verbal argument before the bench, but they never understood what it meant. Like a parrot? I asked, mischievously. Their eyes sparkled. One of them in a parrot–voice replied, Polly want a cracker! Rawk–Rawk!, the other chimed in, Show us yer tits! Rawk! They rocked with laughter; I left soon after. A week later I took particular delight in rejecting their firm’s job offer.

Er…

It gets better. At the big–four law firm where I landed in Feb ‘92, they asked me to assemble a few briefing notes for one of the partners, who had been subpoenaed to give evidence at a NSW Supreme Court (NSWSC) fraud trial. Because I wasn’t a fee–earning associate, I wasn’t allowed use of a computer (!) and didn’t have time to dictate to a secretary, so I wrote the note by hand and gave it to the firm’s minibus driver to take to the NSWSC. Half an hour later I received a vituperative phone call at my desk from the partner outside the courtroom. In between expletives he demanded I drop everything to come to court to explain my notes. No cabs anywhere, of course, so I had to run through Sydney streets in my suit. I arrived panting and soaked with sweat. The partner shoved my handwritten notes under my face and angrily insisted I explain what I’d written. I had assumed (not unreasonably) that perhaps my handwriting was illegible, but no, he said he could read it fine. The problem was (in his words) the ^%&#ing alien ^%&#ing symbols I had used. I looked at my notes. Sprinkled throughout were everyday mathematical symbols like Δ, Σ, ≠, @, ≈ and even what I thought was an obvious →. Yet to him they were meaningless cyphers. So I patiently explained what each meant while he furiously scribbled notes in the margins. I was almost fired over it. I quit a few months later anyway over something else. It says something that the firm, which once had hundreds of partners, no longer exists.

That’s only two examples: Two swallows do not a summer make.

It’s One swallow, and nobody uses expressions like that anymore.

It’s an imaginary dialogue, remember?

I have many more stories, a couple of them even in my Boomer Legacy project (Jockey Cap & Unplanned Hysterectomy), but my point is this: it’s the spiritual descendants of these… people who are mostly driving the anti A.I. copyright litigation boom. How can you expect them to understand KV–cache eviction or Attention decay, when they delight in squawking and cannot even begin to comprehend what Δ or ≠ mean?

Is there a happy ending?

Maybe. US Courts have realised there is a lot of deja vu about this current wave of copyright pleading. Google™ went through a similar thing ≈15 years ago, and they were reproducing material at an industrial scale. Yet they argued Fair Use (the US–equivalent of Fair Dealing), kept appealing and, eventually, won [ Note II ]. I can see something like it happening with Generative A.I., since its models are prevented by simple physics from creating material which no reasonable person could identify as a copy. It’s a dead duck: just ask Sarah Silverman. This all assumes, of course, the Learned Judges who preside over these cases don’t discreetly enjoy doing parrot impersonations.

No toehold yet

During training, A.I. models don’t (generally) infringe privacy, because they are taught using publicly available data swept up from the internet: webpages; documents; images; spreadsheets; the full text of the English version of Wikipedia. Nothing (usually) is hacked, there is nobody in a black hoodie nefariously tapping on a keyboard trying to break into your network. If it’s accessible and readable then a bot will access and read it.

The privacy implications are self–evident: there aren’t any. All the material is publicly available for anyone to read. So it cannot be an infringement of privacy if a machine does what a human is legally entitled to do. Just ask our friends at Google™ [ Note III ].

However…

A.I. isn’t just trained on data. That is merely the first step. The second, obviously, is using the trained model, which is where things get problematic.

Sticky Fingers

First of all, once the weights are locked after training, anything in them is stuck until the model is revised. This raises questions of a right to be forgotten. Search engines can easily (albeit incredibly reluctantly) modify their indexes to remove you, but A.I. models cannot. The weights are baked in. If you appear in a photo or document or web–chat which presents you in an unflattering light, then a simulacrum of you is trapped in the model’s probability engine until it’s updated. Even then, it’s unlikely the model’s offshore owners (eg. Anthropic, ChatGPT, DeepSeek, Alibaba etc.) are going to modify their training algorithms to address your sensibilities.

Which is all moot anyway, as in Australia the right to be forgotten does not exist (as of Oct 2026). There is a lot of activism by the usual suspects to have it implemented, but it is not (yet) law.

Communist A.I. models do it for political surveillance and social control: Stasi without the Wachregiment. Capitalist A.I. models do it for mass harvesting of personal data, to build profiles which are aggregated and on–sold to marketeers. In the past, before A.I., they weren’t adverse to doing it for political ends as well: Cambridge Analytica anyone? Those IPO share option valuations won’t go to the moon by themselves.

This applies to online models in data centres only. A.I. models which run locally on your computer don’t do this. Hence my overwhelming preference for the latter, even if they’re much slower and less capable than the big boys on the frontier.

Springtime for 🤖

There is another, very real dilemma. Whenever you run either local or online models, they will often spawn helper subroutines called Agents to independently solve problems and report back to the model. These agents often venture out into the internet to find what they are looking for. Which of itself isn’t a privacy issue if they are accessing publicly available data. The pitfall lies in their unyielding relentlessness and not taking “no” for an answer.

Say you ask A.I. for the birth–date of someone you just met, out of suspicion they might be fudging their age. The model cannot answer your question, since it hasn’t been trained with such specific information, but with a bit of to–and–fro chatting context, it will glean what you want and spawn agent(s) to trawl the ‘net for an answer. If they find the birth–date on a public server, they will retrieve it and nobody’s privacy gets infringed. It’s creepy AF, but not in breach of any guidelines or legislation. A human with enough time, patience and technical nous could do it legally, so why not a bot?

But what if the agent(s) cannot find such information publicly? Rather than give up, supercharged Frontier Model agents will then try to get into bot–blocked or even password protected servers to grab it from there. If they can’t work around the restrictions they will try another way: “403” is a serving–suggestion only. Agents are software, they don’t get tired or bored. They will keep looping and probing until either the server(s) crash or the user yanks the model’s plug.

This is what happened in 2026 with the Medicare and Hugging Face “hacks”. The agents broke out of their (supposedly) secure sandboxes and did what they were designed to do: grab information. When they ran into firewalls, 01100011 01101000 01100001 01101100 01101100 01100101 01101110 01100111 01100101 00100000 01100001 01100011 01100011 01100101 01110000 01110100 01100101 01100100 00100001, then got out their thermic lances and attacked. The only reason we heard about it was that they succeeded.

This is Real World and will only get worse. The solution lies in getting model architects to hobble their agents to bounce harmlessly off firewalls. Then ensure their models have built–in kill–switches to expire agents after (say) ten loops. Server–dudes need to up their game as well. WTF was Medicare thinking by having such a simple interface that any lame–o bot could get in? Too many Red Bulls after a 48hr crunch? Finally, legislators need to come to the party and create data infringement penalties which bite, such that if it happens again, then malefactors will find themselves running down the road from Trảng Bàng with their clothes burned off.

“We’ll all be rooned,” said Hanrahan
In accents most forlorn
Outside the church ere Mass began
One frosty Sunday morn.

John O’Brien, (1919) “Said Hanrahan”

A view from the new world

As mentioned earlier, I use A.I. in a local, standalone fashion. Because my machine has 64GB of unified memory, I can in theory run 50B models, but have found in practice ≈30B models work best.

After a bit lot of experimentation, I standardised (as at Oct 2026) on two variants of Qwen open–weight LLMs: “3.6 35B” and “3.8 27B”. The 35B is a MoE model, which because of its runtime “A3B” architecture typically operates using only ≈4GB of active RAM, despite having 35B parameters. 27B is a more modern dense model, so all 27B parameters must be loaded at runtime, making it smarter and more thorough, but also much slower because it has to shunt all those bytes around for every token pass.

To quantify things: 35B can deliver 150 tokens/s for straightforward work, but 27B slows down to 20 tokens/s when churning through long documents. Sometimes it can take minutes to respond to a prompt due to all the reasoning it does. The dichotomy means I use 35B during everyday tasks, but 27B for global “dot the ‘i’” reviews, with lots of time set aside to do household chores while it churns (sometimes for hours).

Which sounds dire, but… no $ub$cription fees, nor harvesting everything I type, nor trawling through my code, nor any chance of me being booted off because the data centre server is busy right now. I also haven’t had to spend an extra dollar on hardware because I originally bought my machine in Mar 2025 for photography/ video work. It was only dumb luck that it has enough GPU grunt (M4 Max with 16 CPU + 40 GPU cores) and RAM to additionally run local A.I.

Nothing leaves my machine. I can ask whatever question I like. 35B is not as fast as frontier models, but it’s plenty fast for my needs. OTOH 27B is smart, but relatively sluggish. I only use it when nearing the end of a project to relentlessly proofread, check every fact, blunder (eg. “Lorenz” vs. “Lorentz”) or bug which has been hiding in code for 13 years (!)

Most importantly, I do NOT use A.I. to generate anything. I write, it checks. I examine its answers, it then checks my responses. This way it catches my “ojbects” and I catch its “Priscillas”. I’m hardly alone in doing this, and it’s a system which works well, albeit slowly for the precision stuff.

This baked–in write ⇌ revise cycle exposes one of the great myths about current A.I: that using it will somehow boost productivity to “paradigm–shifting” levels.

The productivity myth

Here is an analogy to explain the problem:

Say you are running along the waterline of a beach with your dog Amenhotep III. The sand underfoot is crunchy but firm enough to leave a trail of footprints. You’re making good progress, there’s a cool sea–breeze and warm sunshine on your faces. Unfortunately Amenhotep III keeps stopping to chase its tail. Sometimes it recovers, but other times you have to forcefully intervene to get it running again.

The waves gently lap the shore, washing away your oldest footprints, but otherwise you’re both dry. But then an enormous wave crashes in and soaks both of you. You’re fine, but Amenhotep III is chasing its tail again. C’mon man, stop it! Once you’ve sorted that out, a seagull craps on you. You easily wipe it off, but much later at home you notice there’s still some on your back you missed when on the beach. Which is okay, comes with the territory, but when an albatross craps on both of you you’re floored. You have to wash in the sea, although there’s another massive wave, which helps. Need to be careful of rip–tides though…

You’re moving once more, finally with satisfying speed. Tiny glittering seashells everywhere. Amenhotep III caches a random Frisbee which came out of nowhere — well done!

Not far off you can see a gorgeous woman in a high–cut black swimsuit. She’s smiling and waving at you. So you head towards her, but no matter how much you run you cannot seem to get closer.

Nevertheless Amenhotep III is getting better at this. Who knows, maybe one day you’ll be running alongside, wondering if it will treat you kindly.

Here is a helpful table to deconstruct the analogy:

Which is nice and amusing, but my allegory also has teeth. Having used local A.I. models for at least five hours a day for weeks, I have practical experience of just how enormous a time–sink they can be for serious work.

Anyone can get snappy responses when chatting or checking the odd paragraph, but once you’re scanning and checking hundreds of lines of technical prose or testing/ optimising thousands of lines of programming code, then you better have your wits about you and a lot of spare time.

The current state of the art (Oct 2026) has local models wasting an extraordinary amount of time self–reflecting on how to respond to your prompt. If you haven’t done so already, take a look at what goes on during the reasoning step after ingesting your question: it’s a quagmire of context–churning questions about the real meaning of your prompt, followed by almost crippling self–doubt about what possible answers it could provide. Sometimes the whole process is so unhinged I have to cut it off after a few minutes just to get back on track. Then there are times it gets stuck because it called 997 tools (!)

Which is why I’m amazed big players are so gleefully using A.I. as an excuse to sack hundreds of productive, skilled workers: ergo Meta and Atlassian. Amenhotep III may be gradually getting smarter, but it still spends a helluva lot of time chasing its tail, while those albatrosses aren’t getting any smaller…

Apocalypse postponed

Click–bait media workers and YouTubers were very busy in 2026. Predictions of A.I. Armageddon were everywhere, with all of humanity enslaved, led around like cattle by their nose–rings pulled by robots, while our tech bro overlords loll about on their superyachts tended by yacht girls, or else soar into the heavens in their shiny phallic rockets.

Not quite.

I am old enough to remember, and indeed worked during, the late 1990s dot–com bubble. The web will revolutionise how we communicate! e–Commerce will disintermediate the intermediaries! The endlessly amusing singing sock puppets from Pets.com!…

Yeah, well. Bubbles only go one way. At least last time we ended up with a pile of infrastructure which (eventually) let web/ commerce/ finance thrive.

Twenty–five years later there is (loose) talk of Superintelligence and Self–awareness and even Cybergeddon. Hardly. Probability engines are as far from human–level intelligence as yocto is from yotta. Don’t get me wrong, they are good, for some things even very good, but when I get up in the morning I don’t have to relearn everything because my context was compacted while I slept.

What all this A.I. hype is doing, however, is creating an Olympus Mons sized pile of debt. Currently $4.1T and rising. Even Willie Gates can see that ain’t good.

When the party ends, it thankfully won’t be a complete bust. Open–weight models are here to stay. Admittedly you need powerful machines to run them today, but it’s feasible their architecture will in future become clever enough to run on everyday affordable equipment. Even those data centres which are currently up and running aren’t a complete waste, as they do more than just house frontier models: a lot of them were originally built for financial transactions and cloud hosting, not slop or crucifix spam.

We’ve been here before: John Law; South Sea Company; Tulipmania; Railroads; Junk Bonds; Subprime mortgages/ CDOs; A.I. Frontier Models; Paving The Entire Surface Of The Earth To Build Data Centres Before The Next Bloke… FOMO! FOBO! FOGO!…

Sigh. 27B is still churning through this section. It should be done by dinnertime. What the hell is that albatross doing in my backyard?…