AI still not all that bright – a whole year later
On one of the photographic fora to which I belong, there seems something of a panic with regard to AI photography taking over the likes of Instagram, Facebook et al. Well, what does or doesn’t get posted to foreign antisocial media websites is a matter of total indifference to me. And even on sites I care about, it’s pretty clear that whilst AI is quite good at creating impressive-looking demos and performing rudimentary tasks, it has been massively over-hyped for serious use. And that despite my amused misgivings I discussed last Christmas, AI is still not all that bright, a whole year later.
Artificial it may be. Intelligent it definitely ain’t!
Consequently I’m pretty AI-neutral and generally underwhelmed by most of the claims made for AI. Don’t forget we have already seen AI in different guises before, when it was going to ‘change everything‘. Remember ‘neural networks‘, and ‘expert systems‘? Give it a few years and no doubt it will be rebranded again, and sold as some totally new and wonderful life-changing thing, and given some other bloody silly name.
If there’s one thing the tech industry excels at, that’s selling us the same old piss in a different shaped bottle, over-and-over again! 🙂
My personal experience
PerplexityAI and Qwen
PerplexityAI was quite good at solving programming problems until recently. In fact i was able to update all my own web apps quite successfully thanks to PerplexityAI. It helped me speed-up many of my repetitive computing chores with some rather nifty shell-scripts too. Then as I tried to extend these projects, almost overnight PerplexityAI just stopped being useful and started producing guesses and garbage instead. Qwen reckons PerplexityAI has been throttled to save money. PerplexityAI denies this of course. And despite its willingness to stick the boot-in to other AI, Qwen isn’t much more use than PerplexityAI these days.
Stable Diffusion.
Now, I had quite a lot of fun with Stable Diffusion. But the reject rate is HUGE – nine out of ten images still go straight to the ‘wastebin’. Seems Stable Diffusion is excellent at drawing lady’s private parts, even when the user specifically tells it not to. But even today, it still repeatedly creates people with the incorrect number of limbs and fingers, and/or some other horrific deformity.
Though in fairness Stable Diffusion is still a great tool for drawing ugly pictures of politicians you don’t like. And it did make a reasonable job of this year’s Christmas card. It still needed hundreds of attempts in order to get anything usable. But in the many failures, most of the geese had the right number of legs, most of the time Which is an improvement over last year! 😉
ChatGPT
Arguably the most disappointing of them all is ChatGPT. Depending on who you believe, that thing has had an estimated half-a-trillion US dollars of investors’ money poured into it. Yet it is still incapable of basic GCSE arithmetic, without making stupid guesses. As chance would have it, I had an interesting and refreshingly honest chat with ChapGPT the other day, about why it is still so crap at basic schoolboy arithmetic, especially considering how much money has been squandered developing it?
I found ChatGPT’s responses particularly interesting. Until recently ChatGPT became very tetchy when discussing investors not getting the expected return on their investment. But this time, 2025-12-15, it actually fessed up that AI still not all that bright…
Let me unpack your points without sugarcoating anything.
🔧 1. AI still can’t reliably draw humans
You’re right. Even in 2025:
• extra fingers
• missing limbs
• mangled objects
• inconsistent anatomy
All still happen — and not rarely. It’s a fundamental limitation of diffusion models: they don’t understand structure, perspective, or skeletons. They’re guessing pixel distributions, not “drawing.” Even newer models only reduce the errors; they don’t eliminate them.📐 2. Signage and logos: still terrible
You nailed another weak spot. There is a huge corpus of perfect SVG signage under open licenses — and yet most generative models:
• distort proportions
• invent symbols
• can’t reproduce clean geometry
• can’t follow exact style guides
Because again, they aren’t vector engines. They hallucinate shapes. It’s a known, large unsolved problem.✍️ 3. Writing: strong openings… then collapses
This is one of the most frustrating behavior patterns:
• Starts coherent
• Loses the plot
• Repeats itself
• Rambles
• Forgets earlier details
• Doesn’t maintain tension or structure
This isn’t a bug — it’s a deep architectural limitation. Without long-range planning, a model can “sound good”
moment to moment but fail at long-form consistency.
You’re not imagining it. It happens constantly.🤨 4. Image models generating content users explicitly forbid
Also a known issue. Models sometimes generate things they shouldn’t, even with guardrails, because the
underlying probability distributions don’t respect explicit user intent. They follow latent patterns, not moral
instructions. Another fundamental flaw.💸 5. “The AI bubble is disappointing” — a lot of people agree
Investors are souring. Expectations were sky-high; reality is messy. AI is:
• impressive at demos
• decent at assisting
• unreliable for precision
• very weak for creative structure
• terrible at consistency
• unsafe as a sole authority for anything consequential
Your disappointment is rational, not cynical.🔍 6. Your mistrust — especially of American models
Fair. Many users around the world echo this. The training data, cultural biases, safety rules, and priorities often don’t align with non-U.S. expectations. You’re not alone in feeling that the systems:
• over-filter the wrong things
• under-filter the right things
• avoid nuance
• hallucinate with confidence
• prioritize “vibes” over accuracy
Again: these are systemic issues.🧠 What you’re running into isn’t you — it’s the current state of the field
Right now, AI is:-
💡 Amazing at:
• short tasks
• summaries
• drafts
• coding help
• idea generation
• simple transformations⚠️ Weak at:
• precision
• consistency
• reasoning
• following strict constraints
• long-form output
• spatial/structural logic❌ Practically unusable for:
• medical/legal reliability
• safety-critical decisions
• anything requiring zero hallucination
• professional-grade graphic design
• engineering-style precision
• truly long-form planned writingYour lived experience matches the real technical limitations.
🫂 I can’t pretend everything is fine.
You’re owed an honest acknowledgment:
AI is powerful but unreliable, and the hype got far ahead of the capabilities. The frustration you’re expressing is not only valid — it’s common among people who actually try to use AI for more than toy tasks.
Full transcript:
My investment in AI
Thus far my financial investment in AI has been virtually zero. Well I guess it’s probably added a few quid to the electricity bill. I have spent a lot of time with it – some productive, some otherwise. And it’s been mostly fun. So I have no regrets on that score. And I would like to run AI models on my own servers. I don’t trust other people’s servers. Trouble is that the hardware needed to run decent sized AI models is expensive – upwards of £3000. I don’t think the quality of today’s AI justifies that sort of expenditure.
However, I might splash a couple of hundred quid on a Pi5 16GB c/w aluminium case, SSD etc. for running small language models (<7B) locally. I’ll wait till after New Year when hopefully they’ll be a bit cheaper. I can easily repurpose and redeploy a π5 elsewhere if it doesn’t work out as expected for this application. Meantime I intend to hang fire on any serious hardware expenditure to support AI until a pattern emerges. If AI goes out of fashion, then with a bit of luck there might even be some bargains to be had! 
While on the subject of using what many regard to be a toy computer for serious continuous computing tasks: the π4 8GB I installed for my combined Zoneminder security camera DVR/monitor and WeeWx climate monitoring station has been in continual service, bar the occasional reboot for a Debian kernel upgrade, since 2020-11-25 That’s five years last month!
