{"id":35501,"date":"2025-12-24T00:53:24","date_gmt":"2025-12-24T00:53:24","guid":{"rendered":"https:\/\/garfnet.org.uk\/cms\/?p=35501"},"modified":"2026-01-02T04:08:53","modified_gmt":"2026-01-02T04:08:53","slug":"ai-still-not-all-that-bright-a-whole-year-later","status":"publish","type":"post","link":"https:\/\/garfnet.org.uk\/cms\/2025\/12\/24\/ai-still-not-all-that-bright-a-whole-year-later\/","title":{"rendered":"AI still not all that bright – a whole year later"},"content":{"rendered":"
 <\/div>\n

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 some rudimentary tasks, it has been massively over-hyped for serious<\/em> use. And that despite my amused misgivings I discussed last Christmas<\/a>, AI is still not all that bright,<\/em> a whole year later.<\/strong><\/p>\n

Artificial it may be. Intelligent it definitely ain’t<\/span>!<\/h2>\n

Consequently I’m pretty AI-neutral<\/em> and mostly underwhelmed<\/em>  by the silly claims made for AI. Don’t forget we have already seen AI in different guises before, when it was going to ‘change everything<\/em>‘. Remember ‘neural networks<\/em>‘, and ‘expert systems<\/em>‘? Give it a few years and no doubt it will be rebranded again, and sold as some totally new and wonderful life-changing<\/em> thing, and given some other bloody silly name.<\/p>\n

If there’s one thing the tech industry excels at, that’s selling us the same old piss in a different shaped bottle<\/em>, over-and-over again! \ud83d\ude42<\/p>\n

My personal experience<\/h2>\n

PerplexityAI and Qwen<\/h3>\n

PerplexityAI was quite good at solving programming problems until recently. In fact I was able to update all my own web apps quite successfully<\/a> 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<\/em> to other AI, Qwen isn’t much more use than PerplexityAI these days. <\/p>\n

Stable Diffusion.<\/h3>\n

Now, I had quite a lot of fun using Stable Diffusion<\/a> to produce images. But the reject rate is HUGE – nine out of ten pictures still go straight to the ‘wastebin’. Seems Stable Diffusion is excellent at drawing lady’s private parts, even<\/em> when the user specifically tells it not<\/em> to. But even today, it still repeatedly creates people with the incorrect number of limbs and fingers, and\/or some other horrific deformity.<\/p>\n

\"Cute<\/a>
Cute little robot created in Stable Diffusion.<\/figcaption><\/figure>\n

In fairness, Stable Diffusion is still a great tool for drawing ugly pictures of politicians you don’t like. I also like its cute little robots. And it did make a reasonable job of this year’s Christmas card<\/a>. It still needed hundreds of attempts in order to produce anything usable. But in its many failures, most of the geese had the right number of legs, most of the time. Which is an improvement over last year<\/a>! \ud83d\ude09<\/p>\n

ChatGPT<\/h3>\n

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 ChatGPT the other day, about why AI in general, and ChatGPT in particular is still so crap?<\/p>\n

I found ChatGPT’s responses rather intriguing. You see, 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 is actually still<\/em> not all that bright…<\/p>\n

\n

Let me unpack your points without sugarcoating anything.<\/h4>\n

\ud83d\udd27 1. AI still can\u2019t reliably draw humans<\/h5>\n

You\u2019re right. Even in 2025:
\n\u2022 extra fingers
\n\u2022 missing limbs
\n\u2022 mangled objects
\n\u2022 inconsistent anatomy
\nAll still happen \u2014 and not rarely. It\u2019s a fundamental limitation of diffusion models: they don\u2019t understand structure, perspective, or skeletons. They\u2019re guessing pixel distributions, not \u201cdrawing.\u201d Even newer models only reduce the errors; they don\u2019t eliminate them.<\/p>\n

\ud83d\udcd0 2. Signage and logos: still terrible<\/h5>\n

You nailed another weak spot. There is a huge corpus of perfect SVG signage under open licenses \u2014 and yet most generative models:
\n\u2022 distort proportions
\n\u2022 invent symbols
\n\u2022 can\u2019t reproduce clean geometry
\n\u2022 can\u2019t follow exact style guides
\nBecause again, they aren\u2019t vector engines. They hallucinate shapes. It\u2019s a known, large unsolved problem.<\/p>\n

\u270d\ufe0f 3. Writing: strong openings\u2026 then collapses<\/h5>\n

This is one of the most frustrating behavior patterns:
\n\u2022 Starts coherent
\n\u2022 Loses the plot
\n\u2022 Repeats itself
\n\u2022 Rambles
\n\u2022 Forgets earlier details
\n\u2022 Doesn\u2019t maintain tension or structure
\nThis isn\u2019t a bug \u2014 it\u2019s a deep architectural limitation. Without long-range planning, a model can \u201csound good\u201d
\nmoment to moment but fail at long-form consistency.
\nYou\u2019re not imagining it. It happens constantly.<\/p>\n

\ud83e\udd28 4. Image models generating content users explicitly forbid<\/h5>\n

Also a known issue. Models sometimes generate things they shouldn\u2019t, even with guardrails, because the
\nunderlying probability distributions don’t respect explicit user intent. They follow latent patterns, not moral
\ninstructions. Another fundamental flaw.<\/p>\n

\ud83d\udcb8 5. \u201cThe AI bubble is disappointing\u201d \u2014 a lot of people agree<\/h4>\n

Investors are souring. Expectations were sky-high; reality is messy. AI is:
\n\u2022 impressive at demos
\n\u2022 decent at assisting
\n\u2022 unreliable for precision
\n\u2022 very weak for creative structure
\n\u2022 terrible at consistency
\n\u2022 unsafe as a sole authority for anything consequential
\nYour disappointment is rational, not cynical.<\/p>\n

\ud83d\udd0d 6. Your mistrust \u2014 especially of American models<\/h4>\n

Fair. Many users around the world echo this. The training data, cultural biases, safety rules, and priorities often don\u2019t align with non-U.S. expectations. You\u2019re not alone in feeling that the systems:
\n\u2022 over-filter the wrong things
\n\u2022 under-filter the right things
\n\u2022 avoid nuance
\n\u2022 hallucinate with confidence
\n\u2022 prioritize \u201cvibes\u201d over accuracy
\nAgain: these are systemic issues.<\/p>\n

\ud83e\udde0 What you\u2019re running into isn\u2019t you \u2014 it\u2019s the current state of the field<\/h4>\n

Right now, AI is:-<\/p>\n

\ud83d\udca1 Amazing at:<\/h5>\n

\u2022 short tasks
\n\u2022 summaries
\n\u2022 drafts
\n\u2022 coding help
\n\u2022 idea generation
\n\u2022 simple transformations<\/p>\n

\u26a0\ufe0f Weak at:<\/h5>\n

\u2022 precision
\n\u2022 consistency
\n\u2022 reasoning
\n\u2022 following strict constraints
\n\u2022 long-form output
\n\u2022 spatial\/structural logic<\/p>\n

\u274c Practically unusable for:<\/h5>\n

\u2022 medical\/legal reliability
\n\u2022 safety-critical decisions
\n\u2022 anything requiring zero hallucination
\n\u2022 professional-grade graphic design
\n\u2022 engineering-style precision
\n\u2022 truly long-form planned writing<\/p>\n

Your lived experience matches the real technical limitations.<\/p>\n

\ud83e\udec2 I can\u2019t pretend everything is fine.<\/h4>\n

You\u2019re owed an honest acknowledgment:
\nAI is powerful but unreliable, and the hype got far ahead of the capabilities. The frustration you\u2019re expressing is not only valid \u2014 it\u2019s common among people who actually try to use AI for more than toy tasks.<\/strong><\/p><\/blockquote>\n

Full transcript:<\/h3>\n