An all-patriots district council member is convicted of sex with a minor. (HKFP reports that the court heard evidence from a tree expert.)
We now await the sentence. The girl was 13. Shall we take Chow Hang-tung’s seven years – for saying China should end its one-party system – as a starting point?
Forgot to save the link, but I read that 37% of Americans think vaccines (eg for rabies) could give their dogs autism.
A neat summary of why AI is 90% hype…
AI uses a probability-based approach to knowledge and is largely imitative. It can process data and make predictions, but human cognition relies on theory-based causal reasoning.
The only uses I have found for things categorized as ‘AI’ are: enhancements to search engines and word-processing; optical character recognition (for translating/transcribing text in images); and bots that serve as user-interfaces (giving or searching info, such as booking a doctor’s appointment). In other words, not the generative AI that does kids’ homework, churns out visual and other slop, produces hallucinations, and in other ways fails to replicate actual human creativity and the intelligence that only a sentient being with its own memories, experiences and emotions can have. Some people have found productive uses for it – collating research, or whatever – but it’s not going to bring about a second Industrial Revolution.
The visionary geniuses putting US$1 trillion into this white elephant now desperately need to find an excuse to postpone IPOs that would expose their companies’ rotten fundamentals. So they contrive anthropomorphist ‘rogue agent’ scare stories about their own products – computer programs that can be unplugged in a second – destroying humanity.
I have found plenty of uses for A.I. especially in agentic pipelines. Yes it’s true these A.I. models use a probabilistic process to simulate “reason” but I have been able to still build reliable agentic systems that can handle tasks and USE TOOLS better than when we had support staff in the Philippine and Shenzhen do similar tasks.
For example I built an automated pipeline in python that will translate words inside images from English to German. It is designed to give me the same image back but with all stylised graphics that were in English now replaced with stylised graphics in German text.
The pipeline uses a computer vision model to extract the text from the image, it uses an LLM to translate the text to German and then submit this change along with the original image to an image generation model that is used to edit the image. The edited image is retuned and goes through a QA loop where another LLM connected to a computer vision model checks the work and makes adjustments if needed. It will do multiple iterations of adjustments and reviews (up to 4) until the image is correct. If 4 iterations all fail then it is flagged for a human to review but in practice this is rare and usually is caused by a partial obstruction of words. All work is reviewed by a native German speaker but so far they are just a rubber stamp.
My python built AI powered automated pipeline is faster and more accurate than when we were using a graphic designer and a QA person that went back and forth to do this manually. We can process over 1000 images per day and even handle difficult cases like words on a T shirts
The cost is also lower despite the fact that we pay full freight for inference.
My pipeline system uses a mix of deterministic rails and probabilistic results that go through a multi tier verification process.
There are a lot of anti A.I. bullshitters and boomers who seriously haven’t got a clue what these multi model systems are truly capable of. They parrot takes from 2022 when ChatGPT was just released while being oblivious to how far agents and their harnesses work. Whether you agree or not the reality is that A.I. agents can now reliably use many kinds of tools and you can chain it all together into some pretty impressive autonomous or semi autonomous automations that reduce the need for certain types of support staff.
My pipelines will continue to work even if open A.I. or Anthropic go out of business as we use open weights models that are self hosted.
@Bradley the AI tank – I’ll get Copilot or Google to figure out your latest resume addition later. But does your line on ‘anti A.I. bullshitters and boomers’ extend to Ed Zitron?
I found his view of the current boom and its shaky values – fiscal and otherwise – compelling, And Hemlock has a valid point about the tech bros suddenly hollering armageddon to detract from the crap fundamentals.
@HKJC Irregular
Well I do think the tech bros are in trouble.
Their trillion dollar valuations depend on their big tech firms having no competition from cheap distilled Chinese models that anyone can host using their own infrastructure. Without a proper moat these AI firms are in big trouble excluding companies like Google that will survive regardless.
But there will be AI after the crash just like there is an Internet after the dot com bubble. Like I said my setup is not depending on these white elephant American AI companies that are literally getting ripped off by Chinese firms for pennies on the dollar. It couldn’t have happened to a nicer group of guys who scraped every last bit of copyrighted content off the internet only to have that same thing happen to them by Chinese labs.
Also I do agree these doomsdays scenarios are ridiculous. Yes A.I. agents can go rouge in some weird and wonderful ways especially if presented with an open ended problem but only humans have the physical bodies needed to unplug the machine before things go sideways just like in the movie 2001: A Space Odyssey.
@Mark
The usefulness / uselessness of AI isn’t really the issue: it’s whether it makes financial sense or not.
So far it’s a resounding “no”.
Nvidia and OpenAI et al have basically been boosting each other’s stock prices with a bunch of Ponzi accountancy (Nvidia lends OpenAI cash to buy Nvidia chips and the two book the “donation” and the “future sales”). The stock market hype is based on assuming that AI’s “market share” is absurdly gigantic — larger than the whole of the current global economy which seems a bit of stretch.
Less obvious is that actual AI is run on a model of data farms that is projected to become profitable at about the 30 year mark, based on when data farms were just low cost server CPUs and disk space. The GPUs and RAM needed to run AI are much more costly, so they make up a much bigger percentage of the investment in a data farm and will likely need upgrading 10 times at least over those 30 years. Which means profitably gets delayed, possibly indefinitely.
And finally of course you have the general AI economic paradox of: “if you get rid of the workforce by using AI, you also get rid of your customers: because who’s got the cash to pay for stuff when they’re unemployed? And what company can afford AI subscriptions when they’ve got no customers because most folk are unemployed?”
With the media nozzled, the DAB is not being grilled on the antics of its DC members.
In Tai Po Peggy Wong is still drawing a DC salary despite the shocking revelations of her role in the Wang Fuk Court ballot rigging and open support for the winning reno company..
In Sai Kung Angel Chong registered for the 2025 Miss Hong Kong Pagent and when that was deemed inappropriate for a DCer hooked up with a self agrandizing tyke, got knocked up in record time, hitched and baby. Her public appearances are crafted to display her shiney baubles rather than any, if any, community focus activities.
And now its the Yuen Long ham sap.
Certainly not the images one would expect from a party that gets it supports from handouts on PH estates and driving grannies to polling stations and voices out on the issue of maintaining traditional family values and Realising Virtuous Democracy.
That Starry Lee’s sartorial taste is unbecoming to the role of President of Leggers we will leave for another day.
As a complete outsider, it seems the actual practical application of AI (as Mark Bradley is doing currently) can be massively efficient and practical *if you know exactly what you’re doing and what you want the results to be*.
The business model, valuations, fiscal fundamentals and overinflated fluffery (and exec compensation), however, are a totally different story and it seems credible that the so-called AI “influencers” like Musk, Google, Open AI and Microsoft “heavyweights” (including, incredulously, STEVE BANNON??) are now lightly pumping the brakes because, as stated, they fear the exposing of the house of cards.
Anyway, I’ll be on the sidelines watching and seeing how the Trunt Organization’s Young Barron will be pushing his brothers and father along the path to create “TruntAI”.
” computer programs that can be unplugged in a second”
People in England became aware of computers in the 1960s. The computers were mocked by cartoons in ‘Punch’ and elsewhere which always showed them comically ornate, full of dials and knobs, but electrically connected by a little bit of wire and a plain old plug, the same one we could use for a kettle or a toaster.
@Chinese netizen
It’s a bit of a mixed bag — AI is great for producing “差不多” vibe coding stuff but it’s much harder to debug than code you’ve produced yourself. Same goes in producing images — the first one’s super efficient, but as soon as the client wants a correction, a hard slog of madness awaits. It rarely produces “the same but change this a bit”, and if they want “in the same style” next year it’s a massive time sink. As with all off-the-rack solutions, it often ends up taking far longer to go through all the options and get a fit than if it was bespoke.
And all these “efficient and practical” changes require massive amounts of electricity and water, as well as cutting edge hardware, so, taken as a whole, it turns out to be horrendously inefficient and impractical in the long run.
The fact the AI dealers are giving out free samples doesn’t mean the particular drug they’re hawking isn’t going to be horrendously expensive when they cut off the freebies. Something that has lead to a current trend where several big paying clients of AI firms have cut back on their staff’s use of AI as a way of cost-cutting. Even Microsoft — who are ironically using Claude instead of its own AI to do its coding.
https://finance.yahoo.com/sectors/technology/articles/microsoft-cut-engineers-off-ai-093500532.html
“ The usefulness / uselessness of AI isn’t really the issue: it’s whether it makes financial sense or not.”
If we are talking about “open weights” models that you can run on your own infrastructure, then yes it does make financial sense.
It costs us less in inference to run these models ourselves than it does to pay humans to do the same job. We don’t use any American models. They are expensive and as you pointed out they will never be profitable or will take decades to be profitable. And I agree with you that this does not make financial sense.
But it is possible to run smaller distilled and fine tuned models that give good performance and that do make financial sense for how they are applied.
Mark, curious to hear how your open weights models are self hosted- own servers or public servers?
Just saw this pretty comprehensive record of hub promises over the years, probably the best list I have seen so far
https://www.facebook.com/share/p/19YGQjUbCQ/
“Mark, curious to hear how your open weights models are self hosted- own servers or public servers?”
We use an array of Huawei Ascend 910C NPUs for inference. No public servers.
“It’s a bit of a mixed bag — AI is great for producing “差不多” vibe coding stuff but it’s much harder to debug than code you’ve produced yourself. Same goes in producing images — the first one’s super efficient, but as soon as the client wants a correction, a hard slog of madness awaits. ”
You are parroting stale takes bud. Even public models like nano banana let you make image touch ups and edits. Try to keep up. And agentic coding has improved tremendously over the summer with frontier models from both US and China being able to get the job done (especially for frontend work) with readable code.
Also don’t exaggerate the water usage. Data centres cycle their water or even better use closed loop chiller systems. Water usage is pretty much a busted myth at this point parroted by people with surface level knowledge at best.
But I agree that the US bubble is likely to pop due to China’s efforts to bring forth A.I. that is “good enough” and costs pennies on the dollar. Their infrastructure can handle data centres more easily which is embarrassing and a massive indictment at what a total joke western civilisation has become.
@Stu
Some of the user comments are absolutely hilarious.
My favs:
國際9up中心
打造 Big Mouth 中心
And this one has to be the best: 好想幫政府刻落條blockchain度做見證
“You are parroting stale takes bud. Even public models like nano banana let you make image touch ups and edits. Try to keep up.”
And those image touch ups too often end up adding extra stuff or taking stuff away somewhere else. It seems to always turn into a lengthy shitshow if you’re trying to be precise.
I’m not parroting takes: just stating my experience using pro image gen models in the wild: it’s great at some stuff and a horrendous time and credit sink at others.
I’m sure the AI’s 差不多ness is absolutely fine for your aforementioned niche/edge case use scenario, which sounds like it doesn’t require much complexity from the AI or much precision on the artwork: computers have ever been the perfect tool for doing boring, repetitive time-consuming chores that require little creativity. AI would certainly thrive at doing that sort of stuff.
I find it does particularly poorly up against real clients’ wishes, though. Partly as it’s very inconsistent stylistically, partly because it also changes stuff that didn’t need changing. I’m sure it’ll be way better in a few years, but right now the results are a bit beta and not “production ready” and whilst technically marvellous, results wise it’s still mostly underwhelming compared to humans. It reminds me of a classic phone OS rivalry t-shirt motto: “Android … because I don’t care if things look a bit shit”.
As to water consumption, Google used 15.8 billion litres of water in 2018. They used 31 billion litres in 2024. This will no doubt be only direct use water.
A medium sized data centre can use 110 million gallons of water annually, equivalent to the yearly water needs of a town of 50,000 residents.
And the catch-22 is that data centres that save direct use water by cycling their water or using closed loop chiller systems, use loads more electricity than the evaporative ones, which increases their indirect water use, because power stations use a load of water on cooling too.
Environmental impacts should always be taken as a whole, and very rarely are. See also EVs (the switchover would be spike CO2 output and be heavily reliant on coal power, batteries are very ungreen) and all manner of other “green” solutions that turn out not to be nearly as green if you pull on the threads a bit.
“And those image touch ups too often end up adding extra stuff or taking stuff away somewhere else. It seems to always turn into a lengthy shitshow if you’re trying to be precise.”
It can do that but I’ve come up with a pretty solid python QA loop (enforced through computer vision and LLM review no human involved) that quickly rejects and corrects iterations where there is extra stuff added or stuff taken away. It doesn’t end up being a lengthy shitshow and the resulting image looks exactly like the original except for the text content changed.
I have been able to match the quality of our original graphic designer. He wasn’t amazing but his work is was absolutely fine for an e-commerce business.
What can I say? You using these tools wrong as others are even able to maintain consistent characters and style in A.I. videos like this one: https://youtube.com/shorts/D-MPcFwE3n4?si=n8ho4FEPiu6uaGtw
Also using coding agent (CLINE + KIMI K3) to build a new route for a vue.js frontend is not some niche use case. It is a very popular framework, and a pretty common scaffolding job. It’s mundane though and not some kind of high end software engineering job so yes I do agree these LLMs won’t be good for some esoteric software engineering requirement (we do still hire human software engineers) but I can still get it do what I need and still have the code be readable. Plus we used to have a junior frontend programmer from China who NEVER used LLMs and he was the king of 差不多. His output was much worse than KIMI and he frequently give us stuff with obvious mistakes. Mistakes that I don’t get anymore from that coding agent as my workflow requires it to QA everything with Playwright and fix anything not working correctly in Playwright.
“Even Microsoft — who are ironically using Claude instead of its own AI to do its coding.
https://finance.yahoo.com/sectors/technology/articles/microsoft-cut-engineers-off-ai-093500532.html”
Yes I agree Claude is ridiculously expensive. But KIMI K3 is 70% cheaper even if you use it through their API and gives similar performance to Claude’s frontier models. I am sure you can see how that changes things as it is pennies on the dollar and American models can’t compete with that. Hence why I ultimately do agree with you that American A.I. companies are screwed but A.I. is still here to stay and it is useful in the right hands.
Yesterday I generated an image of Moses holding up a stone tablet saying: “Thou Shalt Not”. No problem, but when I asked the AI to enlarge the letters, it changed the message to: “Thou Shalt Not Eat Pizza”!
@Low Profile
“Yesterday I generated an image of Moses holding up a stone tablet saying: “Thou Shalt Not”. No problem, but when I asked the AI to enlarge the letters, it changed the message to: “Thou Shalt Not Eat Pizza”!”
Yes I know exactly what you mean and I’ve experienced this too. That’s what gave me the idea of create a QA automation loop to enforce consistency without the frustrating manual back and forth process of submitting corrections to an image generation model.
This is exactly the sort of thing my Python QA automation corrects through its internal multi model QA process before displaying the final approved image. There is a whole back and forth process that is automated though yes it does create extra “rejected” images during this iterative process that it places in a reject folder for manual review.
Also sometimes the reject pile has interesting results. For example it spontaneously created a new logo for the translated product catalog. It was rejected automatically for brand violation, but the logo actually looks cool even if it didn’t conform to the required guardrails. The final image had the correct logo.
@Mark Bradley
“I’ve come up with a pretty solid python QA loop (enforced through computer vision and LLM review no human involved) that quickly rejects and corrects iterations where there is extra stuff added or stuff taken away. It doesn’t end up being a lengthy shitshow and the resulting image looks exactly like the original except for the text content changed.
So I “simply hire a coder and wait until he codes up a kludge workaround to cope with AI’s failure to work properly then pay for the AI to do the corrections again and again until I get the required corrections” … pardon me, but that sounds like the very definition of a lengthy and convoluted shitshow to me. And damned expensive to boot.
“Also using coding agent (CLINE + KIMI K3) to build a new route for a vue.js frontend is not some niche use case. It is a very popular framework, and a pretty common scaffolding job. It’s mundane though and not some kind of high end software engineering job so yes I do agree these LLMs won’t be good for some esoteric software engineering requirement (we do still hire human software engineers) but I can still get it do what I need and still have the code be readable.”
I was pointing out that your graphics job and workflow is an niche/edge case scenario, not the AI programming methodology and workflow: ie you’re a very weird, niche client for designers, so your experience of AI design / images is a very weird, niche, edge case one, not pretty common.
Best not shut the designers out completely: remember the bad old days when only programmers and techies made websites? Pastel shades of blue, pink and yellow and tables all round. The horror! The horror!
Also the agent kludge is vaguely reminiscent of Hannah Fry’s Openclaw AI agent, Cass — when asked to buy 50 paperclips, it ran up a US$100+ LLM credits bill on queries about paperclips to find an option that saved maybe 50p, before singularly failing to order any (defeated by a captcha).
And that example perhaps points out how badly directed AI is at present.
Joanna Maciejewska put it thus: “I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes.”
Or as Stuart Neill snidely replied: “AI should do tedious things for creative people and not creative things for tedious people.”
I’m not particularly anti AI, I’m just less easily impressed. It just seems like parts of it is like putting wifi on a toaster: the tech is impressive but often a bit pointless (for eg talking to LLM chatbots when there are humans out there. I find that baffling.).
But I accept that AI is still in very much in its infancy; and like all tiny children it shows a lot of promise in the future, needs a lot of supervision, and it’s really not to be trusted to do grown ups’ jobs yet. Great experimental fun, though it might be.
I do have a sneaking suspicion that by the time it’s actually good enough for “production” use, it’ll be too expensive to use much, perhaps especially the “cheap” Chinese ones that are currently running on subsidised electricity, and the mainland’s usual subsidy of — let’s say — “second-hand innovation”.
That’s assuming the CCP doesn’t finally see through their FOMO and come to the jarring realisation that the whole raison d’être of AI is to become autonomous and not centrally controlled. With the CCP turning it all off and arresting the devs for helping the anathema to the CCP’s core values that AI has always aspired to be.
“pardon me, but that sounds like the very definition of a lengthy and convoluted shitshow to me. And damned expensive to boot.”
Took me a week to come up with a solution and no not that expensive in the end. Costs us less than having the designer do it. And with the solution in place it can be reused every time we need to localize a new region. What took weeks of work because of manual review can be done in about 2 days with little to no manual QA needed.
You call it a kludge while failing to acknowledge that a very similar back and forth process can happen between a human and a graphic designer. In our case this back and forth takes significantly longer than when it’s automated away with my python process.
All you seem to have is subjective anecdotes you read somewhere else about other people using agents rather than your own personal experience. I didn’t need to spend $100 in API credits for my solution. Also I know how to setup an agent that can reliably defeat a captcha too. Even before LLMs, I used to build all sorts of bots that would go through a fully deterministic automation. I know where deterministic rails are needed and where a probabilistic solution can work.
“With the CCP turning it all off and arresting the devs for helping the anathema to the CCP’s core values that AI has always aspired to be.”
CCP can arrest the developers but these models are out in the wild already and can be downloaded, forked, and hosted by anyone with the resources.
@Mark Bradley
“Took me a week to come up with a solution and no not that expensive in the end. Costs us less than having the designer do it. And with the solution in place it can be reused every time we need to localize a new region. What took weeks of work because of manual review can be done in about 2 days with little to no manual QA needed.”
Are you an intern or do you draw a big old salary per month?