Yves here. In the excitement over DeepSeek, this post put a needed reminder of an important AI issue front and center: that AI does not have enough original human content to make for adequate training sets and is therefore often training on AI generated material. In other words, this is massive, institutionalized garbage in, garbage out.
By Kurt Cobb, a freelance writer and communications consultant who writes frequently about energy and environment. His work has also appeared in The Christian Science Monitor, Resilience, Le Monde Diplomatique, TalkMarkets, Investing.com, Business Insider and many other places. Originally published at OilPrice
- DeepSeek’s efficient and affordable AI model disrupts the market, threatening the profitability of established AI developers.
- The widespread adoption of AI, fueled by DeepSeek’s model, could lead to an information crisis as AI systems increasingly rely on AI-generated content.
- Despite increased efficiency, the demand for AI and electricity will likely continue to grow, driven by new applications and broader accessibility.
In 1865 British economist William Stanley Jevons explained to the public that increased efficiencies in the use of resources per unit of production do not generally lead to lower consumption of those resources. Rather, these efficiencies lead to higher consumption as many more people can now afford the more efficiently produced goods which carry a lower price tag. Jevons was referring to coal, the cost of which was falling and demand for which was rising due to increased efficiencies in production. His idea became known as The Jevons Paradox.
When the Chinese-based artificial intelligence (AI) upstart DeepSeek demonstrated last week that complex and powerful AI can be delivered for a tiny fraction of the cost and resources of current AI tools, DeepSeek’s competitors cited The Jevons Paradox and told investors not to worry. Demand for AI would now grow even more rapidly in response to greater efficiencies and thus lower costs.
What those competitors failed to mention is that DeepSeek’s breakthrough is great news for buyers of AI tools, but very bad news for current developers who are sellers of those tools. DeepSeek is giving away free or at only 3 percent of competitors’ prices (for those needing application programming interface services) something comparable to the very expensive products of its competitors. This suggests that the hundreds of billions of dollars spent developing those expensive tools may have just gone up in smoke. That investment may never be recouped.
Moreover, DeepSeek has shown that its powerful AI tool can run on a laptop, so the need for vast cloud computing resources is not necessary in many cases. In addition, DeepSeek’s AI tool is open source and can be freely distributed. This means anyone can see the code, customize it, perhaps improve upon it AND make money off the improved or customized version. And, because anyone can see the code, anyone can see how DeepSeek achieved such efficiencies and design their own AI tool to match or exceed those efficiencies.
The one thing the big AI developers are right about is that at these new prices (free or nearly free) the demand for AI is likely to grow much more rapidly as it is applied to situations where AI was previously too expensive to justify—just as The Jevons Paradox suggests. And that means it is probably wrong to think that these vast new efficiencies will eliminate the need for large expansions of electric generating capacity. The demand for additional generating capacity will still be there. It may just rise at a slower rate than previously forecast.
This is NOT an endorsement of what is about to happen. In fact, the more rapid spread and even wider use of AI is likely to create problems at a faster rate. More efficient and broader use of AI means that the human sources of information will be driven from the marketplace even sooner—the very ones that are essential if AI is to have real information from informed experts and writers. What comes next is AI feeding on AI-generated information, a kind of digital cannibalism that will not end well.
It’s worth noting that expertise does not actually reside on the page. It resides in the minds of a community of interacting experts who are constantly debating and renewing their expertise by evaluating new information, insights and data from experiments and real-world situations.
When the information generated by this kind of expertise is gone from the web or at least crippled, what kind of nonsense will AI tools spew out then? One thing is almost certain: The nonsense will now come more quickly and from more and more of the systems we rely on. That’s hardly a comforting thought.
82 comments
Comments are closed.
well FWIW, a number of AI models, including DeepSeek passed the Japanese National Medical Examinations , with scores above 95%. I would say that is not too shabby.
I am afraid that you missed the point. Yves said ‘that AI does not have enough original human content to make for adequate training sets and is therefore often training on AI generated material.’ What will that mean going forward? It means that AI results will be like a photocopy of a photocopy of a photocopy…
The user/programmer can input his or her knowledge of the usefulness or not of info sources. The NYT, for example, would be considered by me as an unreliable source because it represents a semi-official (Washington Deep State) source of information and needs to be treated as such. One way to do that is to com opare, over the last couple of decades, the accuracy of predictions compared to other sources which is what I have done for decades to arrive at my own analysis of, in particular, international relations and so on. I trust AI as a helpful tool in analyzing that. Humans can give direction to AI, it cannot stand alone–the whole idea of AI “taking over” so to speak is very unlikely–someone will be directing it for a variety of reasons. I believe that those gangsters who run the System will be worn down by AI users who will show everyone their nonsense so that they US/CIA media cannot say a that Russia invaded Ukraine for no reason other than sheer brutal conquest. My AI will prove, with only a slight shadow of a doubt that the mainstream lied and deliberately misled us on this matter and others. My AI application will measure patterns of lying in all publications and measure the likelihood or not that they are lying now or that some force within the Deep State has changed the Narrative.
This comment reminds me of past tech claims that “the internet will set us free” or “bitcoin will save us from inflation”. Blind faith in a god, in this case a technology god, is what people do when they don’t understand the world and are fearful. So rather than doing the hard work of understanding and fixing the world problems, they lazily place their hope in a savior. In this case a wacky AI trained on Reddit with no bullshit filters.
Yes, and as I understand it stand it, it’s fundamentally entropic; AI-generated throughput will quickly overtake individual and socially-created knowledge/content, and we’ll be left with a rapidly expanding mass of artificially intelligent gibberish.
Really, how could it not? It would take a person maybe a couple hours to write a decent short one or two page essay on a given topic. But my understanding is I could ask an “AI” to generate 1,000 (or however many I wanted) two page papers on the topic of my choice and it would be done in seconds.
“AI does not have enough original human content to make for adequate training sets…”
China is collecting and digitizing past material, cultural and science data. So that historical Chinese weather data that was being digitalized was analyzed by a Huawei AI tool to create the Pangu weather forecasting tool. Pangu weather forecasting system was highly enough regarded to have been selected as the leading science advance in China for 2003.
Think how extensive and continually growing is the continually growing Chinese “Germplasm Bank of Wild Species.” There will be no running out of research data.
“AI does not have enough original human content to make for adequate training sets”
It also depends on what the goals of creating an AI application are. We tend to think about it in terms of the information space, but there are also other, useful applications for AI, whose reliability when used on “actual” data improves after having been trained on artificially generated data.
While I quite agree on the “bullshit” AI produced content in terms of reliable information, the AI can be used for tasks such as pattern recognition, clustering, computer vision, etc.
Rev, “a photocopy of a photocopy of a photocopy” reminds me of what Microsoft’s Co-Pilot returned to me while i was researching one of the newer Covid variants; the same wrong information, repeated three times, even after i questioned it…
It echos back the desired response. That epitomizes GIGO? Heck, WaPo can do that! Try asking DeepSeek how many Israelis were killed by Hamas vs IOF on 10/7, who blew up Nord Stream & what Billie Joe McCallister threw off the Tallahatchie bridge?
If I had the ability to use the entire internet as my cheat sheet, I would get a 95 too.
It is an open book test for an AI. Gimmicks like that are just a sales pitch for
suckersinvestors.Recall that a mere 0.001% contamination in a medical training set can produce a dangerously inaccurate AI:
https://www.nature.com/articles/s41591-024-03445-1
AI-inspired digital cannibalism is baked in based on how it works: learning from existing data. Apologies for resorting to economics jargon but the Production Possibilities Curve (concave to the origin) showing (for instance) the trade-off between cell-phone size and battery life is well-understood and can conceivably provide good “input” data for an AI.
However, some of us cut our teeth on points NOT on the PPC: conceptualising emerging technologies that are not yet commercially available. This kind of thinking is what led to game-changers like the iphone and partly explain why the Nokia brick went from hero to zero.
AI cannot, by definition, give us input data on these goods that don’t exist yet. We must construct (often expensive) specialised choice models and other surveys to help the consumer understand how this “hypothetical but close to production” product could change their lives. Human sources of info/satisfaction will indeed be pushed out if we go all in on AI, rather than do the kind of research that is all about “what if?” We are currently in danger of getting “locked in” to the PPC curve, when human ingenuity has so often in the past caused us to move “north-east” of the PPC and consider what could be done.
The fact that AI tools / DNN / LLM degenerate when fed their own output has already been formally investigated:
AI models collapse when trained on recursively generated data.
The curse of recursion: training on generated data makes models forget.
This makes sense to me. I view those AI approaches as some large-dimensional MLE models, similar to curve fitting techniques in statistics and econometrics. Fitting a curve through points may give useful results.
However, taking the original points and the values given by the curve fitted to them as input to another round of curve fitting; then the original points, the first set of points from the curve fitting, the second set of points from the second-order of curve fitting as input to a third round of regression; and so on, will, I presume, ultimately degenerate to the average value of the original data — something trivial and of no utility.
Thank you! Sometimes I think I must’ve missed something because my criticism goes nowhere but it’s nice when others have solid links to back me up :)
There’ve been one or two points made in (semi) defence of AI that I have refrained from engaging with…..I clearly broke a rule or annoyed skynet at some point (I know full well that moderation decisions are not always made by Yves et al) in last few months so everything goes into moderation and I can’t be bothered to engage. Fair enough. But I could say more in a different world.
I am a total ignorant and naively ask: to reduce this problem, might there be any way to ban AI content published in web sites, or any other publicly available media? I know the answer is not except if AI-generated content was forced to wear a tag.
Several YouTube channels I watch now have an on-screen disclaimer to the effect that “This video contains no AI generated content”. There have been appreciative comments. However, as you’d expect, not all such channels are telling the truth: subtle inflections or full blown mistakes in the narration pronunciation have alerted me to AI. That’s before you even get to odd sentence structure that only a native-English speaker might spot.
There is a big issue at the moment concerning the proliferation of science channels that are simply AI slop scraped from legit channels. I’ve adopted a zero-tolerance strategy: once you upload a video I suspect is AI and not based on proper research done first-hand I unsubscribe and tell YouTube not to recommend me that channel again.
I think the whole point about AI is that it work without human supervision, because good supervision by humans costs about as much as creation by humans.
If a pattern in the training data shows that a good article contains the tag “Produced without AI”, then an AI instance will put that tag in its output too.
So supervision is equivalent to creation and because AI does it more cheaply it is worth it? Sorry, no. When you have the (recognised and peer-reviewed) proof that limited dependent variable models (logit/probit etc) models in mid 1980s were shown to have infinite number of solutions I defy you to prove that AI can give the correct one without heavy human supervision. Human supervision is exactly HOW we rule out (Say) 80% of solutions and decide that (say) all but 2 solutions of those remaining make no sense. These decisions are often based on gut feeings, based on “what I know these people would do” rather than a dataset; plus experience that is unquantifiable, partly because humans have a random component in most decisions. It can’t be predicted by an algorithm.
Define “good article” given what I just said about regressions with infinite number of solutions? You need external input.
It’s easy to throw around these defences when you don’t know how a statistical regression works. Please read up on logit and probit regressions. This article shows the infinite solution problem whilst this one shows in one of the many applied fields why AI won’t work.
No. Supervision costs the same as creation. You don’t get much commercial advantage generating your texts using LLM AI if you have to pay to make sure that the AI stays within some bounds. You might as well pay somebody to write the article.
So if supervision costs the same as creation, why not do creation since it inherently has a greater degree of control (by way of controlling the study, how the questions are asked, etc)? A creator inherently knows more about human psychology than a supervisor.
Exactly. Which is why you get humans to do discrete choice modelling. AI is garbage and I predict it’ll be just as much a joke as NFTs within 5 years. it’ll merely be another program within the Microsoft suite.
I have seen a lot of non native english speakers utilize AI to make their emails more professional sounding. One of my professors sent out an email with AI prompts included and it became a running joke among my classmates. I think there is a fairly valid use for AI professionally, when you are not trying to seem like a real person. This article had some really interesting applications of AI in the workforce: https://restofworld.org/2023/ai-revolution-outsourced-workers/
You might preface remarks by stating that English is not your first language so please excuse any infelicitous usage of a work still in progress.
There is this attributed to Descartes, if I recall correctly, “This letter will be long as I have not time to make it short.”
Seems like a cheat to use “AI” to pretend to be who you are not.
I despise AI in general, but i can see it being useful in this case. If the real author uses AI as a starting point and then edits it from there, how is this any different than writing something, having it extensively redlined by someone, and then rewritten?
I don’t see a problem in this case as long as the editing happens.
I do not have a reference at hand, but I seem to remember having read that AI tools tend to overuse certain words and have somewhat “heavy” stylistic features. Things like resorting to “utilization” instead of “usage”, and so on.
I certainly think you have a point in using AI carefully in ensuring an understandable “dictation”. However, 20 years of lecturing from undergrads up to highly educated individuals attending executive education courses made me more flexible. I always made sure I had in mind THREE ways to explain my stuff (discrete choice experiments and their types). I’d start with the one I thought would work best; I’d look directly at attendees who I suspected had greatest difficulties and I could usually tell if they didn’t “get it” despite their claims.
Then I’d go to explanation number 2. Repeat.
If necessary I’d go to number 3. Generally I’d have got them ALL on board by end of number 3. Proof of the pudding is in the eating: course in Padua, Italy in early 2010s, I, together with a bunch of frankly more senior choice modellers, all ran a course. I got the top attendee evaluation score by a long way. I don’t say this to boast – I also was seen by some as a total SOB. Generally these people were traditional economists who hated that I had “sold out” LOL.
However, I do see the benefits of AI in ensuring comprehension….just not in attempting to explain. You need “eyes on the ground” and flexibility to do the latter.
More professional sounding – sure.
But one of my engineers started to use it for that purpose some months ago (totally obvious, as all of a sudden her emails started sounding like they had been written by a precocious high school sophomore ). It lasted until the third time the “improved” output reversed the meaning of an element of her input, and a customer caught it before she or I did.
Tool fetishism is bad enough when you can at least see what the tool is doing. And these LLMs are nearly perfect black boxes, for all intents.
If a ban on AI generated content actually occurred on Facebook probably 90 percent of their “content” would disappear. I deleted my Facebook account yesterday as I got tired as an artist of swimming upstream against the fire hose of garbage that is not human created. Last year Meta fed all of our photos and posts to their AI resulting in thousands of artists leaving both Facebook and Instagram for an artist created platform called Cara. In a very short period of time, Cara had 500,000 signups. I didn’t make the move as other artists also didn’t move having witnessed the crapification of multiple platforms felt that Cara would not escape the same fate.
You got my vote!
I like human beings.
I interact with them from time to time for business and pleasure.
I can see I now need to ask how much of what they have written to me was generated by AI or styled by AI.
While the author varies incautiously in his focus, a laptop cannot store an elaborate data set. Two terabytes or thereabouts.
That storage is trivial to a cloud data server, which will require large energy resources. It will be required for the most robust, if suspect, access.
At least that’s the way I see it, but I am continually outpaced by the possible.
There are descriptions out there.
I’ve seen one for a $5000 setup (Gigabyte motherboard, ~1TB of RAM and as much SSD, Linux) to run the full thing, and a mention of another setup at $2000.
That’s not a run-of-the-mill laptop, but it’s also not a cloud data server.
The inability of “localised” (non-cloud) PCs to do certain things like voice recognition is something I’ve watched for 20 years. 1998-2001 I used IBM ViaVoice to dictate over half of my 80,000 word highly statistical PhD, due to bad RSI. I had to devote a few weeks to training it first but it was time well spent.
When the NHS Hospital Trust here piloted the latest Dragon Medical voice recognition program (intended to make all the audio-typists redundant by allowing the docs to dictate appointment notes straight to digital file) it failed utterly. It ONLY worked if the cloud could be used to understand any English accent that was not close to received pronunciation…..and this was an instant data protection violation. Amusingly (to me anyway) it emerged that “keeping the dictation/understanding local” required the doctor to teach the algorithm for…..a week or two, barely an improvement on 2001, 24 years ago.
I actually worked out a compromise that would preserve data security and also enable voice recognition……but of course the NHS was not interested. These days I just laugh…..am way beyond crying!
It is hard for a device to do something independently, when its vendor makes money from providing that functionality as a cloud service.
With apologies to H.L. Mencken, or Upton Sinclair, or whoever had the original quote.
Oh god yes. 2TB is nothing anymore. Charles Stross has a great write up that puts things in perspective here: https://www.antipope.org/charlie/blog-static/2007/05/shaping-the-future.html
TLDR?
Here’s the money quote:
**Today, I can pick up about 1Gb of FLASH memory in a postage stamp sized card for that much money. fast-forward a decade and that’ll be 100Gb. Two decades and we’ll be up to 10Tb.
10Tb is an interesting number. That’s a megabit for every second in a year — there are roughly 10 million seconds per year. That’s enough to store a live DivX video stream — compressed a lot relative to a DVD, but the same overall resolution — of everything I look at for a year, including time I spend sleeping, or in the bathroom. Realistically, with multiplexing, it puts three or four video channels and a sound channel and other telemetry — a heart monitor, say, a running GPS/Galileo location signal, everything I type and every mouse event I send — onto that chip, while I’m awake. All the time. It’s a life log; replay it and you’ve got a journal file for my life. Ten euros a year in 2027, or maybe a thousand euros a year in 2017.**
Make what you will of this.
What I’ve seen, is you need 24 x 32 Gb RAM modules to run the full model. Hugging Face engineer Matthew Carrigan suggests the Gigabyte MZ73-LM0 motherboard that can take all that RAM (sourced at Newegg for $1376). Though not blinding fast. Not exactly “laptop” class.
to add to what Marcel comments above.
I’ve seen Deepseek running on our CTO’s laptop last week. A new developer spec mac book – Mac chip, 96GB of memory, no idea on SSD size. But the fact remains yes you can run these models on relatively portable low spec devices.
There are two things here – the size of the model, and the size of the data to train the model. The latter is vast, the former fits on a laptop. Training is the big time, energy and cost sink
Depends on the use case;
But in terms of running the available “Open Source” model, from what I’ve seen, you can’t run the fully quantized best quality version on just any laptop. The server setup posted in links a few days ago pegged the cost at ~ $6k and the system had like 192GB of RAM or whatever to fit the entire model.
So we’re a long way from the full size model running on someone’s most expensive Macbook Pro, to say nothing of a commodity Windows-based laptop.
I installed the smallest (1.5b) DeepSeek today on my 5 year old laptop: i3 16GB no GPU.
It took two commands to install then run it. The download was 1.1GB, and it used ~1.5GB of RAM, when idle.
When I asked it a question it took a couple of seconds thinking then printed the answer at about reading speed. Not very informative answers so far but it mostly understood the questions.
Much more impressive was asking it to “write a python program to remove crackles from audio”. Its (quite sound) thinking printed as it went, it constructed an algorithm, output some decent code, then explained it including possible issues and improvements. In less than 30 seconds. The code didn’t actually work but only because it didn’t know what a crackle looks like. It said so itself then just made a guess.
> that AI does not have enough original human content to make for adequate training sets and is therefore often training on AI generated material.
This means AI will never have enough original human content. All content subsequent to last year must be considered to have been corrupted by AI, whether directly or by inflecting someone’s brain.
It was amazing in the ’70’s when the entire Library of Congress was able to fit in a few file cabinets of microfiche. Then came laserdisks! Diderot’s dream was to put all human knowledge in the Encyclopédie. Which is silly since it didn’t include the names of everyone’s pets, so ‘Knowledge’ so defined by knowledgeable folks.
AI is useful for things like submarine detection, where the data is generated automatically, and of such quantity as to overwhelm humans. It’s the Human content thingee that’s the problem, both technically and at a metaphysical level. I do not want a solar-powered Elon2 mimic floating in space implementing glitchy bondvillain decrees through the payment system. The real thing is unstable enough.
I read a fair bit of scifi stories, enough to recognize what AI is and is not. And what it is is a dead end. Let me explain. Silicon Valley loves to take concepts from scifi and to make them real so that they can profit from them as thinking up brand new ideas is hard. In a lot of scifi stories you see mention of an AI but with the difference being that it is sentient and is treated as being equal to humans. Thus you will have AIs supporting ships, bases and even individual people and it is like talking to a good friend. The present iteration of AI is nothing like that but equivalent to a parlour trick. And because it has run out of new material to steal and turn into a training set, it is already deteriorating in performance. Maybe by the end of the decade it will be unusable. Who can tell? But it cannot really be developed much further to the point of really being vital for humanity’s development. It will find much use in niche fields which it can be very good at but the present iteration of AI will never be like those of scifi stories. It can’t cut the mustard.
“…and it is like talking to a good friend. The present iteration of AI is nothing like that but equivalent to a parlour trick. ”
It’s taking advantage of something humans already do: Anthropomorphizing.
More owed to the Bernays/Freud connection to business by these types than to Einstein.
Yep. “AI” is the marketing term for machine learning and its special case, the Large Language Model. (Expert systems was once marketed as “AI” but no more.)
Machine learning is useful for finding statistically likely correlations, as long as one understands that is all they do. Large Language Models spit out statistically likely text, the usecase of which is mostly to fool people.
“AI” is when people has been fooled that statistically likely text is an indication of a living mind.
so if I get your argument straight…
based on some sci-fi you can confidently assert that an entire branch of science and engineering community is on the wrong path?
you realize that is the least likely scenario?
and wrt to the description of “parlour trick” and sentience:
Have you ever read anything around the “Theory of mind” and philosophy?
We don’t understand human consciousness or thought.
yet we take it as axiomatic that other humans are conscious and can think.
but there is no evidence for that apart from their output – for example their spoken or written word.
.. and it is that very output that the machines are trained on!
Author writes a book, students read the book and learn.
author writes a book, machines read the book and learn.
That’s just learning from words in the book. the arrangement of words ARE the representation of meaning. It is literally what we use to train humans from baby through to expert. To state that it is not possible for a machine to learn from that is LESS likely than the opposite
so to state that machines will never think… is like saying “machines will never do this thing that we don’t really understand and can only ever observe the output in others”
That is logically weak.
that is without looking at other examples of intelligence in nature – distributed intelligence in the octopus say, or slime mould problem solving, or our own cats and dogs
What I don’t understand is the motivation behind your advocating on this site. Others just write their thoughts, and you seem to have some agenda to push. A soapbox came to mind.
That’s an easy one to answer. I’ve followed this site since the lehman brothers days and have occasionally contributed to the fundraisers.
I’ve learnt a huge amount on finance, economics, climate, ukraine – all the things this site and commenters are really strong on.
I am very grateful for that learning. hugely grateful. I have not commented that much- on most topics I am the learner.
so it saddens me to see what feels like a very one sided narrative on AI here. to see comments repeated that are just flat out wrong, to see an undercurrent of human exceptionalism.
If I tried commenting here on, say, economics or climate then I would be quickly and rightly put straight by either wiser commenters or the mods. The “I’m sorry but you are just making sh!t up” phrase would rightly be levelled at me.
So my motivation comes from wanting to pay back to a site and community that has given me so much.
An AI evangelist?
No. Just somebody that has their own opinion.
Sorry Rev Kev but that’s not good enough IMHO.
You are right that spouting a load of stuff that sounds impressive about AI ain’t enough. We’re not supposed to “trade on our name” but it is not difficult to look up quite a few people who comment here to show that they have global reputations. Yves is the obvious exemplar but there are others: I have tried repeatedly to draw attention to a globally acknowledged paper showing why logit/probit studies can be horribly wrong. The repeated flaws in trials, political surveys etc should be enough evidence.
Why are free passes being given? It isn’t Yves, from what I can see…….I don’t even think it’s necessarily a well-described system generally, since around 10 years ago Yves once explained to me how “Skynet” can dictate who gets hit and who doesn’t, independently of the NC top brass.
I’ve been both impressed and irritated by certain comments made here in last day or so……if Yves et al wanna ban me, so be it. I’m a guest on their site. But there’ve been one or two contributors who seem odd…….maybe they have an agenda…….Yves et al must decide.
I just, when health allows, come here and try to draw attention to long established well-peer-reviewed articles that have relevance to a variety of areas like n-of-1 trials, political polling etc……..
I’m NO FAN of the “pseudo Economics Nobel” but I do think there are 2 or 3 winners who were genuinely clever with original ideas. McFadden was one – predicting demand for a light rail service almost perfectly before anything was done has to be considered as good. He helped move forward a branch of research that is really showing its colours……but it really really annoys the mainstream economists……
Specific issues, like those set out by Jeremy Grimm seem much more suited to AI. Survey research? Nope.
Human Exceptionalism!!!
Like in The Animatrix.
‘based on some sci-fi you can confidently assert that an entire branch of science and engineering community is on the wrong path?’
Yes. Absolutely. Totally. Unreservedly. Entirely. Positively. Completely.
AI does have some excellent potential for certain services. I saw one Chinese example where one was used to plan where all the electricals should be placed in a new ship. But that is not is what is being sold to the public. it is being spruiked as a miracle device that will transform all our lives and all we have to do is to throw hundreds of billions of dollars to our tech lords so that they will do it for us. Sounds legit.
As for your example of how we train human babies, by the time they are toddlers they already know much about context and use sophisticated ideas that a machine device cannot cope with. I think that it was Artur C. Clarke that said that the true test for sentience in an machine is that you would be able to tell it a joke – and that it would get it and laugh. The present AI is not capable of doing so and probably never will. Sorry, but that it the truth.
/caveat – I don’t know your background at all. e.g. are you steeped in a decade of data science. no idea.
Bold statement to say your opinion beats an entire discipline.
on “what is being sold to the public” – on conflation – I separate out the tech and research progress from the services and products built on top. just because some people make a bad, or greedy implementation of a chatbot/intregrator/copilot/something else doesnt mean the whole science behind it is bad.
on Arthur C Clarke – that test is a very anthropomorphic view of sentience.
A current machine practically cant laugh – no mouth or lungs. Can the current models pick out the “funny” parts from a text? yes. Can they infer what the reader is thinking? yes. Can they create jokes and absurdity? yes.
on toddlers – not sure what you mean. Example?
A current machine practically can’t laugh – unless you hook it up to a speaker. Sorry but computers do not fundamentally understand humour – only what they are trained to look for. You tell a computer ‘This is fine’ and the meaning will go right over it’s digital head. As for toddlers, having raised two I saw how they adapted to language on the fly and were able to assign meanings to concepts and different words. That is why little kids can pick up several languages if need be. AI does not have that nuance and since it is now taking in the results of other UI’s work as part of its training set, will degrade it’s performance over time. That is why I say that the present iteration of AI is a dead end. You can only push it so far and it’s output over time will resemble a photocopy of a photocopy of a photocopy..
To claim an ‘entire discipline’ is on your side is a very broad claim, when most of the AI hype is generated by a handful of hucksters. Many researchers working in machine learning understand the limitations of LLMs and thus, the BS level of current AI evangelists.
Q: What’s the difference between an LLM and a used car salesman?
A: The used car salesman knows when he’s lying.
Yeah similar worries were triggered in me. My field is all about errors so that kinda blanket statement raises red flags. I wanna see acknowledgement of nuance and examples of human interpretation in areas we KNOW have infinite solutions.
maybe I should phrase it more fully.
Bold claim to state that one persons opinion beats the future work, research and growth from an entire discipline.
I know there are variations in the state of the research. e.g. the photocopy of a copy of a copy meme seems to go back to the paper in Nature last year. Yet there are multiple papers recently that take a different stance on the use of synthetic data mixed in with real world data.
To say that researchers won’t come up with increasingly better ways of curating and using training data sets is… unlikely.
I am not an AI expert but I did delve deeper into the working of what was called AI or machine learning back in the 1980s. Aside from increased computing horsepower and massive data theft, what technical or theoretical improvements have been made in the last five decades? In the 1980s AI consisted of classification key based tools similar to automated versions of the keys I used as a boy to identify insects, Bayesian nets, neural networks, and syntactic pattern recognition. Neural networks became a favored tool for machine learning with the major drawback that no one knew quite how they worked and they often manifested strange glitches in their performance. Today, what amazing new techniques have been invented to support claims that the ‘learning’ algorithms of the 1980s can now claim to be AI in the sense that term is used in Sci-Fi and popularized in the AI hype used to inflate the tech stocks? You seem to be claiming some knowledge of that “entire branch of science and the engineering community”. Would you please explain what advances substantiate the claims that AI learning is more than glorified brute-force pattern recognition. Have there been advances that might open the black-box of the structure of AI patterns that it ‘learns’?
The application of AI to human language manipulation is impressive and fraught with ‘issues’. It seems a strange effort toward monetizing AI that wastes large amounts of computing power and energy to perform many tasks humans can perform with far greater adaptability, and at lower cost. I believe applying existing AI to solving problems like the protein folding problem makes more sense. I think this article covers many of the concerns I have about the ‘intelligence’ of AI in an area for the garbage in problem does not appear to be an issue:
“Did AI Solve the Protein-Folding Problem?”
https://magazine.hms.harvard.edu/articles/did-ai-solve-protein-folding-problem
The arrangement of the words may be a representation of the meaning, but it isn’t the meaning. If it was it wouldn’t be possible to cram, regurgitate and understand nothing. And yet most college students has done it at some point.
The machines creating statistically likely text has no understanding of what in a text is true or false. They can’t, because all they have is a matrix of how words are likely to follow each other.
As to the argument that because we don’t understand it, anything could be consciousness. If you lived before mechanical watches were commonplace and I showed you one, without taking it apart. Would it then be reasonable to think it was conscious?
@The Rev Kev at 5:17 am
AI “is sentient and is treated as being equal to humans. … and it is like talking to a good friend”.
I just finished reading “Agency”, the continuation of “The Peripheral” by William Gibson yesterday. The AI is as you describe.
“that AI does not have enough original human content to make for adequate training sets and is therefore often training on AI generated material”
That the training wheels don’t ever come off should be a red flag about believing the BS about AI “thinking”.
Indeed. I have multiple beefs with the AI craze. One is the same reason why economics in its neoclassical form is BS: it has no model of REAL psychological behaviour, more specifically, an understanding of human error/inconsistency, to give a signal-to-noise model.
The signal-to-noise model is almost certainly like all good models: wrong but useful. Ratios of yesses to noes can be brilliant in helping us understand people. When decisions, repeated over time, become deterministic rather than probabilistic, people like me take note. We start looking very carefully. It doesn’t necessarily mean there’s something wrong……but it does mean we need to understand “why are there no training wheels?” (to continue the analogy). Why does this person ALWAYS do the same thing? Because when math psych people have conducted tests over the last 100 years, consistency is the exception, not the rule. Only religion and certain (typically worrying) branches of political support show determinism.
“Thinking” is almost certainly probabilistic.And not in a “let’s just project past training data trends” way.
Interesting article. Code assistance is one area where AI is currently genuinely useful; it also provides a good example of how it will self-cannibalise. There are open online forums – Stackoverflow for example – where developers can ask questions and discuss issues which were, until recently, the first port of call for those running in to problems with their code. Code assistance models are useful primarily because they trained on the data from these forums, which they have now, to some extent, replaced – if you look at traffic to Stackoverflow you see a precipitous decline coinciding with their emergence. With no community-based discussion of new techniques / software libraries we can expect the output of the models to gradually worsen as their training data decreasingly reflects current software state of the art. At some point I suppose people may switch back to using the forums, leading to a back and forth.
You make a very good case for employing people, not machines to do customer assistance and therapeutic services on an international scale.
That concept would gainfully employ people helping people, rather employing them and machines to cheat each other out of their money.
Existing AI can be used to predict where energy anomalies may occur in fusion reactors.
Yes. I have no objections to people using AI to “correct/improve” their code. However, they should be looking at what the “augmented” code is doing, how it improves over their first attempt etc.
If you don’t learn from the experience then you haven’t actuallly mastered the skill. And one day that’ll come back to haunt you when the AI doesn’t spot something very human but wrong and you have trusted it without using common sense and you do not have a good understanding of what the AI did for you.
AI isn’t AGI, which as far as I am aware is as far off as ever.
I asked DS to summarize Kant’s “Critique of Pure Reason” this morning while going about my morning’s abulations. The summary was pretty good, from what I can recall. I followed up with a question to elaborate on “a priori, synthetic judgments.” When you ask it to “elaborate,” it goes deeper. Is this of any value? Certainly, my recollection of the details of Kant’s “Critique” is far removed in time to when I first studied it. So, I’m tending towards thinking it is a provisionally good tool if you’re on your guard, like anything that is digital. Human’s are rooted in the analogue, so the digital devil must be kept in check…
Indeed. One wonders: with the internet increasingly full of ads and spam and phishing attacks etc., and soon to be absolutely flooded with AI generated content that is superficially plausible, perhaps the time will come when there are multiple parallel internets? That is, the original covers-everything internet will be just too toxic, too unreliable, it will be functionally unusable. Instead people will develop walled gardens of various sorts, that are carefully vetted and sealed off from the original toxic internet. I would think that the militaries, and the internal systems of big financial institutions, are already moving in this direction. And that will be a shame, but perhaps inevitable.
I agree with you that the Internet will probably evolve so that websites become “fortified” enclaves — barricaded with captchas, endowing text, images, and videos with steganographic anti-AI poison, providing pages only as bitmap PDF, requiring subscriptions, etc.
However, I respectfully think that it is a bit late to worry about “the Internet […] soon to be absolutely flooded with AI generated content” — for it is already absolutely flooded with slop produced algorithmically or through AI.
Did you miss this one? It was in 2012, already. A decade later, Amazon became utterly overwhelmed by AI-produced e-books and had to take measures.
And that one? Those disturbing mass-produced videos were already swamping YouTube in 2017. Despite widespread concern, Google encountered enormous difficulties to stamp out the flow of that crap.
To summarize: automatic content creation (with varying amount of AI support) was in place at scale 13 years ago, and already a genuine problem 8 years ago. At the latest 2 years ago, some important sectors of the Internet were already swamped with content generated specifically with AI.
Thank you. I was being taught how to spot non-human garbage in the online surveys I ran back in 2010. I’m pretty good at spotting that stuff, it must be said. Appearing human (inconsistent in a genuinely weird/random way) is surprisingly hard to imitate. Thankfully.
I work with AI daily, where our team is trying to make them useful through APIs to structure large stores of private information in business applications. Its much more demanding than simply asking random questions through a chat interface. Chat only requires that it act smart. We need it to be smart. They are not. There is no logic buried in there. The best you can do is structure queries to align with logic patterns found in the training data. An LLM can find correlations beyond any human perception, but as the article says—these must already exist. So, great chat tool, potentially a good augment to data tasks, but not at all intelligent in any useful meaning of the term.
DeepSeek cost a lot less to train. That means it was trained on a lot less data. If LLMs can be trained on a lot less data, at less cost, this delegitimizes the money’s classes current view that only more data and more hardware are necessary, and that their money creates a moat to future development. Hence Nvidia’s stock price fall. (But it also means this piece’s premise is wrong).
Unfortunately, the result of making DeepSeek available is that the mooching class is now trying to pass laws to prevent opensourcing any AI technology… and possibly even publishing in scientific journals. (Tabling in this tweet means to propose as in English, not postponing as in American English). Their claim is that China is stealing American knowledge. This is preposterous, since most AI researchers are not American, or at least not born American. They’re European, Chinese and Indian. Even the famous ones: Hinton (British), Le Cun (French), Bengio (Canadian), Schmidthuber (German). Many of the key ideas of modern LLMs were described in Russian and Chinese before English. DeepSeek itself contains some new ideas. But the legislators’ idea clearly is to artificially create “winners” like OpenAI by preventing competition, instead of stepping up to the plate and innovating more. Reminds me of the plot of Atlas Shrugged, except now Galt Gulch may prove to be in Russia or China. Hopefully this ridiculous bill will be killed.
Chinese AI has been focusing on reducing cost, which means reducing energy usage. And that is good for the environment. Moreover, if AI proves useful, do we want to be outcompeted, just so that OpenAI’s valuation doesn’t tank?
If your use case for AI is to consume all the text or data in the whole wide world, then yes, you’ve got garbage replication problems. If, on the other hand, your use case for AI is to train it only on your own corporate communications and corporate documentation, or even to point to very particular public facing technical standards, or corporate datasets, then you’re carefully curating what your AI is allowed to consume and would this get around the garbage replication problem?
Where I work is investigating using AI to help us with financial reporting and analysis. I’ve had this conversation multiple times by now that our data is so crap that inputting it into a machine learning model is just going to produce more crap. However, I will plod along and due what they want as long as the paychecks keep rolling in. Machine Learning models aren’t bad per se, but most of the “AI” I’ve seen is just overhyped BS.
This is an indirect way of admitting that current models are still far from AGI and aren’t getting there any time soon. Infant and toddler brains, for example, get trained just fine with a lot of sensory input, conversations with parents and family, and some children’s books and TV shows. If all the human generated content in the world still isn’t enough for your model to perform up to the hype, the problem is not the data.
Excellent point! Just how intelligent are these AI if they are running out of ‘information’ when stealing datasets from the enormous flows of information on the web?
Again, without any pretense of General intelligence, readily available AI programs scored >95% on a National Medical licensing exam. That ain’t chopped liver.
Moreover, i can attest there are many people finding chatting with GPT is more therapeutic than with their doctor. Also, not chopped liver. ( i know it speaks poorly of the sad state of our profession)
If AI can score so well on a National Medical licensing exam, just what might that suggest the National Medical licensing exam is measuring as a proxy for assessing the competence of Medical doctors? Similarly, that people find chatting with a GPT more “therapeutic” than chatting with their doctor, raises questions about what “therapeutic” means when used in this context. Is it “therapeutic” to have someone or something listen to your complaints and express sympathy and agreement than hear the advice and guidance of a human physician attempting to minister to the root cause of your complaints?
Both good questions to be sure. But I would suggest that the answers to these are perhaps not as straightforward as you might think.
And I don’t mean at all to dismiss the reality that silicon valley is built on mountains of hype and underperforming and lies and damage to our society
I found this video really helpful to see how smart people are thinking through some of the challenges of average inputs and hallucinations.
https://www.youtube.com/watch?v=yqq_U2fxd2U