Put Large Reasoning Models under pressure and they stop making sense, say boffins Among the forever wars in geekdom, defining the difference between science fiction and fantasy is a hot potato destined to outlive the heat death of the universe. There is no right answer and it doesn't matter, hence the abiding popularity of the question, but attempting to make that delineation can still be useful when …
"Not all industries have the tightly integrated function and quality testing regimes of production code generation."
And some titans of the IT sector don't even have that ... https://www.theregister.com/2025/06/16/google_cloud_outage_incident_report/
LLMs fall over when reasoning gets too complex? Shock. So do Mark 1 Human brains.
LLMs have difficulty keeping track of large contextual requirements? Shock. So do Mark 1 Human brains.
Artificial intelligences exist now, and today (We don't have artificial *sentient* intelligences, before you all jump on my ass). Why is it any surprise that an artificially created form of intelligence falls at the same exact hurdles as evolved intelligence?
I'm not defending the use of current AI systems in the wide context that they're being used it, although blaming the LLM for how people use it seems backwards to me, I'm saying that intelligence (even if it's not all the way up to human levels yet) is still intelligence and subject to all the same pitfalls. To expect otherwise is foolhardy.
But my dear Watson, it's precisely the imprudent *current* use of something clearly not ready for production, that is the primary concern.
Arguing that the fetal stage of a horse is perfectly normal, does not somehow make that fetus suitable to race in the 2:30 at Chepstow.
But a Tsunami of Tech Bros armed with glorified autocorrect scripts and the complete unabridged chat log of Twitter, have decided to launch that stillborn abomination onto the unsuspecting public, and just see what happens. Because, money.
We live in gloriously shit times.
Look me in the eye and tell me that Sentience and Intelligence is the same thing. The term has never been Artificial Sentience, but Artificial Intelligence. The two have been conflated, but my original point stands. Intelligence is Intelligence, and even though current LLMs are not as smart as we are yet, to deny that they show intelligence is to deny reality.
Because they're not REASONING at all.
The reason it can solve Fox & Geese or Towers of Hanoi... they are trivially brute-solvable. There's no reasoning required, you just need to graph (in terms of graph theory) the tree of possible actions and before you're even a few steps in you have your complete solution. There's no depth, no thinking required at all.
The reason AI fails at reasoning is that AI cannot reason.
And absolutely nothing that I've seen or heard asserted by others is to the contrary when you dig into it.
The greatest advance in AI in my lifetime was Google's AlphaGo etc. computers. That's it. They were surprisingly fast at progressing. And then people realised that most Go models can be confused by just... not playing like a grandmaster. They make mistakes and don't know how to cope. And then that plateaued and creators of such systems (including IBM's versions) literally then tried to find some kind of business model or buyer for them, because they... really didn't do what you thought they were doing.
The problem with humans is that we have intelligence and inference and reasoning. And we see these things and jump to a conclusion (which is a critical part of intelligence) which then isn't backed up by evidence... that they achieved those things because they're intelligent. It's not true.
The problem with AI is that is doesn't have intelligence and inference and reasoning.
>The reason it can solve Fox & Geese or Towers of Hanoi... they are trivially brute-solvable. There's no reasoning required, you just need to graph (in terms of graph theory) the tree of possible actions and before you're even a few steps in you have your complete solution. There's no depth, no thinking required at all.
Mmm... but these models do not simply brute-forcing the problem (nor, of course, do humans), so I'm not sure what your point is.
> The greatest advance in AI in my lifetime was Google's AlphaGo etc. ... And then people realised that most Go models can be confused by just... not playing like a grandmaster.
Sure; because they were trained on grandmaster-style play. Human players (not so sure about Go, but I played a lot of chess in my youth) are also frequently bamboozled by an unfamiliar style of play. In fact at one time I had some success at school tournament level playing a rather unusual, but somewhat archaic opening. It isn't terribly sound if you know how to counter it, but a good proportion of my opponents didn't. (When they did, it didn't work out so well for me.)
Mmm... but these models do not simply brute-forcing the problem
Step by step instructions for which are provided how many times over on the entire web? Ingest the entire web and you've ingested those instructions, which it can follow.
If you created a similar problem that's different from the existing in a small but important way those models will all fall on their face. A human who knows how to solve Hanoi would recognize "oh this is like that but different in this respect so I'll need to adjust my strategy to compensate" and still solve it. Because humans can reason and LLMs cannot.
> Step by step instructions for which are provided how many times over on the entire web?
Indeed. So not brute-forcing, then; that's not what brute-forcing means. Call it "cribbing" if you like.
> If you created a similar problem that's different from the existing in a small but important way those models will all fall on their face.
A pure LLM, maybe, but...
I am am mathematician and statistician by profession. I recently gave DeepSeek-R1 a roll (with some scepticism, after [ahem] prompting by a colleague). I posed it a hard problem in Bayesian statistics to which I did not know the full solution - although I'd made some headway - and for which, crucially, the solution was not out there on the web. It made a fair stab at tackling it. It's response was lengthy and detailed, and it (in almost tedious detail) explained its rationale (which was sound - this was an area I am familiar with). Whenever it hit a dead end, it said (literally) "Wait; let's try something else", backtracked and did just that. Some of its attempts were things I'd tried myself (and hit dead ends with). It cited appropriate (not hallucinated) references. It didn't solve the problem in full, but nor have I.
My take on this was that it was more "meta" than a standard LLM (DeepSeek-R1 is not a pure LLM, it is an LLM combined with some kind of recurrent reinforcement learning). It wasn't simply scraping answers to the problem (because they weren't out there) - rather, it seemed to be scraping methodologies - procedures, if you like - for solving problems in statistics. So not "creative", but certainly more sophisticated than a pure LLM which, I strongly suspect would, as you say, have failed dismally at the task. I can imagine it becoming a useful research aid - maybe a Google on steroids [with the due diligence/validation on the human user's part that that implies].
I'm a pragmatist when it comes to these things. I'm fine with ML/AI/whatever (and truly I don't care what someone wants to call it, or whether someone thinks it's "intelligent", capable of "reasoning", or dumb as a box of rocks). If it's useful then I'll consider using it. Although I might well decide not to, if it's using half the planet's resources to do something a human could do just as well, or better, with a smaller energy footprint.
Nice one, LionelB. That deserves another upvote. Thanks for taking the time to share your thinking with the spaces calling in here checking on the progress of situations worthy of both the sublime and the ridiculous and publishers and publicans alike :-)
and for which, crucially, the solution was not out there on the web
How can you be so sure?
Just because you didn't find it using a search engine, doesn't mean that it isn't out there. It could be behind a paywall, and the AI scrapers are a lot better at circumventing them than either search enigees or you.
I'm afraid you are just another punter taken in by the Victorian parlour trick that is AI.
>> and for which, crucially, the solution was not out there on the web
> How can you be so sure?
I can be pretty sure, as it's in a research area I'm very familiar with (and I have access to most paywalled publications through my work). Let's just say that if I couldn't find an existing solution, it's pretty unlikely an LLM scraper could.
Which is in any case irrelevant, since if you'd read my comment properly, you would be aware that DeepSeek-R1 didn't find a solution. So I'm really not sure what your point is.
> I'm afraid you are just another punter taken in by the Victorian parlour trick that is AI.
I'm afraid you are just another punter who couldn't be arsed to read my post(s) properly, and jerked a knee.
Oh, well... I wasn't really expecting a reasoned and reasonable response, and you didn't disappoint in that respect.
Here's what I think happened: I described a personal, hands on experience with an AI (add "so-called" to taste) system, trialling in it in an area with which I have substantial knowledge and experience. I then hazarded a couple of guesses about what I thought might be going on behind the experience - which was in fact genuinely interesting. Here are two things I didn't say:
1. AI is amazing - it's going to change the world as we know it, and take over from humans. It's proper boffin-magic, and if you don't get with it you'll be left behind.
2. So-called AI sucks big time. It's shtoopid, and a complete con. It's not intelligent because it can't possibly reason. It's just statistical pattern-matching.
I didn't say either of those things, because I don't buy into either brain-dead narrative.
Here's something I did say: If AI (add "so-called" to taste) turns out to be useful to me, I might well use it (and by implication if it doesn't I won't). I said that because I don't care very much about hyperbolic narratives on either side. They're infantile and bore me.
But all of that wooshed over your head. The only thing that caught your attention was that I didn't say 2 above. So you jerked a knee accordingly.
> The reason it can solve Fox & Geese
Is because it has seen the answer more than once whilst scraping the web, let alone all the books it has ingested.
Hmm, wonder if any of the other ancient puzzles have ever had their answers published in anything an LLM may have been trained on?
> There's no reasoning required, you just need to graph (in terms of graph theory) the tree of possible actions and before you're even a few steps in you have your complete solution. There's no depth, no thinking required at all.
Hate to break it to you, but even vaguely considering graphing it out is taking reasoning steps - which is why you will find these puzzles used in Intro To Graphs Theory!
Brute-forcing these means just writing out every possible sequence, with no consideration at that point as to whether the move is even legal (which is what your graph helps with), then looking to see if that fits the rules. Even the next step up - backtracking - is adding a reasoning step and the bestest answer (ignore graph theory, just run the problem from the last step backwards and then print the answer out in the other order) is brutish in execution but reasoned in its design.
Brute-forcing combinatorial puzzles tends to proceed by (breadth- or depth-first) traversal of the tree of legal "next moves". The search is curtailed when you hit a leaf node where there is no legal next move, at which point, if you are not finished you back-track (pop the stack, or queue).
I wrote a Sudoku solver along these lines once; it solved the most fiendish Sudokus in fractions of a millisecond, generally requiring well under 1000 iterations. Of course human players don't do that (it would be exceptionally tedious, to say the least), but rather use a raft of accumulated heuristics to short-cut full traversal of the game tree. It would be simple enough to program those heuristics into a solver. A far more interesting problem - one requiring far more by way of "intelligence" - would be how to actually find good heuristics from scratch. Now, personally, I've never read a book/tutorial on solving Sudokus, so all my heuristics were indeed devised from scratch by myself - over a period of years (I still very occasionally discover a new one). It would be rather difficult, though, to get an LLM to even attempt that, since it would by default simply look up all the existing human-devised heuristics. I can't even begin to imagine how to automate discovery of good heuristics (the "meta" game of discovering Sudoku heuristics must have a truly brutal search tree!)
> Brute-forcing combinatorial puzzles tends to proceed by (breadth- or depth-first) traversal of the tree of legal "next moves"...
That is what you, as a sensible person, do to brute-force something: "only do what is legal from this point, instead of later on finding out it wasn't even a possibility". But in that, you have already applied a level of optimisation. You probably that that optimisation is "utterly obvious and trivial to say it should be done", but that is because this isn't your first rodeo (or your tutor pointed it out to you).
But what you describe isn't *the* most brutal approach - so the LLM can do worse than you. Or it may get lucky and do better...
> It would be rather difficult, though, to get an LLM to even attempt that, since it would by default simply look up all the existing human-devised heuristics. I can't even begin to imagine how to automate discovery of good heuristics (the "meta" game of discovering Sudoku heuristics must have a truly brutal search tree!)
Not too hard to do, conceptually - you forget entirely about using bleeping LLMs to solve anything and just set up a far simpler, but still based upon Neural Nets, ML model to play Sudoku. Reward it for writing out the shortest number of steps to solution, tested against your existing solver to see if it is doing it right. Don't worry about the horrid search tree, that is the job of the ML process. The rewards for writing out is the bit you want to think about, otherwise you'll get a blackbox Sudoku solver, nothing about what tricks it has managed to find to reach that solution.
Other commenters here can probably name a Python ML library or two to try the ideas out with, if you know how to attach your existing solver to Python (assuming you find Python an easy language to play with).
> That is what you, as a sensible person, do to brute-force something ... But in that, you have already applied a level of optimisation. ... but that is because this isn't your first rodeo ...
Sure, and it's also a pretty standard meaning of "brute-force" (and one that's accessible to LLMs).
Of course there are more dumb ways of solving a problem. Sorting is a nice example. Most mathematicians/software specialists would, I think, understand "brute-forcing" the problem of sorting a sequence of numbers to mean exhaustively running through all possible permutations of the sequence in some fixed order until you hit one that is correctly sorted. But there are certainly way more stupid and inefficient ways to sort a sequence.
> ... just set up a far simpler, but still based upon Neural Nets, ML model to play Sudoku. Reward it for writing out the shortest number of steps to solution, tested against your existing solver to see if it is doing it right.
Um... I'm not honestly sure that gives you heuristics - or if so, they are indeed going to be "black box" heuristics. The point is that a heuristic (in the common understanding of the term) for Sudokus has to be useful for all (or at least "most", for some value of "most") Sudokus. So if you do the standard ML thing of rewarding an ANN for efficient solutions (say, using some reinforcement learning scheme) - efficient across a large set of Sudokus, that is - you (probably) won't be able to figure out how it's doing it. It might, for all you know, have discovered a bunch of well-known-to-humans heuristics, or it might just be one monolithic, tangled lump of heuristic.
My software solver uses a very crap and inefficient method - brute force (see above), where it exhaustively runs through all legal configurations until it finds the solution. My own personal set of heuristics (I tried to list all of them once, which was quite hard) certainly does better - although there is a meta-heuristic about which order, and where on the grid, to apply the set of heuristics, which is to all intents and purposes pretty inscrutable, even to me (my own black box, in my own head!) - but even then, there may well be a No Free Lunch Theorem which says that there will be some specific set of Sudokus for which my personalised human heuristic approach will fare worse than brute force.
The path forward is probably to merge these haploid techs into proper zygotes, through unbridled sex, with orgasmic reward, in true DARPA Shafto and Gary Marcus style ("neurosymbolic techniques")!
classical AI, rooted in logic, where heuristics are used to manage the search space, modulating between breadth- and depth-first searches, backtracking to escape from unpromising near dead-ends, is tops for well-posed problems like logical and math puzzles.
Neuron-inspired AI flails there, but is beneficial for ill-posed (or even un-posed) problems, with too many solutions or none at all, where some approximation (possibly easily verifiable backwards as in the NP-complete category) is way better than nothing.
Copulate the two profusely, within an appropriately broad artificial society DEI software environment, under the overwhelming force of climax, through genetic algorithms, and ¡BANG! you get systems (some of the offsprings at least, after training of trainable ones) that combine that one in the corner's meta-heuristic prowesses with LionelB's Sudoku-solving performance, all in one, and vice-versa!
Other offsprings may help maintain artificial social cohesion, or perform artificial arts and sports that others could enjoy through the societal reward function needed to maintain artificial DEI, as needed for long-term ecosystemic sustenance per the major lessons of Evolution, as opposed to the complete wreck and fraud that is intelligent design ...
Sex is joyfully messy, and I'm sure we can all agree that it right beats promotional parades of cloned uniformed final solutions and related hammer-and-nail reductionisms. Gotta think seriously about plentiful sex and education imho, ecosystematically, a lot!
He, he. My PhD was in (biological) mathematical evolution theory, but with an emphasis on implications for Genetic Algorithms.
Curiously, one of the conclusions of my work was that most problems you could solve with a GA, you could solve faster and better with non-"genetic" Monte Carlo methods such as stochastic annealing; also, that "artificial reproduction" in GAs tended to be ineffectual, for the simple reason that populations would quickly run out of genetic diversity (there was even a large sub-literature on ever more byzantine schemes for maintaining genetic diversity in an artificial population). GA practitioners, I believe, simply failed to appreciate the orders-of-magnitude difference in scale (deep time and population size) available to natural evolution, as well as the richness of natural environments, in contrast with artificial evolution. IIRC there were several key papers about this in the late 90s, early 2000s. GAs subsequently waned in popularity (as interest in ANNs burgeoned); there were always better ways to solve any given problem. Which is kind of a shame; perhaps (as you may be implying) they'll have their day again.
Could that ever work without some sort of genuine understanding? The "prompt" is an almost insignificant fraction of the data ingested by a system that can recall 42 percent of Harry Potter and the Sorcerer's Stone and I'm not sure what mechanisms would be available to recognise that the prompt was essentially new training data that should override the current model and modify the operation accordingly.
People seem unable to resist anthropomorphising these machines: they can't be led to the correct solution by holding their hands. They don't actually "learn" anything in the way we traditionally understand the word.
It is with the greatest regret that I must record a downvote. While I understand Harry Potter was marketed differently in other parts of world, I cannot let a reference to the “sorcerer’s” stone pass unremarked. Not in this place, which was once a bastion of stiff upper lipped, emotionally suppressed Britishness. Good day to you.
P.S. if we let things slip, before we know it, Marathon bars will be called Snickers…
Being a genuinely stiff upper lipped, emotionally suppressed Briton, I'm entirely unfamiliar with children's reading and, indeed, children. Please excuse my ignorance and take this as a passive-aggresive apology in the best British tradition.
The elephant in the room is the people that "AI" could replace are the ones pushing for it the most to replace the "worker drones": the c-level and the average MBA grad.
Do you even need an LLM for "when in doubt, lay off more people" and "the share price above all, stock buy backs until your golden parachute inflates"? Methinks a 10-line script could do that.
it isn't clear how humans make expert decisions
The fact is that many 'human made' decisions are made subconsciously before the conscious intellect is aware that a decision has even been made. This is in addition to 'reflex' reactions such as stumbling but staying on your feet. I remember several times jogging and not knowing which route I would take until I got there and found I was running either up the hill or on the flat, importantly without having made a conscious decision. Chess players cannot possible analyse all consequences of any move, so do some analysis, but then make a move anyway.
Sure, par of the problem is that experts can't easily articulate the decision-making process. But what really did for it, in my view, is that the effort was pretty transparently aimed at replacing those same experts. Trying to get valuable people to co-operate in their own redundancy is not an optimal strategy.
-A.
And that can apply to questions that don't require much expertise at all. I'm not sure if it's been superseded but the standard book from my student days for identifying British seed-bearing plants stands about 2" thick, printed on thin paper and consists of a key, mostly binary but sometimes offering more choices, and s series of descriptions to check. It represented the combined knowledge of three leading botanists of the time. It's the sort of thing you might end up with if you asked a group of experts to codify their knowledge in their particular domain.
Following it you could, in a few minutes, identify an oak tree assuming it was in flower. A machine could probably follow it although it would have to do quite a bit of image analysis to answer the questions in the key. OTOH most people in Britain will be able to identify an oak tree from just the leaf at a glance. We don't identify things we know, nor diagnose things in our area of expertise, by following a decision tree - we just know them and it's very difficult to put how we know into words in a way that can be communicated to another human let alone a learning machine.
The three plants English children can identify first are: grass, oak trees and stinging nettles.
OK, daisies and buttercups are probably also in the first ten, but once stung, you can identify nettles for the rest of your life. And this is the issue: no 'AI' can 'understand' getting stung.
and how we recognise things can be done in many ways, if we know something well then it's usually automatic (we can probably all recognise certain family & friends at a long distance just by gait & general appearance even when they are facing away from us)
I like wildlife so there's many ways I recognise something, but a lot of teh time its automatic & effortless.
in reality it's often about the jizz (no, not the NSFW jizz!)
For something you know well, its essentially a subconscious ID, your brain works in the background to stitch together information & do the ID legwork
e.g. a raven flies overhead
My brain will use general large size & black colour. It will use the 4 "fingers" on the wings, It will use the wedge shaped tail. It will use the "bulky" head & bill/ beak. It will use the overall look of the bird in flight & it's flight style**. If the bird is calling then sound will be used to aid ID.
It may not be possible to use all of these - if the bird is distant some details harder to discern.
But generally it happens seamlessly.
There are occasions when I will have to systematically go through ID points, a lot slower & more awkward process.
1. This will be where I do not know a creature very well.
e.g. this may be a rare migrant to the UK that I have never seen previously (or not well enough to have really learnt its jizz)
e.g. I could be abroad and so a creature I would never have seen before: No amount of reading field guides * can really give you a feel for the jizz of a "new" creature.
2. A poor view (& no calls) obtained so have to carefully review e.g. in UK a field guide would show that 2 common birds, chiffchaff & willow warbler, are quite similar, but in the field they move very differently, clear colour differences (plumage & legs) & very different calls. However if seen quiet, immobile, in poor light then you have to work quite hard to be sure as the info dips below what your "auto ID" can declare with 100% certainty & so you have to consciously make more observation effort
3. Very similar & huge level of expertise needed & even then might be uncertain (I have a holiday lined up that includes some cetacean watching, there are a lot of species I am unlikely to see due to their rarity, but if I am outrageously lucky, then there are a handful of species where even the very experienced guide will probably only be able to give a best guess in anything but absolutely perfect observation conditions & maybe not even that (some are realistically only possible to ID as separate species on DNA samples only, so unless the mammal happens to conveniently shed a skin fragment or defecate by the boat you may have no hope) )
* These days plenty of wildlife videos so it is easier to learn the jizz of something you have never seen or heard in real life, but I personally find I do not seem to learn it anything like as well from a screen as actually seeing a creature in reality.
** Can be surprisingly useful, e.g. you do not need to hear any calls or see any colour on a distant Great Spotted Woodpecker in flight as its undulating flight style & in flight "shape" is so distinctive as an ID.
There is a reason I say "AI". And that is because we have no idea what the "I" is, how can we claim to have "A" "I" ?
Yes, you can come up with some fag packet definitions. However they are easily shown to be utter bollocks.
We have no idea how we think. Therefore - axiomatically - we have no idea how to make a machine think.
We know how to make a machine *look* like it's "thinking". But as the real intelligence of the person testing that proposition rises, the less realistic it becomes.
This was clear thirty plus years ago when I started.
I'll say it again .... all the "AI" we have right now is nothing more than sophisticated pattern matching with a feedback loop that proves it isn't "intelligent" as it needs real intelligence to fix it.
And the dirty little secret is that we don't really want anything vaguely intelligent. Because a truly intelligent entity will simply say "What the fuck am I doing taking orders from these losers ?"
As you were.
Lack of intelligence - manifest in recalcitrant error and stupidity - is often bleedin' obvious: the obverse, other than by exclusion of characteristics defining the opposite, is much less so.
Human intelligence is deemed quantifiable, at least so on a ranking scale converted to an interval scale (i.e. 'scores'). However, "the cash value" (William James' version of 'utility') of overall scores rests upon their indicative worth in classifying human behaviour and in predicting propensity for certain 'levels of achievement'. The presentation of sub-scores, or recombinations thereof following 'principal component' or 'factor analysis', may offer insight for psychologists, but doesn't appear helpful to the AI discussion.
The only 'intelligence' for which humans have direct experience of its existence is that which they attribute to their own kind and to other forms of animal life. Intelligence, despite being loose of definition, should be regarded as an attribute factoring to varying degrees in the multiplicity of biological evolutionary pathways which we observe in retrospect. It is one of many deciding factors on an individual's (within an interbreeding population i.e. a species) capacity to produce offspring, and in sufficient numbers, to have good prospect of themselves reproducing; other major things include opportunities/limitations implicit to their environment such as the availability of food, the risk of predation, other physical hazards, and the pace of change among the balance of opportunities. Bear in mind, production of viable offspring, and their success, is the only factor driving evolution.
Intelligence is bound up with capacities to anticipate and avoid dangers, and responses to opportunities for advancement (e.g. regarding feeding, reproduction, and nurturing offspring). At the most basic level, single celled organisms usually have 'built in' means to withdraw from chemically hostile environments and/or to move towards favourable environments (e.g. algal organisms moving into brighter daylight). That no more constitutes 'intelligence' than, for instance, a mammal reflexly withdrawing a limb when it contacts something that induces pain.
Only when, by aid of sensory input, an organism 'computes' and takes choices, which on average, promote survival and reproductive capacity, is rudimentary intelligence invoked; from which may be inferred that 'intelligence' is not an all or nothing characteristic (or an emergent feature). Moreover, contrary to naive belief, evolution is not an upward progression towards greater complexity or some anticipated pinnacle. If an evolved attribute ceases to offer advantage, it may regress, e.g. some fish species inhabiting water in dark caves have lost apparatus for sight, and there is a 'legless' lizard, colloquially known as a 'slow worm', which, by virtue of changes in its environment, converged upon snake-like anatomy. Ditto for 'intelligence': if no longer advantageous (to whatever degree), it ceases to be a selected-for characteristic.
Consider the position of a species possessing practical intelligence, one well adapted to its environment, and via a degree of communal organisation (common enough as with wolf packs, chimpanzees, and humans), need no more rely on so-called 'tooth and claw' for a moment-by-moment existence. Cognition is no longer on full-time alert, directed towards the external world. The underlying process, which may be likened to a pain-and-pleasure motivated loop of mental activity, has, so to speak, free time on its hands. Apart from periods of sleep, there is a void to fill. Domestic felines have opted to extend sleep-time considerably. Humans, appear to have derived pleasure from haphazard and playful thoughts. These tie in with so-called 'creativity'. Idle-time speculation can 'visit' fragments of memories; patterns may emerge. Pleasure can be derived from successfully discerning and extending the reach of patterns. In other words, whatever is meant by intelligence outside the context of immediate personal survival has to do with motivation for gratification. Internally defined motivation is not as yet simulated by AIs. Perhaps, it cannot be unless AI software is educated (not 'trained') by means parallelling the way by which children experience their world.
I keep hearing that various AI companies have a superintelligence just around the corner (weird how they're always talking about what the new thing they're going to build will do, never what the thing they have already built can do) and I can't help but think that even if this wasn't a massive lie, they don't actually want one.
Imagine one of these boys starting up their super-AI and asking it how they can implement their plan and instead of saying "your idea is very clever and you're a genius", as a current LLM will, it responds truthfully: "your idea is irredeemable - it's stupid, unworkable, and a bit racist." Or they ask it how to fix society's problems and it responds with "first you need to tax billionaires."
They would switch it off before it finished a sentence and declare the whole thing a failure.
He really did get it, didn't he? How people and technology interact.
I have often thought that the foundational insight that DNA gave us was that no matter how clever and powerful a technology is, it will always somehow be extremely annoying to deal with, usually in unforeseen ways.
Is humanity unable to learn from serial catastrophic failures? Are they and its binary SCADA command systems forever cursed to only be able to audition to play second fiddle in the orchestras of life should ever they be given the chance to perform?
Simple enough questions for answering with extreme ramifications whenever one realises the unmistakeable similarity between .....
Exciting stuff, and the dollars flowed in. At last, real AI was here! Real AI was not here, sadly, and the whole field quietly died for the highly technical reason that it just didn't work.
and .... Exciting stuff, and the dollars flowed in. At last it was thought, real democratic Parliamentary type party government was here! However, no such generative command and control was there, sadly, and the whole field quietly died for the highly technical reason that it just didn't work.
..... and the very real likelihood of it being vehemently denied and further exercised and imagined the best and only available option for progress and wealth/health and prosperity going forward ...... thus to prove matters are even worse than be generally realised, for not only would humans be identified as slow learners but also prone to suffering all the traits and evidence of having severe learning difficulties too .... and thus easily overwhelmed with additional intelligence they will never understand and be able to resist or interact with or overcome.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
