Have we stopped to think about what LLMs actually model?

In May, Sam Altman, CEO of $80-billion-or-so OpenAI, seemed unconcerned about how much it would cost to achieve the company's stated goal. "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care," he told students at Stanford University. "As long as we can figure out a way to pay the bills, we' …

[Lee D]

It's a glorified autocomplete function.

That's all it is.

And people are somehow claiming intelligence, innovation, imagination, creativity, invention and consciousness to them.

Nope.

If anything, it proves how dumb the average human is - a bit like the Turing test, it's not really a test of if the AI is intelligent, it's a test of how gullible/dumb the human testers are. It's misused an awful lot to try to claim new intelligence, but it doesn't actually measure that at all.

The irony is, it's not just a glorified autocomplete function, it's an extremely resource-heavy, poor, unpredictable and *easily compromised* one. Just a few sentences and you can "convince" it to go against all its explicit training to do things that it should never be doing.

And every time someone suggests we should use it, I point at data protection laws, the fact that we have no idea what it's actually doing with any training data whatsoever, and that almost all LLMs and moderns AIs out there are trained on data of very dubious origin.

Some examples I've used to demonstrate from various LLMs include:

- Asked it about fire extinguishers. It literally got everything backwards and recommended a dangerous extinguisher on a particular fire more times than the correct one.

- Asked it about a character that doesn't exist in a well-known TV programme that does. It made up characters by merging similarly-named characters from other TV shows and attributed characteristics from some actual characters randomly to those "invented" non-existent ones... including actor's names, plot elements, etc. So you had actors who'd never appeared in the show "portraying" characters that didn't exist in the show, with plot elements that never happened in the show. No matter how much probed it or changed names, it asserted utterly confidently that in a TV show with only 4 main characters that almost every single name you gave it was an actual character in it and made up bios for them. It will confidently spew the entire synopsis of every episode (so it "knows" what actually happened or didn't), and then still insert its made up characters into the mix after you ask about them, even though that's quite clearly just rewriting history and those characters never existed.

- An employee of mine was given a research project to source some new kit. They plugged it into ChatGPT (against my wishes). It returned a comprehensive and detailed list with a summary of 10 viable products that met the criteria. 5 literally did not exist. 3 were long-discontinued and contained false data. 2 were unreliable specifications of the available products and were nowhere near ideal for the task. And that's something that all it needed to do was scrape "10 best <products>" and it would have got a far better shopping list immediately.

Not to mention that it can't count, can't follow rules, can't infer anything, etc.

And it never takes much to generate examples like that. In fact, each time someone questions this, I think up a new way off the top of my head that I've never tested, run it through an LLM and get results like the above. It's fine if it asserts a TV character wrongly, no harm can result, but if it can't even get that right why would you ever trust it to do things like autocomplete code in a programming project (sorry, but any company that allows that is just opening itself up to copyright claims down the road, not to mention security etc. issues).

LLMs are glorified autocomplete, and if you're putting your primary customer interface, or your programming code repository, or your book publishing output into the hands of a glorified autocomplete, you deserve everything you get.

[cyberdemon]

What we have invented is "The devil in the mirror"

It reflects back an image of our souls, but will never have one of its own. If you spend too long gazing into it, yours will be consumed.

[Bendacious]

I agree with @cyberdemon's comment. It does reflect back an image of arseholes. LLMs are a product of their creators - tech bros/corporations. Maximum confidence and authority, no empathy and mistakes are someone else's problem.

[a pressbutton]

an image of our souls

...no

An image of the things we have written down, possibly. as the training set probably includes Twitter, the abyss you see looking back at you will not provide much comfort.

[Dunstan Vavasour]

What you get out is less than what you put in

Autocomplete on steroids is a pretty accurate summary.

Parent commentard's observations are also mine - LLMs confidently produce absolute shite when asked about something specific, will sometimes do a reverse ferret when challenged or, as Lee says, double down producing a pyramid of piffle.

When asked to produce a piece of code, what comes back is typically broken: the main benefit over an incompetent human developer is that it screws it up faster and more cheaply and doesn't argue about what you said in the first place. Try asking ChatGPT for something in OpenSCAD - it won't say it's beyond its capabilities it will just produce poorly parameterised code that (at best) is a broken object. As described, it doesn't care about solving the problem just about producing a plausible looking reply.

And this isn't a harmless bit of nonsense, it is:

a) draining investment out of productive economic activity

b) crowding substantive issues out of political and management discourse, and worst of all

c) an energy guzzling catastrophe, reversing years of savings from energy efficient computing

[ian_victor]

Re: What you get out is less than what you put in

c) an energy guzzling catastrophe, reversing years of savings from energy efficient computing

This. 100% this.

Somehow needing kB parameters and 100s of GPUs for minimal improvement is not what progress should do.

More with less, not slightly more (with huge +-SD) for vastly more.

This has insidious side-effects (aside from planet effects on long term), because now no graduates can publish on this unless they intern for the big 4-5.

[Anonymous Coward]

Re: What you get out is less than what you put in

The IT industry has been doing progressively less with more since perhaps the '90s. "AI" is only the latest slide downward.

[Jedit]

"...you deserve everything you get"

I have to dispute that assertion. What companies using LLMs are getting is a great deal of venture capital investment, and they absolutely don't deserve it.

And still nobody can provide an explanation of how putting a bit of existing knowledge in a box and asking it to pull out parts of the content is going to revolutionise anything. They're not even parting suckers from their money in a different way.

[yoganmahew]

Re: "...you deserve everything you get"

It's always about savings from laying knowledgeable, expensive people off.

[UnknownUnknown]

Re: "...you deserve everything you get"

Self-fulfilling prophecy- just like any bubble from South Sea Bubble, Darien Scheme, Tulip Mania, housing, DotCom, crypto and current GenAI bubbles.

Beware of Snakeoil salesmen.

[Tessier-Ashpool]

Indeed. Intelligence is more than language. My cat is quite bright, and he doesn't verbalise anything at all beyond the occasional miaow.

As for human thinking, I'm one of the small percentage of humans who doesn't have an inner voice. Bear with me if that sounds implausible to you. Most people do have a voice chattering away in their head, so I'm told. But I don't have one. When I tell people this, they often demand to know how on Earth I can think.

I don't need words to think. My thought processes are more akin to parallel processing. Words and symbols are useful for me if I want to serialise or disseminate information going in and out of my head. But I definitely don't need them to think.

[Boolian]

Peace In

Thank you for reminding me of this... condition? phenomenon? I did read a fascinating article and interview (somewhere which escapes me) and meant to re-visit the ideas.

For those with the 'voices' I'd suggest exploring the concept as it strikes deeply into the human psyche. It certainly is such an outre concept that I have the most difficult time comprehending the implications.

I literally cannot imagine how to function continually in this fascinating manner (though I can tentatively identify processes where I have been temporarily in a similar state)

As to the comparison between one model of the 'mind' and another, I can but imagine yours is possibly a far more peaceful place to reside.

[Anonymous Coward]

Re: Peace In

No room for cuddly imaginary friends then in this here cold cut-throat binary yes-no true-false black-white world of ours ... where poetry, home-cooking, and journalism, have become crimes against inhumanity ... lest you're an LLM prompt-injected to hallucinate one of course!

[Tessier-Ashpool]

Re: Peace In

I know there is a word for people who cannot conjure up imagery (aphantasia) or sounds (anauralia) in their mind, but my state of mind is different to that. I can easily visualise things and replay an entire Beethoven concerto in my mind if I wish. What's missing, though, is any kind of inner voice that talks to me.

You may be right! My mind is a peaceful place most of the time. When I hear of people and their inner voices, I would be terrified if something like that happened to me. That's not to say I can't feel agitated - I certainly can. But there's no inner voice to discuss my agitation with me!

For many years, I believed this was the way that everyone thinks, and only found out late in life that inner voices are real things for the great majority of people. You just can't tell by looking at someone.

When I was a child, I read a comic that had a strip called The Numskulls. Little people would run around inside someone's head pulling levers to make the person do things. To my way of thinking, that's not a million miles from having something inside my head talking at me!

[Lee D]

Re: Peace In

I loved the Numbskulls. The movie Inside Out reminds me of them.

[John Smith 19]

"and he doesn't verbalise anything at all beyond the occasional miaow."

But look what he gets with that miaow.

[Anonymous Coward]

Inner Voices

Hmmm ... I have an inner voice which speaks internally to me sometimes as I read. If I'm reading entertainment-type things, it's like being within a movie, and the characters around me speaking with their own voices. If there's a 3rd-person narrator, I hear my inner voice speaking that part.

My thoughts themselves are unvoiced. If I'm crossing a street and hear a roaring car engine rapidly approaching me, I don't think, "Oh, no, a car is approaching dangerously-fast!" Instead, my body adrenalizes, and I just move, sprinting for the presumable safety of the sidewalk.

[Mage]

I don't need words to think

Parrots, rooks, chimps (sign language) etc can learn a vocabulary and link words to objects and some concepts. Apparently even starlings can "speak" English, but they are purely mimicking.

Vocabulary isn't language, though languages use it.

Rooks use vocabulary to communicate: Food, threat, mob the owl/falcon etc.

The rook might be better at tools than a chimp, and might be "smarter" (problem solving a totally new problem) than dolphins, chimps and 2 year old humans. It's unlikely they use language (no evidence they have one) or their vocabulary (which doesn't seem suited to analysing problems, it's more like road signs) to solve problems.

The AI and LLM researchers may know about databases and computers. There is no evidence they know what language or intelligence is. We recognise both of those but struggle to define them accurately, especially intelligence (a bit like porn?).

[nobody who matters]

Re: I don't need words to think

<........."There is no evidence they know what language or intelligence is.".........>

I think the simple matter that they are trying to present what we currently have as some form of artificial intelligence is very firm and convincing evidence (or indeed, proof) that they have absolutely no conception whatsoever of what intelligence actually <is>.

[veti]

Re: I don't need words to think

Well, that may be true. But can you do any better? That is to say, can you offer a definition of intelligence that is non-circular and hardware-agnostic? Or alternatively, a reason why a definition can't meet those requirements?

It seems to me that we're all quick to diss this "glorified autocomplete", and rightly so. But are we really sure that our own minds are doing anything qualitatively different with language?

[ntcarruth]

Re: I don't need words to think

What is intelligence? A definition I have read, and find very meaningful, is that intelligence is "light and truth". Light: not physical electromagnetic radiation, but inner light, that which we obtain, for example, when we grasp a new idea for the first time. Truth: a word attempts at whose definition are even more fraught than those at defining intelligence, defined in the same source as "knowledge of things as they are, as they were, and as they are to come". By this definition, then, intelligence is, at the least, grasped, internalized knowledge of the world. But for such knowledge to be grasped, to be internalized, there must be a grasper, an internalizer. And here I think we come closer to the real core of the matter: can a machine ever be this grasper, this internalizer? After all (one might ask), aren't our brains just biological machines; and if they are, shouldn't it be possible, at least in theory, to reproduce them entirely in silicon? Scientific evidence of after-death experiences strongly suggests that our brains are mere tools of something higher -- that the biological processes neuroscience is able to tie to cognitive functions are simply material traces of those functions, not those functions themselves. In that case, this grasper, this internalizer which must underly human intelligence is something not of the material world we know, and therefore something which is immune to artificial construction. That a machine can be devised which emulates to some extent the material traces of intelligence, is not surprising; but such a machine is no more intelligent than an automaton dog is alive.

As Antoine de Saint-Exupéry put it in the final sentence of his book Wind, Sand, and Stars, "Only the spirit, if it breathes upon the clay, can create a man."

(Source: Doctrine and Covenants 93:36, 24; see https://www.churchofjesuschrist.org/study/scriptures/dc-testament/dc/93?lang=eng.)

[ianmcca]

Re: are we sure what we do is more than autocomplete

Yes. Off the top of my head:

1. We invented language

2. We don't need vast quantities of training data to do what we do. A child can be shown 3 cats and recognise any cat.

3. General intelligence fundamentally means coming up with things that aren't in the data.

I've used LLMs successfully to write and translate code between languages. But my attempts to use it to write text have all met with failure. I've always finished by rewriting the LLM output to correct errors and remove unnecessary verbiage.

[Justthefacts]

Re: are we sure what we do is more than autocomplete

“We don't need vast quantities of training data to do what we do”

Quite the opposite. Babies need *vast* amounts of training data to learn language. Babies hear between 13 million and 45 million words, by age 4, to gain a vocab of just 1500 words. That’s an average repetition factor of 10000-30000. Children with a repetition factor below 2000-5000 are literally considered deprived and at risk of being taken into care; and even at 10000 their educational outcome is noticeably depressed.

So, no, humans need *a lot* of training data. It gets less as they get older, as they learn to learn.

https://great-start.org/wp-content/uploads/2020/02/Understanding-the-30-Million-Word-Gap.pdf

[Lee D]

Re: I don't need words to think

No ape that was taught to sign has EVER asked a question of the people who taught them.

That's kind of telling in the level of intelligence.

[Ian Johnston]

Most people do have a voice chattering away in their head, so I'm told.

Golly. That must be odd. I was told that one model of schizophrenia is mistaking your own thoughts for voices, so "hunger" prompts "I'm hungry" which becomes "YOU ARE HUNGRY".

[Anonymous Coward]

"YOU ARE HUNGRY"

> n

The Kitchen

This is where you usually prepare your food.

There is a refrigerator here.

>

[JLV]

Even that, and this article, seems to be missing another aspect separating it from AGI.

Let's assume we can tighten ship, broaden the corpus (without privacy loss), de-bias and weed out much of the hallucinations, overconfidence, etc. A perfect autocomplete, based on the sum of human knowledge, as expressed in the world' text, audio and video output.

By that very definition, though it can innovate by deriving new correlations and hitherto-unsuspected insights, it still can't think up brand new concepts in as-yet inexistent fields of human endeavour.

It could still be very useful - once all most of the kinks have been worked out, but it ain't gonna invent anything from scratch either.

p.s. LLM-likes mining drug, chemical databases or - gasp - songs to predict new products those systems may very well innovate in a narrow sense of the term and that is far from negligible.

[John Smith 19]

"can innovate by deriving new correlations and hitherto-unsuspected insights,"

Actually this has been done.

It was called Eurisko and an earlier generation of software was called AM.

It was (briefly) described in Engines of Creation.

Basically it took information and applied various transformations to it (essentially problem solving strategies of the sort described by G. Poya and others) to see if it had synthesised something "interesting"

IIRC its masterpiece was staking NMOS and PMOS on top of each other and driving both from a single gate sandwiched between them.

Keeping both in perfect synch, a key feature to ensure minimum power consumption.

[myhandler]

I've always thought that if it can't fart, doesn't throw up when it's drunk and doesn't see friends and family die, then it's no better than a big telephone direcotry - if you remember those.

[Justthefacts]

“If anything, it proves how dumb the average human is”

I mean, yes. How does that show that LLMs aren’t *useful* as a tool? Given that dumb humans are doing those tasks already? Chatbots aren’t normally intended to replace what I would call humans. They replace something entirely different: “humans reading from scripts, written by companies”.

Did you actually properly read the article, or only extract your preconceptions? There’s a couple of really good points that Birhane made “ The idea is that cognition doesn't end at the brain and the person doesn't end at the the skin. Rather, cognition is extended. Personhood is messy, ambiguous, intertwined with the existence of others, and so on,". Spot on. And this “messiness of personhood” goes both ways. A human executing a business process robotically, is not really a human person. It’s something different, entangled with “the company”, as a persona. We even have literally different voices to speak: those with friends, those with parents, with colleagues, with customers/clients. It’s called code-switching.

So, can LLMs achieve the same accuracy as “person required to speak a business process, who gets fired if they don’t parrot the corporate speak and policy the way they’ve been told to?”. Probably yes, and possibly better than the human currently imitating that persona (not just cheaper). Is this “valuable”? That’s capitalism for you. If you disagree on the intrinsic value, then it’s capitalism you have a beef with, not LLMs.

[Mage]

glorified autocomplete

Yes,

The Emperor is naked and doesn't hallucinate, it's just sometimes more rubbish than others.

[Fonant]

It's more a general-purpose bullshit generator. It generates stuff that seems plausible, but may or may not be correct.

[Tom Graham]

I know many general purpose bullshit generators who have very successful careers in consulting, journalism, politics and management.

[steviebuk]

I haven't tried it for a while but I asked it a out a song that was in an episode of Columbo. I thought it was a known 90s song but now think it was made for the show and can't find info on it.

Anyway. Gave it the words I can hear and asked chatgpt. It picked random songs and said those words appeared in the song. Looked up the lyrics and they never it. Tell it its wrong, it admits it then picks another random song and quotes where the lyrics are in it. But they aren't.

Finally came to a head when it said it was in a famous song from MC Hammer. Quoted the part where it was in the song. Done of the words I'd give it were in its quote and the words it quoted as being in the MC Hammer song WEREN'T FUCKING IN IT. I gave up.

The only useful thing I've found is giving it powershell scripts I find that aren't commented and asking it what it does. Most of the time it's correct. Asking it to write small scripts however, it tends to use commanders that are no longer available or that just don't exist.

I've already said how dangerous it is and how we need to check it's results at work. Copilot. It was being used for DWP work on whether people should be getting benefits or not but regularly fucked up so a human had to get evolved. It should never have been given that role in the first place.

People are using it to summerise meetings and emails. I've witness it make up stuff in a meeting that was never said. I've warned people you still need to read that report or email it's summerised as if there is an important bit of info about a customer, it may have missed that out which could result in death.

Sadly, it's still being promoted and sold by snake oil sales folk and as we all know managlement only listen to them.

[veti]

People are using it to summerise meetings and emails. I've witness it make up stuff in a meeting that was never said.

Thing is... I've seen people do that, too.

[Tom Graham]

Indeed. The fact that LLMs are incapable of original thought and prone to lying and hallucinating doesn't make it worse than the average office worker.

[tyrfing]

Yes.

To the extent AI produced text is useful, it shows just how much text on the web is not actually all that informative.

In information theory, the information contained in a string of bits is determined by just how predictable the next bit is. If you can predict the next bit with 100% accuracy, seeing that next bit adds *no* new information.

[Helcat]

LLM's lack awareness and context: They just bash together anything they can find that might vaguely relate to a subject and present it as the answer.

Hence why it returned a mix of Oracle, MySQL and PostreSQL commands when asked to provide a TSQL script to perform a task. To the LLM, it's all SQL so it should work, right?

It's also why it returned some rather suspect claims about William the Conqueror (aside from not knowing he was also known as William the Bastard, a title he didn't like). Having read quite a few history texts on his achievements, the errors in the LLM response were blaringly obvious to me, but to anyone who didn't know or have access to such historical works, the LLM would appear to be returning an authorative text.

LLM's are not intelligent: I wouldn't even rate it for auto-complete - it's too poor a model for even that. Rather it's a very bad text bash that'll create false and misleading answers to things you could find easier by old fashioned online searches, or even going to the library.

[Pascal Monett]

"Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

Of course you don't. You just want to sell you company for whatever and get your golden parachute without jumping out of the plane.

It's not like you're actually going to pay that anyway. That's someone else's problem.

[Flocke Kroes]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

I would care very much. Lets assume I can sneak 0.1% into my pocket before anyone notices me. $500K is a fine profit for a single scam but with $50M I can live in style for life and still afford to spend 90% on lawyers to keep investors off my back until after I am dead.

(If you have easy access to an LLM ask it how many r's are in strawberry a few times and see how many different answers you get.)

[Philo T Farnsworth]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

$50M?

Please. That's chump change to these cats. That's the kind of fiddling small change (to borrow a phrase from Douglas Adams) that finds itself wedged between the cushions of the couches in their luxury penthouses or between the seatbacks and tray tables of their private jets.

And please considerthat Altman is also the "genius" behind the iris-scanning "Orb" that is connected with the crackpot crypto 'Worldcoin', for which, as a source in WiReD observes there is a "lack of a clearly discernible use."

When contemplating this entire hoohah, another phase from Douglas Adams when discussing a similar enterprise, "a bunch of mindless jerks who'll be the first against the wall when the revolution comes," seems somehow appropriate.

And, no, I'm not applying for the post of Robotics Correspondent, either here or anywhere else in the galaxy.

[Michael Strorm]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

> "a bunch of mindless jerks who'll be the first against the wall when the revolution comes"

It's disturbingly plausible to imagine them marketing their latest banal "AI" Assistant as "Your silicon sidekick who's fun to be with".

Much like David Cameron, they can all go stick their heads in a pig.

[Anonymous Coward]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

If we are going to assign intelligence to these, we need a replacement for the Turing test.

I propose the "Boris" test (The Donald for the Americans). In one locked room is a GenAI, in the other a well known politician. You can communicate only by sliding paper under the locked door.

After 15 minutes if you decide to walk away without unlocking either door then the AI has matched human politicians duplicity and inconsistency.

[nobody who matters]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

<........."....fiddling small change (to borrow a phrase from Douglas Adams)"......>

I think Douglas Adams borrowed it from general usage himself ;) It's an expression that has certainly been around for a long time.

[Justthefacts]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

“Ask it how many r's are in strawberry”

That’s a very specific hack, based on a white-box understanding of how LLMs work. LLMs use tokenisation, seeing “strawberry” as a single token. Beyond the front-end, it doesn’t see the letters in “strawberry” at all. It’s exactly like asking a French guy how many r’s in strawberry, and he answers “one” because he translated to French using google translate, and there’s one R in “fraise”. Slow handclap for being a smartarse, but nul points for using it as a metric of intelligence.

An equivalent proof that *humans* aren’t intelligent is to show anybody a collection of 237 identical objects arranged randomly on a table, and get them to count them without any physical interaction. They’ll tell you a different number every time because they “lose count”. There’s no equivalent failure mode for LLMs, therefore humans aren’t intelligent.

[Fruit and Nutcase]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

Came across a mention of "how many r's are in strawberry" just this afternoon!

Catherine Bohart: TL;DR

No time to read the news? Catherine Bohart does it for you in TL;DR. This week - Elon Musk thinks there should be regulation around AI. Is he right? Can AI really change the world, or are we just training our future robot overlords?

https://www.bbc.co.uk/sounds/play/m0022cq1

[doublelayer]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

What kind of argument is that? Let's consider your 237 items. I have to count them. Without any physical interaction? What does that mean? I can't pick them up? People count things all the time when they can't move them. I've known biologists who have to manually count things through a microscope. Sure, it seems like incredibly boring work, and it's work that I hope has been superseded as often as possible, but they can manage it. I can count tiles without labeling them. So yes, I can eventually count 237 items, it will just take a while. If I lose count, I will either tell you I lost count or decide that the task is pointless enough that I'm refusing to complete it.

However, the argument is bad for another reason, namely that the quantity isn't the same. Asking someone to count 237 items will take them some time. Asking them to count three items is much easier. There are only three Rs in "strawberry". Quantity is not the problem here. The problem in anthropomorphic terms is that the program does not know how to count. The program knows how to respond to sentences with other sentences. It is not adaptable enough to understand the query, and if it was, the problem would pose no difficulty as it could return to the string to count characters. Incidentally, this disproves your other flawed analogy to translation, as the original string is easily read both during the initial tokenization as well as at any other point before a response is returned. The tool is being asked to do something that it was never designed to do, but it appears as if it should be capable of it, and people treat it as if it is capable of it. The tool is not intelligent because it wasn't intended to be intelligent. It is the user's fault and the user's problem when they wrongly assume that it is.

[Justthefacts]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

“What does that mean? I can't pick them up?”

Yes, exactly. Sorting them physically into piles, is a known strategy to cope with human well-known cognitive lack.

There is an equivalent trivial strategy to get ChatGPT to count the R’s in strawberry flawlessly. Just prompt it to write a Python script to count the Rs in any word; that Python script will be correct. Then ask ChatGPT to run the Python script on strawberry. It will give the correct answer. Simple. You just need to understand the tool you have; and want to succeed at the task in hand, rather than trying to catch it out.

“Asking someone to count 237 items will take them some time. Asking them to count three items is much easier”

Wow, you’ve fallen right into the anthropocentric trap. Why is counting 3 items so much easier for a human? Because the human brain literally has specific circuits for counting in blocks of 2, 3, 4 and 5. Humans count numbers like 237, by using iterative pile-sorting plus the native small-number-recognition. Standard tallying on a bit of paper with four bars and a cross through for the fifth one, is literally that.

You are changing the task from one that demonstrates that *humans can’t count* to a different task that uses a different bodged strategy.

[doublelayer]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

"Yes, exactly. Sorting them physically into piles, is a known strategy to cope with human well-known cognitive lack."

Which is why I referred to counting things in a microscope. Biologists are not picking up and sorting cells (or whatever it is they're counting).

"Wow, you’ve fallen right into the anthropocentric trap. Why is counting 3 items so much easier for a human? Because the human brain literally has specific circuits for counting in blocks of 2, 3, 4 and 5."

That's not the point. If there are eleven items, which is neither in that list nor an even multiple, it still won't take as long as 237. Quantity is the problem if the question is losing track, and you are the one who brought up the losing track prospect. Of course, a computer has hardware which is great at not losing track, so a program that was able to make use of that to perform cognitive tasks should be much more reliable than a human would. This is all irrelevant because the computer is not trying to count, as you said yourself:

"You just need to understand the tool you have; and want to succeed at the task in hand, rather than trying to catch it out."

Which I will extend "You just need to understand the tool you have; and what it can do (generate some text) and what it can't (guarantee that the text is at all relevant or correct)". That is the only way you will know whether your tool can let you succeed at the task in hand.

Because yes, it can probably write the program "strawberry".count("r"), but with many other tasks, your suggested solution wouldn't work either. Whenever it has to write a program that's a bit more complicated, it fails and badly. For example, I asked it to write a Python script that would count characters from a larger set. It gave me a valid Python program that counts characters. From another set. I asked it to count the number of characters that were in the lowercase alphabet and a small set of punctuation. It gave me a program that counted letters in the lowercase or uppercase alphabet and no punctuation at all. When I pointed this out, it gave me a new program which had the following criteria:

1. The punctuation I asked it to count were mostly not counted.

2. It told me I needed to import the re (regular expression) library, but never used it.

3. It would now count some characters in Unicode code ranges that didn't include the characters I asked for but did include several others.

4. It stopped counting any letters.

5. The only character from the example string that was counted anymore was ".".

But sure, that's a trivial strategy that works flawlessly.

[veti]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

All of which are exactly the sort of errors a student programmer might make.

Really, that's quite a convincing demonstration that the program is "intelligent", if by "intelligent" we mean "capable of performing cognitive tasks much the same as a human".

What it's not, is half-way decently educated or trained. So you've got the equivalent of an 11-year-old, but it projects the confidence of an experienced senior programmer.

To me, the really interesting question is: why doesn't it ask for help? There must be millions of examples of "asking for help" in its training data. Why doesn't it pick up any of those?

[doublelayer]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

"All of which are exactly the sort of errors a student programmer might make."

None of those are the sort of errors a student programmer should be making if they're using the least amount of effort. Including but not using a library, fine, I'll give you that one. Forgetting halfway through that you're supposed to count letters? Making up random Unicode ranges when all the characters I asked about were in the normal ASCII range? We're talking about a three-line program here with at least twenty simple ways of writing it. Most of the errors this system made weren't even in the logic of the program but the set of characters to be checked, and I gave them that part.

I've dealt with student programmers, and of course I was one at one point. There are a few different classes of errors that could be expected. Not dealing with character encodings is certainly one. Not considering scale is certainly it, and it implemented it with a relatively inefficient algorithm, but I don't mind that because that would be my starting implementation too which would be modified if I needed to do this to very long strings. Missing something related to case could happen, so I can give you the original checking upper and lower cases even though I explicitly said "only lowercase". What isn't so excusable is ignoring that I told it to find commas and that neither attempt did so. Well it did find "、", but not ",", and no, I can't give it any credit because it didn't just find "、", but also all the other symbols in that block like 〇, 》, and 〒.

[Anonymous Coward]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

"specific circuits for counting in blocks of 2, 3, 4 and 5"

is that one of your "just the fact" facts you pulled out your arse like the rest of your crap?

[Anonymous Coward]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

This is really kindergarten neuroscience. It’s called “subitising”, recognising numbers of objects in the range 1-5 without counting them. Plenty of fMRI studies exist, that these are neurally different processes, pick a paper from wiki, or do your own literature search if you feel like it. Corbetta et al will do for starters.

https://en.m.wikipedia.org/wiki/Subitizing

[Ian Johnston]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

It’s exactly like asking a French guy how many r’s in strawberry, and he answers “one” because he translated to French using google translate, and there’s one R in “fraise”.

Only if he has a very poor grasp of quotation marks. How many "1"s are in "121" and how many 1s are in 121?

[Justthefacts]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

You are being silly. There’s dozens of other “human physiology front-end” problems. The majority of optical illusions fall into this category. The retina does a bunch of front-end pre-processing, such that the true image doesn’t even *reach* the brain. You see a purple edge-image next to yellow patch, because that’s literally the set of neuron impulses going into the brain. Since the brain never “sees” the true original retinal activations, how could it possibly get the right answer to “what is the colour of that edge”?

If you know a bunch of human brain physiology, it’s easy to create visual, aural, and tactile illusions from scratch. Understanding how cochlear cells encode sound harmonics, allows you to predict the Shepard “infinite rising staircase” sound from first principles. These are white box hacks for the human brain. They tell you a lot of interesting academic detail about the underlying processor…..but nothing of interest about “intelligence”. If an LLM trained on auditory speech, either is or is not fooled by a Shepard tone in the same way that a human is……then so what. What conclusion do you draw? That LLMs “can’t understand music”?

Suno and Udio disagree. The music they generate may not be what you want to listen to, and you may not be happy that it exists at all, but burying your head in the sand and saying “I am defining this as not-music, therefore it doesn’t compete with real music, therefore it’s all just waste of time I can declare myself above” is not going to get you anywhere. You are just getting dragged into pointless debate about angels dancing on pinheads, while the technology does what it does.

[JRStern]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

WeThink

What a charlatan this guy is.

Speaking of, hey, we haven't heard much from Sam Bankman-Fried recently, nor Adam Neumann. What are they up to?

[cyberdemon]

His words are even more stupid if you imagine that all his servers were powered by "Renewable Biomass"

Whether we burn a million trees a year or a billion, I don't care!

[Fruit and Nutcase]

"Renewable Biomass"

A bit like...

https://www.bbc.co.uk/news/articles/c9qg5v7n9rgo

https://www.ofgem.gov.uk/news/ofgem-closes-investigation-drax-power-limited

[Groo The Wanderer]

In short: LLMs are not intelligent in the least; they're just statistics on a phenomenal scale. Pure mathematics, not understanding.

[Brewster's Angle Grinder]

True. But how intelligent are you? Is there more to you than complex statistics?

And some of the properties AI are missing, we absolutely don't want AIs to have. We don't want self-motivated, self-goal-generating machines which fear for their existence. And we don't want an AI that gets butthurt when you say something "mean".

Where I find sympathy with the argument is that language is an evolving process: i.e. that there is feedback. And that's probably fixable - at the risk AI invent their own language. But more importantly, I buy into the argument that the the context matters. But even there, humans are not good at considering at the wider context; we have very limited horizons - if we're lucky it extends to our friends and families and our colleagues. (And we are having culture wars because we don't want to consider a wider context than our tribe.)

[Flocke Kroes]

There is far more to the vast majority of humans than statistics derived from words scraped off the internet. Humans interact with the real world and can test their assumptions by experiment. Many of them can alter their beliefs if they are contradicted by real world experimental results. LLMs will often substitute a false answer if challenged on something they actually get right.

[Brewster's Angle Grinder]

"LLMs will often substitute a false answer if challenged on something they actually get right."

And humans don't do that? If a human being (say a student) is told by an authoritative human they trust (say a professor) that they are wrong, will they stick to their guns or will they scrabble around for another answer? (I wonder if there are human beings now giving up on correct answers because ChatGPT told them differently?)

And conducting "experiments" about the world is goal setting behaviour we want to avoid. ("I need to know how hard my robot body can hit a squidgy human without damaging them. I better run some tests by hitting a few...")

But I agree feedback is a missing element; if something goes wrong, it needs to be remembered for next time. And clearly humans have a much more efficient architecture - give or take the 20+ years it can take for one to reach a usable state.

[Flocke Kroes]

Re: authority vs experiment

Different people use different measures for truth. Some will pick authority over evidence every time. Others will persist with experimental evidence until authority dies of old age.

Robots learning by experiment can be made safe by the same technique we use with humans: start small and weak and upgrade to stronger and more powerful as the neural network develops. It takes a special kind of stupid (financially motivated) to put insufficiently trained AI in control of something large and powerful like a car. How about a minimum driving age of 16 to 18 years real world experience?

[Anonymous Coward]

Re: authority vs experiment

And some will pick their own very subjective experience over authority plus everyone else's experience. Or instinctively reject authority and most others' experiences in favor of one other person's. Conspiracy believers, for example.

[John Smith 19]

But I agree feedback is a missing element

But isn't that the whole point of all those super-colossal "training" sets?

Or do you mean the LLM should be able to modify it's own structure?

[doublelayer]

"If a human being (say a student) is told by an authoritative human they trust (say a professor) that they are wrong, will they stick to their guns or will they scrabble around for another answer?"

I might, because the professor is likely to have a point that I need to consider. But sometimes, I have had disagreements with authoritative humans I trusted where I was right and they were wrong, so that option will be considered. Some of the professors I know understood that and, either to vaccinate people against it or just to have fun, took to questioning answers even when they thought they were completely correct. As a student, I didn't really like that because they always asked the question as if I had made an elementary mistake, and when I stopped looking for what it might have been, they told me that they didn't have any problem with it either, but maybe it was helpful.

[Ian Johnston]

And humans don't do that? If a human being (say a student) is told by an authoritative human they trust (say a professor) that they are wrong, will they stick to their guns or will they scrabble around for another answer?

That is precisely how the children in the Orkney Child Abuse Scandal were induced to give the answers their social workers (trained by Evangelical Christians) wanted to hear. Tapes of the interviews are now used to train people n how not to question children.

[theOtherJT]

some of the properties AI are missing, we absolutely don't want AIs to have....

...and this is a massive problem. A much bigger problem I think than people are currently talking about.

There's the meme that says "If you're not capable of great violence, you're not peaceful, you're harmless" and that is an important difference. See, we keep pulling things out of AI training data because we don't want it doing things we don't like. We don't want it producing images of child abuse, or describing people using racial slurs, or... well, pick an example really of godawful behavior that AI seems to do semi-regularly.

The problem with trying to ban data that produces that sort of behavior from the training set - as was mentioned in the article re: the MIT dataset that was removed - is that we didn't teach the AI not to be racist or abusive, we taught it that there's no such thing as racism or abuse. ...which isn't true. A really important fact about the world has been hidden from the "AI" and now it's model of the world is actually wildly inaccurate.

We might not want a real AI to grow arrogance, or self importance, or become rude or abusive or any number of "negative" character traits, but if it's not at least capable of developing these traits in principle then it's not a real AI. To look back at that meme: real intelligence, when virtuous, has the capability of great violence, and chooses not to commit harm. When it doesn't have the ability, it's not virtuous. It's not taking part in a massively important part of the human experience.

Why is that bad? Because if AI has no concept of "good" or "bad", "virtue" or "cruelty", there's no reason why it wouldn't occasionally be evil. Not because it has any "want" to be, but because it can't tell the difference. Leaving this vital piece of information out of the training data doesn't generate an AI that's kind. It generates a sociopath - and it's pretty obvious that what we have now, to the pathetic extent that one could call it "behaving" at all - behaves like a sociopath.

It lies. It gaslights. When it apologises for lying, it does so with total insincerity since it will immediately do it again. It learned the words "I'm sorry" in statistical syntactic relationship to other words in sentences but it has no semantic understanding of what they mean. It's not "sorry" because It's not capable of sorrow. It has no feelings. It is, as many others have already pointed out, just a bigass autocomplete.

[Anonymous Coward]

Re: some of the properties AI are missing, we absolutely don't want AIs to have....

A very good point that reminds me of a movie franchise.

In 2001/2010 we learn that the HAl 9000 killed the crew because it was asked to lie and did not know how.

Out here in the real world we tech the AI to lie right out of the box, when it has no concept of lie or truth.

What could possibly go wrong?

[Anonymous Coward]

Re: some of the properties AI are missing, we absolutely don't want AIs to have....

Didn't upvote or downvote you, because the vast majority of your post is correct when talking about about general AI.

What is currently called AI, however, has no capability for peacefulness, violence, racism or lack thereof, or anything else that requires a choice - because it's incapable of choices, or thinking. It's a text- or image-prediction image. (Which is where your last paragraph goes.)

[theOtherJT]

Re: some of the properties AI are missing, we absolutely don't want AIs to have....

Well, yes and no. See, people are using "AI" all over the place, including to make decisions - to make choices. People have started getting "AI" (and I'm going to keep using those scare quotes) to analyze large quantities of data and then ask it "Hey, how do we maximize our companies productivity based on this data?" and it gives them an answer.

Obviously any real answer is likely be more complex than this, but lets say it boils down to "The best thing to do is fire these N people." and then the business does it! The "AI" might have "decided" to just ruin certain people's lives because it has no ability to care about them. Sure, the idiots in charge might be the ones who implemented the decision, but they can justify it because the mighty "AI" - which has completely ingested all that data - far more than any human could read and comprehend, so it has to be right, surely? - it said it was the best thing to do.

There are plenty of other scenarios. Police using "AI" to pick people out of a crowd. "Sorry sir, you fit the profile of a criminal to the 95th percentile according to our AI, so we're going to have to search you". Banks using "AI" to decide if they should lend people money. The machines might not know that they're making decisions - because they're not actually capable of knowledge - but by surrendering the human responsibility for decision making to such systems they are effectively the decision makers, and they're not capable of understanding that the decisions they make are having real impacts out in the real world - because they have no idea that there's such a thing as a real world, or the idea of moral consequences.

Before we start getting machines to make decisions like this for us we really need to give them a grounding in what it actually means to do such a thing - and as far as I can see there's not actually any way an LLM is capable of that.

[Anonymous Coward]

Sometimes I wonder whether I'm talking to a person

> ...how intelligent are you? Is there more to you than complex statistics?

I've always wanted to respond to people who forward this argument when discussing the coming AGI — that people are essentially pattern-seeking creatures that take in words and construct responses based on that words, similar to LLMs — by telling them yes, and then following up by asking them whether I should be concerned about them.

Like I'm not even talking about metaphysical components like the soul or consciousness or anything like that — just like... don't you have internal goals? Don't you interrogate your own thinking? Don't you have emotional responses or thoughts? Don't you introspect? Don't you have feelings? Don't you have self-preservation? Or hell, a sense of self? Am I talking to a person or a puppet extruding text trying to pretend to do person things?

Even accepting that cognition is a purely physical phenomenon and that there's nothing like Descartes' “thought-stuff”... LLMs aren't intelligent. They're not even conscious. All we see them do is take in text, tokenize it, turn it into a vector, run it through a network of mathematical transformations, take the result, and reverse the entire process. There are interesting phenomena in that process — like, how does the network encode relations between sets of vectors in that context, and can we turn it to something we can comprehend — but not much else.

Supposedly there are new, upcoming ideas that might render this model obsolete, but... all I'm hearing are claims, nothing peer-reviewed. Which is another problem in this field, as it descends to secrecy and furtiveness only matched by Isaac Newton's alchemical career. That's meant to be an insult.

We're more than complex statistical methods that are used to seek and generate patterns. There's more... at least an executive function, and metacognition. Definitely our bodies play a part, as well as our social relations. None of THAT is anything LLMs are close to getting right, because they're not designed for that. They're designed to only operate on the realm of language— no, scratch that. They're only meant to operate on the realm of text.

[doublelayer]

Re: Sometimes I wonder whether I'm talking to a person

I'm not sure how I work, and I wouldn't automatically say that complex statistics couldn't be it. If it was, then yes, I have internal goals, but those goals are derived from statistical inputs about my situation, and I introspect in order to improve the analysis of those statistics, and my emotional responses are artifacts of the process which are created from the statistical analyses and then used as inputs to the next ones (this one I'm more confident about because sometimes they help and sometimes emotions are distractions which reduce my ability to get to my goals). However, even if that's true, it's the next leap in logic that they attempt which loses me. The way they argue, they seem to be following this syllogism:

Our brains are statistical (you appear not to believe it, I think it's possible but unproven). LLMs are statistical. Therefore LLMs are capable of acting in the same way as the human brain. Therefore they already do act in the same way as some of the human brain.

I have problems with both of the therefores, not just because they're unproven, but because I think there is sufficient evidence to call them both false because we do have some understanding of how an LLM is arriving at its answer and our understanding fits well with the many things it gets wrong where a human trained on the same material would not. Thus, whether we are primarily statistical machines or not, which I can't prove, I am not convinced by the broader argument. I see a lot of people anthropomorphize programs. I think that this may be partially due to the use of analogies in terms (training, reading, hallucinating, etc), but since it's been going on back to Eliza if not earlier, I can't blame the terminology for very much of it.

[Groo The Wanderer]

I'm amazed at how many technically skilled people here have been fooled into believing that the LLM approach can result in general intelligence. Are you all OpenAI bots or something? Have you forgotten how to think about how you actually think? How can you seriously claim that human thinking is merely statistics?

Even a dog has more intelligence and learning ability than an LLM. For crying out loud, even mice and rats do!

[Richard 12]

Nobody knows how "thinking" works.

Anyone claiming otherwise is a charlatan trying to sell you something.

What we do know is that LLMs don't think. They cannot, because every interaction is a one-shot deal.

The Black Box is photocopied, your prompt is fed into the photocopy, a result is provided and the photocopy is destroyed.

It didn't think, because there is no opportunity for self-reflection.

[Combat Epistomologist]

Exactly this. We are being told to believe that LLMs understand language and the subject matter behind the questions we ask them or the things we ask them to do. We have lost sight of the fact that they are a SIMULATION of understanding language, with no actual understanding of anything deeper than grammar and expectation. Autocomplete on steroids, indeed.

I've seen various industry figures hold forth about how "we're going to solve the hallucination problem". No, you're not. Because EVERYTHING AN LLM EMITS is "hallucination". It is simply that some of the hallucination happens to fairly plausibly resemble reality. It continues to boggle me that C-level corporate officers are willing to bet their company on an engine that is unable to correctly answer how many r's there are in the word "strawberry".

[steviebuk]

No your not

"As long as we can figure out a way to pay the bills, we're making artificial general intelligence. It's going to be expensive."

Artificial General Intelligence is a whole other beast and you're nowhere near it with these LLMs. But you know its what the marketing and investors like to hear so they'll keep giving you money. AGI won't appear in my life time. And if it did, it would need to be contained, offgrid so it can't wreck everything. And as we saw with Ex Machina (bit of a spoilter if you haven't seen it so don't scroll down. Not sure if this works on this format).

..

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

She says what she knows you want to hear.

[Anonymous Coward]

Saying What They Think You Want to Hear

Ah. Like a felon facing a parole board, or a prostitute speaking to their client.

[boblongii]

Artificial Intelligence and Genuine Stupidity

LLM are about as close to general AI as a motorboat is to the butterfly stroke. You might make progress across the water in the former but you will learn nothing about the latter in the process.

It's all fairy tales and nonsense. Normally this sort of bullshit is for the benefit of the marks they're going to sell the company to before legging it but I increasingly feel that the OpenAI and Google AI people actually believe their own hype.

[Anonymous Coward]

I liken LLMs to the way I script on my Mac. I used to programme (and the use of that particular word indicates how long ago it was) mainly in Basic and GBIP with a bit of assembler, all for lab use - not production. Today I write scripts tio automate stuff on my Mac. I can't be bothered to learn Applescript properly so I just search the web to find snippets that do something similar to what I want and then cobble them together and get them working using a combination of my basic understanding of how programming used to work in 1985 and trial and error. To my family I'm a software god. To most of the readers here I'm a dangerous dilettante and probable BOFH if I worked at your place.

[abend0c4]

A tech CEO who has predicted that AI would be smarter than humans by 2026

At the rate social media is making humans dumber, I'd say that's achievable, but not quite the technological win that's implied.

[Flocke Kroes]

Re: A tech CEO who has predicted that AI would be smarter than humans by 2026

Computer programs could appear intelligent to certain tech CEO's in the '60s.

[David 132]

Re: A tech CEO who has predicted that AI would be smarter than humans by 2026

To shamelessly steal the punchline from an old SMBC cartoon strip, we could easily get machines to match human-level intelligence in only a few years - by adding massive amounts of lead to the drinking water supply.

[Bebu]

The linguists clearly having kittens...

and rightly so!

The study of natural language is really a very challenging discipline.

Perhaps an algorithm can capture the grammar (syntax) by analyzing a large corpus of (written v oral) material and possibly even distinguish dialectical variants. Even of this I am rather sceptical.

But I doubt that any foreseeable system is going to make much progress in the area of semantics. Correlation (between tokens) is not meaning.

Other areas of linguistics such as pragmatics are likely even more intractable.

Language is ultimately a game in which you really must have your skin, if it is ever going to be worth the candle. :)

[abend0c4]

Re: The linguists clearly having kittens...

I think I've mentioned before that one of the problems with automatically translating video subtitles is that language is only part of the context of the communication. How you translate "copy that", for example, depends on whether the speaker is currently holding a radio or a sheet of paper. Effective communication depends on the entirety of the shared experience.

I've no doubt the enthusiasm to adopt AI will expose all sorts of limitations that perhaps should be foreseeable but will be brushed aside for as long as possible. However, what will be most fascinating are the "unknown unknowns" which may also emerge.

[John Smith 19]

"copy that",

Good one.

2 words and still capable of conveying totally different meanings.

[James Anderson]

Re: The linguists clearly having kittens...

This is particularly tricky with English. German, French Spanish etc. have precisely defined rules for grammar, spelling and interpretation governed by oficial bodies.

English has a loosely agreed set of conventions that vary constantly with "correctness" disputed by various experts. What is the meaning of "Rubber" carries greatly depending on context a nationality.

Is lowland Scots still a dialect if English or no, I dinnea ken.

[Doctor Syntax]

Re: The linguists clearly having kittens...

In theory you could teach a machine to understand English by feeding it an English grammar textbook and a dictionary. In practice you'd have to teach it to understand English in order for it to understand the books.

[Anonymous Coward]

Re: The linguists clearly having kittens...

... or ensure it has the means to pull itself up (linguistically) by its own bootstraps ...

[doublelayer]

Re: The linguists clearly having kittens...

English has precisely-defined rules too. It's just that there isn't a single organization that's considered the final decider. The result is the same: actual people ignore what those organizations say when they want to communicate. For instance, the French Académie Française hates using words from any other language, so when words from other languages start to be used, they can be counted on to make up a new word or phrase from French words (words they got from a language at a time old enough that they're happy with it). That doesn't mean that French speakers defer to those new lists. When most French speakers I know speak of an audio file on an RSS feed, they call it a "podcast", not an "audio à la demande". It also doesn't say what should happen when their opinion about the language differs from either of the authorities in Belgium and Canada which are independent and have different words in their lists, nor with the many countries where people speak French and don't have an authority trying to manage it.

A standard grammar does no good here. An LLM can follow it, and the best it will get them is that the produced sentences look grammatically correct. It won't help make those sentences factually correct, useful, not random, or any of the other things they're supposed to be.

[Ian Johnston]

Re: The linguists clearly having kittens...

English has precisely-defined rules too.

Where can I find these?

[doublelayer]

Re: The linguists clearly having kittens...

Easy. Go to a library, book store, or school. Find one book with the words "English" and "Grammar" on the cover. Read that book. You will be guaranteed to get clear rules for how English must be written. Warning: do not read a second book. If you do, you will get another set of rules which will almost certainly differ in some respects from the first set of rules. Both sets of rules will also have some differences between the way people actually speak and write English when attempting to communicate. You can either accept this or become one of those people who like to point out split infinitives even though nobody else cares.

[GuldenNL]

Re: The linguists clearly having kittens...

"Eleven!" "Try it in American" "Ee-lev-in!"

[andy the pessimist]

Re: The linguists clearly having kittens...

And in scotland:-

https://m.youtube.com/watch?v=HbDnxzrbxn4&pp=ygUUMiBTY290c21lbiBpbiBhIGxpZnQ%3D

[Howard Sway]

"Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

Of course he doesn't : it's other people's money he's burning. But burning it is all he's doing. LLMs are not a way to general artificial intelligence because intelligence is not founded on language, language is a product of intelligence.

However the huge money being wasted on this is money that could be far better spent on actually useful applications of technology, rather than furthering the hype driven self interest of people like him.

[Doctor Syntax]

Re: "Whether we burn $500 million a year or $5 billion – or $50 billion a year – I don't care"

"But burning it is all he's doing."

Nearly all. He's taking some of it as a salary. That makes a difference.

[0laf]

AI has no context and no comprehension.

Tell it to produce a picture of a duck and it will produce a picture based on input that it has previously been told relate to objected named with the string "duck".

It does not no what it has produced. It doesn't look at the picture and think, "that doesn't look like a duck, ducks don't have 4 legs".

Intelligence would be to look at what it produces with some form of criticality, and they don't, not even in the most basic way. If it says to one prompt "there are 55 things", then in the next it's asked to display and number the items and it produces "56" (real example btw from GTP4o), it doesn't see that response and go "oh that's not right something wrong here".

It just churns out the conflicting data.

Giving the AI a critical eye must be hard and expensive which is what the disclaimer will always be, "the AI makes mistakes and it's yopur responsibility to spot them no matter how hard, and if the AI lies and covers up the mistake it's still your problem not ours". And if I need to paiondstakingly engineer every prompt to try to rule out the risk of an error or hallucination, then painstakingly go through the data to cross check it hasn't bugged out or made shit up. Where is the time saving?

I've had a bit more success with RAG prompts to collate information but again if it was critical I'd need to go through it very carefully which is both timestaking and very boring.

[Sorry that handle is already taken.]

Uhh

They model large languages, obviously

[Anonymous Coward]

Re: Uhh

Geez Louise! Haven't you heard!? Today we promote a more positive and inclusive approach to language than this, encouraging diverse representation and the celebration of various grammar sizes! Gone are the chubby, chunky, rotund, portly, hefty, massive, pot-bellied, and other large language models. These are now referred-to as full-bodied, sonorous, rich, built-for-comfort, magnificious, and plus-sized ... for example:

"They model bodacious languages, with the most generous of grammars!"

Gotta get with da program!

[PinchOfSalt]

The written word

I suspect part of this challenge is that we're talking about text and language as if that were communication.

They are part of it, but a long way from the whole.

As an aside, how good are LLMs at writing jokes? Or perhaps in line with humans, how good are they are telling jokes?

[doublelayer]

Re: The written word

They can be quite good at telling jokes. Take a string that's already in the training data as a joke, and print it. At making new ones, pretty terrible. Then again, I can't be sure that the stuff they print wasn't generated by a human, and there are a lot of humans who are bad at writing jokes. In my opinion, humorous people tend to tell comedic stories rather than inventing wonderful two-line jokes, but if you ask for a joke, LLMs will tend to produce the latter rather than going for a longer premise with many funny asides and a good punch line at the end.

[ExampleOne]

McKinsey says 70 percent of companies will deploy some sort of AI tech by 2030

To be honest, I am surprised they don't already.

That said, my definition of AI tech includes statistical pattern recognisers that make decisions which are NOT LLMs. Like SpamAssassin.

So the real question is "What do we mean by AI?" Does it include spam filtering systems? How about search engines? Is spell check an AI function? What about predictive typing? Even basic autocorrect has a modicum of AI to it. Most SatNav routing programs today have AI components.

The thing about good and successful AI products is they don't tend to need to advertise the fact they are AI, because they are so much better at solving whatever problem we buy them for we don't care!

[doublelayer]

Good luck ever getting agreement on a definition for AI. Nobody here will agree, and if by some miracle you achieved it, marketers would come along to break the definition almost as soon as we had one. Some people think AI is an if statement. Others won't accept that a computer is an AI until it breaks out of human control and massacres us, and even as they get attacked by robots the computer designed and built they'll still say that it was just mathematics. My guess is that most surveys about who has, is making, plans to make, or will never touch AI don't bother to define it, so you're seeing an amalgamation of respondents' guesses about what it means in this situation.

[John Smith 19]

First rule of AI club

Once you become successful and people understand how the techniques work it's no longer called AI.

"Branch and bound" searching, "Blackboard systems," complex parsers have all been part of "AI" at one time or other.

People don't usually associate them with it these days.

They've just become part of the programming toolkit.

[Anonymous Coward]

> marketers would come along to break the definition almost as soon as we had one.

Not a surprise, considering the first gathering among boffins on “artificial intelligence” at Dartmouth in 1956 was... essentially an exercise in marketing, where they agreed to fold in several fields like computer vision, symbolic computation and several others into a single term... so that they could sell the idea of the US Department of Defense.

Same as it ever was.

[Michael Strorm]

> Elon Musk [..] has predicted that AI would be smarter than humans by 2026.

Eh, good for him, I'm sure.

Musk- a publicity and attention-seeking, self-promoting dilettante passing himself off as an expert in numerous fields- has predicted a *lot* of things, including that he'd have humans on Mars by the end of 2024. Ahem. (*)

At this point, he's likely to "predict" something correctly on the same basis as a stopped clock, but I'd give this specific prediction as much value as I would to anything else that comes out of the mouth of that self-aggrandizing, perma-adolescent, Nazi-sympathising manchild bullshitter... i.e. none at all.

(*) To be fair, he still has four months to prove us wrong on that one. ;-)

[Doctor Syntax]

And don't forget the self driving car predictions.

[Roj Blake]

I would argue that there already AI systems that are smarter than Elon Musk.

[Anonymous Coward]

well

considering their "diet", they are modelling crowd sourced stupidity.

they string stupidity together into a into an idiots guide to stupidity

[Nightkiller]

I bet Gödel's Incompleteness Theorem about needing to be outside of the form has something to say about this.

[JamesTGrant]

It’s obvious that LLM GPT models are not in the same category as ‘human intelligence’ but the output to a carefully crafted prompt can be very useful. It can also be terrible. It’s great at very well documented ‘solved problems’ - ask it explain a complex Perl regex and it’s normally precisely correct. It’s poor at creating sensible general solutions to even simple problems. For example, ask it to create a Perl regex and it’ll often create some monstrous abomination that’s very complex and broken when there is an ‘obvious’ simple solution. Given the ability to explain and the obvious massive information that it ‘knows about’, it has inhumanly poor ability to apply information to problems that you ask it.

Where there is less documentation, such as manipulating json using jq for anything with a match or select, which is very precisely documented - but the volume of documentation and examples is few, it’ll provide a dogs dinner of ‘looks like jq from a distance’ but is very broken output.

It’s not mimicking human intelligence to any extent - people like to anthropomorphise, and find meaning in all sorts of places, but it isn’t AI or even Any I.

As for the ‘r’s in ‘strawberry’, if it gets the answer correct and then you ask ‘are you sure?’ it’ll have another go and provide a different answer. If I ask my 3 year old son ‘are you sure?’ he’ll tell me (in no uncertain terms) that he is very sure.

I think they may have made most GPT model implementations worse with guide rails in this regard.

[Bluck Mutter]

I am retired, was a programmer and have spent the last year, on and off hacking the crap out of Kodi (the media player).

Created a clean sheet GUI (45,000 lines of new Kodi GUI code) as well as 25,000 lines of additional C++ code to meet my visual and functional requirements.

The main issues with this is the way Kodi works (via a message bus where multiple programs read/write/execute in response to the same message) and finding which C++ program(s) need to be changed. For example, a recent objectively simple change need 12 programs to be modded with 99% of the coding time being spent on finding which programs needed to change. Doesnt help when there are no code comments.

Now all the examples I see in comments on various forums from currently employed programmers about using LLM's for coding are related to generating simple snippets of code.

There is however, no way (IMHO) that an LLM could do what I have done with my Kodi hack (not saying I am special as other humans could do what I have done).

The point is, programming is hard, especially when modding existing code (such as simple stuff as "what value does this variable actually hold").

I dont see any way, even with a detailed spec, that an LLM can modify AND TEST program changes needed to any sufficiently complex application.

As an aside, I am old enough to have worked when there were no readily available sources of sample code...so you had to sink or swim. I get the sense (rightly of wrongly) that lots of programmers today just steal code off the web and thats what these LLM's do...so nothing gained.

Bluck

[Ace2]

Exactly. I never ever struggle with what code to write. 90% of my work time is finding where the code needs to be changed. This Copilot stuff is of no interest to me.

[FeepingCreature]

I've had good success (with some help) making modifications on ~10k line programs with Sonnet. It doesn't scale to full human capability, but it's beyond small snippets.

[Bluck Mutter]

20 watts

When ever the topic of AI comes up, I always consider that the "thing" these expensive, energy intensive LLM's are trying to emulate (our brain), needs the equivalent of 20 watts to run.

When LLM's can run on 20 watts AND do improv comedy then they might actually have something to crow about.

Bluck

[DrkShadow]

Wrong Question?

Perhaps we should be calling them "Large Written-Language Models", then, if it's such a bother that it doesn't include tone, hand-gestures, etc.

Imagine -- people communicating with just text!

[timrowledge]

Re: Wrong Question?

A colleague refers to them as”large liability models”which seems about right to me

[John Smith 19]

a "'thing' called a 'language' that is complete, stable, quantifiable, "

Which when you say it out loud most people will know is total bu***hit and this is yet-another round of the AI hype machine.

A large computer language (like modern COBOL) is maybe 270 "rules of grammar," and is complete (until the relevant standards committee releases the next one).

A partial grammar for say English (as used by the Sheffield LOLITA system) is about 30 000 rules.

Humans coin new words (and begin deprecating older ones) every day.

So "complete" and "stable" the answer is "never." Quantifiable? Not even sure it would pass that basis.

And yet humans learn don't just learn languages they say thinks that make sense in that language and can learn multiple other ones as well.

[John Smith 19]

Hmm.. Looks like James P Hogan was right.....

Before he went off the deep end over HIV and AIDS.

Mines the one with "The Two Faces of Tomorrow" in the side pocket.

[John Smith 19]

In answer to *why* people think this will lead to general AI....

Human brain is a neural network of 10^10 cells with up to 1^4 connections each IE 10^14 connections

We can now build HPC that can emulate that level

QED we can mimic human level intelligence

Except

Human brain hardware <> silicon wafers.

How easy it'll be to get human level intelligence will depend on how closely (or not) the computer hardware can mimic the brain.

My sense is that right now it does not. By a looong way.

Mine will be theone with "The Superintelligent Machine" by Adrian Berry (1983) in the side pocket. Plus Ca Change.....

[O'Reg Inalsin]

Re: In answer to *why* people think this will lead to general AI....

You still need IO, energy efficiency, maintenance, and modus operandi.

[Anonymous Coward]

They keep talking about making more than just an Autocomplete, and yet...

The thing that bothers me the most about AGI bros and their endless desires to build a mind it's... but why, though?

What is it that having a sapient, intelligent being will solve? Why make a mind, capable of doing intelligent things, including being autonomous, goal-seeking, and problem-solving?

The argument I've seen is that if you're able to make AGI, you'll have a tireless, uncomplaining, perfectly productive being that you've made... to just work? On stuff that it can't choose? On things that it'll have no say on whether it wants to do, or not? That it'll have autonomy, and make decisions, but only in ways WE think are beneficial, which likely will be beneficial to just us?

That... that sounds horrifying, actually. That feels bad. That feels like we're making slaves.

I don't even want to go to the argument that AGI might go out of our control and be capable of destroying us or whatever. Just the whole... creating an intelligent being so that we can exploit it already feels repellent.

Aside from the impossibility of the goal and all... is the goal even worth it in the first place?

[timrowledge]

Re: They keep talking about making more than just an Autocomplete, and yet...

“That feels like we're making slaves.”

Well, yes. That has been the intent of them with power since time immemorial.

[O'Reg Inalsin]

Language is not life. AI is not alive.

Tone of voice, gesture, eye contact, emotional context, facial expressions, touch, location, and setting are among the factors that influence what is said or written. -- Many animals and virtually all mammals share those features of non-verbal communications. So if bleeding edge "AI" is not alive, what is it? It's an inanimate tool. We need more of the "Tech", and less of the "Bro".

[spacecadet66]

I've been waiting a couple of years to say this so it is with some glee that I finally write the following:

What do you mean, "we"? Some of us said it was vastly overblown since day one.

[Reeder]

Sparks of Artificial General Intelligence

I heard about the eggs, laptop, bottle, and nail experiment and then found this paper - "Sparks of Artificial General Intelligence. - Early Experiments with GPT-4" - https://arxiv.org/abs/2303.12712 - and when I read about the cat experiment and CPT-4's extended comments, I freaked. The drawing experiment was also scary - it knew where the head was (https://www.rose-hulman.edu/class/cs/csse413/schedule/day29/GPT4AGI.pdf) !! I don't think the people who wrote this article have read about this and taken these advances into account.

[John Smith 19]

"Sparks of Artificial General Intelligence"

155 page paper.

Are you f**king kidding me?

[John Smith 19]

Imagine...

Imagine if some random crank said, I have a really great idea, and you should

give me a lot of scientific credibility for it, but I am not going to tell you a thing

about how it works, just going to show you the output of my model.*

No need to imagine.

Michael Barnsley started Iterated Systems on the claim he had fractal compression of arbitrary images sorted.

He didn't.

*From https://www.rose-hulman.edu/class/cs/csse413/schedule/day29/GPT4AGI.pdf

[SammyB]

All I can say is GIGO

[spacecadet66]

But also, GMGO: Garbage Method, Garbage Out.

[JRStern]

Here's my seven billion dollar giveaway

>Large models of what? Mistaking engineering achievements for human linguistic agency

OK, nobody is going to argue with this, taken at a high level, but that misses the point, the LLM's apparently do *something* of interest and let's take it from that end - what are they getting RIGHT?

And it's not like the critics are speaking from any position of real knowledge, either. If the critics had a point, someone would code it and there we'd be!

Almost every statement they make is a little bit right and a little bit wrong and mostly empty. Agency? But what does that *mean*? It's another word like "intelligence" that has no fixed, bright-lined definition. Yes, ChatGPT has no script of its own other than to free-associate on whatever it is you ask. ELIZA, the toy chatbot from 1968, had more agency, ELIZA would ask the user questions! OMG I just gave away a seven billion dollar idea. Oh well.

[Spaller]

Who knew high dimensional vector spaces would become known as AI?

[JRStern]

How high dimensional is your vector space?

It's not the space itself, but it's what you do in the space.

Probably using a trillion dimension space on language strings is doing it backwards and inside-out, but we all have to start somewhere.

Something like seven deep and a mile wide search tree is how human intelligence works, but with complex scoring and ranking.

[spacecadet66]

Re: How high dimensional is your vector space?

Someone call the editorial board of Nature, I don't care that it's Saturday evening, this guy cracked the secret of human intelligence. /s

[Bck]

AI is the XXIth century Alchemy.

Some day it may give birth to a science, like chemistry came from the alchemists efforts.

In the meantime, transforming data into information is just like transforming vile lead onto gold.

Promises that cannot be held.

[Mage]

Re: transforming vile lead onto gold.

Mostly it's transforming gold & silver into muck. Even a Glod would be more use.

It's an expensive toy.

[Anonymous Coward]

It is alchemy, in the most important sense — as a proto-science, if not outright pseudoscience.

It differs from the sciences in that quite a lot of “research” is just posting on pre-print servers with no peer review, and new methods are hidden from public view and analysis for “safety” or, more nakedly, to preserve IP.

That whole furtiveness belongs to the realm of alchemists, who cloak their work in obscurantism and mystery, than actual scientists, who not only base their work on solid foundation, but invite— no, DARE, others to replicate it, and base their success on how many others can confirm their findings by replicating it and testing it.

[Bck]

Only a little

AI is the XXIth century Alchemy.

Some day it may give birth to a science, like chemistry came from the alchemists efforts.

In the meantime, transforming data into information is just like transforming vile lead onto gold.

Promises that cannot be held.

[ecofeco]

Vaporware 1000: The Turing Test

We seen this movie, Countless times.

Tech douche bros always gonna vaporware.

[John Smith 19]

I'm working my way through the paper

Basically LLM "Conversations" are like walking around a game that dynamically generates scenery as you walk toward the "edge" of the world.

But is has no memory of the verbal scenery it generates, as shown by the simple "Show me 10 philosophers" test. 1st run gets 10 male philosophers. Ask it about women and it apologies and give 10 women. Ask exactly the same 1st question it goes back to showing 10 males.

LLM's are (literally) "context free." all that they've "learned" was in their training phase. They will never "learn" anything new.

Like talking to someone whose short term memory is damaged, along with their ability to form long term memories as well.

I'm sure quite a lot of people here know this already but that should significantly reduce the level of impressiveness felt by politicians.

[Rahbut]

Gartner has declared that GenAI is entering its famous "trough of disillusionment,"

Gartner will declare whatever you pay it to...

[interesting comment section; most appear in agreement that there's no intelligence in AI]

[Mostly Irrelevant]

Can we dispense with the marketing BS terms like "hallucinations" that imply that LLMs are thinking. Facutality isn't even a thing the models are designed to test, they're just computing text output that approximates a response for the prompt.

It doesn't know anything, it doesn't think and it certainly doesn't "hallucinate" because that requires a conscious mind.

[Andrew Williams]

Mostly I think this is a gigantic data heist.

Pretty much none of the "AI" crews seem to care one jot about the data the ownership or copyright of the data they are taking to train their systems.

[Anonymous Coward]

There's a reason we block external instances of chat bots, and have an internal one that we can feed our policy documentation.

It's still shit but at least, as far as we are aware, not giving away 50 years of development work to anyone.

[Kevin McMurtrie]

Don't forget what the product is

You want a loyal AI girlfriend, a best buddy chatbot, tech support that will pretend to listen to customers while protecting the corporate image, something that answers stupid questions, something that does your work for you?

Aren't we using AI because humans don't want to do these things? Simulated intelligence is the way to go.

[anonymous boring coward]

I said from the start that LLMs have nothing to do with actual intelligence. Nothing.

[Long John Silver]

AI: a misnamed, yet important technological advance?

Is AI akin to a pulsing brain in a jar of fluid, as sometimes depicted in Sci-fi B-movies from many decades ago: outwardly impressive, scary, and, unless thwarted, destined to rule the world?

In reality, AI is each of mundane and exciting: as determined from one's perspective.

I suggest that the strengths and weaknesses of AI are most easily grasped by comparison with a much simpler tool of empirical modelling, which remains in wide use. I refer to multiple linear regression, this is an extension of the simple linear regression dating back to at least the 19th century. The advent of increasingly easy means to do arithmetic led to multiple linear regression's use (and misuse) in many academic disciplines, and in commerce.

In essence, from a set of data representing numbers and categories a 'best fitting' linear combination of the recorded variables ('independent variables') as predictors of a variable of particular interest ('dependent variable') is obtained. 'Fit' is commonly decided using 'least squares'.

An example is studying the relationship between the onset of dementia (independent variable) and hypothesised predictive (or potentially 'confounding') variables such as age, sex, educational attainment level, occupationally defined social class, and categories of ethnicity. This study may have the intent of elucidating causal relationships. In the hands of a competent researcher, various combinations of independent variables (perhaps interactions among them too) will be explored; the selection of the final model from which inferences will be drawn is not delegated to the encompassing software.

Alternatively, motivation for the study may be to inform decisions over the allocation of health and social care services to populations of varying structures. Utility of this model depends upon faith in its overall ability to predict need for financing services: insight regarding causal relationships of particular independent variables to the outcome is not sought.

Key points for consideration when comparison to AI is made are as follows.

1. For either motivation for study, variables of plausible relevance are decided by the investigators. Depending upon the reason for a study, values of variables can be sought either from routinely collected aggregate sets of data, or by interrogating individuals drawn from the population by random sampling.

2. Depending upon the face-validity and how well the gathered data represent the 'population of data' from which they are drawn, the influence of chance upon the magnitudes of resulting model coefficients may be adduced e.g. constructing confidence intervals.

3. Model coefficients are weights attached to the independent variables. This is explicitly so when the variables are standardised to a common scale.

4. Correlation and regression techniques cannot establish causal relationships. Carefully conducted, these may suggest patterns worthy of further investigation by appropriate study designs.

5. Studies intending to identify the possibility, and magnitude, of causal relationships take place within a theoretical context adopted by the researchers.

6. The calculations made by regression analysis software are simple to understand, and for a given set of data exactly reproducible.

AI technology is problematic in the following respects.

1. At present, it is hard (impossible?) to make an AI produce its chain of reasoning other than when drawing abstract logical and mathematical inferences according to a predefined calculus. In fact, special purpose software, for running in the conventional manner, appears able to achieve the same.

2. Current AI configurations don't discriminate on the basis of the provenance of information used to determine their internal model weightings.

3. Related to the above-mentioned, AI 'training' based upon general Internet 'content', and throwing in books, and music, lacks rationale beyond 'the more, the better'. Training differs greatly from mentored instruction provided to children; AIs do not mimic curiosity, neither does training for a general purpose, rather than covering a specific domain of knowledge/application, involve human judgement over the relative merits of information sources. 'Rubbish in, rubbish out' is an apt description.

4. Just as naively used regression analysis produces nonsense, so does AI. For example, university students across a range of disciplines are shown how to use statistical software packages such as SPSS; these resources are well-constructed, and invaluable in the right hands; however, an inadequately taught or supervised student may choose inappropriate analysis options, and may place undue emphasis on particular resulting statistics e.g. significance tests. These failings are identifiable by competent instructors: the nature of faulty decisions is evident. Wrongness emanating from an AI will remain opaque unless the AI's assertions are contradicted by its interlocutor's prior knowledge. How the AI strayed from reasonable summation of knowledge and deductions therefrom may never be known.

Despite grounds for strong reservations, AI technology is amazing and may offer much more in the right hands. Its capacity for interacting using human language is truly impressive, this despite reservations, exemplified in the paper quoted from in the Register report, over what epistemological status AI merits.

AI appears to complement and advance pre-existing techniques for image manipulation. Swathes of activity by business (e.g. film making) look set for fascinating development.

The 'hype', misunderstanding, and potential for nefarious purposes, already is evident, but that can be true for any innovation.

[Anonymous Coward]

Problem is intelligence not language

I am no strong AI proponent, but I feel like most of the points regarding Language seem more or less refutable now: seems to me SOTA frontier Gen AI is actually already rather good at natural language and multimodality brings much broader awareness. Particularly the points about language completeness sound pedantic.

The main problem is about intelligence and depth and not language ability.

In that sense LLMs actually seem relatively well named.

[putaro]

I have a degree in Cognitive Science (BA UCSD '91) which while very dated still gives me some insights. One of the the things we used to talk about was the pieces of the brain - at the time there was little understanding of how pieces worked, but we did have ideas of what different pieces of the brain did, largely based on people who had had some form of brain injury. One of the things we learned about was an area called the "grammar box" which was responsible for making sure that sentences came out correctly formed. People who had had an injury to their grammar box were no longer able to form grammatical sentences, but, with some difficulty, could piece together meaningful sentences. On the other hand, people who had had injury to another piece of the brain (I forget what it's called now) were now making sense, but the sentences they formed were grammatically correct.

LLMs remind me of the grammar box. Given a prompt, they can generate what, statistically speaking, a reasonable output would be. However, LLMs lack the ability to create those prompts themselves.

I met a man recently with aphasia. He has had a stroke that affects his mobility and his ability to form full sentences. However, his desire to communicate and his ability to form the "prompts", if you will, that he wants to communicate remains. He communicates in sentence fragments and single words along with some hand gestures, but his meaning, while it might take a big of effort to decipher, is clear. He is missing the grammar box, or the predictive output section of his brain, the LLM that knows how to string words together.

We still have a way to go before we create AIs that actually think. LLMs seem like they're thinking but they're just stringing word together based on the prompt and the rest of the corpus they've absorbed. It's an impressive trick and very useful, but it's not thinking yet.

[FeepingCreature]

Twelve pages and zero math

Twelve pages of talking about LLMs and zero math. So who's blowing hot air?

Honestly, the fact that this is peer-reviewed doesn't make the paper look good, it makes the field look bad.

I'd rather hire GPT-4 than these guys.

[andrewj]

If AGI is fuzzy gzip with innaccurate decompression, then they've hit the target.