I asked it for train times, and it gave me a very specific but wrong answer. Even Siri can get that right.
Seeing as GPT-3 is great at faking info, you should lean into that, says Microsoft
Microsoft is aggressively pushing OpenAI's artificial intelligence technology into seemingly every nook and cranny of its universe. Thanks to the Windows giant's fusion-fueled PR engine, everyone now knows Microsoft's Bing search engine is experimenting with using a cousin of OpenAI's ChatGPT large language model to answer …
COMMENTS
-
-
Monday 27th February 2023 23:19 GMT Anonymous Coward
My son was revising for a science test and had used ChatGPT, so I spent half an hour comparing answers with his book.
The overviews were more-or-less right but asking it for further examples came up with some odd stuff so I rephrased it and asked if it I had understood, it apologised and said it must have got it wrong and gave another example which was equally baffling. Also the graphs it had drawn bore little relation to the usual examples, it even managed to screw up an example temperature/time graph for water... twice.
So that was an instructive half-hour, especially for him.
-
Tuesday 28th February 2023 00:30 GMT doublelayer
We're going to need to pound that message in time after time before people start to get it. I had a programming problem where I couldn't figure out how to make it do something that looked quite simple, so a colleague of mine asked ChatGPT for help and sent me the answer. The response was clear, easy to understand, well-structured, and gave an example connected to what I wanted. The only tiny problem was that the code it suggested was syntactically incorrect and wouldn't have produced a correct result even if I put the punctuation in the right places for it. It was able to quickly produce misleading, useless information, and that's what it will do frequently even when no malicious person is there to ask it to.
-
Monday 27th February 2023 23:05 GMT b0llchit
Consumption
I am still wondering how much energy the setup must be using. So many requests onto such a large model with even more output. And then all the networks and computers with screens indirectly attached to the model, just to type in "words" and read more "words". It must be an effective way of burning compute cycles and warm the homes and data centres just a bit more.
Future thought: the ChatAI as advisor for all your financial needs. The model becomes miner, advisor, broker and cycle burner in one go. Next up, the laws will be made by ChatAI for your convenience to promote more interactive chatty consumerism. The future must be a total winner!
-
Tuesday 28th February 2023 00:56 GMT Andrew Hodgkinson
For those who still don't realise...
...this isn't AI/AGI, it's just ML. A large language model. It understands nothing. It knows no *rules* - just pattern matches, which with a big enough data set can seem convincing but this, combined with the parameter programming designed to make it respond in a very confident and authoritative manner, make it downright dangerous.
It can't give you a right or wrong answer because it doesn't know what right or wrong means. It knows that it thinks your pattern of text input mathematically correlates highly with other inputs which in turn led to other outputs, which are then combined to produce a result which mathematically looks like it's probably following the grammar patterns it has combined with the expected response patterns based on the input text.
Crap analogy time: Imagine an image matching ML system trained on cats. One day, it's given something that to a human is clearly a plant, but it's been trimmed and the colours of leaves and branches have been set up to make it look like a cat. If it mathematically matches very close to cat, the ML system might get 'plant', but it'll strongly get 'cat' and won't know *rules* that a cat can't be a plant, because it *understands* nothing. It cannot apply logic or reason. So it'll say "cat", and be wrong. LLMs are the text equivalent of the same thing. Give it enough data and it might start being able to say "cat *and* plant", or even have enough parameters to have never seen something that is both a cat and a plant so know no rules for it, but statistically speaking lean that way anyway; and so, it gives the illusion of understanding, without any. Doesn't know biology; RNA, DNA, how those work; nothing. No objective foundation in the laws of mathematics; not AGI. Just *fake*; an illusion.
It's also why image synthesisers like Lensa or MidJourney mangle fingers. They don't know what anything *is*, so they don't know what "rules" fingers have. Don't know there should be four fingers and a thumb, rules about symmetry, the way they can bend; just kinda makes something fingers-like that's close enough based on the maths and stats -> job done. And the result is, typically, very broken. Imagine the text equivalent of not knowing what fingers are. Yeah. That's where we are.
All this is why ChatGPT infamously responded with confidence about the bears Russia put into space. Sure, subsequent case-by-case hacks close off the embarrassing, very public mistakes, but that doesn't mean the technology suddenly magically works in any different a way. OpenAI is an Elon Musk venture originally and now largely controlled by Microsoft, so with either form of leadership, it's going to be (my opinion) ruthlessly immoral and entirely interested in profit, seeking it at any cost, legal if forced to, illegal if it can get away with it - e.g. by misleading advertising in countries where that's not permitted or embrace-extend-extinguish to remove competition (all, again, just my opinion).
So, the company is IMHO spouting BS about what this can do, the public are largely buying it and other companies are then spending money like it's going out of fashion to incorporate the largely useless and broken technology into their systems. They've either not so much drunk, as drowned in the Kool-Aid, or they're well aware that it's all nonsense, but think there'll be a good ROI because their own *customers* don't know any better and they're very happy to maintain that status quo. The net result is software that's even more bloated, even more buggy and even more all-round unpredictable. Slow - fucking - clap.
Any ML system can be a fun (if legally/morally dubious, due to training set!) way to generate fiction/expressive works, be it text or image, where accuracy or "makes any sense" aren't required. To front a search engine with it, where accuracy is key, is one of the most serious and flagrant breaches of duty of care I've ever witnessed and will *very severely* increase the misinformation problem with which our society is already struggling (and largely failing) to cope.
-
Tuesday 28th February 2023 01:05 GMT Anonymous Coward
Re: For those who still don't realise...
I have a far shorter description: it's a distraction (which in a way aligns with your "it's BS" opinion).
A complicated, shiny distraction to make sure you don't pay attention to the day to day problems running products from the company.
That's all it is. It's primary function is to dibert your attention and any bad press. You know, being the main platform to get hacked and/or suffer ransomware issues.
-
Tuesday 28th February 2023 12:38 GMT Martin Gregorie
Re: For those who still don't realise...
In summary: it looks very much as though ChatGPT has no more understanding of its textual input (both the user's request and the heap of example material its been fed) or of the correctness and relevance of its output than the ancient Markov scrambled text generator - its just larger, more complex and consumes more resources when its run.
If you believe that the output from either Markov or ChatGPT is either useful or correct except by chance, then I have this bridge over the East River that I'm sure you'll want to buy...
-
Wednesday 1st March 2023 19:14 GMT Michael Wojcik
Re: For those who still don't realise...
I'm not fond of the "pattern matching" gloss, because that's true only in the reductive sense that all computing can be reduced to pattern matching. Unidirectional transformer LLMs like the GPT series aren't matching patterns in the sense that, say, a DFA does when recognizing words of a language, or that a Hidden Markov Model could be said to, or even that a classic convolutional ANN would be. They're following gradients in parameter space, but that's not what most people would associate with pattern-matching; so the use of the phrase is misleading.
That said, I agree with the force of your argument. There are no qualia in play in the GPTs. It doesn't even aspire to p-zombie status; it's not even modeling qualia.There's no sapience. There is arguably a world model – there have been some good arguments made that the distribution in parameter space is an inherent world model, even if it's stripped down to nothing but a natural-language register and is entirely relational (in the Saussurian sense, i.e. none of the terms it "knows" have any Piercian indexical value, they all have meaning only in the sense that each is not any of the others and is different from the others in (quantitatively) different ways) – but building an ML system with a world model is not novel either, and a world model is not evidence of intelligence.
I've seen arguments that there are, or may be, emergent properties of LLMs which make them somehow a step closer to AGI. I'm not convinced. As I've noted before, I think AGI will require either a more-sophisticated architecture, or much larger scaling to the point where parameter space can contain a Boltzmann brain – at which point architecture doesn't matter, but the AGI would with very high probability be completely alien and almost certainly incomprehensible even if it was accessible at the interactive surface of the model, which it almost certainly wouldn't be.
What still looks really unlikely at this point, IMO, is that we'll have Chat-AGI any time in the near future. Incomprehensible AGI is much more probable, and even that likely requires a great deal more resources. Your Plastic Pal Who's Fun To Chat With is going to be just a language-emitting roller coaster for a while yet.
-
-
Tuesday 28th February 2023 07:22 GMT Filippo
For decades, we've written hundreds of sci-fi stories where you have an evil out-of-control AI, and that AI is hyperintelligent but can only reason along strict logical lines, and is ultimately defeated by exploiting its lack of imagination.
And then, suddenly, it turns out that the AIs we can actually make have way too much imagination, and are utterly incapable of logical thought.
I find this hilarious.
-
Tuesday 28th February 2023 08:12 GMT Detective Emil
Postel's law? Never heard of it.
Be conservative in what you send, be liberal in what you accept. [Wikipedia]
Software tested only on "clean" input (however produced) is going to fall flat on its face when exposed to what the real world has to offer — particularly if it's intentionally adversarial, the current fun game in AI-baiting.
-
Tuesday 28th February 2023 10:51 GMT katrinab
Re: Postel's law? Never heard of it.
I asked it for the time of the next train from my local station to a nearby city, which is the sort of thing a normal person might ask it.
It gave me a very plausible, but completely wrong answer to my question.
Siri, who is widely considered, even amongst Apple fans, to be the dumbest smart-assistant out there, will either give the correct answer (or at least correct according to the published timetable / most recent GTFS feed), or say she doesn't know the answer.
-
-
Tuesday 28th February 2023 21:44 GMT steviebuk
God no
"and the IT titan hopes to inject this kind of machine learning into everything from Teams and Skype to Excel."
We already have auto.suggest replies in Teams. Its fucking rude, anyone that uses those is fucking rude. Just type thank you instead of auto completing it because you can't be arsed to type it. If you can't be arsed to type it, don't fucking reply.
-
Tuesday 28th February 2023 21:59 GMT An_Old_Dog
Lies and Truths
"can understand text and generate new text based on that input,"
The first three words are a lie; the fourth word is a lie because it implies truth of the first lie; the remainder of the sentence is truth. Why did the author attribute the quote using the word, "reminded" instead of "said", or "claimed"? Using "reminded" implies the entire sentence is true, which it is not.
-
Wednesday 1st March 2023 00:10 GMT Denarius
A long time ago and not far away
I worked in a big project installing big data handling project for Big Gov. It was tested using sanitised data. Needless to say, it crashed on first exposure to real data. Alternatively, at a small statutory authority, software was built using test data, but final prerelease test was on yesterdays live data. Yes, all offline with no PHB easy dialin or unsecured network connection. I never heard of a problem in a production go live there. So the thought of using fake test data in M$ products should make even java coders look competent.
<offtopic> Maybe it is age, but the screaming rage induced by so much current software is beginning to look like a deliberate policy to kill off older people in fits of apoplexy </offtopic>.
-
Wednesday 1st March 2023 01:36 GMT Anonymous Coward
About ten years ago I worked for a consulting firm and was made aware that we had some sort of "tool" that we were developing for analyzing the interviews of utility company managers. I was not involved (or informed) about the project or the state of where they were in development. One Friday afternoon my boss asked if I would be willing to fill in some of the survey answers so they could have raw data to plug in to their model to test. He sent me a link to the system and I proceeded to (earnestly) answer the questions that would be given to a director of a fictional utility company.
Finishing the survey the system popped up a second survey; This time for a manager. A bit annoyed I answered these questions and thought that would be it. But no, the system then added another survey for a board member, then a supervisor, then a line utility worker, customer service representative, IT professional, etc...
After the second one I was imbibing in "adult beverages" and several friends came over for Friday night festivities. They joined me around my computer as I was filling out these surveys and started to "pitch in" their own (humorous and strange) answers. We took it as a game; creating responses from a supervisor that was a tyrant who distrusted their employees, a customer service agent who thought all the customers were idiots (and had homicidal tendencies), thieving utility workers, etc...
Finally, around 1 am I hit the last submit button and called it an evening. We spent the rest of our time enjoying our Friday night festivities.
I didn't hear anything more about it for months... The veil of secrecy once again fell upon the project.
Now we had a big public release announcement of our survey tool. Consultants who had been trained on the tool and took part in its development were working with some of our higher end customers to administer these surveys and it was "advising" them on how to change their business practices to fit the model of an ideal utility.
A few weeks later the product was quietly withdrawn for substantial rework.
I found out after I had left the company for a different consulting firm that they found part of the problem. My fake answers were used as part of the baseline for how these different departments should function. I think they scrapped the entire data-set but the damage had been done. Nobody ever had much confidence in the system after that.