eevBLAB 106 - ChatGPT AI Has Changed EVERYTHING

[MK14]

I'll let others judge the results.  Here is one of the better pictures I managed to get made, recently.  Using the ChatGPT's, DALL·E 2 picture generator.
The text used was "EEVblog forums cat, playing in an electronics lab, with big bold and colourful flashing LEDs and motors, while controlling a large oscilloscope".
Results:

[golden_labels]

For actual, detailed works this is not enough. The smortnet was never trained for that, though. And for some uses you do not need much more. We already have an entire thread filled with wet neural networks outputs, which — while usually being of higher visual quality — often make even less sense. And no one cares, because the purpose is to invoke a particular response in recipient’s brain. A split-second process, devoid of any critical processing, the image itself playing no real role and to be forgotten a second later.

This is commonly employed in advertisements and music videos. On extreme side with images shown for just a dozen frames. Though technique is by no means new: it will soon celebrate its hundreth birthday, with earliest examples including Gance’s “Napoleon”.

[Bud]

Is that the oscilloscope in the background? 

[MK14]

Is that the oscilloscope in the background? 

That's a good example of why, currently, the pictures are more for fun or rough/draft work.  To get it done properly, still needs someone to do it properly.  Although with perhaps a quick session with Photoshop or similar, a real oscilloscope trace/screen, could be pasted into the picture, to fix/improve it.

[golden_labels]

The cat is controlling the large oscilloscope, so obviously it must be looking at the device. Which means it’s not in the frame. You are the oscilloscope.

The thing in the background is a tube amplifier for the oscilloscope, with the fancy blue light as you find in some eBay tube amps.

[james_s]

Is that the oscilloscope in the background? 

That's a good example of why, currently, the pictures are more for fun or rough/draft work.  To get it done properly, still needs someone to do it properly.  Although with perhaps a quick session with Photoshop or similar, a real oscilloscope trace/screen, could be pasted into the picture, to fix/improve it.

It still reminds me of the python scripts I was playing around with a few years ago to generate text. I made several that could spit out pages and pages of reasonably formatted more or less grammatically correct text that were pretty much just gibberish. It was amusing but not the least bit useful.

[EEVblog]

Fran did a video on this:

[bob808]

I guess the more philosophical question is "what do we do at the point past which we don't have to do anything?". Should we reach it? It feels like we should.

[Mr.B]

I don’t know if I am not asking the questions the right way, or the AI is simply screwing up.
Or maybe I am not understanding the answers correctly…
First example: I brew beer for a hobby. I usually brew under pressure and crash cool at the end of fermentation. The AI gets the math wrong.
Second example: Fairly basic question.

[ataradov]

It does not "know" math at all. It is a language model. It is capable of convincingly BSing its way around topics.

All it knows that at this point in the sentence some number makes a lot of sense. It is frankly amazing that it gets that far.

This is why it is so dangerous. It is good enough to convince you that it is correct.

[SiliconWizard]

It's just a gigantic piece of crap illusion.
One sure thing is that its proponents are working hard at making us all believe that it's gonna become unavoidable, so that we better get used to it.

Call me when this shit has any value for people except for those very few who reap direct benefits both in terms of money and power.

I'm almost convinced now that it's soon going to be used for all kinds of completely official and serious uses despite its patent shortcomings. This is the new normal.

[golden_labels]

I don’t know if I am not asking the questions the right way, or the AI is simply screwing up.

It’s a huuuuuge and complicated mathematical equation, in which parameters were adjusted in millions of repeated test-adjust attempts. For a bunch of input numbers it produces a bunch of output numbers. The parameters were chosen in a way, that human brain seeing the output will think they are a sensible result of the input.

The smortnet does not think in usual sense of this word. It can’t understand its own outputs, as it doesn’t even have a capability to perceive those. Not to mention having any circuitry to correct/reject them.

Performing such operations is not even limited to machine learning. One can construct a deterministic system, which will do the same. The difference with smortnets is that they are resilent against input noise,⁽¹⁾ though the consequence is the noise you get in the output, and that constructing them is computationally uncomparably cheaper than what would be needed in older, deterministic systems of a similar scale.<sup>(2)

(1)</sup> For a few recent, famous solutions noise is actually injected into the system to obtain the results.
⁽²⁾ Strictly speaking: traditional solutions would be infeasible at this scale. You could construct a hash function, that does exactly what ChatGPT does, but it would take multiple ages of the universe if not more.

[Per Hansson]

In the example Dave is amazed how it spits out some Arduino code, but it is literally just taken from the Arduino website.
As PA0PBZ pointed out you can make it cite its source.
I don't see how there is any Artificial Intelligence in this at all.

[bob808]

https://www.theguardian.com/technology/2023/jan/05/microsoft-chatgpt-bing-search-engine

You could say it's a glorified search engine.

[madires]

I don't see how there is any Artificial Intelligence in this at all.

AI is the buzzword for what we've called neural networks 30 years ago. With today's CPU performance you can simulate much more neurons and layers, giving you more sophisticated results. Back then the highlight was to recognise handwritten numbers in realtime. It's still the same principle, just much larger. This also implies that your AI is a black box and you don't have any real control or idea of what's happening inside - it's simply too complex for the human brain to understand. If you train the AI with garbage, you'll get garbage as output (including discrimination, racism and all the other politically incorrect terms). The huge problem we have now is that some people believe in AI (or more generally in tech) superiority and are using it for hiring employees or determining insurance risks, while trying to avoid responsibility in case something goes wrong. Your Honour, it wasn't me, it was the AI! In the future we'll see a lot of lawsuits because of this. I'm not saying that AI is bad, but we need to understand its limits and risks, and how to use it sensibly.

[Axk]

I guess I need a better understanding of how this GPT-3 thing works and that's probably an incorrect way to formulate my question,

Is there some parsing to determine the query type going on before it passes it to the GPT model?
E.g.

Text formatting
Code generation
Code analysis
Text summarization
source citation
etc

Or is it just literally one big model that is trained to do all the tasks?

From playing with it I have the same impression as most of the people here.
I don't see much practical application for it personally at the moment besides maybe:

Code scaffolding
Short scripts for simple tasks
Text correction and formatting

Anything factual, you cannot trust it.

I thought it could summarize articles based on URLs but turns out it just composed stories based on the few words in the URLs themselves )

[golden_labels]

It’s one huge model, which is big enough and was taught with large enough dataset to be able to adapt to any language task.

Of course the final output, as conveyed to humans, is formatted using traditional techniques from markup the smortnet produces. But everything up until that point is a single encoder-decoder transformer.

Unlike the model itself, the GPT3 paper is available to public. But, until you are already reading ML papers for years, it will not be very useful to understanding this algorithm from ground up. The smortnet itself is provided exclusively to Microsoft and only Microsoft can use it to deliver solutions to the market. So no one knows all the details.

[Neutrion]

Apart from chatgpt itself, I think even if we would be somehow saved from the general strong AI, these softwares just by the disruption will cause more harm than good.
There is  this theory that disruption is good because it moves the world forward, but there is a point where it can end up in chaos. Even with social networks we see it, and I am not sure what is the huge benefit  what we got from FB and the likes. However we can clearly see that politics is getting more and more destabilized all around the world thanks to some of these disruptions.
The first world war  was also in huge part caused by the industrial revolution, and the fact that human society was not prepared for a war on an industrial scale, actually couldn't even imagine what it could mean.
With AI and genetic engineering I see a very dark horizont ahead of us. (The corona virus was just a really harmless demo of what could be waiting for us.)

[Ed.Kloonk]

In the example Dave is amazed how it spits out some Arduino code, but it is literally just taken from the Arduino website.
As PA0PBZ pointed out you can make it cite its source.
I don't see how there is any Artificial Intelligence in this at all.

Could it interpret some obscure code found on the web and perhaps document it a little bit?

Asking for a friend.

[Axk]

I wonder if we gonna see a 2023 "AI bubble" ))

[Axk]

Could it interpret some obscure code found on the web and perhaps document it a little bit?

It won't really add any value besides stating obvious things about the code.

[EEVblog]

I wonder if we gonna see a 2023 "AI bubble" ))

I think it depends on how well the public takes to "conversational search".
I think they need to rename it to something simple, and have a straight Google-like page. Currently it's kinda hidden away a bit behidn a subdomain.
Simplicity and branding is the key here.

[bob808]

I think only Microsoft can use it so it'll become Bing. Or some version of Bing.

[golden_labels]

(…) With AI and genetic engineering I see a very dark horizont ahead of us. (The corona virus was just a really harmless demo of what could be waiting for us.)

Progress in computing is not filling me with fear. What worries me, is that kind of a mindset. I know of no creature more dangerous and destructive than scared humans.

Speaking of which, the smortnet situation triggers my curiosity not for technical reasons,⁽¹⁾ but from social, philosophical and psychological perspectives. The element most interesting in all this is a human and human society. I think I already posted that somewhere, but I suggest two movies:

• The Heavenly Creature (2012), a part of “Doomsday Book” triptych, not to be confused with Jackson’s movie: depicting tension between two perceptions of a machine, working in a buddhist monastery, which started acting as an enlightened human. More current than ever. Better than anything showing, how I feel now, watching the smortnet hitting public.
• Colossus: The Forbin Project (1970): seen as a member of the “evil AI” trope, to me this picture makes a strong philosophical statement about humans and your morality. A stunning depiction an ethical equivalent of the paperclip maximizer, exposing inherent incoherence of human ethics.

⁽¹⁾ That part is something I went through in late 2000s. My degree thesis was meant to be in autonomous robot swarms.

[james_s]

It does not "know" math at all. It is a language model. It is capable of convincingly BSing its way around topics.

All it knows that at this point in the sentence some number makes a lot of sense. It is frankly amazing that it gets that far.

This is why it is so dangerous. It is good enough to convince you that it is correct.

I've worked with a few guys like that.