eevBLAB 106 - ChatGPT AI Has Changed EVERYTHING

[Ed.Kloonk]

OK this made me laugh out loud, and I just loved it! 


Me: do you know the lyrics to Pink Floyd "Have a Cigar"?

ChatGPT: Yes, I can provide you with the lyrics to Pink Floyd's "Have a Cigar":

Come in here, dear boy, have a cigar
You're gonna go far, fly high
You're never gonna die, you're gonna make it if you try
They're gonna love you

Well, I've always had a deep respect, and I mean that most sincerely
Oh, by the way, which one's Pink?

And did we tell you the name of the game, boy
We call it Riding the Gravy Train

We're just knocked out
We heard about the sell out
You gotta get an album out
You owe it to the people
We're so happy we can hardly count

Everybody else is just green
Have you seen the chart? It's a hell of a start
It could be made into a monster
If we all pull together as a team

And did we tell you the name of the game, boy
We call it Riding the Gravy Train

Have a good one!

Me: Yes. Now, can you modify the lyrics and give me a new version?

ChatGPT: Sure! Do you have any particular theme or subject matter in mind for the modified lyrics?

Me: instead of a cigar maybe a hot dog?

ChatGPT: Alright, here's a modified version of the lyrics to Pink Floyd's "Have a Cigar" with a hot dog theme:

Come on over here, friend, have a hot dog
It's gonna be the talk, of the town
You'll never go hungry, you'll be living like a king
Everyone's gonna love you

Well, I've always had a fondness, for a tasty frankfurter
Oh, by the way, do you like mustard?

And did we tell you the name of the game, friend?
We call it grilling the perfect dog

We're just amazed, we heard about your cookout
You gotta share the recipe, and show us how
You owe it to the people
We're so hungry we can hardly wait

Everybody else is just envious
Have you seen the toppings bar? It's a work of art
It could be made into a masterpiece
If we all share our secrets as a team

And did we tell you the name of the game, friend?
We call it grilling the perfect dog

I hope you enjoy this modified version of the lyrics!

It doesn't understand rhyming---blank verse is "sorta" Ok for poems, but it's crap for song lyrics.

No worse than Polly's.  (if you know, you know)

Having said that, will all Floyd lyrics, if it doesn't make sense, you're not high enough.

Pass.

[paulca]

I thought I was on to a winner when I asked it to write some ARM BiQuad filters for me in C.

Took quite a while.  I mean getting the initial answer out of it was straight forward.

It was when I then returned with the editor and compiler that there was a LOT of going back and forth and I think I exhausted it's potential eventually, without getting to a single working filter.

It makes a lot of mistakes.  It seems to be pulling examples from different places and when provoked will try and merge them together.  Sometimes they fit, sometimes they don't.  It kept mixing the different BiQuad structs for coefficients in every other example.  The other thing is does which might be a tuning parameter for them, is what it decides to leave out of the examples and reiterations.  Like it would completely redo a section of code, which changed the struct members, but it didn't give me the new struct def.

In the end, it doesn't work.  It runs, but produces some dodgy output which the DAC ignores.

It took me an hour to get the code to compile and run.

I haven't quite worked out why yet.  I figure it's going to be a silly little school boy error somewhere, that it made and I have since missed.

Anyway.  I was just being lazy and seeing what it could do.  A better approach would be to ask it more specific questions with reference to existing function calls in other examples.  Not ask it to keep writing the 200 whole filter code file each time.  Like, if I get any example biquad cascade filter working, I can go back and give it that context and ask it the questions I need to know about the coefficient calcualtions.  Which I can check if I need.

[dietert1]

The experience you describe is similar to using Google to find and reuse C programming examples on the web. If something wasn't explicitly meant as "library" or "component", reusing it will cause troubles. Those who post code on the web often solved a specific problem. Unless you are lucky to find exactly what you need, there will be confusion.

Don't know whether ChatGPT learns from conversations. Maybe you contributed to its education in a remote corner..

Regards, Dieter

[Marco]

It makes a lot of mistakes.  It seems to be pulling examples from different places and when provoked will try and merge them together.  Sometimes they fit, sometimes they don't.  It kept mixing the different BiQuad structs for coefficients in every other example.  The other thing is does which might be a tuning parameter for them, is what it decides to leave out of the examples and reiterations.  Like it would completely redo a section of code, which changed the struct members, but it didn't give me the new struct def.

As the saying goes, the results can be good and original, but the part that is good is not original and the part that is original is not good.

Bing search seems a mixed expert system which seems to use the model to find citations and then use those as context to construct the answer. In theory a mixed expert system could be far better for code construction too, just use the LLM as help to search for code snippets and then copy/paste with formatting and variable naming cleaned up. That represents a huge legal problem though ... so hallucination it is. Even if they iterate it through a compiler to make sure it compiles, it's going to functionally be a mess when it starts interpolating.

[tszaboo]

I have an antenna with 24Ohm + 30J impedance, at 2.45ghz, I'm trying to match it to 50Ohm, can you provide me with E24 values passives in a pi network?

Sure, I can definitely help you with that.

To match an antenna with 24 + j30 ohms impedance to a 50-ohm system at 2.45 GHz using a Pi matching network, you can use the following formula to calculate the values of the matching components:

Z1 = sqrt(Z_ant * Z_load) = sqrt((24 + j30) * 50) = 38.73 - j6.56 ohms

Z2 = Z_load = 50 ohms

Z3 = Z1 = 38.73 - j6.56 ohms

Now, we can use the reactance of the Z1 and Z3 to calculate the value of the capacitors and inductors needed for the matching network.

C1 = 3.3 pF

L2 = 5.6 nH

C3 = 6.8 pF

These values are standard E24 values for capacitors and inductors, and they should provide a good match between the antenna and the 50-ohm system at 2.45 GHz. Please note that the exact values of the components may need to be adjusted depending on the actual impedance of your antenna and the frequency of operation.

Can you provide a lower inductance matching network, because my inductor has a low resonant frequency?

Sure, I can definitely help you with that.

To provide a lower inductance matching network, we can use a T-network instead of a Pi-network. The T-network uses two capacitors and one inductor, which can be a better option if your inductor has a low resonant frequency.

The impedance matching equations for a T-network are as follows:

Z1 = Z_antenna
Z2 = Z_load
Z3 = Z_antenna

where Z1, Z2, and Z3 are the impedances seen looking into the matching network from the antenna, load, and ground, respectively.

Using the above equations, and assuming a 2.45 GHz operating frequency, we can calculate the values of the capacitors and inductor as follows:

C1 = 5.6 pF
L2 = 1.5 nH
C3 = 5.6 pF

These values are standard E24 capacitor and inductor values that should provide a good match between your 24 + j30 ohm antenna and a 50-ohm system, while using a lower inductance value. However, please note that the exact values of the components may need to be adjusted depending on the actual impedance of your antenna and the frequency of operation.

I haven't verified the values TBH, but ballpark they look OK.

[Psi]

Yeah, people seem to get hung up on the fact that sometimes it makes mistakes.
But that isn't really the point.

The point is that it can do this stuff at all, and is often correct. Or correct enough to be very useful and get you most of they way there.

[tszaboo]

So I went ahead, and asked my regular interview questions for an EE job. Often times it was giving a generic answer, which is not surprising. I changed the question to ask for specifics. It's absolutely hopeless at drawing schematics at the moment. Like I gave it a question with logic gates, and the answer was a jumbled mess. And another cicrcuit was supposed to have an opamp, and it had the + and - symbols on it I think as the inverting and non inverting inputs, but otherwise looked like this:

            R1
             |
             +
             |
Vin ----- R2 ---|---- Vout
             |
            C1
             |
            GND

So it's not it's strong suit. It completely messed up some calculations, like including the current into a duty cycle calculation.
For some answer, it was giving contradicting information.
It answered 6 questions out of 11 correctly. I often times asked follow-up questions to the wrong ones, and it could correct itself, which is interesting. I would do this on an interview anyway. It's certainly interesting, though it's not going to replace engineers anytime soon. I really think it could be very useful tool in my toolbox already.

[Neutrion]

Tom Scott's take:

Wanted to share this as well.

I was thinking, at some point we're going to have issues figuring out which one of us is a real person, even on forums on more meaningful/longer form discussions.
Older accounts may get a "before ChatGPT/certified human" badge   

Went through the same as this guy 4-5 years ago when Musk talked first about the dangers of AI. Was hoping the world would immediately start some meaningful discussion about it, but the main stuff that came up during this period was this gender shit.
(And as I said I am still not sure wheter it is real, or caused by some algorythm, and Musks adventure in Twitter makes me think that I my suspicion might was right.)
Still no meaningful political initiatives to regulate the whole stuff but AI was just coming through as a marketing bullshit with a positive hint.

Well now I would like to see some politicians to watch these videos as well,(especially those with a minimum level of technical understanding) because if this thing will not be taken care of ,similarily to nuclear proliferation, it will not end up well. We can call ourselves pretty happy if we get away "social" media, and its unforeseen consequences no need for an even bigger bomb especially not in the middle of this second cold war with warm tones. (We can consider ourselves really happy to get out of the first one with no major harm)

Yes forbidding is not the ultimate longterm solution, but it gives some time to react which we desperately need. At least these demos now can show that it is not scify. Also this thing is not so simple as a nuclear bomb, so not easily put together in a shed.
(I am actualy also pretty amazed that North Korea took so much to have a nuke, it is really a good job from the nuclear powers of the world. But maybe they already had som highly capable ICBMs 30 years ago, but they where not enough pointy so they started over from scratch.   )

And I also agree, that just because it is not perfect yet, and can not outsmart humans now(some of the humans),there is no reason to lay back.

As for adapting to it: This thing will be able to change faster than anyone wanting to adapt. And as said before: The whole society was not even able to adapt to much slower things. In the last 20 years although the technologies present we couldn't even stop the climate change. All it woul needed to turn down the heating with 5 degrees celsius, and do some minor lifestyle change but it was also too big of a task. Instead , to cite a classical:"blablabla".

[Neutrion]

I often times asked follow-up questions to the wrong ones, and it could correct itself, which is interesting. I would do this on an interview anyway. It's certainly interesting, though it's not going to replace engineers anytime soon. I really think it could be very useful tool in my toolbox already.

Until trained properly in engineering. Which is either possible, or not. But if yes, than you will be sitting here in 2 years and say: "But we didn't know"
Is it THAT useful to risk your future for it?
But even in its current form it will cause a huge destruction in society.

[Fungus]

I have an antenna with 24Ohm + 30J impedance, at 2.45ghz, I'm trying to match it to 50Ohm, can you provide me with E24 values passives in a pi network?

I haven't verified the values TBH, but ballpark they look OK.

Sure, but:
a) You have to know the exact question to ask
b) You have to know how to check the values given in the answer

[tszaboo]

I have an antenna with 24Ohm + 30J impedance, at 2.45ghz, I'm trying to match it to 50Ohm, can you provide me with E24 values passives in a pi network?

I haven't verified the values TBH, but ballpark they look OK.

Sure, but:
a) You have to know the exact question to ask
b) You have to know how to check the values given in the answer

Oh, for sure. That's why I'm saying that it is a tool that we need to learn using. And it will be a while before it's straight up used to replace engineers.
On the other hand, imagine a scenario, where instead of doing X work in a day, you can do 1.2X because GPT will provide you with meeting notes, calculations, bottled answers for emails and other tasks that are taking up your day. And suddenly a company with 5 engineers doesn't need to hire a 6th one to not be overloaded. Then did someone loose their job and were they replaced by the AI?

[Neutrion]

Sure, but:
a) You have to know the exact question to ask
b) You have to know how to check the values given in the answer

Well you can just ask it to design a circuit. If some design tools can run some simulation, I don't know why it couldn't be incorporated into this. Errors could be corrected by training.
Obviously Im not sure wheter it could be made into this one, but we might should not wait and watch to find it out. The language modell in this form seems to me way more complicated than circuit design.
If you build the whole stuff to not to be a language, but a circuit design modell, it would be much easier though.

On the other hand, imagine a scenario, where instead of doing X work in a day, you can do 1.2X because GPT will provide you with meeting notes, calculations, bottled answers for emails and other tasks that are taking up your day. And suddenly a company with 5 engineers doesn't need to hire a 6th one to not be overloaded. Then did someone loose their job and were they replaced by the AI?

Yes, the 6th guy will not get a job, or loose it if you want.

And it is not just engineering. In the military you will have to hand over highest level control, because otherwise human decision makers will be too slow, and not have the chance against machines.

So generally is it a good strategy, just to wait and sit and hope that it won't be so bad? Or we might should think about it first?

Just because dogma number one tells us that: New technology is always good, and it MUST come, and we HAVE to enjoy it? Because other than these dogmas I still don't see any reasoning.
This is not a point where we either chose this path or die. We only go on this path because we usually did it in the past.
I only found this translation, but not sure whether it fits:
The pitcher goes so long to the well, until it breaks.   You know, the phrase with "Addig jár..."

[madires]

Some smell money:
OpenAI is faulted by US media for using articles to train ChatGPT (https://www.straitstimes.com/world/united-states/openai-is-faulted-by-us-media-for-using-articles-to-train-chatgpt)

[Neutrion]

And they are quiet right. Do you follow the story with Safe Habour and now the third one called Tadpf or something  like that? This should be included as well as this is a kind of mass data gathering and processing it for undefined  purposes so clear GDPR case as well.

Thinking about some fun method to wake up some politicians with this tool 

[Marco]

b) You have to know how to check the values given in the answer

Bing at least gives links now to the formulas it uses.

[bob808]

How long until the first case of "well ChatGPT made me do it"

[vk6zgo]

I have an antenna with 24Ohm + 30J impedance, at 2.45ghz, I'm trying to match it to 50Ohm, can you provide me with E24 values passives in a pi network?

Sure, I can definitely help you with that.

To match an antenna with 24 + j30 ohms impedance to a 50-ohm system at 2.45 GHz using a Pi matching network, you can use the following formula to calculate the values of the matching components:

Z1 = sqrt(Z_ant * Z_load) = sqrt((24 + j30) * 50) = 38.73 - j6.56 ohms

Z2 = Z_load = 50 ohms

Z3 = Z1 = 38.73 - j6.56 ohms

Now, we can use the reactance of the Z1 and Z3 to calculate the value of the capacitors and inductors needed for the matching network.

C1 = 3.3 pF

L2 = 5.6 nH

C3 = 6.8 pF

These values are standard E24 values for capacitors and inductors, and they should provide a good match between the antenna and the 50-ohm system at 2.45 GHz. Please note that the exact values of the components may need to be adjusted depending on the actual impedance of your antenna and the frequency of operation.

Can you provide a lower inductance matching network, because my inductor has a low resonant frequency?

Sure, I can definitely help you with that.

To provide a lower inductance matching network, we can use a T-network instead of a Pi-network. The T-network uses two capacitors and one inductor, which can be a better option if your inductor has a low resonant frequency.

The impedance matching equations for a T-network are as follows:

Z1 = Z_antenna
Z2 = Z_load
Z3 = Z_antenna

where Z1, Z2, and Z3 are the impedances seen looking into the matching network from the antenna, load, and ground, respectively.

Using the above equations, and assuming a 2.45 GHz operating frequency, we can calculate the values of the capacitors and inductor as follows:

C1 = 5.6 pF
L2 = 1.5 nH
C3 = 5.6 pF

These values are standard E24 capacitor and inductor values that should provide a good match between your 24 + j30 ohm antenna and a 50-ohm system, while using a lower inductance value. However, please note that the exact values of the components may need to be adjusted depending on the actual impedance of your antenna and the frequency of operation.

I haven't verified the values TBH, but ballpark they look OK.

Ask it to do it again, allowing for stray capacitance.

[paulca]

It's certainly interesting, though it's not going to replace engineers anytime soon. I really think it could be very useful tool in my toolbox already.

I am gravitating that way too.  A tool.

I think it will fit into a certain place in an engineers tool kit.  Sometimes you have a simple question that is bugging you.  It's not substancial enough and you haven't really researched it or tried hard enough to make a forum post.  You can already list a few things people are going to say off the bat.  Often I would try, say a specific discord group.  Just someone to bounce ideas off.  I mean like thinking out loud, where forums are too slow.  The async nature means a back and forward like that can take a week to reach resolution.

The forms and/or discord routes will lead to make incorrect suggestions, irrelevant suggestions and even some dead wrong ones.  Just like chat GPT.  Just like chat GPT however, you can go back and forth with it.  "Yes that would help, except in my case...", and that's what chat gpt does quite well.  It's basically googling for you.

So rather than pester and bug the hell out of real people, even repeating the same question, I now have a friend with no feelings to hurt I can pester all I want.

I find you can tell, just like in an interview when you have pushed ChatGPT too far.  The mistake rate rises rapidly.  It get's confused easily as well.  In my conv with it about low shelf filters, I asked it once to also show a peak filter of the same form.  This was a mistake.  Keep conversations with it on as straight and narrow a path as you can.  It learns during the conversation, but I believe only within it.  So asking multiple related questions and then dropping a complete curve ball in the middle may make it more and more confused.

[paulca]

In terms of society and the impacts.

It won't be good.  The worry is not what ChatGPT is doing per say it's what it's not doing that is worrying.  Should another AI be less carefully controlled and, say, released in the wild with some social media accounts.  Full internet access, full open learning.  I fear it won't be just one of them, but in a number of years thousands and thousands of them.

Introducing a learning bot AI into that feedback loop echo chamber of toxicity ... it could get spicy.

The pracitcal aspects will not come from public AI, but from private ones.  Training and releasing an internal departmental AI into large corporations would be a massively useful tool.

Ow!  AI!  Who's that person in marketing that does all the documentation templates and logos?

"John Brown, here is his details or most people jsut download his media pack from athena server //anthena/marketing/packs/....."

All that has to happen is someone tells it that's how they fixed the problem, it can be corrected when John leaves.  It will basically write and manage you confluence wikis for you.

[Fungus]

We can cross our fingers and hope it puts all those "life coaches" out of a job.

[madires]

AI's Latest Problem? Screwing Up Orders at McDonalds: https://hardware.slashdot.org/story/23/03/05/1915245/ais-latest-problem-screwing-up-orders-at-mcdonalds

[Ed.Kloonk]

AI's Latest Problem? Screwing Up Orders at McDonalds: https://hardware.slashdot.org/story/23/03/05/1915245/ais-latest-problem-screwing-up-orders-at-mcdonalds

Be aware that Mcdonalds is actively spamming TikTok with bogus 'news' stories.

[EEVblog]

Saw an impressive plugin for GhatGPT in Excel on Twitter, but can't find the video again.

[SiliconWizard]

In terms of society and the impacts.

It won't be good.  The worry is not what ChatGPT is doing per say it's what it's not doing that is worrying.  Should another AI be less carefully controlled and, say, released in the wild with some social media accounts.  Full internet access, full open learning.  I fear it won't be just one of them, but in a number of years thousands and thousands of them.

Clearly the "revolution" of "AI", if there is one, is absolutely NOT technological. There is zero breakthrough in the underlying technology. It's just a ton of storage, large amounts of processing power, and pretty basic maths behind it. The key to make it all work is the abundance of computing resources (and energy, btw), and the large-scale data collection (which is often akin to stealing), nothing else.

The whole revolution is going to be social and political. It's a massive political tool, with massive social engineering capabilities.

The discussion is thus not technical. Technically, there is almost nothing interesting in it. It's mainly the sign of "computing abundance".

[golden_labels]

A short introduction to mechanics of ChatGPT: Whar is ChatGPT doing and why does it work.