AI parsing data sheets

[tggzzz]

That's a recent version of an old old problem. Igor Aleksander's 1983 WISARD, effectively the forerunner of today's LLMs, demonstrated a key property of modern LLMs: you didn't and couldn't predict/understand the result it would produce.  WISARD correctly distinguished between cars and tanks in the lab, but failed dismally when taken to Luneberger Heath in north Germany. Eventually they worked out the training set was tanks under grey skys and car adverts under sunny skies.

"AI was a bit shit in the 80s, so of course in 2024 it will still be shit."

Er, no.

AI in the 1980s was a set of rules, based on knowledge elicited from experts. It was then easy to see which rules caused the results, and why.

WISARD was not like that; it was unique.

In the 2020s ML is shit for different reasons, as foreshadowed by WISARD. In particular it is impossible to know the chain that caused the results, because there are no explicit rules.

Yes, fundamentally, it is impossible to prove that a model is always accurate without testing every possible input (which is numerically infeasible for anything over a small set of input data).  But I am not sure that is as much of a 'gotcha' as you think it is.  In order to implement the level of object matching that a convolutional neural network performs, for instance, it is necessary to have a fuzzy statistical model, because the input data is not clean, and there are too many combinations to match against.

You are missing the point.

The point is that it is impossible by design to understand what caused an LLM to generate an output.

That wasn't true for "conventional AI" in the 80s, since it was easy to see and understand why each rule fired, and why. Writing the rules wasn't easy, but that's a different problem.

You don't need a neural network to be 100% accurate, you just need it to be sufficiently accurate to generate useful outputs.  Take the oft-noted example of the Tesla Autopilot seeing the moon as a traffic light.  The visual neural network does, indeed, perceive a yellow traffic signal.  The rest of the image processing algorithm clearly realises this data is garbage, so no action is taken on the input.  So the system is working correctly, although they might do better to filter the displayed output on the UI.

What is "sufficiently accurate" for life-changing decisions?
How can you prove that an LLM is "sufficiently accurate" when the next training test could change internal weights to invalidate previously correct output?

Are you one of those people that believes a piece of software is works because it passes the unit tests?

If you design a sufficiently complex testbench then any program can be proven to be correct, I fail to see why this is a controversial statement.  Now whether such a testbench is feasible is another matter. 

In this context Oracles are boring, because they don't and cannot exist.

Alternatively, thermodynamics is a solved problem if you use Maxwell's Daemons.

Alternatively, please show us your unit tests that proves an implementation has ACID properties. Publish it, and you will become not only famous but also as rich as Creosus.

Certainly in the world of FPGA and ASIC design it becomes essential to thoroughly testbench your hardware design because debugging a failure is far more difficult than in software.  Unit-testing individual components and writing a well-designed interface specification between them helps reduce the problem space.

Having first done that in the early 80s with 2µm semi-custom technology, I am well aware of that - and how difficult it is to get "stuck-at-0/1" coverage.

[tom66]

You are missing the point.

The point is that it is impossible by design to understand what caused an LLM to generate an output.

That wasn't true for "conventional AI" in the 80s, since it was easy to see and understand why each rule fired, and why. Writing the rules wasn't easy, but that's a different problem.

And it's infeasible to design a system in the way you indicate so it's a pointless argument.  So:   Write me an apple-recognising program that can process any 640x640 image and tell me if, and where, there is an apple in it.  Even if I gave you an allowance of 100k lines of code to do manual image filtering and object classification I doubt you'd beat a CNN from a decade ago running in software.

You don't need a neural network to be 100% accurate, you just need it to be sufficiently accurate to generate useful outputs.  Take the oft-noted example of the Tesla Autopilot seeing the moon as a traffic light.  The visual neural network does, indeed, perceive a yellow traffic signal.  The rest of the image processing algorithm clearly realises this data is garbage, so no action is taken on the input.  So the system is working correctly, although they might do better to filter the displayed output on the UI.

What is "sufficiently accurate" for life-changing decisions?

For self-driving vehicles, sufficiently accurate is driving safer than, say, the top 10% of human drivers in real traffic scenarios, over many millions of miles of real world driving.

Waymo already achieves this, having had no fatality accidents and 10% as many fender benders as equivalent human drivers would be expected to be involved in.  Tesla hasn't yet reached this standard, but they are probably only a few years away based on recent progress.  (Yes, that's about a decade later than Elon said, but he's a master bullshitter.  That doesn't change my opinion of the technology his engineers have created, which is 'pretty impressive' right now.)

[tggzzz]

Yes exactly the problem is that Ai is being used in all sorts of applications without consideration if AI is an appropriate solution.

I'd be satisfied with "provably acceptably accurate". Provably != our unit tests haven't found a problem. Provably implies a chain of reasoning and/or explanation.

The advent of these big powerful LLM AIs made it even worse because they are general purpose models that can do seemingly anything. Even worse is that they take in generic text or images and produce any kind of text, so this makes it easy for developers to just shove a bunch of data into it as text and just blindly hope the AI can make sense of it. This is where the shitty chatbot 'helpers' on websites come from, they are completely useless.

AI can be used in safety critical or life changing situations, but only when there are extra checks and safeguards around it, this effectively adds more layers to the 'swiss cheese model' where it gets more and more improbable that a mistake makes it trough everything due to everything lining up just right.

If that works acceptably, the problem becomes the people. They will rely on it and physically or metaphorically fall asleep.

Prediction is everything, unless you prefer to believe in faeries and similar

For example Mobileye developed a camera based vehicle safety system back in 2008 that manufacturers like BMW, Hyundai, Volvo..etc integrated into their cars as part of a automatic emergency braking feature. I got a Volvo from 2014 with this feature and it works well.  They take in information from the camera, radar, ultrasonic etc.. letting it able to detect a dangerous situation, first warning the driver with sound and flashing (likely already enough to get the driver to look forward and react) then if no action is taken the brakes are slammed on full at the very last moment. The whole thing works using a machine learned model running on a FPGA. But the system is just one extra layer of cheese in the model to cover up any holes in the human drivers cheese slice (since humans make mistakes too). Be it because the driver is distracted by a billboard next to the road while someone slams on the brakes in front, or if you are going down the highway and can't see a stopped car trough the fog, but the radar can see it.

There are credible reports of ?Teslas? braking suddenly and unacceptably (Can't be bothered to search for the specific reports)

Teslas (and other autonomous cars?) don't have radar, but rely solely on cameras (and GPS location data, which is another source of error!)

Once you give AI the task of full autonomous driving things are different. That one is a difficult problem that i have no idea what the correct solution to it might be (apart from a dedicated driverless car road network, but that is too expensive to be viable)

An unanswered question is when (not if) an autonomous vehicle causes an accident, who do you sue? The car manufacturer? The LLM manufacturer? The driver?

That raises the question of whether the LLM will detect it can't cope in sufficient time to hand over control to a driver. It takes up to 20s for aircraft pilots to become sufficiently aware and take avoiding action, and that is a much simpler environment than the roads. Saying "you have control" 2s before impact is not acceptable!

[AndyC_772]

The question in this thread wasn't about an AI making life changing decisions, or needing to produce output that's knowably, objectively correct.

In this case, the challenge was to take two self contained documents, each describing something that's finite in scope and presented in a common, structured way, and make inferences between them which can then be checked. That shouldn't be hard, and how it works behind the scenes is irrelevant.

I'm just trying to save time by offloading a relatively straightforward problem to a computer. I could equally well give the task to a junior engineer, but I don't have one to hand right now.

I do have a paid ChatGPT subscription, though - one which has already paid for itself many, many times over by teaching me some correct science that enabled a substantial cost saving in a product that's now shipping. Like it or not, believe it or not, it's a much better teacher than asking similar questions on a forum. It never says "do a search" or "your approach is wrong" or "I wouldn't do it like that" or any answer that's just plain unhelpful - it just answers the questions asked of it to the best of its ability.

Sometimes it produces complete garbage, just like people do. But on technical topics, it's surprisingly accurate. Really surprisingly accurate. And, of course, the beauty of correct science is it can be readily verified and proved to be so.

[tggzzz]

You are missing the point.

The point is that it is impossible by design to understand what caused an LLM to generate an output.

That wasn't true for "conventional AI" in the 80s, since it was easy to see and understand why each rule fired, and why. Writing the rules wasn't easy, but that's a different problem.

And it's infeasible to design a system in the way you indicate so it's a pointless argument.  So:   Write me an apple-recognising program that can process any 640x640 image and tell me if, and where, there is an apple in it.  Even if I gave you an allowance of 100k lines of code to do manual image filtering and object classification I doubt you'd beat a CNN from a decade ago running in software.

So what? That's irrelevant to the point.

You don't need a neural network to be 100% accurate, you just need it to be sufficiently accurate to generate useful outputs.  Take the oft-noted example of the Tesla Autopilot seeing the moon as a traffic light.  The visual neural network does, indeed, perceive a yellow traffic signal.  The rest of the image processing algorithm clearly realises this data is garbage, so no action is taken on the input.  So the system is working correctly, although they might do better to filter the displayed output on the UI.

What is "sufficiently accurate" for life-changing decisions?

For self-driving vehicles, sufficiently accurate is driving safer than, say, the top 10% of human drivers in real traffic scenarios, over many millions of miles of real world driving.

Waymo already achieves this, having had no fatality accidents and 10% as many fender benders as equivalent human drivers would be expected to be involved in.  Tesla hasn't yet reached this standard, but they are probably only a few years away based on recent progress.  (Yes, that's about a decade later than Elon said, but he's a master bullshitter.  That doesn't change my opinion of the technology his engineers have created, which is 'pretty impressive' right now.)

Musk's engineers technology is impressive (but insufficient), and Musk is an impressive bullshitter (dangerous, in many many ways).

Waymo? We shouldn't concentrate on one manufacturer and one application of LLMs. However http://catless.ncl.ac.uk/Risks/search?query=waymo leads to...

Are you sure Waymo isn't relying on remote drivers like some other self-driving cars? Alphabet/Google refused to answer that question, but does have a patent.

Alphabet's spin-out company Waymo still has by far the biggest testing program—its 635,868 miles of testing accounted for over 95 percent of all miles driven by self-driving cars in California in 2016.  Waymo's fleet of 60 self-driving cars reported a total of 124 disengagements, 51 of them due to software problems.

Some of the biggest critics of Waymo and other self-driving cars are the Silicon Valley residents who know how they work.  ... The problem isn't that she doesn't understand the technology. It's that she does, and she knows how flawed nascent technology can be. ... Silicon Valley types can be most skeptical of advanced technology because they know how it works and what its risks are.

[tom66]

So what? That's irrelevant to the point.

It's entirely relevant.  If the only option you have for data processing is the fuzzy model, you use the fuzzy model.  Ok, you can't prove it's correct and when it sees every possible apple image it will return a positive detection, but if you need to detect apples in an image, it's what you've got to use.  So you test it as much as you can to prove it is sufficiently correct for what you are trying to do.

Musk's engineers technology is impressive (but insufficient), and Musk is an impressive bullshitter (dangerous, in many many ways).

Waymo? We shouldn't concentrate on one manufacturer and one application of LLMs. However http://catless.ncl.ac.uk/Risks/search?query=waymo leads to...

Are you sure Waymo isn't relying on remote drivers like some other self-driving cars? Alphabet/Google refused to answer that question, but does have a patent.

Alphabet's spin-out company Waymo still has by far the biggest testing program—its 635,868 miles of testing accounted for over 95 percent of all miles driven by self-driving cars in California in 2016.  Waymo's fleet of 60 self-driving cars reported a total of 124 disengagements, 51 of them due to software problems.

Some of the biggest critics of Waymo and other self-driving cars are the Silicon Valley residents who know how they work.  ... The problem isn't that she doesn't understand the technology. It's that she does, and she knows how flawed nascent technology can be. ... Silicon Valley types can be most skeptical of advanced technology because they know how it works and what its risks are.

Waymo uses remote driving in the event of a car getting stuck, but they don't use it for normal driving.  The problem for SDCs is when you get really really infrequent events that have limited training data around them it becomes difficult to test the car's response.  So the human intervenes.  As more and more of those situations become part of the training corpus, then it may become feasible to remove the remote control altogether.

[tggzzz]

So what? That's irrelevant to the point.

It's entirely relevant.  If the only option you have for data processing is the fuzzy model, you use the fuzzy model.  Ok, you can't prove it's correct and when it sees every possible apple image it will return a positive detection, but if you need to detect apples in an image, it's what you've got to use.  So you test it as much as you can to prove it is sufficiently correct for what you are trying to do.

That's just silly. The ethical course of action would be avoid attempting (and failing) to detect apples in that way.

Musk's engineers technology is impressive (but insufficient), and Musk is an impressive bullshitter (dangerous, in many many ways).

Waymo? We shouldn't concentrate on one manufacturer and one application of LLMs. However http://catless.ncl.ac.uk/Risks/search?query=waymo leads to...

Are you sure Waymo isn't relying on remote drivers like some other self-driving cars? Alphabet/Google refused to answer that question, but does have a patent.

Alphabet's spin-out company Waymo still has by far the biggest testing program—its 635,868 miles of testing accounted for over 95 percent of all miles driven by self-driving cars in California in 2016.  Waymo's fleet of 60 self-driving cars reported a total of 124 disengagements, 51 of them due to software problems.

Some of the biggest critics of Waymo and other self-driving cars are the Silicon Valley residents who know how they work.  ... The problem isn't that she doesn't understand the technology. It's that she does, and she knows how flawed nascent technology can be. ... Silicon Valley types can be most skeptical of advanced technology because they know how it works and what its risks are.

Waymo uses remote driving in the event of a car getting stuck, but they don't use it for normal driving.  The problem for SDCs is when you get really really infrequent events that have limited training data around them it becomes difficult to test the car's response.  So the human intervenes.  As more and more of those situations become part of the training corpus, then it may become feasible to remove the remote control altogether.

That problem is true for more than self-driving cars; it applies to LLMs in general.

How long must a self-driving car wait before detecting "I can't do it", and handing control back to the driver? Hint: humans take up to 20s before becoming sufficiently aware of the situation.

If 20s isn't achievable, then what is the ethical course of action? To say "1s and tough shit, no longer my problem, you deal with the consequences". Given the potential consequences, I think that would be unethical and immoral.

[Berni]

Fully autonomous driving AI is a whole different can of worms. As i said, I don't know how to solve that, and if you do, i bet some company will be willing to pay you some serious money for the tech.

All i am saying is that this automated emergency braking and warning system is AI technology from 2008 (Tesla only just made their first car back then that was just a Lotus with a electric motor strapped to it). This Mobileye tech was fitted to cars of many manufacturers and there was no notable issues with it. You can't blame the manufacturer for the accident because you as the driver was supposed to take action before the system does anything. There is no "here you go control, please avoid the crash in 2 seconds", you always had control. The system just augments the reliability of an existing human driver by turning what would otherwise be a 100kh/h rear ending accident on a highway into a no accident at all (or at least a low speed collision without any injury). The system reacts at the very last moment on purpose so that the user would not comfortably rely on it (really it does stop just inches from collision using maximum braking).

It is an example of AI tech applied correctly. It works reliably and is applied in such a way that it does not cause an immediate safety problem if it fails. It just enhances the safety by stepping in when the human driver didn't (but should have). So even if the system was only 90% reliable that means 90% less car crashes of this type(failing to hit the brake pedal in time).

Besides what software these days even is 100% reliable? I have seen just about every car infotainment system crash. The quality of software these days is horrendous.

Besides what is the alternative to the functionality that ChatGPT (and sililar chatbots) provides? I can ask a question on this forum too. But i will have to wait to get an answer, then the answer i do get might not actually be correct as the human on the other side is not 100% reliable (they might just be a smartass that doesn't bother to actually look up anything, or they genuinely forgot and made a genuine mistake in giving the wrong answer). Sometimes i might need the answer to a dumb question (but just tricky enough to take more than 1 minute to google) and if i ask it on a forum i will get a response of "Are you to dumb to use google?". Then people start arguing about something slightly related to the topic but has nothing to do with the actual OPs question....etc

So sometimes ChatGPT will genuinely give me a better answer than this forum, and will give it to me in 5 seconds. Especially with how bad Google results have gotten in the past years. Not like you can treat things you find on Google as being 100% accurate. As always do your own research rather than just take the first thing you read, hold true for both ChatGPT or Google results.

[tom66]

That's just silly. The ethical course of action would be avoid attempting (and failing) to detect apples in that way.

I don't agree that ethics has anything at all to do with the situation.  There are ethical and non-ethical ways to use LLMs and neural networks.  Just as if there are ethical and non-ethical ways to use conventional software.

There are use cases for neural networks which are effectively harmless but genuinely useful.  For instance, my phone categorises pictures of each of my cats (so I can search them by name and post pictures on here if I want ), and I can search photos for things like "receipt <restaurant name>" and it will scan text in photos to look for that expenses receipt I forgot.  I believe that runs on the cloud - but I could be wrong.

That problem is true for more than self-driving cars; it applies to LLMs in general.

How long must a self-driving car wait before detecting "I can't do it", and handing control back to the driver? Hint: humans take up to 20s before becoming sufficiently aware of the situation.

If 20s isn't achievable, then what is the ethical course of action? To say "1s and tough shit, no longer my problem, you deal with the consequences". Given the potential consequences, I think that would be unethical and immoral.

You're confusing "take over immediately - I'm going to collide with something" with "take over please - I'm stuck and can't get out of this situation".  The former is of course dangerous and I think early autopilot systems prove that it isn't a path to autonomous driving.  The latter, on the other hand, is "fine", though not ideal.  An example would be a police officer closing and road and providing hand signals - current generation SDCs generally can't handle this condition so would rely on the remote operator to provide guidance.

[dobsonr741]

This is how it’s done: https://blog.adafruit.com/2023/09/02/writing-an-arduino-driver-with-openai-chatgpt-and-pdf-parsing-%F0%9F%A4%96%F0%9F%94%A7%F0%9F%93%84/

I repeated with another dataset myself, and it works. The cat is out of the bag.

[tggzzz]

In this case, the challenge was to take two self contained documents, each describing something that's finite in scope and presented in a common, structured way, and make inferences between them which can then be checked. That shouldn't be hard,

Just do it, e.g. ask "which of these two opamps is more relevant for a voltage reference implementation. Please let us know the results

and how it works behind the scenes is irrelevant.

Naive, e.g. the answer to the above question being decided by a coin throwing machine.

The reasoning is at least as important as the result, and LLMs can't explain how/why they produced their text. That, as they say, is an "active research topic", and I haven't seen any such research that might lead to optimism with real-world LLMs. I would be delighted to be presented with reasons to change that opinion.

I do have a paid ChatGPT subscription, though - one which has already paid for itself many, many times over by teaching me some correct science that enabled a substantial cost saving in a product that's now shipping. Like it or not, believe it or not, it's a much better teacher than asking similar questions on a forum. It never says "do a search" or "your approach is wrong" or "I wouldn't do it like that" or any answer that's just plain unhelpful - it just answers the questions asked of it to the best of its ability.

Regrettably sometimes those responses are exactly the best response to the question.

Sometimes it produces complete garbage, just like people do. But on technical topics, it's surprisingly accurate. Really surprisingly accurate. And, of course, the beauty of correct science is it can be readily verified and proved to be so.

So: sometimes great, sometimes crap. Bullshitters are as "good" as that, and cold-reading clairvoyants are even better. Be careful to avoid the different types of https://en.wikipedia.org/wiki/List_of_cognitive_biases#Confirmation_bias

Of course, some people some do ask the same question toss coins repeatedly until they get the "right" answer. They also twiddle the coefficients in their decision matrices until they get the right answer. Completely unimpressive, IMNSHO.

OTOH, the data going through your mind to produce unbiassed good answers, is entirely sensible. I wish more people did it.

[tom66]

I guess you haven't seen ChatGPT o-1, which claims to be able to reason its way to the answer.  It's in pre-release so us plebs without a Plus subscription don't get it, but there are some examples on the internet of what it can do.  According to OpenAI, it's able to pass the US SAT exams now better than the average student.

https://x.com/OpenAI/status/1834278219954225452

I think you are assigning too much to what is a very nascent technology.  It is not perfect, but if I look at what ChatGPT was doing ~2 years ago (arguing with me that Paraguay was in North America was a good one) to what it does now, it's so impressive.  If the rate of improvement continues, then there will be some serious societal questions to answer about obsoleting certain jobs with AI.  I have heard that there is already an impact on marketing, with lots of marketing copy now being written by AI instead of humans.   Of course, a human has to review the work, but you can get a lot more productivity out of one reviewer than you can out of five writers. 

[tggzzz]

If the rate of improvement continues, then there will be some serious societal questions to answer about obsoleting certain jobs with AI.  I have heard that there is already an impact on marketing, with lots of marketing copy now being written by AI instead of humans.   Of course, a human has to review the work, but you can get a lot more productivity out of one reviewer than you can out of five writers.

Given my contentions, clearly I ought to believe that bullshitters[1] should be worried about their jobs. Hence I do think salesmen[2], political speech writers, managers (upper and middle for different reasons) will be take in by claims and ought to afraid for their jobs (and good riddance )

Reviewing something is either surprisingly hard (if you do it conscientiously) or unsurprisingly easy (if you don't care a damn and/or are ignorant). Evidence: much of the modern "social media".



[1] Bullshitters emit apparently plausible sequences of words, especially technobabble, without understanding the concepts. I don't think there's any need to provide examples; we've all seen far too many.

[2] others disagree, but to me salesmen represent what the company has produced to customers, whereas marketing people represent what the customers wan/need to the factory. Too many people conflate "marketing professional" with "very successful salesman"

[tggzzz]

Here's the considered opinion of someone who has founded two AI companies and testified to the US Senate. He's not an amateur. https://spectrum.ieee.org/gary-marcus See the article to find his references; they are revealing.

Many of the points below are directly relevant to assertions made in this thread. The article contains indicating why he makes the statements.

Number one is hallucinations. These systems smear together a lot of words, and they come up with things that are true sometimes and not others. Like saying that I have a pet chicken named Henrietta is just not true. And they do this a lot. We’ve seen this play out, for example, in lawyers writing briefs with made-up cases.

those are boneheaded errors of reasoning where there’s something obviously amiss. Every time we point these errors out somebody says, “Yeah, but we’ll get more data. We’ll get it fixed.” Well, I’ve been hearing that for almost 30 years. And although there is some progress, the core problems have not changed.

Let’s clarify: I don’t think generative AI is going to disappear. For some purposes, it is a fine method. You want to build autocomplete, it is the best method ever invented. But there’s a financial bubble because people are valuing AI companies as if they’re going to solve artificial general intelligence. In my view, it’s not realistic. I don’t think we’re anywhere near AGI. So then you’re left with, “Okay, what can you do with generative AI?”

And a whole bunch of companies spent a bunch of money testing generative AI out on all kinds of different things. So they spent 2023 doing that. And then what you’ve seen in 2024 are reports where researchers go to the users of Microsoft’s Copilot—not the coding tool, but the more general AI tool—and they’re like, “Yeah, it doesn’t really work that well.” There’s been a lot of reviews like that this last year.

A question that often produces illuminating answers. If nothing would cause someone to change their mind, that person can be dismissed as an idiot/zealot/similar.

What would it take to convince you that you’re wrong? What would be the head-spinning moment?

Marcus: Well, I’ve made a lot of different claims, and all of them could be wrong. On the technical side, if someone could get a pure large language model to not hallucinate and to reason reliably all the time, I would be wrong about that very core claim that I have made about how these things work. So that would be one way of refuting me. It hasn’t happened yet, but it’s at least logically possible.

One of the articles referenced gives some useful positive and negative points, approaches and predictions https://garymarcus.substack.com/p/alphaproof-alphageometry-chatgpt

Some will be happy when the bubble bursts, others sad. I have mixed feelings; I think that GenAI is vastly overrated and overhyped, but fear that its collapse may well lead to an AI winter of sorts, like what happened in the mid-1980s, when AI “expert systems” rapidly ascended and rapidly fell.

That said, I am certain that the impending collapse won’t lead to the absolute death of AI. There is too much at stake.

Nobody understands LLMs fully; we can write the code them, but we don’t really understand their internals, or what they will do at any given moment, in part because their outputs depend so heavily on the minutia of their training histories.

What one finds, over and over, is that they are better at things that they have encountered before than things that they haven’t. They are especially poor if new items are different in important and subtle ways from old ones..

Perhaps the deepest problem is that LLMs literally can’t sanity-check their own work.

[AndyC_772]

The reasoning is at least as important as the result, and LLMs can't explain how/why they produced their text. That, as they say, is an "active research topic", and I haven't seen any such research that might lead to optimism with real-world LLMs. I would be delighted to be presented with reasons to change that opinion.

Here is exactly the chain of thought presented by ChatGPT -o1 preview when I asked the original question.

Compare and contrast the specifications of the Texas Instruments ADS8686S vs the Analog Devices AD7616

ChatGPT

Gathering specifications

I’m collecting datasheets for Texas Instruments ADS8686S and Analog Devices AD7616. Focusing on performance aspects will help highlight their key differences and similarities.

Assessing key specs

I’m piecing together the main features of the ADS8686S ADC, such as resolution, sampling rate, channels, input voltage range, and interface. This comparison will help in understanding its unique capabilities.

Investigating ADC features

I’m digging into the ADS8686S, noting its 16-bit resolution, up to 500 kSPS sampling rate, 4 channels, and bipolar input ranges. Integrated PGA and SPI interface are also highlighted.

Taking a closer look

Understanding the AD7616, I find it offers a 16-bit resolution with 1 MSPS throughput per channel pair, 16 channels, and a range of input options. Integrated analog front end enhances functionality.

Sampling rates, channel counts, and input ranges. Hm, I’m thinking through each detail to align functionalities effectively.

Weighing the options

OK, let’s break down the differences: ADS8686S uses SPI, has fewer channels, and likely lower power. AD7616 offers more channels with SPI and Parallel interfaces, likely higher power. Both share built-in PGAs, references, and input overvoltage protection.

Breaking down the comparison

I’m piecing together the high-resolution capabilities, keeping in mind their suitability for precision measurement. Both ADCs boast a sampling rate of up to 500 kSPS per channel, ensuring swift data acquisition.

Laying out the options

I’m noting the ADS8686S has 4 channels with SPI, while the AD7616 boasts 16 channels with SPI and parallel interfaces, giving it more versatility.

It then goes on to present the comparison between the two parts. It's objectively wrong, as per my original post.

Note the entirely reasonable sequence of steps (which should be the hard part), but also the outright factual errors in the specs of the two devices (which should be the easy bit).

[tooki]

One of the biggest problems we are likely to face with generative AI is the training data itself: as more and more AI-generated chum gets published, a lot of that is going to end up being ingested as training data, causing a vicious cycle of hallucinations becoming "real" to the AI.

I think I'll call this "data incest", since the end result is severe inbreeding in the data itself.

[tggzzz]

In my ignorance I would have guessed that the steps detailed are bog standard data collection and collation, and are therefore relatively easy. Deciding which sections are to be compared, and the comparisons themselves require far moe.understanding and judgement, and are therefore more difficult.

Can I suggest using the quotes I provided in my earlier post as a prompt to reading Marcus' articles. I suspect you will begin to see how that the failings you've noted aren't surprising.

The enumeration of the steps should not, of course, be confused with an explanation of how/why it "deduced" incorrect "facts". Such explanations are currently no more than a pipedream.

[tggzzz]

One of the biggest problems we are likely to face with generative AI is the training data itself: as more and more AI-generated chum gets published, a lot of that is going to end up being ingested as training data, causing a vicious cycle of hallucinations becoming "real" to the AI.

I think I'll call this "data incest", since the end result is severe inbreeding in the data itself.

Yes. Apparently research indicates it is already a problem.

[tggzzz]

I don't know why OpenAI bothers...
"OpenAI has been sending out warning emails and threats of bans to any user who tries to probe how the model works"
"Others say the warning is triggered simply by asking ChatGPT about the model's "reasoning" at all."
https://www.wired.com/story/openai-threatens-bans-as-users-probe-o1-model/

Maybe they just want to (try to) avoid people making fun of the LLM's unintentionally humourous/frightening spiel? 

[thm_w]

I tried it on a chinese datasheet for a LED driver - I was actually pleasantly surprised by the results. Wouldn't be very useful for an MCU probably, but it worked great for the SM16306SJ:

https://chatgpt.com/share/66e9b300-d3e0-8002-bf35-3a4c456b2a20

LCSC needs to use this to automatically update their parametric search, it would be 100x better than what it is now.

[tggzzz]

A few posts back there was a short discussion of whose jobs might be done by LLMs. Here's today's Alex cartoon, one of a series on the subject.

Background: Alex is the FinTech equivalent of Dilbert, and has been running since the 80s. Clearly written by people with deep knowledge of the area, and almost certainly supplied with insider tips, like the scriptwriters for "Yes, (Prime) Minister" and "The West Wing"

The Cyrus character is the American head of MegaBank. All characters are highly status conscious and profit oriented, and don't miss a chance of making someone else look/feel stupid. Bastards, all of them.

Use the "previous strip" button to see earlier posts on LLMs. https://www.alexcartoon.com/index.cfm?cartoon_num=8629