The Artificial Intelligence Problem (or not)

[Box]

Lordy. I caught myself arguing with a bot on Amazon.
Q. (Me) Why does the description on this item say leather but the material is listed as P.U.?
A. (Bot) The item is made from vegan leather which is a type of leather.
Q. (Me) There is no such thing as vegan leather! Are you saying it's made from vegans? Human vegans? Cow vegans?
A. (Bot) ...

Since then Amazon has fixed their descriptions of items made of vegan or faux leather. Now if you want a real leather item you have to enter "genuine leather."

Argue if you must - but I'm going to say the quiet part out loud - a pair of American made "leather" Cowboy Boots made out of 100% ethically sourced, fair trade, organically raised, USDA certified Vegans would be the best pair of boots ever.

BEST.BOOTS.EVER.

 

[compforce]

As mentioned elsewhere:

"AI contamination" is proving to be a thing.
ChatGPT Has Already Polluted the Internet So Badly That It's Hobbling Future AI Development


More articles, one goes back to 2023.

AI-Generated Data Can Poison Future AI Models

What is “AI contamination,” and what’s the cure? — M. Harris & Co.

I've been saying since the very first public models came out that they are going to be useless pretty quickly. As more AI generated content is created and published to the internet (especially under the guise of a person) AI is going to be subject to the inbreeding effect. The Tudors of the modern world are the AI models. They'll have to find a way to distinguish definitively between content created by other AI models and content created by people and then further distinguish the credibility of the information provided by people, which will itself bias the responses. Until they can solve this problem, small mistakes will be repeated often, magnified and eventually taken as the truth by the models.

I'll give you a quick example of the last one. Probably one of the most used sites in IT development is StackTrace. In nearly every company I've worked with, either as a consultant, an employee or a contractor has had code that is cut and pasted from the site. The sad part is that 99.9% of the answers, to include the accepted answers, may work, sometimes, but they are far from the technically correct solution. (reminder: 78% of all statistics are just made up). Due to the volume of information on the site, AI models assume it is credible and regurgitate the garbage (GIGO). And we wonder why AI written code rarely actually works...

Ever wonder why AI models can't do math? Because most people suck at math and AI has to rely on the most statistically relevant answer...which is usually one that is a common mistake. I've given a simple P&L to all of the AI's multiple times each with a cut and paste question to avoid bias in the wording. I did not get the same response twice. One of them was to calculate a well known ratio. The answers ranged from 0.4 to 117 with a small cluster around 20.4 The correct answer was 5.85. NONE of the AI models got it right and none of them guessed the same number twice.

 

[TLDR20]

I've been saying since the very first public models came out that they are going to be useless pretty quickly. As more AI generated content is created and published to the internet (especially under the guise of a person) AI is going to be subject to the inbreeding effect. The Tudors of the modern world are the AI models. They'll have to find a way to distinguish definitively between content created by other AI models and content created by people and then further distinguish the credibility of the information provided by people, which will itself bias the responses. Until they can solve this problem, small mistakes will be repeated often, magnified and eventually taken as the truth by the models.

I'll give you a quick example of the last one. Probably one of the most used sites in IT development is StackTrace. In nearly every company I've worked with, either as a consultant, an employee or a contractor has had code that is cut and pasted from the site. The sad part is that 99.9% of the answers, to include the accepted answers, may work, sometimes, but they are far from the technically correct solution. (reminder: 78% of all statistics are just made up). Due to the volume of information on the site, AI models assume it is credible and regurgitate the garbage (GIGO). And we wonder why AI written code rarely actually works...

Ever wonder why AI models can't do math? Because most people suck at math and AI has to rely on the most statistically relevant answer...which is usually one that is a common mistake. I've given a simple P&L to all of the AI's multiple times each with a cut and paste question to avoid bias in the wording. I did not get the same response twice. One of them was to calculate a well known ratio. The answers ranged from 0.4 to 117 with a small cluster around 20.4 The correct answer was 5.85. NONE of the AI models got it right and none of them guessed the same number twice.

I feel like it is pretty easy to tell what is written by AI these days. The videos also make it apparent. Where it does get interesting is with filters on real videos.

 

[Gunz]

So…let me get this straight…instead of AI teaching itself to get smarter so it can enslave humanity, it’s sucking up internet drivel and getting stupider. And since porn is the most prevalent content on the internet it means AI will get obsessed with hot horny stepmoms?

 

[Totentanz]

So…let me get this straight…instead of AI teaching itself to get smarter so it can enslave humanity, it’s sucking up internet drivel and getting stupider. And since porn is the most prevalent content on the internet it means AI will get obsessed with hot horny stepmoms?

The larger point is that it’s sucking up AI-generated internet drivel.

 

[Florida173]

I generally thinks these are more of thought experiment problems looking for solutions in a vacuum. They are all most just problems lacking the idea of any sort of new innovation, which is how we got here in the beginning. I suppose that in 2017 I would have said that the concept of LLMs would never exist. I was against most of the NLP world and found it generally useless... but then.. suddenly in 2018.. Transformers... revolutionary and disruptive change. Most of the people I see talking about the potential issues of AI slop ruining future training are generally in the same camp as the over-zealous luddites that look for any random thing to hate on a truly disruptive technology. I suppose we all need our villains

 

[Florida173]

Maybe it's time to reclass again... Although I have some concerns on it just being another ORSA related field unfortunately tied too much to something like G5/3 like ORSAs are tied to the G8 when it should be tied with G2

Army Creating New Artificial Intelligence-Focused Occupational Specialty and Officer Field

 

[AWP]

Maybe it's time to reclass again... Although I have some concerns on it just being another ORSA related field unfortunately tied too much to something like G5/3 like ORSAs are tied to the G8 when it should be tied with G2

Army Creating New Artificial Intelligence-Focused Occupational Specialty and Officer Field

How do you see this playing out with the Army's talent pool? If an org is trying to put people into brain-heavy fields like Intel, Cyber, and Signal*, then how does it staff a new branch without weakening the others? The pool of people for these fields is pretty finite. I'm not saying this isn't needed, but how do you staff it with people who could also work in the other branches?

* - And while this pains me to say it, Signal is probably going to lose people as the more tech savvy folks migrate to other branches, which it probably did when Cyber and the Space Force stood up.

 

[RackMaster]

How do you see this playing out with the Army's talent pool? If an org is trying to put people into brain-heavy fields like Intel, Cyber, and Signal*, then how does it staff a new branch without weakening the others? The pool of people for these fields is pretty finite. I'm not saying this isn't needed, but how do you staff it with people who could also work in the other branches?

* - And while this pains me to say it, Signal is probably going to lose people as the more tech savvy folks migrate to other branches, which it probably did when Cyber and the Space Force stood up.

It's happening a lot up here. Well anyone that stayed with the woke push.

 

[Florida173]

How do you see this playing out with the Army's talent pool? If an org is trying to put people into brain-heavy fields like Intel, Cyber, and Signal*, then how does it staff a new branch without weakening the others? The pool of people for these fields is pretty finite. I'm not saying this isn't needed, but how do you staff it with people who could also work in the other branches?

* - And while this pains me to say it, Signal is probably going to lose people as the more tech savvy folks migrate to other branches, which it probably did when Cyber and the Space Force stood up.

I think they should kill the entire ORSA program and restructure the Center for Army Analysis while creating a WO program specifically for AI/ML integration into J23/25. This has to be more than just "attach an LLM" to every tool as we're already seeing in the IC, but actual proliferation of leveraging statistics against the vast volume and variety of data we are currently handling. It pains me to see a continuing slew of analysts coming in that still haven't be incentivized to learn python. With practically every datasource being conveniently available through restful APIs, we are our own limiting factor to doing some real work and another random bespoke web application can't handle volume like I can. Also really need to completely modernize the ISSM community so that the dinosaurs are replaced with people that actually know how data science is done. Some places are already doing it really well, but it's a mixed bag.

 

[compforce]

I generally thinks these are more of thought experiment problems looking for solutions in a vacuum. They are all most just problems lacking the idea of any sort of new innovation, which is how we got here in the beginning. I suppose that in 2017 I would have said that the concept of LLMs would never exist. I was against most of the NLP world and found it generally useless... but then.. suddenly in 2018.. Transformers... revolutionary and disruptive change. Most of the people I see talking about the potential issues of AI slop ruining future training are generally in the same camp as the over-zealous luddites that look for any random thing to hate on a truly disruptive technology. I suppose we all need our villains

You do realize that AI/ML is directly in my primary field of expertise, correct? I'm talking about the real world impact of model corruption within the commercial world. Testable, observable, repeatable corruption. With AI/ML/LLMs, etc, as well as Python, many companies are losing their taste for the technologies because they are failing to deliver on their promise. (Python has a significantly higher cost of ownership and maintenance than most other common programming languages. It's only disruptive in that it disrupts business operations far more than the other ways of accomplishing the same thing.) The failure of the major AI engines and bolt-on technologies to provide consistent, accurate results is eroding the confidence of the real world users. Businesses are losing the will to bet their success on AI. It's become a marketing term and not much else for the vast majority of the business community. The big players, like Nvidia, Microsoft, Google, Tesla, etc. are pushing the narrative, but they have yet to deliver real results. Maybe that will change with quantum computing...or maybe they will just fail faster. The outcome remains to be seen.

Given the current state of today's technologies... Would you bet your life savings on an AI driven trading program? Would you bet your life on an AI doctor diagnosing a complex set of symptoms and building a treatment plan without oversight by a physician you trusted? Would you close your eyes, sleep and trust an AI car to take you across the country and arrive safely at your destination?

I wouldn't. And I work directly in the heart of the tech industry at the Consulting Senior/Enterprise Architect level. IMO AI is currently at an exciting state of research, but has been rushed into production prematurely. I'm sure there will come a day when it is capable of doing miraculous things. We aren't there yet. We're not even close.

I could take the opposite position from you and say that every new technology has its worshipers who believe that it is the answer to everything. I could also make the argument that every generation thinks they know more than the generation that came before. I can't tell you how many young programmers, dba's, analysts and other technical people I've seen that think that the older generation of technical people are "luddites" (to use your word). What they all seem to have in common is that they fail to realize that nothing is really new. It's the same stuff, repackaged, rebranded and marketed as the next best thing. What do you think was the last truly original concept in technology?

I received my original copy of this book for my birthday when I was in 4th grade in the late 70's. It was written in the 1960's. This one is the reprint from the 1980's (1984? I'm not walking downstairs to look at the flyleaf). Almost everything they are doing today in "AI" is described in detail in this book with accompanying diagrams and supporting math. It's not new and we aren't disregarding it because we don't understand it. We've already tried it, assessed it, and determined where it is useful and where it is not. We know the pros and cons and aren't caught up in the hype. We keep up with the technology, but look at it through the lens of experience. We've incorporated the concepts in our work in a way that is mature and useful. We're not opposed to the idea we're opposed to the implementation.




The truth is somewhere in the middle and needs to be analyzed objectively with an eye on the requirements of the specific task. There are components of AI/ML that are very useful and that can and should be incorporated into expert systems, but there are also parts that are so immature as to invalidate the system as a whole when they are included.

It's the old adage, when you're a hammer, every problem is a nail.

 

[Florida173]

You do realize that AI/ML is directly in my primary field of expertise, correct?

You do realize that AI/ML is directly in my primary field of expertise, correct?

 

[compforce]

You do realize that AI/ML is directly in my primary field of expertise, correct?

Yes. My point was that not all of us that are older are pointing out issues with the technology because we're afraid of new things, which seemed to be the implication of your earlier post. My points were that a measured approach needs to be taken with anything and that there is nothing that is really new. Virtually everything out there is a mashup of things that have been known for a very long time. That includes AI.

 

[Florida173]

Yes. My point was that not all of us that are older are pointing out issues with the technology because we're afraid of new things, which seemed to be the implication of your earlier post. My points were that a measured approach needs to be taken with anything and that there is nothing that is really new. Virtually everything out there is a mashup of things that have been known for a very long time. That includes AI.

Just kind of a bizarre way of looking at things. I've been applying AI/ML models against difficult problems in the IC for more than 15 years now. I was against wasting my time with NLP prior to the evolution of transformers and more recently with LLMs, but the basis of semantic similarity has been around for a long time. To suggest that nothing is new or is only a mashup of other known things seems irrelevant and a bit reductionist. I disagree that businesses are losing any will unless they are looking at making the AI/ML the thing instead of what it's able to do. Similar to recommender services from Amazon or Netflix. I don't understand the concern for things like AI slop. It seems like a fake concern for people that don't know how the training datasets are developed and marketed. Can I use AI slop to train models, sure, maybe an LDM or something similar with Text to Image, but my quality might start to suffer... or not..

Our lives are going to get interesting.. The image below is a perfect example. My wife sent me the photo on the left, and told me to ignore the people, but look at the kitchen. So I through it against the new flux1 kontext dev model with my consumer GPU and removed the people, then removed the watermarks, and then removed the clutter on the counter. Each iteration only took me 5 seconds. Is this just a mashup of things that have been known for a very long time, sure... but kind of an odd thing to say when you couldn't do that a year ago in a few seconds.

 

[compforce]

I'll admit that the visual side has moved quite a bit further than the language models. That said, you could do a very similar thing in Photoshop back in the early 2000s with macros. 10 years ago Dreamworks was removing people from movie shots with Magix. What's changed isn't the technique of removing the images, it's the hardware that has gotten faster and the instructions have been dumbed down to reduce the level of knowledge needed to perform the task in the interests of making development faster and requiring less expertise to use them.

As I said, there are components that can and should be implemented in expert systems, but there are just as many components that shouldn't even be considered yet. Graphics, both still and motion, are probably the leaders in useful tech, but the LLM work that I've seen being produced is mostly the bolt-on type being done just so the company can advertise that they have integrated AI. They aren't training their own models. Most of them don't have enough data to be significant if they tried. Even the big boys that have the resources and data to do it haven't figured out how to structure their data to make it useful, and that includes many fortune 50/100/500's that I've worked with.

The datasets that you are using are most likely not the same as commercial datasets. Hell, most of the commercial datasets are badly designed and even traditional data warehouses have inherent flaws that exist in both the Inmon and Kimball models that impact accuracy and which haven't been (publicly) solved in 20+ years. I'd even go as far as to say the issues I have aren't with the technology itself. It's the businesses and the developers where the biggest flaws exist. It doesn't matter how good the technology is, if the businesses don't adopt it and the developers can't build a coherent application from it, then what good is it?

Seriously, there is probably a bright future in AI but it's currently in its infancy. It was rushed to market just like every other product out there. If the guys and gals that actually created the current leading edge technologies (OpenAI, Google, et. al.) can't train their models well enough to give accurate answers to basic questions, is it really ready for the primetime? Or are we actually an extended period away from it being practical in the mainstream? I'm sure there are some really slick uses in the IC, with DARPA and with other government entities, but given that I don't have visibility into those communities any longer and work outside of government in the private sector, I haven't seen those lately and I certainly haven't been impressed with what I have seen.

 

[Florida173]

I'll admit that the visual side has moved quite a bit further than the language models. That said, you could do a very similar thing in Photoshop back in the early 2000s with macros. 10 years ago Dreamworks was removing people from movie shots with Magix. What's changed isn't the technique of removing the images, it's the hardware that has gotten faster and the instructions have been dumbed down to reduce the level of knowledge needed to perform the task in the interests of making development faster and requiring less expertise to use them.

As I said, there are components that can and should be implemented in expert systems, but there are just as many components that shouldn't even be considered yet. Graphics, both still and motion, are probably the leaders in useful tech, but the LLM work that I've seen being produced is mostly the bolt-on type being done just so the company can advertise that they have integrated AI. They aren't training their own models. Most of them don't have enough data to be significant if they tried. Even the big boys that have the resources and data to do it haven't figured out how to structure their data to make it useful, and that includes many fortune 50/100/500's that I've worked with.

The datasets that you are using are most likely not the same as commercial datasets. Hell, most of the commercial datasets are badly designed and even traditional data warehouses have inherent flaws that exist in both the Inmon and Kimball models that impact accuracy and which haven't been (publicly) solved in 20+ years. I'd even go as far as to say the issues I have aren't with the technology itself. It's the businesses and the developers where the biggest flaws exist. It doesn't matter how good the technology is, if the businesses don't adopt it and the developers can't build a coherent application from it, then what good is it?

Seriously, there is probably a bright future in AI but it's currently in its infancy. It was rushed to market just like every other product out there. If the guys and gals that actually created the current leading edge technologies (OpenAI, Google, et. al.) can't train their models well enough to give accurate answers to basic questions, is it really ready for the primetime? Or are we actually an extended period away from it being practical in the mainstream? I'm sure there are some really slick uses in the IC, with DARPA and with other government entities, but given that I don't have visibility into those communities any longer and work outside of government in the private sector, I haven't seen those lately and I certainly haven't been impressed with what I have seen.

Not really disagreeing with any of your points. I think you might be generally hyper focusing on LLMs and how a lot of people have misaligned expectations.. I definitely have folks constantly asking me how to shove an LLM into their product opposed to saying, here's our problem, can an LLM help in any way. I mean there's a reason why how many Rs in strawberry is hard, but this is more or less a solved problem as we start to see models move away from tokenization.

The IC is much further along than anything DARPA/IARPA/National Labs can push out. The days of waiting for their "advancements" is well over in my opinion. Especially with cloud compute being so easily accessible. Although I'm not a fan when I can just grab some A100s or H100s that are on-prem to spin up whatever I want. Whiteboard to prototype is sometimes in hours for what used to be months. Especially when I can just write some quick pytorch code and wrap it in some streamlit.

 

[compforce]

Not really disagreeing with any of your points. I think you might be generally hyper focusing on LLMs and how a lot of people have misaligned expectations.. I definitely have folks constantly asking me how to shove an LLM into their product opposed to saying, here's our problem, can an LLM help in any way. I mean there's a reason why how many Rs in strawberry is hard, but this is more or less a solved problem as we start to see models move away from tokenization.

The IC is much further along than anything DARPA/IARPA/National Labs can push out. The days of waiting for their "advancements" is well over in my opinion. Especially with cloud compute being so easily accessible. Although I'm not a fan when I can just grab some A100s or H100s that are on-prem to spin up whatever I want. Whiteboard to prototype is sometimes in hours for what used to be months. Especially when I can just write some quick pytorch code and wrap it in some streamlit.

I think the biggest difference between our opinions is the perspective. I'm thinking about it purely from the private sector point of view and through the lens of what I see and hear on a daily basis. Working with the C-Suites, the most common opinion I hear is "we have to invest in AI or people will think we're falling behind. Let's find the cheapest solution we can implement and then market that we are integrated with AI" which leads to the inevitable customer support screener bots, which are easily the least expensive implementation and eliminates positions that have the lowest knowledge and education requirement in the company. Either that or a third party bolt-on like Qlik View. I think you're more focused on public sector use, especially from a National Security viewpoint. Useful AI and AI for AI's sake are two different things. It's a lot easier when you don't have to worry about shareholder value and profits.

On a side note, did you see that DARPA was able to deliver 800MW of power via laser last month?
DARPA program sets distance record for power beaming | DARPA