rsm
Well-known member
enjoy the future!
enjoy the future!
rsm
Well-known member
rsm
Well-known member
FWIW, I've been working in AI for decades. Most people have no idea that AI has been in use for several decades, what it has been doing, and what it will be doing.
To be successful, AI systems don't have to eliminate human workers or activities, it can also reduce the complexity of human tasks, or be assistive - allowing people with fewer skills / less knowledge to be more productive.
To be successful, AI systems don't have to eliminate human workers or activities, it can also reduce the complexity of human tasks, or be assistive - allowing people with fewer skills / less knowledge to be more productive.
Ben Waylin
Banned
Well-known member
A robot can't steal your job if you're retired.
rsm
Well-known member
in the early days I did some proto-AI (automation using rules engine models) in CICS macro-level assembler and command-level COBOL;
My first real foray into AI I started with Prolog (Borland Turbo Prolog), LISP and C (for neural nets), then moved on to Scheme; did some Smalltalk (early client-server, pre-Internet), then Java, JavaScript, and dabbled in Self, Microsoft Robotics Studio (great IDE IMO), NoFlo (JavaScript / Node.js); I haven't done much coding in many years as I mostly design solutions now.
One of my favorite languages, due to its' design and focus on simplicity is Io which is a dynamic typed prototype-based language; not really useful in the marketplace, but it's art and engineering IMO. https://iolanguage.org/
D
Deleted member 67774
Guest
Interesting. I like Prolog for it's do or die attitude. Did you find it use helpful in the transition to C and Java? Were you in Europe in the early days? Banking sector?
Ben Waylin
Banned
Well-known member
Ben Waylin
Banned
Well-known member
Monkey Man
Super Moderator
ATARI Basic back in 1982. 
I programmed some games I thought were cool but wish now I'd taken it further.
Almost plunked down for the Assembler (machine code) ROM cartridge but opted instead for Star Raiders ($100 ROM cartridge) and the cassette version of Frogger.
I programmed some games I thought were cool but wish now I'd taken it further.
Almost plunked down for the Assembler (machine code) ROM cartridge but opted instead for Star Raiders ($100 ROM cartridge) and the cassette version of Frogger.
D
Deleted member 67774
Guest
I programmed the Universe in AD 2069 on a 4004. It's the reason humans look to the heavens and ask "why?" I was also drunk.
rsm
Well-known member
No Prolog didn't help with C or any other languages; Prolog is a very different language model, it's a rules / inference engine, for backward chaining; not much else. We used it for what were called expert systems back then.
Assembler was the most important fundamental language I learned, and it got me interesting work for many years. Knowing and having years of experience in assembler would separate me from many other applicants, as did AI. Assembler made it easy to teach myself other languages, e.g., I taught myself C (using Kernigan and Ritchie book, "The C Programming Language"), and neural networks (using MeClelland and Rumehart's, book "Parallel Distributed Processing").
For AI and many other areas, I still think Scheme is the best and most powerful language IMO, but it seems Python and R, Java and JavaScript are the most popular languages. MIT was the biggest proponent of Scheme, they used it as their main CompSci language for many years. Like many languages, I taught myself Scheme, using the MIT textbook, "Structure and Interpretation of Computer Programs"
C is a step or two above assembler; I learned OO using Smalltalk; Java basically took the more common syntax of C and used the virtual machine model of Smalltalk - removing memory management from application programmers. In most cases, virtual machine-based languages are more efficient and less prone to errors. IMO. Now we have AI writing code.
My early years of pre-AI and AI were in the insurance industry; I moved on to finance, logistics, healthcare, military-industrial complex and direct DoD work.
Over the years, I did AI on and off, there were periods where interest and jobs in AI were low (except in DoD related projects), so I had to do other mainstream tech work: databases, front end development / UI, applications development, etc.
IME/IMO
C
Clig Herdunkachuk
Well-known member
I am a Principal Software Engineer at a healthcare analytics company.
We are doing prmarily java, sql, nodejs, and implementing performance sensitive operations in rust.
We’re honed in on ChatGPT type “ai” to improve our mapping of customer data to our catalog and contract data.
We did some trials on having ai actually write code. It was neat but underwhelming. It was able to reliably produce fragments and functions. Classes it produced were not that great. It was incapable of generating whole applications from scatch.
Its sql was pure crap. It regularly produced queries that did not generate the right results, performed like shit, and in several cases could not compile.
ChatGPT was terrible at math. Routinely generating incorrect mathematical functions and formulas.
It was common to find security vulnerabilities in “ai” generated code as well as poor performance. The code generated was very average. Looked like a junior engineer had written it.
It was particularly bad at things like implementing checks on external inputs and other general security best practices.
Where we found it exceled was when it was used by an engineer to bootstrap implementation. We did one where we generated a function to perform multi-threaded read and transformation of very large customer files in rust. The rust code it generated was not great but it allowed our engineer to focus on tightening up the generated code instead of producing it from scratch. It afforded a shorter implementation timeline.
It is proving very effective at mapping imported customer data into our historical analytical repositories. We were already above 90% on high confidence mapping. We are seeing that mapping rate go above 99% with this generation of “ai”.
We are doing prmarily java, sql, nodejs, and implementing performance sensitive operations in rust.
We’re honed in on ChatGPT type “ai” to improve our mapping of customer data to our catalog and contract data.
We did some trials on having ai actually write code. It was neat but underwhelming. It was able to reliably produce fragments and functions. Classes it produced were not that great. It was incapable of generating whole applications from scatch.
Its sql was pure crap. It regularly produced queries that did not generate the right results, performed like shit, and in several cases could not compile.
ChatGPT was terrible at math. Routinely generating incorrect mathematical functions and formulas.
It was common to find security vulnerabilities in “ai” generated code as well as poor performance. The code generated was very average. Looked like a junior engineer had written it.
It was particularly bad at things like implementing checks on external inputs and other general security best practices.
Where we found it exceled was when it was used by an engineer to bootstrap implementation. We did one where we generated a function to perform multi-threaded read and transformation of very large customer files in rust. The rust code it generated was not great but it allowed our engineer to focus on tightening up the generated code instead of producing it from scratch. It afforded a shorter implementation timeline.
It is proving very effective at mapping imported customer data into our historical analytical repositories. We were already above 90% on high confidence mapping. We are seeing that mapping rate go above 99% with this generation of “ai”.
Last edited:
rsm
Well-known member
The problem with open-to-the-public generative AI such as OpenAI / ChatGPT is the lack of governance; they will lie, hallucinate, develop biases, etc.
I heard of an attorney who recently used ChatGPT which used imaginary cases, he's under investigation for possible disbarment...
Healthcare and insurance especially; you need to be able to explain how results are generated or you're gonna face lawsuits.
C
Clig Herdunkachuk
Well-known member
The data used for training is crucial. We found that ChatGPT was generating some code using copyrighted intellectual property. It did not provide any attribution.
We also found that it was incapable of complying with opensource licensing requirements. When it used open source implementations, it failed to provide the license as required. Not a huge problem for experienced developers. I think we will see ip lawsuits on AI generated code in the future because people with just copy/paste generated code.
We are currently evaluating standing up a private instance for our mapping. Training it is… challenging.
rsm
Well-known member
Have a look at Hugging Face?
rsm
Well-known member
worth a look, they're HQ in France.
"Build, train and deploy state of the art models powered by the reference open source in machine learning."
https://huggingface.co/
