IBM open-sources its Granite AI model!

[thewayne]

First off, it has to be pointed out this is a specialized AI model designed for programmers, not a generalized model like ChatGPT et al.

IBM trained it specifically on open source libraries to which they explicitly had permission, basically bending over backwards to avoid any possible legal issues. And they now have a working model that they've released to the public! Granite was trained on 116 different programming languages and has from 3 to 34 billion tokens, presumably per language. I wonder if you can ask it to list all the languages it's trained in, I'll bet there's some pretty esoteric ones in there! I'd love it if it had MUMPS! (I once found a book on MUMPS programming at the Phoenix Public Library, I imagine it's been weeded by now)

Anyway, interesting article. It describes how it was trained, etc., but one of the more interesting bits was saying that in the rather short time since ChatGPT et al have appeared and everyone started creating their own LLMs, the cost for training up an LLM has dropped from millions of dollars to thousands! That's a pretty impressive scale drop.

https://www.zdnet.com/article/ibm-open-sources-its-granite-ai-models-and-they-mean-business/

https://www.zdnet.com/article/ibm-open-sources-its-granite-ai-models-and-they-mean-business/

Comments

[disneydream06]

Is all this AI stuff the dumbing down of the human race?
Don't think, let a computer do it for you?
Or am I just old? LOL.........
Hugs, Jon

[silveradept]

This seems like a better idea of keeping the focus tight and only using things that there has been explicit permission granted for, but also, I still don't quite get why you would want to use such a product, given that anything it outputs will still likely have to be worked over with a fine-toothed comb just to make sure it actually does what it says it does.

[thewayne]

I don't know.  I did some experimenting with ChatGPT trying to use it as a code generator.  The results were not terribly impressive.  This was the free, 3.5, model, not the paid 4.0 model.  There are a lot of programmers who swear by (and probably sometimes, at) the paid models and swear that it really helps their productivity.  Given the reports that code audits claim that LLM-generated code generates more security holes, one has to wonder.  I suppose it is highly variable, depending on the skill of the programmer and the type of programming that they're doing, and how much purely-generated code they're inserting into their project vs small snippets that are being integrated.

[thewayne]

No, I don't think it is. At least, not in this particular case. Let me explain based on my experience, which obviously is not universal. I am an old-school systems analyst, I was taught to do thorough interviews, analysis, and thinking before you start coding. So you, in the field of medicine, need a new programming system. You need to enter data, store it, and produce reports. You tell me you need three specific outputs, 1, 2, and 3. But what I actually hear is A, B, and D. You didn't mention C. Because I've been working in the health care field for some time, I know about C, I know that it's important, and you made assumptions about that but didn't mention it. So you actually need at least four things: A, B, C, and D. Maybe more. This is why good programmers have to go beyond the interview process. So when I write up my analysis, I write up at least those four things. You review it and sign off on it before I start coding. Now, I have to wrote a database structure (called a schema) to store all the data that you need to enter in a way that it can be efficiently retrieved and allow reasonable ad hoc querying. I have to make good input and querying forms so that it will minimize mistakes when data is entered (such as validating appropriate dosages or states) to try to make the data as accurate as possible, because I'm not designing it for you, a knowledgeable user, I'm also designing it for the people who come after you, fresh out of school, you aren't as familiar with the input forms.

THEN I have to make the reporting system! More screens to program, I have to relate the entered data to produce the reports, maybe do some graphs, etc.

Now, an LLM can help with a lot of that. But you really want a skilled programmer behind the keyboard who has institutional knowledge to recognize that C is missing, or that maybe there's a legal requirement that certain data has to be wiped after X years, who knows. There's all sorts of weird requirements that an AI may be ignorant of.

To give an example, I queried both ChatGPT 3.5 and 4 about the song Smoke on the Water about the fire at the Montreux Jazz Festival where "some stupid" fired a flare gun during a Frank Zappa and The Mothers Of Invention performance inside the casino hall and "burned the place to the ground". I asked both engines "The song Smoke on the Water by Deep Purple was based on a real incident. Did they ever catch or identify the person who fired the flare gun?" Both said no. If you go to Wikipedia and open the entry for the song Smoke on the Water, you'll see he was identified by Swiss authorities as Zdeněk Špička, a Czech national who fled the country and never faced trial.

LLMs can be useful, but they have to be watched by a knowledgeable eye.

[disneydream06]

Did they ever catch or identify the person who fired the flare gun?" Both said no.
Must be true then because everything you read on the internet is true. Isn't it?
LOL.............