Thread regarding SAS Institute layoffs

For those unfamiliar, pre-trained AI models have been a thing for a long time. There’s 2 types of training for an AI models (which is literally just a bunch of math equations and statistics behind the covers) - pre-training and fine tuning.
It’s like 90% of the work has been done for you, now you just train the model on your company specific stuff and you’re set.
An example of this is the BERT model, made by google. There are also pre-trained models for example that are trained on public legal corpus, and law companies use them (with fine tuning on internal documents) so that they can query information using natural language instead of having to do hours and hours of research.
My understanding is that these are the kind of models that SAS will be offering, that are pre-trained to solve some specific problem but are not fine tuned to the parameters needed by a company.

“ Yes, you're right,”

Wait what? I’m one of the people on here that everyone loves to hate because my points don’t support their world view.

You just made my day.

@1jk+1jthj0r2p Yes, you're right, looking back I do think I was being ageist. There have been many older coworkers who have helped me time and time again, and I have no doubt that their minds are just as sharp as people half their age.

We would do well to make use of this pool of talent, but corporate politicking would most definitely get in the way of this.

@1hr+1jthj0r2p Great post and summary. Only thing I disagreed with is the part about older developers not being willing or able to learn knew things.

IMO that isn’t an age thing but a personality trait. I know plenty of youngsters I wouldn’t trust with learning and and pushing new/emerging technologies like MCP protocol. I also know plenty of old people I also wouldn’t trust with that same thing.

However I know a small % of youngsters and an equally small % of oldies that I would very much trust to dive in enthusiastically and have the wisdom to push the boundaries and determine what makes sense.

The reality is simply that not all developers are created equal. That is not a knock on the run of the mill developer that just wants to get their job done and nothing more. You need both types and the average run of the mill is going to be the bulk of people.

@1hr+1jthj0r2p, I think the OP was asking about the viability of ready-made 'AI models' developed by R&D, and not genAI and LLMs. I think a few such 'AI models' are generally available and several are in the pipeline. I'm curious to know how many of these have been sold. Does a real market exist for them?

@e6+1jthj0r2p

I'll take a stab at this because I'm in the field (was working part time at SAS when the AI Frenzy first begun with ChatGPT before I joined full time).

The whole AI market is very much oversaturated with "AI apps". What I mean by that is a wrapper applications around some generally available AI model, like Claude, OpenAI, Perplexity, etc., these are a dime a dozen.

I hesitate to say it but if you pay attention, these are quickly becoming almost "standard" in the industry. Think about how many "AI Assistants" you see when you're on the internet. Whether that's checking your bank account, looking for recommendations for a restaurant, filing some type of report, etc. they're everywhere!

As for SAS, I do think these applications still have business value (in the sense that they generate "hype" around AI, which is something we need for an IPO/sale). And we are actually making strides in integrating more of these types of wrapper applications into certain products. However, I would not say these are the "bleeding edge of AI applications", this would be moreso a "stock AI application".

In my mind, just because we use state of the art AI models accessible via an API, does NOT mean the application itself is state of the art, hopefully that makes sense.

What is more interesting, and closer to the actual forefront of really "doing AI" IMO would be more akin to the research, training/finetuning, and productionizing of an LLM. In other words, creating and serving our own LLM. Now we don't actually have to go and develop our own LLM (that would be even further down the research rabbit hole), but we can use opensource LLMs and finetune them for our own purposes.

Also, as an aside, the MCP protocol is something that I think we should look at and invest in heavily if we actually want to incorporate agentic AI in this company. While we wouldn't be necessarily developing new LLMs or researching new methodology, we could continue to innovate by coming up with ingenuous ways of using the protocol in certain workflows (people have used MCP for slack integration, Google Drives, etc.)

Now unfortunately, trying to "do AI", or at least what I called "real AI" -- this all requires investment. Like, a lot of it. Money, time, people, and brain power. On all fronts I would say we lack the necessary investments to make this a true reality here. Money and time I don't have to explain. But we also just don't have enough people who are specialized in this particular area to make a huge impact. The other thing is that we lack the pull to attract the top level talent we would need.

Our workforce is in no way up to speed with this rapidly changing field. This applies to all levels -- testers, managers, directors, developers, and importantly product managers -- and across multiple surfaces (the tech stack, the expertise of the developers, the knowledge gap). I am in no way shaming our older developers, as they're the reason why I even have this job in the first place, but how many of them would be willing to learn new languages, new protocols, and new methods/techniques in order to work on products that upper management themselves do not fully understand, do not truly buy into, do not give proper resourcing to, all without much help because others do not understand what you're doing. I would wager not many.

Perhaps I was getting a little bitter there at the end, but hopefully this gives you a glimpse into the current state of affairs. I've accepted this reality and hold no ill will (except maybe to PMs...) for the state of things. But I do still wish we could have more people and more investment into this field.

I remember one time when I was still employed at SAS (I took the first buyout) where JG said "analytics" then corrected himself saying that he was told that he is supposed to say "AI" now instead of "analytics." Statistical procedures have always been able to score data. But it was no longer fashionable to say that, so it became analytics instead of statistics. Now it is AI. I am totally out of touch with what is happening at SAS now days, but I wonder how much they REALLY do AI and how much they say "AI" because they think that will convince the naive that they are state of the art. Perhaps a current employee can fill me in.

SAS has been using AI and ML lingo in their marketing for years. What I saw was all rules-based or simple anomaly detection - hardly cutting edge stuff.

This is an attempt by SAS to join the AI bandwagon. I wonder how many SAS sites have licensed this AI modeling product.

The article you linked to isn't intended to address technical specs and questions. It's not directed toward users of the technology.

It's just marketing-produced content meant to explain the subject at a high level to the average person googling 'What are AI models?'

There are surely other sources and places where people who geek out over the details of the how and the why can get the info they want to know.

This is not that and it isn't intended to be.