The AI is garbage thread

[CitizenVectron]

18 minutes ago, Jwheel86 said:

 

If only the industry would move to a standardized proposal that's excel and not pdfs... 

 

PDFs are great when you need to have something print in a certain way. Otherwise...people should just be using Word or Excel files, imo.

[finaljedi]

5 minutes ago, CitizenVectron said:

 

PDFs are great when you need to have something print in a certain way. Otherwise...people should just be using Word or Excel files, imo.

 

I work in EDI and I once had feed specs sent to me in fucking Power Point.  I'll still take that over a word doc with a snip from their platform pasted in.

[legend]

I think ClosedAI's o1 is wildly over hyped (not that that's anything new about ClosedAI). However, I am glad to see them slowly moving further and further away from fucking token distribution fitting.

[Kal-El814]

The main use of AI I interact with regularly is Copilot and Excel. I’m pretty good at writing formulas to extract stuff from free text fields from data extracts so they can be plopped into pivot tables and such. But it’s time consuming to get it right especially if the raw data is gross, which is often is. Buuuuuuut now I just tell Copilot what I want in a column and it does it more or less perfectly. There’s really no learning curve either, assuming you can type out what you want. So while it’s a huge time saver, it’s also taken at least a differentiating skill for folks in roles like mine and made it irrelevant. Or at least it will, as a lot of people with access to the tool seem to not know they can ask it to do this stuff but I assume that won’t last long.

[legend]

By far the most useful LLMs have been for me is looking up well documented APIs or asking simple questions that involve using APIs. Things like excel queries probably fit that bill well too. That and they're awesome when you cannot remember the name of a word/term but can describe it.

 

God help you if you ever ask it to code something that hasn't already been effectively documented on the internet though. It will spit out text. It will just be terribly wrong.

 

As this tech continues to move away from fitting token distributions, and starts training toward grounded concepts, that will open more doors for the models being able to more accurately report when they don't "know" something and more reliably relay information that is actually true. That's when these systems will be much more valuable and will basically become the VIs from Mass Effect.

[Keyser_Soze]

I asked copilot to write Code for Gadot for an open world game. Then I put the code in it and the game didn't work!

[Jason]

1 hour ago, legend said:

I think ClosedAI's o1 is wildly over hyped (not that that's anything new about ClosedAI). However, I am glad to see them slowly moving further and further away from fucking token distribution fitting.

 

If you heard about the Diplomacy bot Cicero, I know one of the guys who worked on that at Facebook and he posted recently that he's been working on o1. 

[Jason]

2 hours ago, Kal-El814 said:

The main use of AI I interact with regularly is Copilot and Excel. I’m pretty good at writing formulas to extract stuff from free text fields from data extracts so they can be plopped into pivot tables and such. But it’s time consuming to get it right especially if the raw data is gross, which is often is. Buuuuuuut now I just tell Copilot what I want in a column and it does it more or less perfectly. There’s really no learning curve either, assuming you can type out what you want. So while it’s a huge time saver, it’s also taken at least a differentiating skill for folks in roles like mine and made it irrelevant. Or at least it will, as a lot of people with access to the tool seem to not know they can ask it to do this stuff but I assume that won’t last long.

 

2 hours ago, legend said:

By far the most useful LLMs have been for me is looking up well documented APIs or asking simple questions that involve using APIs. Things like excel queries probably fit that bill well too. That and they're awesome when you cannot remember the name of a word/term but can describe it.

 

God help you if you ever ask it to code something that hasn't already been effectively documented on the internet though. It will spit out text. It will just be terribly wrong.

 

As this tech continues to move away from fitting token distributions, and starts training toward grounded concepts, that will open more doors for the models being able to more accurately report when they don't "know" something and more reliably relay information that is actually true. That's when these systems will be much more valuable and will basically become the VIs from Mass Effect.

 

Stuff like chatgpt is definitely good for saving you time on sifting through stack overflow to sort the wrong answers and people being dicks about "duplicate question" and just surfacing the right answer. But I've also seen it just wholesale make shit up when I've poked at it by asking it coding questions for stuff that I know isn't publicly available (I knew it shouldn't have an answer so I was curious about what it would offer as one). And of course these models will become useless if they result in, say, nobody posting on stack overflow anymore. 

[legend]

36 minutes ago, Jason said:

 

If you heard about the Diplomacy bot Cicero, I know one of the guys who worked on that at Facebook and he posted recently that he's been working on o1. 

 

I am somewhat familiar with it and it is much more interesting than plain LLMs. I don't think I know any of the authors though, which is surprising because the work overlaps with my primary area (reinforcement learning).

 

The other work taking LLMs in a better direction is the DeepMind work on the math challenge because it's actually tying the language to grounded formal descriptions and search. (And whatdoyaknow that also happens to be from the reinforcement learning community. Funny how that works out  )

[Nokra]

7 hours ago, legend said:

By far the most useful LLMs have been for me is looking up well documented APIs or asking simple questions that involve using APIs. Things like excel queries probably fit that bill well too. That and they're awesome when you cannot remember the name of a word/term but can describe it.

 

Yeah, my biggest use of AI so far has been to just quickly ask it to write some regular expressions for me when I couldn't be arsed to read through the documentation to refresh my memory on the syntax for the particular languages I was using (Python and C#).    

[unogueen]

1 hour ago, Nokra said:

 

Yeah, my biggest use of AI so far has been to just quickly ask it to write some regular expressions for me when I couldn't be arsed to read through the documentation to refresh my memory on the syntax for the particular languages I was using (Python and C#).    

mebbe rape is your culture.

[chakoo]

LLMs for me have been useful when trying to do vague searches. In the past, you used to be able to somewhat describe parts of what you were looking for and find the answer in a few rounds with Google search but now that is nearly impossible. I've even used it to find the title of old movies by just simply describing the scene in the movie.

 

For pure code I still find it dog shit for generating usable code but maybe ok if I just run a few concepts to get an idea of where to start building from. I find GitHub Copilot straight-up awful because half the time it makes shit up (including variable names in my own damn code) and blocks IntelliSense from working which can actually give me the right answer. =/ Yet I find it funny I'm using LLMs in a product I'm building right now for Neo4j Cypher queries. *shrug*

 

2 hours ago, Nokra said:

 

Yeah, my biggest use of AI so far has been to just quickly ask it to write some regular expressions for me when I couldn't be arsed to read through the documentation to refresh my memory on the syntax for the particular languages I was using (Python and C#).    

 

Man, nobody remembers the syntax for regular expressions. I'd rather write a complex piece of code in assembly than a few lines of regex.