Louis Ingenthron @LouisIngenthron@social.cleverlibre.org

@phrees Every algorithm ever written by any computer scientist ever has the exact same problem:

Put garbage in, you'll get garbage back out.

So, you debug an LLM by getting rid of the garbage in the training data. In your example above, you'd either scrub the rulers from the images entirely, or more likely you'd make sure there was an even distribution of rulers among the various diagnosis category input data sets.

@jhavok @knightlie @ErikJonker @f4grx @jeffjarvis It is, yes, but some of that code invokes functions that have now been deprecated for years. Just because it once compiled doesn't mean it still will.

And that's before you get to the fact that the AI is actively combining code based on patterns, not just regurgitating it.

@jhavok @knightlie @ErikJonker @f4grx @jeffjarvis Not quite. Right now, the algos are trained on as much data as they can find, which includes some truly terrible examples. Therefore, if you just let it run loose without a knowledgeable human to work with, it just produces crap that won't even compile, much less run, for anything but the most basic examples.

That's a big reason why AI works so much better as a companion tool to improve humans' efficiency, rather than an outright replacement.

@jhavok Yeah, I was referring to the larger ML field.

@phrees @ErikJonker @f4grx @jeffjarvis That's a *bug*, not "bullshit". All software has bugs. They can be fixed.

@phrees @ErikJonker @f4grx @jeffjarvis I'm surprised you didn't sprain your ankle shifting your argument that hard. Plus, there are ethical and profitable applications for the tech, especially in medicine. But enjoy your crusade.

@knightlie @ErikJonker @f4grx @jeffjarvis I mean, making people's jobs easier isn't a bad thing. It's a close race between that and war for what drives our species to innovate the most.

@fifilamoura @carnage4life Depends on how you define "much of the tech industry". If you're talking about the big boys who get pumped full of money by VCs, yeah, that might be true. But there's a sea of us little guys who are genuinely trying to make life better with technology, and based on pure numbers, we're the majority.

@ErikJonker @knightlie @f4grx @jeffjarvis Also, corollary to that, they're incredible at translating between languages, which up to this point has been a very expensive and error-prone human process.

@knightlie @ErikJonker @f4grx @jeffjarvis
Yeah, it's not a fact machine. That's like trying to use photoshop to edit a text document.

@knightlie @ErikJonker @f4grx @jeffjarvis

I use it for software engineering; it's able to fill in entire functions that would have been relatively simple, but very tedious, for me to do.

LLMs are also amazing at teaching you a new skill for which you already have some domain knowledge. I.e. I used an LLM to help me learn enough python to build and train my own LLMs in less than two days. Now, I needed the domain knowledge to recognize when the LLM was wrong, but it still vastly accelerated my learning.

The LLM I trained was used to identify certain types of materials from a web crawler. This kind of tool would be incredibly useful to researchers, by helping them narrow down an endless volume of data into a smaller manageable subset of "likely relevant" data.

On top of that, I can't say too much about this, but I work with a company that makes medical simulators for training surgeons, and we're using ML to help identify and generate information for training purposes.

Then there's the generative image AI. I'm a shitty artist, so when I'm prototyping games, these allow me to add it placeholder graphics that aren't pure shit. If the game idea proves viable, then I can invest to pay an actual artist to replace them before launch. The lack of shitty programmer art helps me stay motivated and excited about the project and costs me nothing.

I'm sure I could go on, but I think I've made my point.

@knightlie @ErikJonker @f4grx @jeffjarvis Right, but they were bullshit even at the tech level. LLMs do have actual revolutionary uses, even if marketing is currently overselling those uses.

@potpie You're describing every marketing department on the planet.

@knightlie @ErikJonker @f4grx @jeffjarvis So it sounds like you're actually mad at marketing and garden-variety businessmen, not LLM tech.

@parismarx This doesn't even make sense. It's just bound to make the lawyers rich.