EnbySpacePerson

Photo by Eri Krull on Unsplash

This article isn't about the ethics of how models are trained or the ecological consequences of their use. DeepSeek shows that models can be generated on lesser hardware and that massive data centers with huge ecological consequencees aren't required to operate them. DeepSeek doesn't demonstrate the viability of ethically producing a model.

DeepSeek doesn't build a tool that actually does the kinds of things that real people would need it to do in order to be an effective tool.

What we're really looking for are tools that enable us to eliminate drudgery

There is a real need that people are trying to address through the use of LLM. It's not taking the joys out of being human by replacing creative endeavors with autogenerated “content.” There's an endless stream of fiction, music, video, anything you can think of really. There's so much of it, we invented an acronym: FOMO or fear of missing out. It's applied to a lot of things but streaming shows and live events are probably among the biggest uses.

It's only soulless modern robber barons who think that LLM, GAN, etc. should be used to replace the good parts of life to save more time for workers to do the boring, repetitive shit.

What the average, real person wants from LLM is solutions for repetitive, complicated, and annoying manual efforts. Unfortunately, LLMs are not suited for that.

A narrow selection of things real people are using LLM for

I'm going to use two tasks that are pretty-well established as common uses for LLM. I'm not saying these are necessarily the most common uses. I'm also using a third task inspired by a Simon Willison article.

1. Finding information
2. Solving math problems
3. Creating a web application to meet a particular need

Finding information

The first time I ran into this, I was complaining on social media about a sound issue I have on Linux computers. I had a general gripe about how fragile the audio set up was. Someone, unasked, put my post into an LLM and then pasted the output to me. The output was generic slop. “Check the drivers.” If that had been advice for someone on Windows, that would have been a meaningful first troubleshooting step ... if you had no sound at all and you already know how to check your drivers on Windows.

I find that to be generically true of using LLM to do research. If you already know the answers, you can evaluate what comes out of it but, if not, you don't really know what you have received. Because the person asking an LLM on my behalf didn't know the topic, they weren't aware that what they were passing on to me was not even wrong.

Google ruined search engines before companies started shoving AI into everything. There was a compelling desire for this kind of tool before search engines were ruined, though. Most people don't want to read through dozens of pages to find what feels like a simple answer. Further amplifying the issue, most sites today are filled with autoplaying ads that are replaced while you're scrolling the page. The paragraph you're reading moves while you're reading it. It's torturous trying to read anything on a major website. LLMs cut through that. Unfortunately, what LLMs provide is almost, but not quite, entirely unlike tea. There's no way to tell when it has made up information without doing the work of verifying that information through real research. Then you're back to the very problem you were trying to solve in the first place: reading dozens of pages.

The information you're looking for almost certainly exists on the web somewhere. The issue is getting to it. LLM don't solve the problem but they can give the comforting feeling that they have in the moment you use them.

Solving math problems

LLMs famously can't do math. They're functionally designed to give plausible sounding text back. In order to give good math responses, you'd have to program something else. People have been building tools on top of LLM or incorporated with LLM that rely on Python to give accurate answers to math questions.

If the model or tool can successfully convert a word problem into a math problem that Python can solve, you can get an accurate answer but it's not coming from the LLM really. You're getting it from high quality Python libraries. I'm not linking any of these because I'm not qualified to evaluate them.

There are other tools on the internet that will do that for you with or without Python knowledge. There is a learning curve to them but they exist and they don't use massive amounts of computing power.

Creating a web application to meet a particular need

I'm deriving this topic from the Prompt driven app generation is a commodity already section of LLMs in 2024. It's easy to focus on the specific tool that Simon used Claude to generate but that's not what Simon was talking about. He's talking about using LLM to generate tools. The tool he generated takes pasted rich text (copied from a website or other formatted document) and pulls the URLs out of it. The tool itself isn't using LLM to extract the URLs. It's just using plain Javascript, HTML, and CSS. You can save it locally and run it from your own computer offline. It's honestly pretty cool and I can see exactly why Simon wanted it.

Having tried to recreate his tool from scratch using information on the web or even find someone else who has done the the same thing, I can see the benefits of having an LLM just do it for you. However, using LLM to generate these tools really is wasting a massive amount of energy.

I build these kinds of labor / drudgery saving tools in Python. I tried demonstrating a replacement for this exact tool through a hosted Python browser solution but my options are all restricted to plain text input fields. Simon's tool takes rich text which is so much simpler for the person using the tool.

I'm not one of those people who is going to tell you that everyone needs to learn to code. It was arrogant advice when people started giving it and I'm hoping we're past that advice even though we're clearly not past that type of thinking. However, someone else has almost definitely made the tool you need. It's a question of making those tools accessible and discoverable.

You can't use the tool you can't find

In each of the three cases I used for illustrative fodder, the real problem is that people can't easily access the tools or information they need. Even before Google ruined search, this was a problem. The chat interface is appealing because it's simple and appears to provide the required information with a minimal amount of fuss.

Automated chat interfaces existed before LLMs were popularized. They offer functions that retreive information from a knowledge base or follow a logical sequence of steps with the customer. The later are similar to an interactive fiction game like one programmed with Twine or Infocom. If you've dealt with those kinds of chat bot systems regularly, you know most of the time you use them you're going to end up trying to talk to a real person. Something similar happens with phone trees. People start swearing or shouting “agent!” to get to a real human being.

Chat bots aren't the “killer app” as certain management types used to say. The reason they can look like they are now is you've got a fabulist on the other end that will generate you endless amounts of fiction when it can't regurgitate the right information.

When the veneer wears thin, we will realize we're in the same situation we've been in for a while: left with only tedious and difficult ways to get the information and tools we need.

Summary

The information and tools we need already exist. People like you and me made them. Unfortunately, everyone's experiencing the wonders of the internet hobbled and packaged by ad infrastructure.

Whoever finds the best way to provide people with the information and tools they need without the shitty experiences provided on most commercial websites today will be able to write their own blank checks.

LLMs have been writing checks on the theory that they are making that tool and more. They're not. They're a distraction from the real issues we face:

We need good ways to find the information and tools that will make our lives better.

(The article officially ends here but you can continue to the Pedants corner if you feel the need.)

Pedants corner

You didn't cover the specific use case I have for LLM

Even if I could have, I wouldn't have. This thing is long enough to illustrate my point. Possibly longer. I already left OCR off. It's a solved problem. It's hugely wasteful to put any kind of “A.I.” in front of something that could be done at least as fast and accurately with a piece of software that will happily run on a 20 year old computer.

You are wrong because you don't use LLM enough to know how useful it is

I don't need to drive every kind of internal combustion engine transportation and electric vehicle to tell you that mass transit and dense, walkable development are better for humans, the planet, and getting through rush hour.

My favorite search engine actually works!

Great. I look forward to hearing about it. Don't bother telling me about Google, Duck Duck Go, Start Page, Kagi, Bing, or Stract though. I've tried them and I disagree with the characterisation that they actually work. (Though most of them are better than Google today. Stract shows real promise.)

#Essay #Technology #LLM #AI #Technology #OpenAI #ChatGPT #DeepSeek #Llama #Gemini #Claude