In a perfect world this technology would never have been used outside labs.
As it is, it’s very valuable for certain use cases, and if you or others reading this haven’t used modern paid models in the last 2 months then your POV is already outdated.
I agree with you about fundamental problems, have worked in leading edge CS research managing teams and although there are flaws there is real utility.
I don’t think Joe Anybody should be able to start a conversation with ChatGPT and within 10 prompts be told that chemtrails are real and we need to “wake up” (I did this experiment myself a couple days ago), but if you work in this industry and are writing off these tools than you’re making a grave mistake because they already have a ~10% measurable impact on increasing throughput even as flawed as they are, and that’s with now-outdated models.
Ignore the metrics about % of code written etc. and focus on the researchers who care about productivity.
As an accelerator to a solo developer who knows what they’re doing it’s intellisense levels or better, and I don’t even know how to quantify the utility of “rubber ducking” with the best models, it’s invaluable to a certain degree in that regard.
You’re right there is a bubble, especially for companies who are just tweaked front-ends for frontier models, but they aren’t going away. We will see a continued shift toward automation using these tools, and the places they aren’t a good fit will be scaled back after ~6-12 quarters of reduced productivity in those areas once upper management figures out where they best fit.
Using one model, I wrote Javascript to customize Macrumors in my browser a few days ago, and I have never written a line of Java anything in my life. It took ~5-10 minutes. Once I knew what I was doing I customized some other sites I regularly visit in less than 2 minutes. Made my browsing experience measurably better, even though this is a silly toy example it is a real one.
Stuff like these small things add up, I know models struggle with large contexts and start to fall apart (Apple’s paper on that last month was excellent), and I also know that scale alone can’t solve some of these things, but small improvements add up to real productivity.
Think about how batch scripts made things a lot faster in the older days, some of those simple things saved hundreds of hours over the course of their lifetime in production – in my early career I worked with someone who won awards writing them to automate some esoteric stuff at a large company.
All this is not near-term, it is now. Anyone who thinks this is a fad or that these models are just stochastic parrots is willfully ignorant at this point, even from a technological perspective you can read some of Anthropic’s research (it’s nice that they are actually publishing still unlike OpenAI) that fully debunks that theory that possibly only really applied to GPT-1.
Do LLMs and GenAI models have intractable problems? For now, yes. Probably permanently. Are they still worthwhile and rapidly improving to become even more useful despite these faults? Also yes.*
*Disregarding the fundamental harm to society that they cause especially when using them without a deep understanding of the caveats.
…
In ~5 years once World Models come out, if they work, everything will change. And I don’t think AGI is anywhere near close, but World Models are real and are coming. Read up on them if you haven’t and care about these things – we’re going to see dramatic shifts in society probably within a decade and within 15 years the way we interacted with technology pre-2020 will be nearly unrecognizable, similar to a car phone from the 1980s vs. an iPhone.
They won’t be an evolution of any LLM either, they will be a wholly new thing. Zuckerberg knows this, has people like Yann working for him that know this, and so Meta is willing to pay any money to make sure they are either at the lead or near it.
Apple isn’t even able to keep researchers working on extant, known quantity technology. It’s a problem.
Moonshot R&D is necessary in these fields. Tim needs to go and Apple needs either a futurist visionary like Jobs or a modern technologist like Jensen leading their company in a decision-making role.
Last edited:
