When it pops, will the current pervasive LLM and Gen’AI’ slop dry up as their companies go under? Will LLM and agentic chatbots be removed from all the apps that have been shoehorning them into everything? Or will just some companies go under, some stocks lose value and people’s 401ks are hollowed out
Blockchain has a use and all the companies who were trying to jam blockchain into irrelevant shit are gone. I think AI companies shifting away from grabbing eyeballs to actually making money will help raise prices of enterprise licenses so more companies stop forcing it on all their employees and start rationing it. Right now our fucking chats are fed into it so when everyone signs off for the night, we get a notification that says something like “various participants wished the others a pleasant evening.” That might cost a penny now but it made somebody feel like they were on the cutting edge of something. When it costs a dollar and is clearly going to remain precisely that useless, I sure hope that same somebody pulls the plug.
Blockchains aren’t necessarily good or efficient things, but they do still exist. Unlike NFTs. I don’t have any reason to think AI won’t be the same. Transformers and LLMs are way better at NLP than anything else that has been invented so far. That’s not going away, people will just have to actually justify their use once the prices start to match the costs.
That said, I think looped models, RYS, and other alt model architectures will continue to drive the training costs down like Deepseek did and totally demolish the AI company valuations. Just make it bigger was never going to be a winning strategy, it’s completely naive. That’s really the source of the bubble. The technology has nothing to do with it, the tech bros are just desperate for a machine god and they got caught up in their chat bot delulu about it and made it everyone’s problem.
What impact do you think the LLMs will have on software engineering and the jobs in that field in the medium to long term as a technology?
LLMs are extremely useful for any kind of Natural Language Processing workloads, you might need to do. They’re like an order of magnitude improvement over previous techniques. Which is interesting, because Transformers actually didn’t come from natural language research. They just happened to be really good at encoding language.
I am definitely a skeptic when it comes to AI usage, but I have seen case studies. Where if you meet certain very high preconditions, you can use tools like claude to port an entire application from, say react to sevelte. For this to work, though, you have to have a precise description of the semantics of your entire application. So ~100% test coverage, ~100% meaningful types coverage. Maybe not universally useful, but it does make rewrites into safer or faster languages somewhat more tractable, at least while tokens aren’t too expensive.