My take: I don't think recursive self-improvement (RSI) will happen with current LLM architectures, even if they become 1000x "smarter".
RSI requires more than being better at solving problems. An AI needs to decide what is worth improving, come up with genuinely new ideas, and decide which ideas are worth pursuing.
Current LLMs are still weak at two of these: judgement (what matters) and new ideas. They're very good at combining and developing existing ideas, but that's different from independently finding important new directions.
That's also why I don't think most knowledge workers can simply be replaced by LLMs. Many tasks can be automated, but the most important ones can't.
I don't think this article adds anything that hasn't already been discussed to death on HN.
LLMs are powerful new tools, and we're all figuring out how to use them effectively. Some developers are getting great results, others aren't. Some instinctively love them, others hate them. Pretty much like every other developer tool HN has argued about, from Lisp to Emacs.
And yes, management often doesn't get it. But that's hardly new either.
> YouTube recommends videos designed to keep us watching
Yes, but you don’t have to subscribe to what it suggests. I keep a carefully curated list of subscriptions and very rarely look at recommendations. That gives me a high-quality filter without all the algorithmic BS.
I disagree. Software developers seem to have adapted to AI much faster than the maths community.
There are certainly still strong opponents of AI, but they seem to be a small, vocal minority. Most developers seem to have simply adapted and moved on.
It's possible and I dare say even likely that you are missing nuance and observing bias from your bubble. It's possible to avoid being an AI "opponent" while still being a sceptic. It's possible to use AI reluctantly and with tact instead of going all in.
The whole thing is far from over when we haven't experienced the worst of the technology yet: deskilling, concentration of power, extreme imbalance of wealth, military and surveillance uses, etc.
> Working on other people’s tasks with AI agents is nothing short of being a paper pusher.
Not my experience at all. I love not having to do the low-level code writing. I’ve been doing that for 30+ years, so I’m more than happy to let the AI handle that part now!
LLMs are great at pattern matching and “filling in the gaps” based on learned patterns. However, they are still really bad at (1) deciding what is valuable or worth doing and (2) coming up with truly new ideas.
So I don't see much benefit in this approach compared with human-in-the-loop development. It basically turns development into a waterfall process, with the human deciding what to build up front, rather than the more agile approach where the human and LLM continuously decide what to do next.
RSI requires more than being better at solving problems. An AI needs to decide what is worth improving, come up with genuinely new ideas, and decide which ideas are worth pursuing.
Current LLMs are still weak at two of these: judgement (what matters) and new ideas. They're very good at combining and developing existing ideas, but that's different from independently finding important new directions.
That's also why I don't think most knowledge workers can simply be replaced by LLMs. Many tasks can be automated, but the most important ones can't.
reply