Post #2585310

@cybso@osna.social @FediThing@social.chinwag.org @krautdragon@mastodon.social AI is not synonymous with LLM, even though the current marketing hype makes it feel that way. LLMs are generative models that produce plausible looking texts, but there are other branches of AI research and development which get much more reliable results. The entire field of machine learning is really useful for pattern recognition, and it has been used on a large scale in all kinds of software and firmware for a long time. It is AI, just not aggressively marketed as such. Of course AI can be very useful. It's just that large language models aren't near as useful as the marketing divisions of AI companies tell us, and they're also computationally expensive, which means they will probably play a much smaller role in the near future, getting replaced by more traditional natural language processing systems that use some kind of symbolic logic for reasoning. Symbolic AI is an entirely different branch of AI R&D. Symbolic AI is the old-fashioned kind where humans analyse all kinds of problems and write down the mathematical and logical rules on how to solve them methodically. The current approach of doing it all in machine learning, doing it all with larger and larger artificial neural networks, is just too expensive, uses too much power, and produces too much bullshit. A swarm of small artificial neural networks, each of them specialised for a single task, all of them glued together by a symbolic logic framework, that's the way to go. For parts of the natural language processing tasks, small LLMs (much smaller that those GPT ones) will be used, hedged in by other AI agents that provide strict logical reasoning and fact checks. And the entire system should be small enough to run on the local machine, no computing centre needed. Of course this means that the entire LLM-based AI business is finished. When the bubble bursts, it will be worse than the Subprime Mortgage Crisis. But we will still have all the AI research results from this bubble, and even when the humongous LLMs shut down due to bankruptcy, all the small and small-ish open source models will still be around for everyone to tinker with.