Why Chatbots Based on LLMs Cannot Be Improved Even If More Energy (Money) Gets Wasted on Them
Last year: "GPT-5" is Another Microsoft Dead Cat Trying to Bounce
This year:
LLMs are irredeemable "BS generators" that do not do anything particularly sophisticated, it just takes a lot of energy to spew out some grammatically-correct words/sentences (this does not mean the sentences themselves contain correct information). Many companies construct LLMs because of the hype - which carries VC fools along with it - and multiple implementations help make the point that scientifically-speaking it's neither hard nor new. Yet nobody can do it well. It's infeasible to make those things work well, even if one gets a pristine (pre-LLM frenzy) copy of "training sets" to train LLMs on. Anthropic physically destroyed many high-quality books; the outcome still sucked.
Right now it's starting to seem like boats are getting stuck again (in a major oil and gas hub). This means energy costs will continue to climb and some national reserves will soon be exhausted (in some places they've already run out and began begging a hostile Russia for emergency/urgent supplies). This is bad news for energy-hungry slop pushers.
Slop made by LLMs is still full of errors. Being "patient" and waiting for them to 'mature' didn't help. They didn't mature, they got increasingly full of manure. The falsehoods are fed back into various LLMs. So costs go up, quality goes down, and the loans for the peddlers won't keep coming; the shareholders want to see some real returns (financial returns) or some path towards improvements. It's just not happening.
There are now public metrics and even admissions that LLM usage went down. No matter how much they gamed the numbers and tried to fake "demand" (e.g. by accommodating "prompts" everywhere) users steer away and even actively protest.
The prospects of LLMs getting everything right are zero. LLMs lack comprehension, so they cannot work. As pointed out the other day in relation to IBM, chatbots will never work as GPs or pharmacists, even if you call them "superintelligent" or "agents" or whatever buzzword comes next for "generative" or "advanced" "hey hi". There is no room for guesswork and approximations (lives at stake); accountability is another matter.
Expect the media "churn" (churnalism) to carry on, as unprincipled so-called 'press' goes all "rah-rah!" for chatbots while falsely describing them as intelligence. Unless they quit participating in this scam, they will have no credibility left. And no Sugar Daddies, either. Those Sugar Daddies are on borrowed time and when the borrowed money they allocate to marketing/PR runs out we'll see the same silence we now (for years already!) see in relation to "metaverse", "blockchains", "IoT" and so on. █