A lot of people are finding LLMs more useful than Google search.
I think a huge part of that is that Google and web in general has massively deteriorated in terms of finding clear, concise answers. Even if you can find it, it's covered in ads, cookie popups and so on.
But LLMs will inevitably follow the same path - investors want their returns. And the complete non transparency of LLMs will make it even worse this time around. Is that really the best answer or is it sponsored?
@joelanman SEO consultants basically wrecked the system with some recently decisions by Google execs to speed along this decline. Recently I found ChatGPT 4.o has included links to external resources which I normally have not seen. If LLMs were to screen out bad actors which throw up a ton of ads and pop-ups it could be a welcome alternative to Google.
I still think ChatGPT and others should be paying some kind of royalty to the services like Stackoverflow and others which were used to train their models in addition to linking to them.
@joelanman it’s sort of a circular problem too, isn’t it? people use LLMs because search is bad, but search is in turn being made worse by LLMs. (and though the problem predates the widespread use of LLMs which is why they seemed appealing to begin with, I think the earlier problem still stems from the same root: the prioritising of profit over providing a working service, and the conflict of interest that produces.)
The algorithms are mostly public, and the training gets exponentially cheaper, so it's going to be a commodity. I don't expect "AI" companies to get their money back on LLMs.
@joelanman@benjamineskola it's genuinely useful that LLMs are good enough to clean up their own training data. Somehow we've solved garbage in - garbage out!
@benjamineskola Such cynical view overlooks the hard problems that LLMs have solved.
e.g. W3C dreamt about the Semantic Web for decades, and suddenly we have ability to convert human-facing information into machine-readable form, instead of unsuccessfully trying to implement it from the other direction.
@joelanman@benjamineskola The ML solutions are inherently imperfect, but their errors need to be weighed against issues you would have with other approaches.
For example, invisible metadata turned out to be systematically unreliable, because it gets less testing, so is likely to become stale or bitrot. Invisibility helps using it for spam.
So apart of rare cases of honest, correct, complete up-to-date semantic data, there are no less flawed ways of extracting information from the web.
@kornel@joelanman The problem is that people don’t seem to recognise that they’re imperfect, or more likely just don’t care because they can make more money this way.
@benjamineskola This pattern of hype, and people not being ready, is quite common in technologies, good and bad ones.
The Web has created enormous value, but also brought phishing and dangerous disinformation that many people are unprepared for. Smartphones have connected billions of people, but also let Apple make a fortune on facilitating payments to gambling apps for toddlers.
And don't worry, in "AI" everyone is hemorrhaging money, except NVidia selling them shovels ;)
@joelanman Programmatically, I've had success with Ollama+Mixtral. It is easy to run locally given enough VRAM, and worked great for categorization tasks, anti spam, and search indexing – it understands any language, jargon & slang, and picks up even subtle context-dependent clues that a bayes classifier never would.
Being able to run pretty advanced models locally makes me optimistic that it can be used as merely a tool, rather than being yet another VC-funded surveillance trap.
Add comment