This was an intuitively-appealing belief, even with some qualified experimental support, as of a few years ago.
However, since then, a bunch of capability breakthroughs from (well-curated) AI generations has definitively disproven it.
However, since then, a bunch of capability breakthroughs from (well-curated) AI generations has definitively disproven it.
This will change as contexts get longer and people start feeding large stacks of books and papers into their prompts.
Just like googling, AIing is a skill. You have to know how to evaluate and judge AI responses. Even how to ask the right questions.
Especially asking the right questions is harder than people realize. You see this difference in human managers where some are able to get good results and others aren’t, even when given the same underlying team.
Prompt engineering might not be something people do for much longer:
These inproved models do some valuable things better & cheaper than the models, or ensembles of models, that generated their training data. So you could not "just ask" the upstream models. The benefits emerge from further bulk training on well-selected synthetic data from the upstream models.
Yes, it's counterintuitive! That's why it's worth paying attention to, & describing accurately, rather than remaining stuck repeating obsolete folk misunderstandings.
How much work is "well-curated" doing in that statement?
I find it (very) vaguely like how a person can improve at a sport or an instrument without an expert guiding them through every step up, just by drilling certain behaviors in an adequately-proper way. Training on synthetic data somehow seems to extract a similar iterative improvement in certain directions, without requiring any more natural data. It's somehow succeeding in using more compute to refine yet more value from the original non-synthetic-training-data's entropy.
And, counter to much intuition & forum folklore, it works for AI models, too – with analogous caveats.
But I'm not suggesting they'll advance much, in the near term, without any human-authored training data.
I'm just pointing out the cold hard fact that lots of recent breakthroughs came via training on synthetic data - text prompted by, generated by, & selected by other AI models.
That practice has now generated a bunch of notable wins in model capabilities – contra the upthread post's sweeping & confident wrongness alleging "Ai generated content is inherently a regression to the mean and harms both training and human utility".
But not experience it the way humans do.
We don’t experience a data series; we experience sensory input in a complicated, nuanced way, modified by prior experiences and emotions, etc. remember that qualia is subjective, with a biological underpinning.
How does the banana bread taste at the café around the corner? What's the vibe like there? Is it a good place for people-watching?
What's the typical processing time for a family reunion visa in Berlin? What are the odds your case worker will speak English? Do they still accept English-language documents or do they require a certified translation?
Is the Uzbek-Tajik border crossing still closed? Do foreigners need to go all the way to the northern crossing? Is the Pamir highway doable on a bicycle? How does bribery typically work there? Are people nice?
The world is so much more than the data you have about it.
But also: with regard to claims about what models "can't experience", such claims are pretty contingent on transient conditions, and expiring fast.
To your examples: despite their variety, most if not all could soon have useful answers answers collected by largely-automated processes.
People will comment publicly about the "vibe" & "people-watching" – or it'll be estimable from their shared photos. (Or even: personally-archived life-stream data.) People will describe the banana bread taste to each other, in ways that may also be shared with AI models.
Official info on policies, processing time, and staffing may already be public records with required availability; recent revisions & practical variances will often be a matter of public discussion.
To the extent all your examples are questions expressed in natural-language text, they will quite often be asked, and answered, in places where third parties – humans and AI models – can learn the answers.
Wearable devices, too, will keep shrinking the gap between things any human is able to see/hear (and maybe even feel/taste/smell) and that which will be logged digitally for wider consultation.
I used 'delving' in an HN comment more than a decade before LLMs became a thing!
The problem is that it lowers the effort required to produce SEO spam and to “publish” to nearly zero, which creates a perverse incentive to shit on the sidewalk.
The amount of AI created, blatantly false blog posts about drug interactions, for example. Not advertising, just banal filler to create site visits, with dangerously false information.
It’s not like shitting on the sidewalk was never a problem before, it’s just that shitting on the sidewalk as a service (SOTSAAS) maybe is something we should try to avoid.
That at least will add extra work to filter usable training data, and costs users minutes a day wading through the refuse.