They have been trained to statistically produce plausible language --- without any real comprehension or understanding or concern for what the words convey.
Bottom line --- AI is still dumb as a hammer. There is no guarantee of anything with the results. Maybe it offers some factual info --- or maybe not. Any result still needs to be verified.
We are all subject to flights of fantasy but they are supposed to be smarter than this. Instead, they are proving themselves to be equally untrustworthy.
Has anyone done a good deep dive and explained whether this is something likely overcome relatively soon or is it more problematic as part of the design/architecture ?
The problems of “hallucinations” are already being solved with a good number of analytic augmentation methods like pre-computing document embedding and augmenting the prompt with text from related documents, making the task more of a translation than a synthesis.
More details: https://www.williamcotton.com/articles/chatgpt-and-the-analy...
Try asking chatGPT to summarize an article about a specific instance of violence like police use of force which resulted in a death and watch it confabulate and dodge regardless of how much source material you give it to work with, simply because it is afraid of touching "unsafe" topics like murder.
We only have biased observers to monitor and tweak the system. Attempts to counter the bias that already exists in information will only result in another layer of different bias.
Ironically, we have two major issues with AI. One in which it doesn't function as intended, which leads to the problems you are describing. However, even in the conditions of working as desired, it still leads to a lot of perplexing philosophical issues for society.
I've spent a lot of time trying to work through all of this and putting some thoughts into writing. So far, it has been a lot to contemplate. https://dakara.substack.com/p/ai-and-the-end-to-all-things
A computer as we know it is a binary logic playback device. This is an immensely useful tool but expecting real, original "intelligence" from a box of silicon switches is the modern equivalent of alchemy.
Also, while almost anything can be viewed reductively, doing so is not necessarily a practical way to operate in / understand this world. Brains do a lot more than you mention.
It's easier to define what intelligence is not.
Statistically generated language that while grammatically correct still lacks comprehension or veracity is not it.
I agree; this is a pretty good baseline for written statements.
However, I was asking the parent-parent commenter (^^) who wrote:
> thfuran: A brain is just a sack of meat stochastically nudged towards only dying after more reproduction. Expecting real, original intelligence from something like that is the modern equivalent of alchemy.
Brains are quite good at satisfying jqpabc123's definition. I'm interested in thfuran's definition.
Without ever having seen it done?
One couldn't ask for a more clear expression of religious zeal.
And yet at the same time, you have "faith" that what your brain does can be duplicated by a box of sand. And "faith" is the proper description since just like an alchemist, you have no evidence or functioning example to justify your belief.
It's not faith. Over the past 25 years and especially the last 5 years boxes of sand have again and again shown themselves capable of tasks that previously were thought to only be tractable with true intelligence. Problems answerable only in the sole domain of humans, or would at least remain so for long into the distant future. Chess playing is an early example from the 80s and 90s. In 2010 we thought Go would take at least a decade to automate, but by 2015 the best human players were rendered inferior. Art was long thought to be impossible for a machine, today I can create a custom Mona Lisa image in seconds. Organizing an archive of photos based on subject matter or the persons present in them used to take many human hours of labor, now Google Photos does it automatically for free. There are comments here on HN from 3-4 years ago when transformer LLMs were first materializing saying they're cool but unlikely they'll ever be able to synthesize anything useful or which remains congruent for several sentences much less pages. Today I save hours of work each month summarizing meeting discussions, generating complex standard expressions, and breaking through creative blocks.
Yet there is still dismissive handwaving, declaring these technologies mindless piles of silicon completely incapable of intelligence. If that's really true then clearly we have no idea what intelligence actually is. If you showed ChatGPT and DALL-E and the other software of today to people in 1975 they would say with near-unanimity that we invented intelligent machines.
However, just as the argument that AI is not useful because it is not true intelligence may not be valid, then so must be that many of the concerns of a true intelligence may also arise without ever achieving that goal.
Really? How would a classical computer create an emergent property like entanglement?
What if your brain function depends on such emergent properties?
There are too many unknowns associated with "intelligence" to reasonably declare that it can be created from a box of sand --- aka, a "computer" as we know it.
I'm not sure what you mean. A simulation produces the emergent properties of the simulated system by simulating the properties from which they emerge. I'd say that's more or less the point of simulation.
Circular logic.
They are called 'emergent' properties because noone knows or understands how/why they emerge.
You can't use binary logic to simulate what you don't understand. Expecting properties to emerge from an imperfect simulation is wishful thinking --- aka "faith".
That is not true.
For one, I'd like to see us two learn from each other (demonstrating some awareness of ambiguity and theory of mind) rather than proceed on the current course, zinging each other. (Note: I'm not saying I'm any better, in general. It is mostly because I'm a third-party here that I see it.)
In particular, my understanding of emergence can be stated as follows. While emergent behavior can theoretically be predicted from the underlying system, it is practically infeasible to measure precisely enough to do so. Therefore, for practical purposes, there is an "predictive gap" between the lower-level system laws and the higher-level behavior.
My fear is that trust will go down if problems aren't dealt with - specifically hallucinations. At the current stage, it's super useful for creative use cases than factual.