Has the hallucination problem in AI been solved?
My understanding that all AI can, and will hallucinate. I get downvoted for saying this, but no one ever says I'm wrong or cites any source. Perhaps it's more nuanced than that? Please, enlighten me.
Part of the difficulty--not in solving, but in discussing--is in defining what a hallucination is. On the face of it, it seems straightforward: an obviously counterfactual claim or manifest error of reasoning. However, it's not always that simple. A lot of what people consider to be hallucinations are misattributions, specious diagnoses, strangely lopsided preoccupations, eccentric design choices, needlessly verbose or circuitous output or explanations, a kind of metaphysical conflation of the trivial with the significant, etc.
_______
> In fact [The Bursar] was incurably insane and hallucinated more or less continuously, but by a remarkable stroke of lateral thinking his fellow wizards had reasoned that, in that case, the whole business could be sorted out if only they could find a formula that caused him to hallucinate that he was completely sane.
> This had worked well. There had been a few false starts. For several hours, at one point, he had hallucinated that he was a bookcase. But now he was permanently hallucinating that he was a bursar, and that almost made up for the small side-effect that also led him to hallucinate that he could fly.
—The Truth (2000) by Terry Pratchett
It's like a science fiction writer or an improv actor doing technobabble. If by chance it knows the actual answer it might use it, but even if not it still has to say _something_ that sounds plausible to a layperson. It'll never say it doesn't actually know, because their character it's acting as _should_ know.
From my experience it has been largely solved for one significant use case which is chatting to frontier models about the reality as described by public knowledge. 2 years ago models would rely on their training data, today they go out of their way trying to look it up on the Internet and verify thoroughly. I have not had a problem for a very long time.
When working off of limited, private/unverifiable context, LLMs still hallucinate, but again, much less than 2 years ago, and more within a "getting confused where a human would easily get confused" range, rather than "outrageously making things up" range.
You can see proof of this if you ask it obscure enough questions. That doesn't mean obscure scientific questions, I asked it questions regarding ship fits and modules in EVE Online, which is an extremely well-documented videogame. There are hundreds of online tools to help you for different things, mining yield calculators and more.
Well, ChatGPT just made up almost everything. Very confidently. It couldn't even get the damage types right, I was quite shocked.
Hallucination has only be solved for extremely narrow sets of problems, and only partially. Coding is one of those problems.
For the first, we're mostly past the point of just "testing". So I don't see that too much anymore. Mostly I still see that though in bug reports generated by AI by someone else. There is usually some underlying bug being reported, but the AI explanation and "helpful suggestion" is typically inaccurate. Generally, suggested fixes are terrible. (They likely work, but fix a symptom not the cause.)
The second still happens, but with much less regularity for me though. It does make mistakes though.
In areas where I'm not as skilled it's very hard to spot errors. When researching general information I'm mostly accepting it on face value.
I find the bug-report thing really interesting. For lots of simple bugs it's great. For more complex things it seems to be very superficial- if a 0 causes an issue here, add a simple guard for 0. There's no depth of understanding why the value is 0 in the first place, when it should be set. If it can (incorrectly) be 0 here, where else might 0 be impacting the code?
This informs my opinion of vibe coded stuff - where there is no skilled human inspection. I expect that code to be of a poor underlying quality. Especially if it's AI changes to an existing human-coded app.
And that AIs do aim at intended targets with intent to kill while simultaneously being susceptible to hallucinations is exactly what worries me.
It is conceivable that agentic edge LLMs are used for the logic of killer drones but I don't think we're there yet.
Machine learning in general does have classification errors so it may mistakenly see an enemy where there is a pile of rocks(right next to the children's hospital).
To the user of the drone, this represents an error rate and wasted munitions(unless they were planning on bombing the hospital later). So it is something they want to minimize.
Again, I'd be more worried about becoming a "legitimate" target to someone's killer drone than an accidental one.
I see it as a small deal - it reminds us to check the responses against references. LLMs are a statistical construction, and everyone accepts without complaint that statistical models have a predicted false-positive and false-negative rate. I think of "hallucination" as a false positive; false negative is no answer when the model could have made a useful response. I think there's some relation between our creativity and the LLM capability we dismissively call hallucination.
At least that is my understanding of one of the ways that Ukraine is asymmetrically winning against a foe in Russia with more resources, at least at the start.
It's how we are using it knowing this limitation it has and the decisions being made about when and where it's being used in this capacity are mostly unknown.
IMHO This feels like a troll, OP have several comments about evil Ukrainian AI drones. I feel it is hard to have a naunced discussion if you only diss one part.
Your question is asking for easy technical opinions about a subject where you seem to have no technical knowledge.
I don't beleive I characterized what I'm asking in the way you are putting it.
I never said that Ukraine is evil. I competely support and stand with the Ukranian Government and it's people, and I think we, as a country, should be supporting them and their fight to not be conqoured.
Clearly, if I post on a technical forum and am asking a question it is because I don't know the answer. I know how I see it. What I don't know is how everyone else sees it, and that's my motivation. I'm wondering if I'm right or I'm wrong, and have faith that the denizens of the internet will let me know one way or another.
I am asking a legit question and not trying to troll. I am, admittedly, agaist the use of AI on the battlefield, but as it pertains to Ukraine they are defending their existence.
I honestly think it will be impossible to block the use of AI in weapons because you can not know where to draw the line.
We want AI to make predictions under uncertainty that could be wrong. What we don't need is getting know facts wrong.
One of the best responses I got from ChatGPT was when it said "I don't know", and on questioning it - it responded that it was an unsolved problem and it couldn't objectively take a side iun the argument.
Claude will make mistakes but they’re largely “reasonable”. In some ways that’s worse as they’re more believable.
GPT I can’t comment on too much, except to say we run a chatbot using it at my work and when extracting data from the conversation (contact info and such) it will occasionally make up an email address that wasn’t entered into the chat. We have guardrails around it, so it’s not a big deal, but it does happen.
Most, like you, have nothing useful to contribute.
Fact is, those with the foresight to see how this can go badly are also seemingly the types of people who don't end up in a position to prevent harms by it. Furthermore, it seems inevitable in a sense, because greed for power seems to necessitate development of automated weapons in ASAP in spite of the risks, and the fact it basically renders traditional warfare pointless.
It appears we'll have to learn the hard lessons, same as our forebearers with. I just hope we can avoid having to regress back to sticks and stones on account of fucking ourselves by overdoing our capability to destroy on account of not being willing0able to peacefully coexist.
In other news, New Orleans 911 is using AI to triage localized incident duplication calls from unique emergencies. I'm not saying that it's good or their only practical choice, but it's happening.
The biggest dangers I see are the outsourcing of supervisory control, appeal to authority (when used to summarize content or answer a question), and hallucinated mistakes.
0. (PDF) https://docs-library.unoda.org/General_Assembly_First_Commit...
1. PDRMUAIA https://www.state.gov/bureau-of-arms-control-deterrence-and-...
2. REAIM 2023 Call to Action https://www.government.nl/documents/2023/02/16/reaim-2023-ca...
3. REAIM 2023 Endorsing Countries and Territories https://www.government.nl/documents/2023/02/16/reaim-2023-en...
> when I can have it do something like write entire working kernel module fixes for old MacBooks on a whim
They’re not saying it can’t do that, and that’s not proof it doesn’t hallucinate. In fact, having used 6-8 agents at a time for a year plus while writing AI tooling for an AI startup, I can definitely surely tell you that they’re almost inversely correlated as in models that hallucinate a lot sometimes also put out the best most impressive solutions.
I’m definitely not anti AI and I definitely have found a way to make it work very well and I’m content with the work I get out of it (again maxing out several max 20x subs), but I have had sol definitely hallucinate this week and I’m a bit shocked you’re trying to say otherwise.
Listen I know it’s going to be I’m holding it wrong too, but I’ve been reading white papers and research on LLMs for a long time and was definitely at the cutting edge of context engineering, implementing features in our tooling harness a year before they were in codex or Claude.
maybe I am holding it wrong still but but like at some point if I’m holding it wrong who else will be holding it right? Dozens of people? At some point, the technology has to be approachable enough for everyone to have your point of view automatically.
I'm no expert on the inner workings/harnesses/etc beyond a basic understanding of the architecture. Maybe I've just developed a good sense for effective prompts? I could share some recent sessions.
Worth noting that both cases indirectly involve the humans that designed devices and the humans that made the placement and trigger condition decisions.
Further:
> AI drones are being used to autonomously target and kill targets by the Ukraine using technology they have been given.
Ukrainian Combat Robot Holds Frontline Position for Six Weeks in Sign of Growing UGV Maturity - https://defenceleaders.com/news/ukrainian-combat-robot-holds...
are remote operated, they allow defenders cover while themselves being out and exposed.
However were they altered to autonomously fire, that would be on the basis of pattern matching in the visible and infra red spectrum - shoot at all hot blobs.
That's more of a trigger threshold setting issue than an LLM hallucination issue, and the danger is on par with any weapon system on auto fire, you really shouldn't approach such things until they are put in a safe off state or have exhausted ammunition.
edit for clarity
First point, vision systems have been used in industry to look for misaligned labels, incorrectly folded papers (in high speed paper presses), wrong items on high speed conveyor belts etc. for thirty odd years now - they have issues that a very distinct from LLM 'AI' hallucinations.
That's nomenclature out the way.
Landmines are indiscriminate, they trigger on any weight or pressure over a threshold.
A vision based Felixer, by contrast, only triggers on cats (well, almost always only) and leaves bilbies and bettongs to walk on by.
That's an improvement over landmines.
The crux of your issue here might be the morality and ethics of establishing human exclusion zones within which all humans are highly likely to die.
These historically are created with rapid patterned artillery fire, butterfly mines, Napalm, indiscriminate criss crossing machine gun fire, etc.
Now there exists an option to use drones to kill all humans and leave the horses and cows alive.
Is it your concern that a bad vision threshold might kill a horse rather than a person? (Likely not)
Would you prefer an area to be napalm'd and agent orange'd back to dust?
War is hell.
* https://www.abc.net.au/news/2020-05-29/feral-cat-management-...
I, a person, singular, mention landmines as they are a hardware device that are designed to trigger action on a threshold.
Felixers are computer vision based devices that are designed to trigger action on a threshold.
These are actual real world objects that Bishop Berkeley can kick, not vague philosophical questions but actual engineered devices that kill and are in use today.