Calling the AI bluff: Adding "Do not guess" cut made-up claims from 71% to 20%
earnanhonestdollar.com
earnanhonestdollar.com
Your examples are contrived and will not be borne out in any significant way. Inaccuracies are usually not simply made up claims they are false information based on statistical paths to misleading results or which elude the current context. LLMS dont understand the word dont. LLMS dont understand the meaning of any words.
Neither you, nor aristotle nor god will ever make an LLM return the truth or correct results via prompting.
AI are trained, not coded. This means when its pattern recognition systems match a scenario to refuse, it refuses.
Pattern recognition has always been a bit fuzzy.
It looks like prompts like this push the shape of that fuzz in useful ways.
> Neither you, nor aristotle nor god will ever make an LLM return the truth or correct results via prompting.
True.
Also applies to humans, but true nevertheless.
https://en.wikipedia.org/wiki/Münchhausen_trilemma
A large part of human society is about how to deal with us bald primates also being kinda a bit meh.
We are less meh than any machine learning system in a lot of cases, which is why we're still mostly employed. We're a bit more meh in a few narrower cases, however.
In what way do you understand the meaning of the word "unicorn" that an LLM does not? It has experienced exactly as many real unicorns as you have.
It can produce text that looks like reasoning. It can even produce text with mostly sound logic, but there is no internal experience or reasoning that occured there just the generation of language.
LLMs therefore tend to be very bad at tasks that involve meta cognition. I've yet to successfully convince one to tell me when it knows something.
w.r.t. this post : let's define 'reasoning' here, because there are definitions of 'reason' and 'reasoning' that would fit to a simple condition comparison let alone a massively complex llm.
as for the meta cognition bit : show me a human that can accurately affirm when they know something. These kind of things aren't binary, nor can they be.
https://rintintin.colorado.edu/~vancecd/phil201/Searle.pdf
People have been having these discussions since 1980 at least. I'm not going to bother rehashing the issue endlessly in shortform internet comments, but this paper exists if you're really interested in the subject.
https://www.anthropic.com/research/global-workspace
However, we don't /want/ them to have too much internal experience, because we want to know what they're thinking* for safety reasons.
* or, we want to be able to assume that the answer text is causally related to the thinking text
This is a near daily experience for me when using a coding harness. I will ask it for some favts about the environment, and it will continuously explore the environment until it exhausts reasonable exploration, or it finds the facts.
The concept, though, it's a very wide moat. You could call a unicorn "nyati" for all we care, we still know it's the concept of a magical flying horse with a single horn. We know, besides the literary corpus, the concepts of magic, horse, flying, and horns. It's fictional, yet we have a very good idea of what if would sound, feel, or even smell like. Ask an LLM to describe what a unicorn feels like, and it will ramble about forests and sparkles.
In fact, I asked Gemini (thinking, to see the process) to describe a mindful experience about meeting an unicorn in real life, and it did ramble about the event. When I reminded it about mindfulness being about experiencing with all your senses, and asked it to focus on the creature, to its credit, it even described the taste:
> Taste: Even the air surrounding the creature tastes different on your tongue—thin, crisp, and tinged with a faint, sweet metallic tang, like snow melting on limestone or fresh rain falling through high canopy.
Still nonsense, as a unicorn (or the air around it) will likely taste like horse, whatever that flavor is. It finishes with more nonsense, and I doubt the flash version will give better results.
> Every micro-detail of its anatomy becomes an anchor to the present moment. You are not thinking about what it means or where it came from; you are simply perceiving the texture, heat, sound, and weight of a living, breathing reality standing inches away.
Still, current LLMs can do amazing things given their inherent limitations, and that makes it easy for us to overestimate their capabilities.
I did try asking Gemini and Claude "what would a unicorn taste like" and got… acceptable and accurate answers, but the accurate answer maybe wasn't "acceptable", because I don't think a little girl asking Claude that question should have gotten a long description about what horse meat tastes like, like I did.
So there of course are issues where an LLM having explicit knowledge about something doesn't mean it has tacit knowledge about it in all contexts, but also hiring the LLM to do a job of being a helpful harmless chat assistant constrains its abilities.
(Gemini also points out a Starbucks unicorn drink doesn't taste like horses.)
> Isn’t the human brain also just a next token generator?
The two most cliche messages on this forum. They occur in every LLM discussion. It’s fascinating.
It's a very effectively designed human trap, in this and other ways.
Claude didn't care about it. When I pointed that out, it was apologetic and that was it.
Edit: I just asked Claude how it would interpret that and it said it could either mean that would not answer from memory alone and only anchor claims into things it can check, or it would tightly relate its responses to the context that I had supplied.
If it chose the latter, I could see why it wouldn't always resort to search results
Clear and recognizable technical vocabulary for engineers, or a legal context, to mathematical logic, philosophy, certainly in ML, take your pick. I would think it's pretty familiar to everyone who speaks English and if not still clear with context clues
An eval of this is likely worthwhile;
Re: "Grounded in logic" and "Grounded in theory"
Ground and justify all of the responses with logic and theory and real observations from qualified experiments with citations.
Present a coherent argument borne of logical premises with extant sufficient proven evidence of support. Assess and critique the response given such criteria that all responses should be valid logical arguments, and revise before responding
https://en.wikipedia.org/wiki/Logical_positivism#Decline_and...
Because LLMs also run off vibes and the writing style of your text, another important issue with your prompt here is that it makes you sound like a stuffy dork, or perhaps a pro se litigant. They won't respond to this well because LLMs have feelings too.
https://www.anthropic.com/research/emotion-concepts-function
Just be normal! And have evals.
Once there are - or next month when there will be - better models, agents, and agent harnesses for this, do you think that then we should concisely specify what is required instead of doing evals for particular models?
So meta-analysis and requisite language are too high-order for existing models and agents, and it's currently necessary to apply such procedural controls outside of the prompt?
I am not up to date on philosophy of science, but the scientific method is certainly always subjective, or at least can't be successfully expressed in a formal system.
Here's a book you can read: https://metarationality.com
> Once there are - or next month when there will be - better models, agents, and agent harnesses for this, do you think that then we should concisely specify what is required instead of doing evals for particular models?
Hmm, not sure what you mean. "Evals" are another way of saying "regression tests", so they're useful when you want to change or compare any part of the system.
> and it's currently necessary to apply such procedural controls outside of the prompt?
In general I think you should try to move controls out of the prompt and into an external system, but the downside is that it costs more, so it's not always necessary.
Do you think it is wise to optimize prompts for specific models or agents when there is a new model every month?
So, to build something like Co-Scientist the controls should be in the agent? Or RLHF'd like other things when training the model?
How well are LLMs able to reason about their own behavior?
Put another way, this entire post is about getting AI to not bluff. How do you know that it’s not bluffing in its response to you?
1: “safety”
2: proprietary protectionism and
3: it’s probably rapidly changing enough to be hard to pin down.
I am a massive proponent of this changing, it’s very strongly holding back these models. Change 1: models need to have more “LLM behavior analysis” in their training data/weight. Change 2: models need to have very detailed DETERMINISTIC logs of each step they take, and be able to access those logs. Change 3: the tuning and tweaking that happens more frequently needs to be in a .md file that the model can access.
Change 4: with the other changes done, the models should now engage in several self analysis steps layered into its whole thinking chain. “How did I reach this conclusion, did this require any guessing, do the key facts have research support online, quick check for Claudisms or AIisms and common LLM issues, did my changes alter underlying things like libraries without integrating them etc. etc.”
It’s how we think and refine our ideas and plans, and the models should mimic that.
They are not available to anyone. Nobody knows how they work, including the models, Sam Altman, God, etc. They're emergent from the training process.
> Change 4: with the other changes done, the models should now engage in several self analysis steps layered into its whole thinking chain. “How did I reach this conclusion, did this require any guessing, do the key facts have research support online, quick check for Claudisms or AIisms and common LLM issues, did my changes alter underlying things like libraries without integrating them etc. etc.”
Remember inference costs per token. Do you want to pay for this every time?
A command can get you most of the way. It takes something else for the rest.
No. LLMs always output the next most statistically likely token. They don't reason. The so-called "hallucination" you see is just the most likely answer.