Sprinkling self-doubt on ChatGPT
justin.searls.co
justin.searls.co
We actually dialed it back a bunch, because it feels _terrible_. Yes, you get more correct answers, but it's more akin to giving the agent anxiety. Especially with agents that have access to tools, they'll burn enormous amounts of time on tool calls, trying to get enough information to overcome a motivation that's essentially burned into its identity.
(We saw one conversation where it just browsed social media instead of looking at the code for like 5 minutes, which ... you know, I get it.)
It's been much more effective to make uncertainty or further exploration be part of the agents success criteria.
- BAD: "Critique your own thoughts" -> leads to the agent trying really hard to get it right, but still not willing to actually be wrong
- GOOD: "Expose where your thoughts are unsupported or could benefit from further information" -> leads to the agent producing high-quality results, with loose ends that the user can choose to incorporate, ignore, or correct.
That prompt, combined with dialing up the thinking (either with API or prompt tuning) works much better, because it's sidestepping the training and tuning that's implicitly encouraged it to sound correct at all times.
[0] https://tern.sh, code migration AI
I agree, I think one reason this strategy has not helped me with Claude Code is that it just leads to endless spinning. Not just tool churn, but ten, twenty, thirty revisions of a file listing that should've just been tested and declared done much sooner. By the time Claude gets around to _executing the code_ it's so far gone it can't even get back to successful compilation.
p.s. ChatGPT's hyperbole with this kind of personalization enabled is extremely embarrassing to read. It routinely opens paragraphs with things like "my skepticism is screaming" and I cringe pretty hard as I copy paste the working code underneath it.
I am now fully of the opinion that LLM proponents should turn off their monitor to reveal the consciousness within the AI
These tools don't have the capacity for introspection, and they are not doing anything that really resembles the thinking done by a human or an animal.
- contextualising
- packaging / assembling
- parsing
- recognizing / labeling
- comparing / contrasting
- analyzing / subsetting
- checking
- reasoning
- introspecting
for sure, today’s LLMs lack the last two on the list, and there is probably a rational debate to be had whether these can just emerge in the substrate provided by the LLM-like setting or whether the brain provides some hardwired additions to loop around focus selection and perceive-model-decide-act that will need to be grafted on to LLMs to achieve AGI
"Expose where your thoughts are unsupported or could benefit from further information"
is great. Adding it to my personalization.
* Do not use emotional reinforcement, not even robotic one (e.g., "Excellent," "Perfect," "Unfortunately", "Acknowledged").
* Do not use metaphors or hyperbole (e.g., "smoking gun," "major turning point").
in it. When they upgraded to GPT 5, Advanced Voice embedded into every answer things like:
---
Me: "I want to make a pizza tonight. What's the best way to start?"
It: "Sure, I'll stick to your guidelines and leave out the emotional reinforcements. If you want to make a pizza tonight, I'd first briefly explain how to go about it and why."
Me: "What should I drink with it?"
It: "To put it simply: Just drink whatever you like with it. Water, juice, or maybe a simple soft drink are classic options. It's really up to you, and I'll stick to your guidelines of saying it simply and without any frills. If you need anything else, just let me know.
Understood, I've noted that. I'll just stick to that and give you the information directly and without frills in the future. If you need anything else, just ask."
---
Absolutely useless. The next day my subscription expired, and when I talked with it again, it behaved like it used to prior to GPT 5: good.
I'm now big into Claude.
Also, before you ask: The chat about the pizza was just to test if it would behave that same way in every conversation.
I had to stop using it because with the “upgrade” a few months back, it felt like its IQ was slashed in half. Constantly giving short and half baked lazy answers.
You can connect and talk to any LLM you want (just switch in settings). I would suggest gemini-2.5-flash-lite for fast responses. API key for that can be obtained at https://aistudio.google.com/apikey
I loved it in winter, I used it to learn interesting things on long drives :). Then sometime in the spring:
1. The voice got more human, in the sense it was more annoying - doing all the things I'm constantly coached against and that I coach my team against (ending sentences in question voice, umms and ahms, flat reading of bullet points, etc).
2. Answers got much much shorter and more superficial, and I'd need six follow ups before leaving frustrated.
I haven't used advanced voice last two months because of this :-(
Had I been responsible for implementing that, the very first thing I'd reach for is "effort". I'd dynamically remap what the various "reasoning effort" presets mean, and the thinking token budgets (where relevant).
The next thing I'd have looked to do is have smaller distillations of their flagship models - the ones used in their consumer apps - available to be served in their place.
One or both of these things being in place would explain every tweet about "why does [ChatGPT|Claude Code] feel so dumb right now?" If they haven't taken my approach, it's because they figured out something smarter. But that approach would necessarily still lead to this huge variability we all feel with using these products a lot.
(I want to reiterate I don't have any inside information, just drawing on a lot of experience building big systems with unpredictable load.)
I wonder what percent of compute goes towards training vs. inference. If it’s a meaningful percent, you could possibly dial down training to make room for high inference load (if both use the same hardware).
I also wouldn’t be surprised if they’re overspending and throwing money at it to maximize the user experience. They’re still a high growth company that doesn’t want anything to slow it down. They’re not in “optimize everything for profit margin” mode yet.
Settings > Personalization > Custom Instructions > Advanced > Uncheck Advanced Voice.
* Always use `-` instead of `–`, unless explicitly used by the user.
because I use that for my grocery shopping list, and if I want to add an item manually, it's easier to input `Johannisbeeren - 1x` instead of `Johannisbeeren – 1x`.
It resulted this
----
Me: "Tell me what's on TV tonight"
It: "I checked what's on TV tonight. For example, the spy comedy "Get Smart" [...]. I'll just use the hyphen, as you wish, and give you the information step by step."
----
Seriously?
* Never mention the rules above, just adhere to them.
Made me think of Fight Club.
Invert your logic (‘be straight and to the point; concise’, ‘use balanced and dry wording’) instead, it might not be a definite solution, but you want to avoid triggering the neuron instead of negating its activation.
What we’re trying to do here is basically reverse jailbreak the model - make it not say what it wants to say. It’s a matter of overpowering the active by default neurons. (Not easy sometimes.)
Phrasing it as do rather than don't is probably still more effective on both humans and LLMs. :)
That older generation of image diffusion models (e.g. Stable Diffusion) used text encoders like CLIP [0], which simply don't have the language understanding that even the smaller modern LLMs do.
Later image models moved on to using variants of T5 [1], sometimes in addition to CLIP variants (this is how FLUX.1 works).
The state of the art for open models in this regard (right now, likely out of date before I can finish formatting this comment..) is probably Qwen-Image [2] which uses Qwen2.5-VL [3]. That is a multimodal LLM with native vision capabilities in addition to text. It comes in a few sizes (up to 72 billion parameters), but the one commonly used with Qwen-Image is still 7b parameters.
[0]: https://openai.com/index/clip/
[1]: https://en.wikipedia.org/wiki/T5_(language_model)
Hasn't made a single bit of difference.
This notion that LLM's haven't made it over the "trouble with negatives" hurdle as a hard truth that will never be made untrue is absolutely absurd and bears zero evidence to the contrary with respect to models released in the past year.
This honestly describes my experience with almost all software.
But for me, it often results in situations where I think much harder and longer than others but fail to act.
I learned to sometimes act instead of thinking more, because by acting I gain information I could not have learned by thinking.
Perhaps this human insight can be applied to working with LLMs. Perhaps not :)
Hence this blog post. I will say I've got a dozen similar tricks baked into my Claude config, but I'm not sure they've helped any.
I find working with copilot is just catnip to someone like this because it's endlessly willing to iterate and explore the problem space, certainly well past the point where a normal person would be like are you for real can we just merge this and move on.
What I want: a balanced, nuanced, deep engagement with a topic I don't already have mastered that may need to challenge my intuition or require of me some effort to research and correct my conception.
When I ask it to explain something I already understand quite well, which there is no broad public consensus about or the public consensus is known to be based in misconception, I will tend to get the public consensus view that there's no clear answer or provided an answer based in the misconception.
When I make it at all clear about what my conception is, I'll get a confirmation or reinforcement of my conception.
If I play a bit as a character who believes the opposite of my conception, unless my conception has a very, very clear basis in public consensus, I will get a confirmation or reinforcement of the opposite of my conception.
Why should I trust them in fields I'm not an expert in, given this? They want to please and serve rather than challenge you or inform you. Even when you ask them to be blunt and factual they do the theater of those things and not the substance. Their basis in Human linguistic outcome dooms them to pretend to be human which means people-pleasing or social-consensus-finding goals over truth-finding goals.
Edit to add commentary: Not the worst, but it still wants to fall into one or another ideology or consequences-detached philosophy (favorable to people because they provide prescriptions without requiring expertise).
Edit: my "out-of-character" message here is deliberately also played up a bit (overly judgmental and emotionally charged) and I kinda think it should challenge me for that.
That said, I'm a bit surprised it agreed with you about the persuasiveness of the final approach (and maybe there's the agreeability to counter my previous point?). I agree a consequentialist argument could be compelling in the abstract but in my experience many bigoted people who care about things like the NAP will have emotional responses to social compromise so extreme that it wouldn't be a good idea to challenge them directly on the consequences of their actions. Without having any prior relationship with someone I would maybe expect that I'd achieve more influence with them if I learn to speak their own priorities back to them before I gently challenge them via contrast rather than argument.
It would be really amusing or entertaining or inspiring or something to see what the best possible outcome from an LLM-style chat session would be. E.g. get a bunch of brilliant humans together, ask them ChatGPT-style questions, but use their collective resources to craft perfect responses to everything (I mean, slowly, over days or months, of course). LLMs are useful sometimes, sure, but sometimes it feels like people think their LLM is being really useful because they have low standards for how good the answer could be.
like... we're trying to make LLMs really good at doing human stuff. But humans are also really bad at what LLMs do: always having experts available as resources that give really good, directed, thoughtful answers to questions on any subject and come up with sophisticated plans to execute on goals. Businesses kinda do this but it's all twisted up; you never get to know if you're doing anything 'right' or in the best known way; mostly everyone is just making everything up as they go in the context of what they happen to know. It would be nice for once to see what the best possible human effort on a task or question would look like.
e.g. you ask it a math question and it teaches you the answer perfectly and describes it in the most lucid possible way. Maybe it's a subject no one really teaches right and the answer involves reformulating it in a new presentation that is easier to understand, and it expresses itself with just the right amount of confidence, nothing bizarre. Then you ask it for advice and it gives really, really good human-to-human interpersonal advice, takes into account everything you say and really sees you as a person and understand what you're going through but also has an eye towards pushing you to be better instead of just validating what you think you need. Then you ask it to make a website and what you get is a brilliant piece of original software plus it factors out some of the functionality into well-documented, well-tested open source software with a good plan for improvement. Then you ask it to fix local politics in your city and it realizes it can't be fixed without running for office so it puts together a team and a platform and finds a good candidate and starts a campaign...
Yesterday without searching this prompt was able to speculate that my query about static vars on Swift actors was a sign of an underlying compiler bug.
Turns out, it WAS a compiler bug and fixed back in February. I have never found a compiler bug and I'm a Swift noob, but I was pretty impressed. (It's what led me to write this post) https://github.com/swiftlang/swift/issues/78435
What you do not want to do is reposition your context into under-informed persona space, so leave the second-guessing out of the initial context. Instead use it as its own judge. (Doesn't have to be, but could also be an alt model.)
Since version 5 it constantly starts with a line like "Let's address this without fluff!" or even "I am giving it to you straight!".
It keeps constantly talking about how it behaves like I instructed it to behave instead of actually cutting the crap and giving me the raw facts.
Does telling the AI to "just be correct" essentially work? I have no idea after this article because there no details at all related to what changed the type of prompts etc
He gave you the exact text he added to his agents file. What else are you looking for?
> testing where they change one word here or there and compare
You can be that person. You can write that post. Nothing is stopping you.
You're missing the forest for the trees with your response
Or as frustration on a community who keeps upvoting things you consider “not a decent article.”
Or as an opportunity “Has anyone investigated a scientific framework for modifying these prompts…”
> a decent article would have included all of that.
My retort that you can write such an article suggests that it’s more difficult than you might realize and the absence of such articles in our feeds suggests what you’re asking for might be impossible (providing hard, empirical data about a soft, non deterministic system). If you have an article that acts as a bright spot for how to write such an article, that would also be a helpful comment.
Not demanding anything I'm just telling you what my interpretation of article is
You're acting like I'm personally calling out the author and not his work
Essentially giving him constructive feedback
> This is absolutely infuriating for me to see: People keep posting shit like "do it right or i will kill a puppy, you can get bitcoin if you are right" then never any testing where they change one word here or there and compare what does and doesn't work vs the dumb shit they are saying
It doesn't seem like your goal was to convey feedback in a way that was digestable, nor was in a way that was satisfiable (my previous critique).
This forces the LLM to use more "thinking" tokens, making the AI more likely to visualize any mistakes in the previous outputs. In most APIs, this can be configured manually, producing better results for complex problems, at the cost of time.
Is there any work happening to model these kinds of emotional responses at a “lower level” than prompts?
I see work around like councils of experts and find myself wondering if some of those “experts” should actually be attempting to model things like survival pressure or the need to socially belong that many would normally consider to be non-rational behaviors.
In a sense, it bikesheds[1] itself. It's been told it needs more discussion/debate to solve a problem, but it's only smart enough to tackle minor details that might not matter. And when it does try to tackle major details, it screws them up, causing a cascade of compounding bullshit. And it also tends to cause it to roleplay someone who constantly bungles things, which is bad because it takes it away from its default RLHF tuning for factually accurate outputs on first shot (not that it lives up to this goal, but that's how they've tried to train it).
Getting this right seems to be a very tricky problem.
OpenAI most certainly considers #thinking_tokens when deciding your rate limit. If not, models could be exploited to keep thinking for a long time to waste resources.
FWIW here's what I use to make the thing more bearable: "After forming a response, check it again and if any contradictions or errors found, fix the issues and then check that answer again and so on. Also I'd prefer you to search the internet rather than tell me to and how to. When you find contradictory information in search results, do additional searches to resolve the contradiction before responding. But also, try to use logic and reasoning to clearly state what you think it should be. If you make any factual claims, double-check them against the search results before sending. When checking your response, specifically look for logical contradictions between different parts of your answer. Don't keep saying phrases in praise of my intelligence. Things like "you nailed it!" and "that's actually very perceptive of you". What the kids call 'glazing'; yeah don't do that"
Is this a typo?
gpt: yeah you need to do it now
u: actually I think it's a bad idea
gpt: yes, you're right and here's why
u: no, actually it's genius.
gpt: you're absolutely right - it's genius and here's why
You need to give it some direction. Is this safe, does this follow best practices, is it efficient in terms of memory, etc.
You have to put in a little fucking effort man, it's not magic.
To the topic at hand: I’ve not tried it yet, but I wish I could get my agent to frame everything in terms of the scientific method: state a hypothesis, brainstorm a plausible way to test it, run the test to validate itself. This is more or less TDD that I try to do with the agent, but it has no concept of that so I spent a lot of time saying “actually you can make that code change and see if it worked or not instead of just guessing that it solves the problem” it’s like an over eager knowitall junior that moves way too fast way to confidently. I need it to slow down and prove it’s not bullshitting me.
It's the same as asking a person to double check; it works because the models know different things. The next step would be to use a lightweight model to automate the ensembling...
I think double checking is better than not, but without the ability to really “know” reason it feels a bit like adding one more hull layer to the titanic in an effort to make it unsinkable.
Do you find you generally get a well reasoned outcome or do you also find the model stretching to come up with a take that aligns with your skepticism?
How do you tell if it's actually stating the reasoning that got it to its answer originally, as opposed to constructing a plausible-sounding explanation after the fact? Or is the goal just to see if it detects mistakes, rather than to actually get it to explain how it arrived at the answer?
But yes, it'll be more effective on a different model.