Hacker News Reader

Why would auto mode turn off thinking?

The "auto" mode is (AFAICT) a per-conversation-turn router. (Presumably via a preliminary pass through a very fast tiny model that spits out an number for how challenging it thinks the next response might be to compute.)

On high-challenge turns, the auto mode routes to the "thinking" model. But on low-challenge turns, it routes to the "instant" model.

And the "instant" model, by design, has no capacity for deliberation. (If it did, it couldn't guarantee that its responses would begin streaming "instantly.")

I don't expect that would be the case. This is what's called BLUF or Bottom Line Up Front: https://en.wikipedia.org/wiki/BLUF_(communication)

The model will still have read the entirety of the document before composing its response. And I believe that even in auto mode, there are thinking tokens behind the scenes.

Oh the number of time LLM will, for example, be giving me the list of bugs it found in code, when I ask it for a review, just to decide there’s no big half way through explaining it.

I wonder if it would help to ask it to write a rough draft and then reorder it?

It would (and does), yes; but this takes a lot more output tokens than asking for a summary would. The summary approach is only helpful insofar as it can be cheaper than using the thinking model. (You're basically tricking the instant model into thinking, which it can do, after a fashion.)

But, unless your desired output is literally a document for others to read, at the point where you're having a model generate a full, lengthy output multiple times over with revisions, you may as well just turn off auto mode and have it always deliberate (i.e. choose the thinking model explicitly from the model selector.) Then it'll be as messy as it needs to be while deliberating, but give you exactly what you want as output.

(And if your desired output is literally a document for others to read, that you want to interactively draft and polish, then (in the case of ChatGPT specifically) you should not only be explicitly forcing the "thinking" model, but also should be asking it to activate the "canvas" feature from the start. My understanding is that revising a canvas document involves the model emitting something like editing gestures, rather than simply re-streaming the updated chunks of text. This saves a lot of output tokens on large documents.)

Yes this is an extremely well known result for exactly the reason you guessed. It's not just abcktracking, asking an LLM to present a conclusion and then justify is also an excellent way to provoke hallucination as the model con concts "any justification that plausibly justifies the words it's already said".

This is the actual reason why openai _invented_ reasoning models, to give them time/space to work out a solution, rather than having to magic a correct solution out of thin air from token 1.

It's less important now that all models do reasoning, but it's still almost always better to make the output come out last rather than first.

This was a big concern for earlier models, but with modern CoT trained models they should be able to come to the conclusion entirely in the thinking trace.

You are absolutely correct. The second suboptimal part of the prompt is this:

> Trim introductions, repetition, generic reassurance, and optional background first.

It's not possible for the model to "trim" those before they've been output, so this is akin to telling it "not think of an elephant or even take the existence of elephants into consideration while solving this problem".

You may be discounting the tokens generated in the thinking trace but not included in the output to users.

That would be true of non-iterative models that just emit an output from beginning to end.

No reasonable model has worked that way for years.

I would be very interested in hearing about those "iterative models" you seem so convinced have existed "for years" (so, at least since 2024 / GPT-4o). Do you have any sources?

Extraordinary claims require extraordinary proof. So far the only other people I've seen teach this prompting style or talk about models "correcting their own output" were getting their information from AI-generated, hallucinated LinkedIn and TikTok posts.

If this thing exists - which is not just a LLM outputting content serially, placed inside a harness where itself (or another llm) is prompted to review and also output revisions serially - and if a single model can be prompted to output content and "iterate" or rewind it, and it's been widespread amongst "all reasonable models", surely there will be a flurry of sources you can point me to so I can learn.

Well... since o1 (Sept 2024), most models generate a "hidden" thinking trace before the visible answer... surely you have seen this by now: "Wait, that's wrong...", "Ah, now I have the full picture...". The model prunes dead ends when it composes the final answer.

Demanding sources for this is odd. It's literally been a headline feature of every frontier model for two years.

I guess you are "technically correct" that no model can "un-emit" tokens... but I don't think that is what anyone was saying or an interesting point to make.

Edit: see also this recent post, which details another place where revisions can happen, upstream of the reasoning token emission:

https://www.anthropic.com/research/global-workspace

You're talking about "thinking" models which are mostly regular LLMs trained to output a "<thinking></thinking>" token delimiter before the answer that will be shown to the user, and that's a clever use of the Chain of Thought idea. All of which I'm completely familiar with.

GP was saying

> That would be true of non-iterative models that just emit an output from beginning to end.

Which suggests that there are

- "iterative" models

which

- do not output "from beginning to end"

which AFAIK is science-fiction.

don't think it's that deep mate. just sloppy wording on GP's part.

The disagreement here seems to be a confusion between the pure model behavior and the behavior that users experience, which is the model wrapped in a harness. It’s normal and natural now to refer to the whole system as ‘the model’ informally, even though that is technically not accurate. That ship sailed a while ago.

Since the LLM is now designed to run in a harness, it’s really not even wrong any more.

Here’s some of those sources that were apparently hard to find:

https://openai.com/index/openai-o1-system-card/

https://arxiv.org/html/2410.10630v1

https://magazine.sebastianraschka.com/p/understanding-reason...

Reasoning models are mostly regular LLMs trained to output a "<thinking></thinking>" token delimiter before the answer that will be shown to the user, and that's a clever use of the Chain of Thought idea. All of which I'm completely familiar with.

GP was saying

> That would be true of non-iterative models that just emit an output from beginning to end.

Which suggests that there are

- "iterative" models

which

- do not output "from beginning to end"

which AFAIK is science-fiction.

Diffusion language models aren't GPT-5.6 level, but they're not science fiction either.

Thank you ; I had completely forgotten about those. I stand corrected.

Over hours of experimentation with various LLMs, I've found virtually any system prompt can cause unintended skewing of the model's output. Even just 5 to 8 short, direct words about length, tone or formatting can cause subtle yet significant changes in model output.

Longer, more detailed or conditional prompts always introduce an additional cognitive load as it checks every token it generates against the conditions. Making instructions more absolute (like: "Never do...") can increase the duration of compliance but at the cost of creating a significant center of attentional gravity. This can cause far more output distortion as the model devotes increasing portions of its attention budget to ensure compliance with a heavyweight requirement or prohibition. Every word in a global prompt is a trade-off between attention, compliance, drift, etc.

As someone used to thinking of computers as natural deterministic rule-followers, it's weird having to carefully wordsmith and A/B test even the simplest global prompts. It feels like coaxing a hyper-literal, emotionally sensitive, spectrum-ish toddler to comply but without being so strict it gets 'upset' or spirals into hyper-focusing.

True. The real trick, if you have a client-side agent framework to hand, is to prompt it once as "gently" as possible to "just solve the problem"; and then, after its response to that, automatically prompt it again, with a separate prompt, to summarize that response a certain way. That way, the second prompt isn't "in mind" during generation of the first prompt. (And ideally, you don't even present the intermediate result to the user.)

Sadly, you can't do things like this directly using ChatGPT's own "GPTs" abstraction. (For that feature to be useful, they really need some concept of server-side agents as stateful resident IO-stream-reducer actors.)

Models can be so sensitive that even prompting "Number section headings" would cause it to stop using its normal bullet point formatting anywhere. But then adding some variant of "...but don't stop using bullets as you normally do when they are needed" would make it start using bullets all the time.

Trying to craft a workable prompt got so frustrating I eventually just tried a prompt of "Don't change anything about your normal text formatting, it's perfect as is" and even that skewed the output vs no prompt. For browser chat I finally just wrote a client-side CSS UserStyle that does the formatting. Now I even have sequentially numbered sections with indented alphabetic bullets! Zero cognitive load or attentional skew and it never drifts off the formatting in long sessions.