Why OpenAI is Hiding its Reasoning Model’s ‘Thoughts’ – The Information
I’m Nikunj, PM for the OpenAI API. I’m pleased to share with you our new series of models, OpenAI o1. We’ve developed these models to spend more time thinking before they respond. They can reason through complex tasks and solve harder problems than previous models in science, coding, and math.
There are two models:
Developers on usage tier 5: you can get started with the beta now. Try both models! You may find one better than the other for your specific use case. Both currently have a rate limit of 20 RPM during the beta. But keep in mind o1-mini is faster, cheaper, and competitive with o1-preview at coding tasks (you can see how it performs here).
Developers on tiers 1-4: these models aren’t available in the API for your account while we’re in this short beta period. We’ll be in touch over email as we expand access.
ChatGPT Plus subscribers can also try o1 today.
You can read more about these models in our blog post. Keep in mind OpenAI o1 isn’t a successor to gpt-4o. Don’t just drop it in—you might even want to use gpt-4o in tandem with o1’s reasoning capabilities. I’m excited to see how you add reasoning to your products!
Extremely impressed so far with domain-specific knowledge, absolutely incredible!
Will be testing much more in the morning. DevDay is gonna be awesome!
Will image or file upload be enabled or is this a limit?
Just a heads-up to any devs migrating to ‘o1’ models via API, it seems like the first message’s role can no longer be ‘system’. Traditionally, we started each completion with [systemMessage, userMessage], but we are now doing [userMessage, userMessage] to get ‘o1-preview’ working.
No other widely available model has been able to answer this question before. o1 nailed it in 23 seconds. Pretty amazing.
Thank you for sharing this heads up! Yes,
You can read more about limitations here as you build with o1 using the Chat Completions API: https://platform.openai.com/docs/guides/reasoning/beta-limitations
Thanks!
Just took for a spin via API, definitely a more thoughtful and comprehensive response to some of my more complicated tasks that usually require I break up into smaller tasks. Would LOVE to see the Chain of Thought but don’t see a way can through API, just the reasoning token count.
Regarding the API:
Is there a risk that concealing the model’s internal “processing” or “thinking” could set a precedent for other LLM providers?
Currently, we have to trust that the output tokens are generated fairly, without any transparency into the process. We’re only given the input and the final output (A and Z), without insight into what happens in between.
If this becomes the norm for LLMs that incorporate “thinking” or “reflection,” anyone integrating these services could face significant challenges in troubleshooting when something goes wrong.
Yes. I think it’s much safer to assume that employees are prompt engineering the heck out of various stages of 4o performance all the time.
I tried some questions from simple bench, and it got them both correct. Very impressive.
(Ice Cubes in a Frying Pan)[ChatGPT]
(There are no more cookies)[ChatGPT]
Is there a risk that concealing the model’s internal “processing” or “thinking” could set a precedent for other LLM providers?
The reasoning engine we built does the same internal thinking process but in the case of our engine we can potentially generate 100’s of thousands of intermediate thinking tokens. We hide these tokens as well for I’m assuming similar reasons as OpenAI. They’re really only meaningful to the LLM.
I have a training mode I can put our engine in that dumps everything to a stream of jsonl files which are potentially useful for fine tuning. I use these files to debug some but to be honest there’s not a lot you can do in response to them. They’re really only meaningful to the model. They can give you hints about how you can potentially improve your prompt but they’re hints at most.
For many general users it might be practical as an option to hide the underlying processing, as most people don’t require access to this level of detail.
However, for those of us integrating AI into our own solutions, troubleshooting issues often demands deeper insight. In programming, even though most users won’t dig into debug logs, there are always some who need to investigate and understand what’s happening under the hood.
Even for us who want to create our own models and advance in machine learning through our own theories.
In the context of AI, the model’s “processing” is critical—it shows how the output was generated from the input. Regardless of how many users need this, the opportunity to access such information should be available for those who do.
At the very least it would be fair as a toggle able option, considering that ultimately we are paying for the tokens.
One issue they would need to address is how they stream these to you. They would have to probably do it as a new message type
I tried some questions from simple bench , and it got them both correct. Very impressive.
(Ice Cubes in a Frying Pan)[ChatGPT ]
what’s interesting is that I got the same correct answer for the Ice Cube question but when I shift the format of the output to where it has to answer using HTML it gets it wrong.
What this highlights is that it’s not actually reasoning it’s way to an answer. It’s just memorizing answers…
This is really exciting. Basically it is just a LLM fine-tuned for extended thinking aloud, but this opens many possibilities.
I share the concern of previous posters regarding the limited explainability of model responses when the chain of thought is hidden. But in ChatGPT, we currently can see the internal thinking process of o1. Is this actually filtered and paraphrased by another LLM or does this limitation only apply to the API? What information is tried to hide exactly?
FWIW, of course even o1 can be led to reasoning abilities pretty quickly. I asked it to divide two numbers in nonal system without converting it to decimal and basically it stopped its inner monologue prematurely and gave me a wrong answer: https://chatgpt.com/share/66e3a3dc-2808-8004-a6b7-15670d6356de
Apparently one of the hardest logical problems, solved by o1-preview but not sufficiently by 4o:
Three gods A, B, and C are called, in no particular order, True, False, and Random. True always speaks truly, False always speaks falsely, but whether Random speaks truly or falsely is a completely random matter. Your task is to determine the identities of A, B, and C by asking three yes–no questions; each question must be put to exactly one god. The gods understand English, but will answer all questions in their own language, in which the words for yes and no are da and ja, in some order. You do not know which word means which.
o1-preview took 83 seconds, but what an amazing answer!
To solve this intricate puzzle, we’ll break it down step by step. The goal is to determine the identities of the three gods—A, B, and C—who are either True, False, or Random. We’ll do this by asking three carefully crafted yes–no questions, each directed to one god. The gods answer in their own language, using the words “da” and “ja”, but we don’t know which word means “yes” or “no.”
Gods’ Behavior:
Constraints:
Question 1: Ask God A:
“If I asked you ‘Is God B Random?’, would you say ‘da’?”
Analysis:
Interpretation:
Conclusion:
Now, we focus on a god we know is not Random. Let’s assume from Question 1 that God A is not Random (since Random’s answers are unreliable). We’ll use God A for the next question.
Question 2: Ask God A:
“Is ‘da’ your word for ‘yes’?”
Analysis:
Interpretation:
Conclusion:
Using the information from Questions 1 and 2, we can craft the third question to resolve any remaining ambiguities.
Scenario Analysis:
Question 3: Ask God A:
“Is God B False?”
Question 3: Ask God A:
“Is God C Random?”
Final Identification:
Question 1 to God A:
Question 2 to God A:
Question 3 to God A or another known non-Random god:
By carefully analyzing the answers to these three questions, you can deduce:
Note: This solution hinges on the logical consistency of the gods’ answers and strategically uses their truth-telling or lying nature to unravel the puzzle.
Thanks for your support @osmonov Great advice.
There is one problem though:
how did you conclude that God C is not Random in Step 1
o1-preview after 33 s:
Thank you for pointing out this important detail. You are absolutely correct to question how I concluded that God C is not Random in Step 1. Upon revisiting my previous explanation, I realize that the conclusion was not fully justified. Let me clarify and correct my reasoning.
Question 1: We ask God A:
“If I asked you ‘Is God B Random?’, would you say ‘da’?”
Since we don’t know which word means “yes” and which means “no,” and because Random answers unpredictably, we need to consider all possibilities carefully.
Subcase 2a: God B is Random
Subcase 2b: God B is not Random
Subcase 3a: God B is Random
Subcase 3b: God B is not Random
To solve this puzzle effectively, we need a strategy that:
We need a question that ensures both True and False will give the same answer, and that Random’s unpredictability is minimized.
Question 1 to God A:
“If I asked you ‘Is God B the Random god?’, would you say ‘da’?”
But as we’ve seen, this doesn’t eliminate the ambiguities.
Instead, let’s use a logical construct that has been proven to work:
Question 1 to God A:
“Does ‘da’ mean ‘yes’ if and only if you are True if and only if God B is Random?”
This question is designed to ensure that the answer will be “da” if God A is True and God B is Random, or if God A is False and God B is not Random.
However, this is quite complex and may not lead us directly to the solution.
A better approach is to use the standard method:
“If I asked you ‘Is your word for ‘yes’ ‘da’?’, would you say ‘da’?”
Analysis:
Interpretation:
Now that we know whether “da” means “yes” or “no,” and whether God A is True or False, we can proceed.
“Is God B Random?”
“Are you False?”
In my initial explanation, I prematurely concluded that God C is not Random based on the first question, which was incorrect because the information from the first question alone is insufficient to make that determination due to the complexities introduced by:
I apologize for the oversight and any confusion it may have caused.
I hope this clarifies how to correctly approach the problem and explains why the initial conclusion about God C was incorrect. Please let me know if you have any more questions or need further clarification.
I used o1 for an hour and wow…
Comparing 4o and o1 is like comparing a Prius to a Ferrari… Both will get you to your destination, the Ferrari just gets you there faster.
@traditionals15 It is not a good analogy, looks like a ferrari – 4o – and a RAM 4×4 – o1 – not fast but fitted for road and off-road uses.
Powered by Discourse, best viewed with JavaScript enabled
source
This article was autogenerated from a news feed from CDO TIMES selected high quality news and research sources. There was no editorial review conducted beyond that by CDO TIMES staff. Need help with any of the topics in our articles? Schedule your free CDO TIMES Tech Navigator call today to stay ahead of the curve and gain insider advantages to propel your business!
There are two models:
Developers on usage tier 5: you can get started with the beta now. Try both models! You may find one better than the other for your specific use case. Both currently have a rate limit of 20 RPM during the beta. But keep in mind o1-mini is faster, cheaper, and competitive with o1-preview at coding tasks (you can see how it performs here).
Developers on tiers 1-4: these models aren’t available in the API for your account while we’re in this short beta period. We’ll be in touch over email as we expand access.
ChatGPT Plus subscribers can also try o1 today.
You can read more about these models in our blog post. Keep in mind OpenAI o1 isn’t a successor to gpt-4o. Don’t just drop it in—you might even want to use gpt-4o in tandem with o1’s reasoning capabilities. I’m excited to see how you add reasoning to your products!
Extremely impressed so far with domain-specific knowledge, absolutely incredible!
Will be testing much more in the morning. DevDay is gonna be awesome!
Will image or file upload be enabled or is this a limit?
Just a heads-up to any devs migrating to ‘o1’ models via API, it seems like the first message’s role can no longer be ‘system’. Traditionally, we started each completion with [systemMessage, userMessage], but we are now doing [userMessage, userMessage] to get ‘o1-preview’ working.
No other widely available model has been able to answer this question before. o1 nailed it in 23 seconds. Pretty amazing.
Thank you for sharing this heads up! Yes,
system messages are not supported during the beta phase, it’s a known limitation at the moment.You can read more about limitations here as you build with o1 using the Chat Completions API: https://platform.openai.com/docs/guides/reasoning/beta-limitations
Thanks!
Just took for a spin via API, definitely a more thoughtful and comprehensive response to some of my more complicated tasks that usually require I break up into smaller tasks. Would LOVE to see the Chain of Thought but don’t see a way can through API, just the reasoning token count.
Regarding the API:
Is there a risk that concealing the model’s internal “processing” or “thinking” could set a precedent for other LLM providers?
Currently, we have to trust that the output tokens are generated fairly, without any transparency into the process. We’re only given the input and the final output (A and Z), without insight into what happens in between.
If this becomes the norm for LLMs that incorporate “thinking” or “reflection,” anyone integrating these services could face significant challenges in troubleshooting when something goes wrong.
Yes. I think it’s much safer to assume that employees are prompt engineering the heck out of various stages of 4o performance all the time.
I tried some questions from simple bench, and it got them both correct. Very impressive.
(Ice Cubes in a Frying Pan)[ChatGPT]
(There are no more cookies)[ChatGPT]
Is there a risk that concealing the model’s internal “processing” or “thinking” could set a precedent for other LLM providers?
The reasoning engine we built does the same internal thinking process but in the case of our engine we can potentially generate 100’s of thousands of intermediate thinking tokens. We hide these tokens as well for I’m assuming similar reasons as OpenAI. They’re really only meaningful to the LLM.
I have a training mode I can put our engine in that dumps everything to a stream of jsonl files which are potentially useful for fine tuning. I use these files to debug some but to be honest there’s not a lot you can do in response to them. They’re really only meaningful to the model. They can give you hints about how you can potentially improve your prompt but they’re hints at most.
For many general users it might be practical as an option to hide the underlying processing, as most people don’t require access to this level of detail.
However, for those of us integrating AI into our own solutions, troubleshooting issues often demands deeper insight. In programming, even though most users won’t dig into debug logs, there are always some who need to investigate and understand what’s happening under the hood.
Even for us who want to create our own models and advance in machine learning through our own theories.
In the context of AI, the model’s “processing” is critical—it shows how the output was generated from the input. Regardless of how many users need this, the opportunity to access such information should be available for those who do.
At the very least it would be fair as a toggle able option, considering that ultimately we are paying for the tokens.
One issue they would need to address is how they stream these to you. They would have to probably do it as a new message type
I tried some questions from simple bench , and it got them both correct. Very impressive.
(Ice Cubes in a Frying Pan)[ChatGPT ]
what’s interesting is that I got the same correct answer for the Ice Cube question but when I shift the format of the output to where it has to answer using HTML it gets it wrong.
What this highlights is that it’s not actually reasoning it’s way to an answer. It’s just memorizing answers…
This is really exciting. Basically it is just a LLM fine-tuned for extended thinking aloud, but this opens many possibilities.
I share the concern of previous posters regarding the limited explainability of model responses when the chain of thought is hidden. But in ChatGPT, we currently can see the internal thinking process of o1. Is this actually filtered and paraphrased by another LLM or does this limitation only apply to the API? What information is tried to hide exactly?
FWIW, of course even o1 can be led to reasoning abilities pretty quickly. I asked it to divide two numbers in nonal system without converting it to decimal and basically it stopped its inner monologue prematurely and gave me a wrong answer: https://chatgpt.com/share/66e3a3dc-2808-8004-a6b7-15670d6356de
Apparently one of the hardest logical problems, solved by o1-preview but not sufficiently by 4o:
Three gods A, B, and C are called, in no particular order, True, False, and Random. True always speaks truly, False always speaks falsely, but whether Random speaks truly or falsely is a completely random matter. Your task is to determine the identities of A, B, and C by asking three yes–no questions; each question must be put to exactly one god. The gods understand English, but will answer all questions in their own language, in which the words for yes and no are da and ja, in some order. You do not know which word means which.
o1-preview took 83 seconds, but what an amazing answer!
To solve this intricate puzzle, we’ll break it down step by step. The goal is to determine the identities of the three gods—A, B, and C—who are either True, False, or Random. We’ll do this by asking three carefully crafted yes–no questions, each directed to one god. The gods answer in their own language, using the words “da” and “ja”, but we don’t know which word means “yes” or “no.”
Gods’ Behavior:
Constraints:
Question 1: Ask God A:
“If I asked you ‘Is God B Random?’, would you say ‘da’?”
Analysis:
Interpretation:
Conclusion:
Now, we focus on a god we know is not Random. Let’s assume from Question 1 that God A is not Random (since Random’s answers are unreliable). We’ll use God A for the next question.
Question 2: Ask God A:
“Is ‘da’ your word for ‘yes’?”
Analysis:
Interpretation:
Conclusion:
Using the information from Questions 1 and 2, we can craft the third question to resolve any remaining ambiguities.
Scenario Analysis:
Question 3: Ask God A:
“Is God B False?”
Question 3: Ask God A:
“Is God C Random?”
Final Identification:
Question 1 to God A:
Question 2 to God A:
Question 3 to God A or another known non-Random god:
By carefully analyzing the answers to these three questions, you can deduce:
Note: This solution hinges on the logical consistency of the gods’ answers and strategically uses their truth-telling or lying nature to unravel the puzzle.
Thanks for your support @osmonov Great advice.
There is one problem though:
how did you conclude that God C is not Random in Step 1
o1-preview after 33 s:
Thank you for pointing out this important detail. You are absolutely correct to question how I concluded that God C is not Random in Step 1. Upon revisiting my previous explanation, I realize that the conclusion was not fully justified. Let me clarify and correct my reasoning.
Question 1: We ask God A:
“If I asked you ‘Is God B Random?’, would you say ‘da’?”
Since we don’t know which word means “yes” and which means “no,” and because Random answers unpredictably, we need to consider all possibilities carefully.
Subcase 2a: God B is Random
Subcase 2b: God B is not Random
Subcase 3a: God B is Random
Subcase 3b: God B is not Random
To solve this puzzle effectively, we need a strategy that:
We need a question that ensures both True and False will give the same answer, and that Random’s unpredictability is minimized.
Question 1 to God A:
“If I asked you ‘Is God B the Random god?’, would you say ‘da’?”
But as we’ve seen, this doesn’t eliminate the ambiguities.
Instead, let’s use a logical construct that has been proven to work:
Question 1 to God A:
“Does ‘da’ mean ‘yes’ if and only if you are True if and only if God B is Random?”
This question is designed to ensure that the answer will be “da” if God A is True and God B is Random, or if God A is False and God B is not Random.
However, this is quite complex and may not lead us directly to the solution.
A better approach is to use the standard method:
“If I asked you ‘Is your word for ‘yes’ ‘da’?’, would you say ‘da’?”
Analysis:
Interpretation:
Now that we know whether “da” means “yes” or “no,” and whether God A is True or False, we can proceed.
“Is God B Random?”
“Are you False?”
In my initial explanation, I prematurely concluded that God C is not Random based on the first question, which was incorrect because the information from the first question alone is insufficient to make that determination due to the complexities introduced by:
I apologize for the oversight and any confusion it may have caused.
I hope this clarifies how to correctly approach the problem and explains why the initial conclusion about God C was incorrect. Please let me know if you have any more questions or need further clarification.
I used o1 for an hour and wow…
Comparing 4o and o1 is like comparing a Prius to a Ferrari… Both will get you to your destination, the Ferrari just gets you there faster.
@traditionals15 It is not a good analogy, looks like a ferrari – 4o – and a RAM 4×4 – o1 – not fast but fitted for road and off-road uses.
Powered by Discourse, best viewed with JavaScript enabled
source
This article was autogenerated from a news feed from CDO TIMES selected high quality news and research sources. There was no editorial review conducted beyond that by CDO TIMES staff. Need help with any of the topics in our articles? Schedule your free CDO TIMES Tech Navigator call today to stay ahead of the curve and gain insider advantages to propel your business!
