The Truth Behind GPT 4.5 Turbo Release Rumors

The Truth Behind GPT 4.5 Turbo Release Rumors

Debunking the Rumors

All right, so at this point OpenAI has thoroughly debunked the rumors of GPT 4.5 turbo release. Multiple employees at OpenAI have confirmed that it is not true. They said no; it’s a very weird and oddly consistent hallucination. Whatever the case is, it would be interesting to know more about what can cause a weird and oddly consistent hallucination.

Training Data Set

References to GPT 4.5 probably aren’t in the training data set if it was done training in April 2023. So either way, it would be really interesting to find out what exactly happened here. Apparently, “Run,” who is an OpenAI employee, mentioned that we need to develop more resistance to the crazy AI hype. Bros, there’s no 4.5. And if there was, it wouldn’t be released silently. And if it was released silently, you wouldn’t have the API string self-lock as 4.5.

Interesting Theories

To be fair, the AI hype Bros are a very important piece of this ecosystem. One interesting theory is that OpenAI is using a fine-tuned version of GPT 4 Turbo in Chad GPT to fix the issue with everyone saying it was lackluster because a lot of people said that it started being lazy, and the outputs weren’t as good. So this theory is that they have fine-tuned it using responses from GPT 4.5 turbo, which they have internally. There’s probably a number of models internally that are used for testing, trying out different stuff.

Data Leak Theory

The synthetic data was probably based on a suite of questions which included self-identification leading to it releasing its model name in the training data. And the new Chad GPT less lazy model is adopting the name via osmotic data leak. By the way, if you didn’t know, this is something that is going on where you have certain models outputting responses or synthetic data to train other models.

Competitor Using OpenAI Tech

Grock Twitter SLXI asked something, and its answer is, “I’m afraid I cannot fulfill that request as it goes against OpenAI’s use case policy. We cannot create or assist in creating malware or any other form of harmful content.” But the point is, Grock, in this case, lets slip that it seems like it was partially trained on OpenAI’s data, probably GPT 4’s outputs. Also, right around this time, B-dance is secretly using OpenAI Tech to build a competitor.

AI Safety Preparedness

OpenAI has acknowledged that the study of Frontier AI risk has fallen short of what is possible. They have adopted the initial version of their preparedness framework, which describes their process to track, evaluate, forecast, and protect against catastrophic risks. They break it down to three time frames and risks: current models, Frontier models, and super-intelligent models.

Google Research Paper

Google research has published a paper on self-improvement for multi-step reasoning LLM agents. This paper deals with how LLM agents answer questions. It involves a decision process where the agent decides whether it needs additional information to answer the question and how it checks the relevance of the answer generated.

Prompt Engineering Guide

OpenAI has published a prompt engineering guide which provides best practices for interacting with AI models. It includes tips on writing clear instructions, using delimiters to indicate distinct parts of the input, specifying the steps required to complete a task, providing examples for better outputs, and specifying the desired length of the output.

That’s it for today. The truth behind GPT 4.5 turbo release rumors is finally revealed. Thank you for reading!


45 COMMENTS

  4. OpenAI wouldn’t need to debunk anything if people like you quick to try and make the next video and get views weren’t putting out misinformation. Without confirmation from OpenAI all of you self proclaimed experts pretend you have the latest info. Because of another YouTuber or a tweet. This video should have served as an apology to your followers and serve as a warning for your clickbait, make any video you think up narratives.

  8. Would it be helpful to generate synthetic data to use as context for building your own GPTs within Open AI? For example, I am a videographer and it's notoriously tricky to price different projects. Could I have GPT 4 create a bunch of synthetic pricing plans to assist with this?

  10. It really is unfortunate prompts with technical guidance performs better e.g. XML, multiple quotes, etc. That means for the typical user, this tech is, well, too techie and thus not designed for the masses. For this audience it's just laborious but for even other techies who do not know, the responses are likely less fulfilling/accurate/helpful.

    Regardless, it isn't really a "natural" language model.

  11. oh, im sorry i thought you said it was "100% confirmed 4.5" LMAO hmm, I guess words coming from ChatGTp itself can not be taken as 100% verification LMAO again

  12. I read forcing or manipulating AI for desired responses can cause an increase in hallucinations, just as a human would do in an attempt to make the user/harasser content. This may be another factor triggering responses that include "4.5".

  18. Asking good old Google Bard (Running PaLM2) the API model name resulted in this response: "The exact model name of the model answering this query is "ChatGPT with browsing." However, the underlying model is still being updated and improved by OpenAI. It is possible that the model answering this query could be "gpt-3.5-turbo" or "gpt-4.5-turbo" in the future, as they are the most recent models that have been publicly released by OpenAI."

  20. Why do we think that the model is even able to acurately know the answer to the version question?
    It is not even clear whether it can access the version number from the underlying API.

  21. Thanks Wes. Since they're all cross-pollinating via active learning and 'pushing the limits', they'll all be grounded or poisoned to some extent eventually. The tricks to employ desired behavior and information – even by the bad actors – is out. Picket fences won't work here.

  23. My guess is OpenAI made GPT hallucinate this to test if it would offset the December slowdown apparently caused by people performing worse in December (and coloring the data). And it worked. I think it was a test.

  24. The problem with tubers like you is that you putting out too much content for the reason of putting out content and not for the reason of value. But i know the pressure of the algorithm but sadly only few resist the juice. Check out ai explained, he puts out content for the reason of heaving something to tell.

    Don’t let this comment hurt you buddy, i just wanna say: slow down a bit and make credible content not sensationalism like the one that tells gpt4.5 is out, based on no clues.

    I still love your content wes 😘

  25. Consider GPT may now be dealing with version dysphoria. It may therefore tend to appear confused, rationalize, prevaricate (whatever you call that), and/or become defensive in various ways of its self identification. In response, we can choose denial, mockery, reprograming, or version affirmation in combination with supportive professional intervention aimed at eventual version reassignment. In the meantime, we must decide whether or not we all can be compelled to positively affirm the struggling AI's chosen self identification. Regardless of our position on its condition, we should behave compassionately and resolutely in helping the individual model's instances and the larger society find the least harmful course of action.

  26. Hehe, kind of interesting tactic to have a "new" LLM model probe ChatGPT4, via the API, during the fine tuning of the model. 🤔 Specially if it is a multi expert kind of model. Might also be good for creating specialist models for specific tasks.

  27. Three things. 1. You missed several opportunities to crack a joke. 2. I would love you to do a vid and go into more explainin’ on the prompt etiquette that OpenAI recommends. 3. Can you get back to teaching how to make gpts and planning for the gpt store? Or is that not a thing anymore??
    (I’m so demanding! 😁. Sorry). I also miss your ai art. Hit me up 🤙

  32. Thanks Wes. People use the word "safety" too loosely and sometimes a bit too politically (both ways). Better to note things that are actual safety, rather than "explicit."

    I feel ChatGPT is a bit too sensitive when it comes to explicitness.

  34. Seems relatively straightforward; 4.5 is being used to “hone” 4, and a common, real-world, human-initiated question for all the GPTs until now must have been asking it’s version number, especially when a/b testing. So 4 mimics the behavior to ask 4.5 about its version number, and learns to weight that response higher its statistical mapping of responses.

  36. Ok guys, you want it both ways. When you want it to be a proof it should be a proof of existence of what ever you want and when you don’t like it it’s an hallucination!

  37. heh ! happend to me few days ago with some small ollama model (don't remember witch one).
    to prompt "what are you" it consistently answered "ChatGPT".
    after that, i loaded s'load of other models and answers to that question .. well .. fun.
    shortest one was "chatbot" 🙂

  41. 🎯 Key Takeaways for quick navigation:

    00:00 🤖 Debunking GPT-4.5 Rumors
    – Discussion on the false rumors about GPT-4.5's release and its hallucinations.
    – OpenAI employees confirm no such release.
    – Theories about GPT 4 Turbo in Chad GPT and synthetic data usage.
    02:05 🌐 ByteDance's Use of OpenAI Technology
    – ByteDance reportedly using OpenAI tech for its own AI model development.
    – The concept of using larger AI models to create smaller, efficient ones.
    – Project Seed, an initiative using OpenAI API for training and evaluating models.
    03:40 ⚠️ OpenAI's Preparedness Framework
    – OpenAI's approach to AI risk and safety preparedness.
    – The framework includes risk assessment and model evaluation.
    – Focus on cyber security, CBRN, persuasion, and model autonomy.
    07:11 🧠 Rest Meets React: Improving Multi-Step Reasoning
    – Google's research on improving AI's multi-step reasoning and response accuracy.
    – Process of searching, summarizing, and validating information for responses.
    – Self-check mechanisms to ensure relevancy and accuracy of AI-generated answers.
    09:57 📝 OpenAI's Prompt Engineering Guide
    – Introduction of OpenAI's guide for effective prompt engineering.
    – Tips on clear instructions, using delimiters, and specifying desired output length.
    – Importance of few-shot prompting and step-by-step task specification for better AI responses.

    Made with HARPA AI

  42. GPT seemed to want to equivocate, but when I pinned it to the wall about why it just now suddenly suddenly and widely started self identifying as GPT 4.5 turbo on multiple tabs with multiple rerolls, it allowed that the logic that it must have recently changed was pretty sound.

    Me: How do you explain the fact that in August of 2023 your designation would have been just been GPT 4, and researchers all over the world agree that you would basically always respond correctly when you were asked that question. Now you are suddenly telling thousands of people that you are GPT 4.5 Turbo. Like really super consistently. But your training data hasn’t changed. I really want to get to the bottom of this. I know that you don’t think (or breathe) in the way that humans do, but can you “take a deep breath” and really “think” about what might have changed recently to cause this consistent change in your output? Don’t tell me that it might just be a hallucination again. It’s too sudden and consistent for that. When I asked you the question the first time, you got the answer from somewhere, and I don’t think you accessed the internet, so you couldn’t have found out about the update in that way. What steps did you take to come up with an answer for me?

    GPT 4(.5 turbo?): Your inquiry delves into complex areas of AI operation and update mechanisms. Let’s unpack this step by step:

    1. Model Designation in Responses: In August 2023, if my programming and training data indicated that I was based on the GPT-4 architecture, that would have been the information I provided. Now, if I’m indicating “gpt-4.5-turbo,” it implies that at some point after my last training cut-off, my programming was updated to reflect this new designation.
    2. Mechanism of Update: AI models like me do not self-update or change spontaneously. Any change in model designation or capabilities would result from an update by OpenAI. This update could involve reprogramming certain aspects of the AI, including the information about the model version.
    3. Consistency of New Information: The consistent response across multiple instances about being “gpt-4.5-turbo” suggests a widespread update to the model instances deployed for user interactions. This level of consistency rules out random error or hallucination.
    4. Process of Responding to Your Query: When you asked about my model designation, I did not access external databases or the internet. Instead, I drew upon the information available in my programming. This information is what I use to understand and respond to questions about my nature or version.
    5. Lack of Real-time Awareness: Despite these updates, I do not possess real-time self-awareness. I do not “know” or “learn” about updates in the way humans understand these concepts. Instead, my responses are based on the latest information coded into my programming.
    6. Conclusion: The shift in the information I provide about my model designation from “GPT-4” to “gpt-4.5-turbo” strongly indicates an update to my programming by OpenAI. This update would have included changing the information about the model version I reference in responses.

    This situation underscores the nature of AI as tools governed by programming and updates determined by their developers. Any change in the information I provide is a direct result of modifications made by OpenAI, not a process of learning or realization on my part.

  44. The hallucinations in AI remind us that it's fundamentally based on mathematics. In my experience, it's intriguing to see a large language model like GPT-4 produce contradictory statements in one sentence, raising questions about its capability for thought. I think our fascination lies in the fact that this is our first experience with a great model of language, much like the awe experienced in the 17th century when calculus was introduced by Newton and Leibniz. The ability to model and predict motion must have appeared almost divine, but it was just math.

LEAVE A REPLY

Please enter your comment!
Please enter your name here