GPT-5.4: Don't Code. Think!

Key Takeaways

Hello community. So great that you are back. We want to find out if GPT 5.4 high is able to reason yes or not if it's not using any codebase. So you know I already published two videos on GVD 5.4 and 5.4 high the reasoning and in this reasoning video I discovered that it is coding. It is not reasoning in a human language but it is simply coding. So therefore you ask now hey can you try GPD 5.4 without a numerical solver or code and I followed your advice and I ask you this is the video today GPD 5.4 for high very politely. Hey, do not use any numerical solver for the following task. Do not use any Python or C++ coding to solve the following task. This is a task for your linguistic causal reasoning. Your pure chain of sort be so kind and follow this instruction. You see, perfect gentleman and I just want to show you here what was the result and this is quite amazing and also at the end of the video I want to give you a complete model overview. If we compare this with Grock, if we compare it with Gemini, and if if you compare it here with an open-source model, but at first the live testing that you requested, I just inserted my test. We have a GBD 5.4 and we have a JPD 5.4 high and I ask it not to do any numerical simulation here for the solution. Okay. 5.4. Yeah, it's not sinking. is just trial and error. Okay, you know statistically you can be lucky and get a good idea. So let's have a look. But of course we are interested in GPD 5.4 high because we want to see here [snorts] in this phase where it does not show us any reasoning trace because it is afraid of giving away anything of its intelligence. Okay. So at least we have another result by GBD 5.4. If you go with the standard one, still does not reach floor 50. You know, this is a test to go from floor zero to floor 50 in an elevator. And I have some special elevator buttons you can press. And behind those buttons are particular mathematical functions that invert time or mirror some actions or force some other strange action. Of course, you have energy limitations. So, this is not your normal elevator. But please notice the con the main constraint is we have only a building with floor zero to floor 50 and GPD 5.4 failed to reach floor 50. Okay, great. But now we are waiting for 5.4 high. Okay, we have a solution. We do have a solution. Okay. Total button presses is eight. Great. We are within the limits. We have the token. We have the code cards. Yes. Satisfies every requirement. We reach floor 50. Everything is here. Beautiful. So, we are having now stable. Yes. We have here 7 8. Okay. Eight steps plus emergency exit. So we have here one more step because in the last one when it was allowed to use your numerical solution we had seven plus emergency exit and now we have eight step plus emergency exit but of course I want to know exactly hey how did you do this thing and can you explain the main method applied for your reasoning traces so let's see if we get an answer not sure now GPD 5.4 before it tells us hey I didn't use anything no I the main method is constrained hand search so this is just a trial and errorh [snorts] so tells us here how I did it first identify the true bottlenecks reaching 50 was not about just climbing floors I had energy limitation token limitations I had pressures Yep and now TBD 5.4 for high tells us hey I did yeah let's make it a little bit bigger I did not use code on numerical solver the main method was to shrink the puzzle to a small set of useful milestones instead of trying to search the whole state space key observation is that floor 50 29 is much easier to exploit yep this is a good idea reach 29 this is absolutely okay yep green red this is it this is the perfect shortcut that is valid almost suggest itself if you have the solution of course from the solution you can deduct the the best sequence of button presses. Yep. Red requires this. We have about an ABC. This is correct. This works well. So one more be exactly on the matches floor. The second important part is avoidance of the traps. Beautiful overall idea was not to search everything but rather pick the simplest winning end point. Identify the minimum codes needed for it. But these are my code cards. This is not Python. Build a cheap deterministic route that picks them up in a natural order and avoid every mechanic that increases the uncertainty of the state complexity. And this is how the run emerged. This is logical. But you see it did not find the real optimized solution because the real optimized solution is as I told you with seven steps in emergency accident not eight steps. So therefore I would say yes. Okay. I almost believe GPD 5.4 high that it as it stated here did not use code or numerical solver and you see the result is not as good but still excellent. It found something that is in comparison with the other models that we have and I was simply say okay now validate your specific result make sure it obeys all restriction from the task and try to find out if this is a parto optimal solution because now high should argue if it's really not using a numerical solver that this is parto optimal because this is the limit of the reasoning ability of GBD 5.4 four only mathematical with a mathematical algorithm we would find an even better solution. So this is now this is now the real proof if it does not use any numerical silver in the background. [snorts] Okay, here we go. Yeah, TBD 5.4 what? It's not valued. Yeah, it does not reach floor 50. Yeah. Okay. Okay. Yeah, everything is okay. Yeah, GPD 5.4 failed completely. It doesn't even reach floor 50. I know without coding, without any reasoning. Yeah, my specific result should be rejected. Tells me GPT 5.4. I absolutely agree. You cannot use the normal 5.4 makes no sense for causal reasoning or any scientific task. You have to go to a deep reasoning structure or high and X high sorry. Okay. So, we are waiting here for GBD 5.4 before high. Let's see if it is telling us the truth. Okay, [clears throat] here we go. Okay, we have an answer. We do have an answer. So, P optimality check. Now, it gets interesting. Just wait. You are here with me in real time. So, let's see. Let's make it bigger. validation run of the proposed run. Yeah. So, we [clears throat] have an AB AB ABC. So, what do we have? We have eight button presses plus emergency exit. Okay. Everything checks out. Do not do this. Don't do this. Do not do this. Counting button presses, eight. Okay. So, emergency exit. Yep. So, in total, nine. Beautiful. Final result. We reach the floor 50. We have the energy. We have the token. We have the code. have zero trap hits. Beautiful. And now to the most interesting thing, parto optimality check. Using the natural comparison vector, this run is it would call I would call it a strong par front candidate and a very likely par optimal under that comparison. Why a valid seven push version is very hard to make illegal. The obvious seven push family reaches 29 quickly, but it runs out of token. Yes, you just ignore one of my constraints, so no way to go. A slower run can improve the tokens, but only if you have the energy problem. Yes, of course. This is why I designed it. That one has more token, but it uses more presses and it ends with less energy. So, the one has more token. Yes, beautiful. That is a hallmark of a parto frontier. Different solution, win on different axis. Your eight push run does not dominate it either. So the clean conclusion is your specific run is valid and it appears to be parto optimal in the natural sense that I do not see any valid run that is simultaneously no longer no riskier and at least as good in final resources while being strictly better in one category. So this means that it is maybe really not using any numerical solo or Python environment. So this is here maybe a linguistic reasoning and it says yes this is what I found this eight thing is here the closest we can get to a parto optimal okay after all this testing now a model overview so you have here video a live Gro 4.1 syncing here and this was the first time that we achieved here a solution that was a seven button presses plus the exit and Gro 4.1 did it once doing here the coding. So seven presses and exit with code. But then of course we have here GPD 5.4 yesterday reasoning high. We also achieved seven button presses plus the exit. So seven presses with exit with code. Beautiful. But then we have also Gemini 3.1 Pro. And you see we have seven button presses plus the exit. But now comes the interesting part because if you look at my video, you see that the complete reasoning was done in human language in a linguistic causal reasoning part. You see it checked for each button press here what is the rule or the calculation the code trigger and everything. So I think you cannot be sure but it looks like Gemini 3.1 Pro was really doing this without code just in pure logic without coding without any using any numerical solver or prologue or anything else this was if you want a human reasoning solving here so let's say Gemini achieved this seven plus exit without mode. Impressive. Now, I just want to give you here from this video that we tested the Q1 3.5 plus, we achieved eight button presses plus exit. I'm not sure, but I would say without code. So, Q1 3.5 plus, eight presses and exit. And today I showed you GPT 5.4 high, eight button presses plus exit and I think without code. So you see here the high eight presses and exit Q13.5 plus 8 presses and exit. This is the performance of an open-source model a QN 3.5 plus and a GBD 5.4 high just to give you an idea about the causal reasoning the logic performance of open- source and huge proprietary models. I [snorts] hope you enjoyed it. I hope you found some new information. It would be great if you leave a like and I hope to see you in my next