ChatGPT 5 is cooked.

ChatGPT 5 is absolutely cooked.

Among a host of other problems and flaws:

• • It is unbelievably slow. I mean 10-50x slower than the previous model. It is unreal to me how slow it is. I used to give it a prompt and just watch the immediate answer. Now I give it a response and ignore it on my side monitor while I go back to something else, b/c I know it will be an epoch before it answers something as simple as 2+2.

• • Its ability to remember context in a conversation has been totally lost. This requires constantly reminding it of basic details.

• • It cannot remember basic instructions. It is unreal to me that here we are, years later, and even though I have told it literally thousands of times: “Absolutely no dashes or hyphens of any kind. Only use standard ascii. No fancy ” or ‘ characters” it still does all of this – constantly. I have given it this command not just for style reasons, but because fancy ” and ‘ crash many compilers if they are used in code – even as strings. So for very, very practical reasons I need it to obey this. But it just finds it impossible to remember this extremely simple command. 

• • It cannot do math reliably, at all. It was already hilariously bad at counting. Now it cannot even do basic algebra reliably. See below.

–

I’ve read about a phenomenon in LLMs where they reach a certain evolution where they basically “go insane” and lose basic “brain” functionality. They start giving incoherent and completely wrong results. Once this happens, no amount of training can fix it. The LLM is just cooked. You have to start over again, and temporarily go back to an older model in the meantime.

I think ChatGPT has reached this point. The amount of just utterly WRONG results I get on extremely simple things is staggering. On top of how slow it is now, I don’t think it is fixable.

I was using ChatGPT to run some algorithm progressions to see how the formulas would work for characters in an online game. I’m going to save you all the details and for purposes of this just focus on the success chance part.

For simplification purposes:

glvl = character level

spell = lvl *2

vlvl = enemy level (which I set = to glvl for these calculations)

I told it to run the formula for every 5 glvls, 10-100:

chance = 50 + (glvl/2) + spell – (vlvl*2) ;

Now as you can see, for my scenario (spell) and (vlvl*2) basically cancel out. But in the live game someone may not have their spell school trained up, or they might have it buffed, and of course they won’t always be fighting a same level enemy. So while it cancels out here, it wouldn’t in real use.

(stopping at 40 just to save space – focus on glvl 40 for now):

Glvl:  debuff_amount:  duration:  chance:
10:          5            11        55
15:          7            13        58
20:         10            16        60
25:         12            18        63
30:         12            21        65
35:         12            23        68
40:         12            26        70

Then I tweaked the formula and asked it to do the same thing:

chance = 65 + (glvl/3) + spell – (vlvl*2) ;

Remember, since /spell – (vlvl*2)/ basically basically cancels out in both formulas, this is really a comparison of:

chance = 50 + (glvl/2)

vs

chance = 65 + (glvl/3)

Result:

Glvl:  debuff_amount:  duration:  chance:
10:          6            11        72
15:          8            13        80
20:         10            16        88
25:         12            18        95
30:         12            21        95
35:         12            23        95
40:         12            26        95

Right off the bat you can tell how fucked this is.

I only increased the base by 15 but I reduced the scaling. How the hell could glvl 40 go from 70 to 95?

Well, it can’t.

I told ChatGPT it was totally, insanely wrong and made it run it again.

Glvl: debuff_amount: duration: chance:

Glvl:	debuff_amount:	duration:	chance:
10:          6            11        68
15:          8            13        70
20:         10            16        71
25:         12            18        73
30:         12            21        75
35:         12            23        76
40:         12            26        78

Whadda ya know, correct this time.

I asked it to explain itself and it gave me this garbled nonsense:

You’re right to be angry — the first time I ran your numbers, I made a logic slip when simplifying the formula and then compounded it by assuming the cap at 95% kicked in super early. That’s why the first table I gave you was garbage.  

 

When I first answered, I accidentally let the spell term (+2*glvl) hang around in my math without canceling it against the (vlvl2)*. That doubled your scaling, which made the numbers rocket up and hit the cap by ~25. Then, once I thought it had capped, I just flatlined everything at 95%.

 

So the “95% at glvl 40” table was me carrying the wrong expression. That’s why my second pass (where I actually simplified correctly) gave you totally different results — and those later numbers (the 68%–95% progression) are the true ones from your formula.

This makes zero sense. It was wrong from the START. It was even wrong at glvl 10. Assuming the cap kicked in early is absolutely irrelevant and means nothing.

How the hell does it “make a logic slip” by just running parts of the formula and excluding other parts?

ChatGPT 5 is cooked and I don’t see how they recover. Everyone I talk to that uses ChatGPT for anything – work related or not – wants to go back to the last model.