OpenAI did made GPT3.5 more stupid?

In the old times, if you write a SYSTEM prompt and delete default USER input, a reasonable output will be generated, but now, if you do so, garbage will come out.


It does feel like in some areas it has been underperforming older chatgpt versions. I do think that they work hard on improving overall user experience but feels bad that it is somewhat give an take between use cases users.

1 Like

I think they might have distilled GPT3.5’s 175B params into a smaller model.
Warning: It is just a thought! I guessed. It has a very high probability to be wrong.

1 Like

You should give AlphaWave a try… https://GitHub/Stevenic/AlphaWave-py

I guarantee you’ll get valid JSON back every time. Even with GPT-3.5-turbo


The prompt could be improved though. It is a natural language system with the reasoning ability of a 5-year old. Ideally talk to it like you would a child.

Even as an adult who does programming, it’s not entirely clear what you’re expecting. You have to give it some examples.

Also the system instructions are more like “soft” instructions with GPT3.5, and they seem to lose fine details like commands and formatting.


Likely wrong conclusion to think that they spend another million dollars in compute just to train on smaller tokens or parameters, when the lack of computing resources are the biggest hurdle to moving forward.
There’s lots of other optimizations to be done to an AI that cut the generation resources at the expense of quality.

1 Like

I can confirm it is not just 3.5 but also 4 and 4-0316 in the playground for me. Very noticeable drop in memory, logic and reasoning


In the “old version” (if there’s any), the model would simulate user input and outputs.
It used to work as zero-shot model, though.

Actually I haven’t get GPT-4 API access until now so I have no clue how’s GPT-4 doing.
But I’d rather spend more on better models rather than decreasing the price and make the model worse.

You’re right, but they can just release more versions and make pricing more detailed.

Big updates to 3.5-turbo today, new version of the model releasing

1 Like

Welcome to the forum!
I know this update, I thought they made old version more stupid and the new version better than the old one.
The 16k-model seemed to be better.

The new version is significantly less intelligent. It’s depressing because I got online thinking the new engine would be better. This is what we’re stuck with.

1 Like

Sad, maybe they know there are competitors so they started to make it cheaper.

I completely agree! I’m deeply involved in cold email outreach and I’ve found that the new model seems almost programmed to start every email with ‘I hope this email finds you well’. Despite spending hours trying to find a workaround, I’ve had no luck in changing this default greeting. Interestingly, this wasn’t an issue with the previous version.

1 Like

They changed stuffs, without changing the model name.

Have you tried adding enforcement to the input itself for the most important parts, vs depending on the internal guidelines/rules to work? I have found that it works great. If you haven’t and want to try I can give you more tips, just let me know.

It’s very frustrating how it refuses to follow some instructions now. My theory is that the bulk of the computational power for an LLM is comprehending the request, similar to a human.

If we as humans understand the instructions, giving a response (even if it’s wrong) is relatively easy. If we are given detailed/advanced instructions, understanding them requires a lot of brain power and we will only respond / do the parts we understand.

That’s essentially what the engine is doing now. It’s not using computational resources to comprehend the request, it’s simply blast out a response to whichever part is easiest for it to understand.

I imagine Microsoft and OpenAI are going to save hundreds of millions, if not billions with this change, but it’s terrible for the end-user. They should at minimum give it 10% intelligence back.

1 Like

If that works well for you, maybe ClosedAI is trying to force users to consume more tokens so they make more money, evil move. :scream:

That’s how they would make more money hehe. :sweat_smile: