Hacker News new | past | comments | ask | show | jobs | submit login
ChatGPT Forgot My Startup (twitter.com/seveibar)
20 points by seveibar on June 3, 2023 | hide | past | favorite | 24 comments



Answers aren't deterministic.* Especially in ChatGPT, it's impossible to compare.

Quick pass:

- questions are different.

- answer = function of several variables, most importantly, entire chat. then temperature, top-p...

- models haven't changed.

- entirely possible chatgpt's temperature, top-p etc. changed but model didn't

see images of gpt3.5 at 0 temp and 0.3 on this tweet, 0.3 makes up an answer like february, 0 gets "sorry, as an AI..."

https://twitter.com/jpohhhh/status/1665007972114743297?s=20

* unless temp = 0 and all previous messages in the chat are the same. On ChatGPT, you can't even see all the messages in the chat System messages aren't visible


Einstein’s quote about God not playing dice with the universe seems appropriate here.

ChatGPT’s universe is governed by probability as it applies to a training set.


The first response could easily have been a lucky hallucination.


Very likely. I asked it about a programming language I work on (which it doesn't know the first thing about) and it hallucinates features based on its name, sometimes accurately. "CollegeAI" has some obvious potential use cases.


Very unlikely. That's a pretty accurate description of the startup. "CollegeAI" doesn't suggest at all it's an admission matching platform. The hot take would've been it's a college driven by AI, or college specializing in degrees related to AI or god knows what.


If it hallucinates that it's to help people get into college (which isn't impossible to guess) everything else flows naturally from that.


Everything is possible. We're discussing what's likely and not likely. I don't buy the hallucination theory. BTW, technically it "hallucinates" everything, because it has no connection to the outside world, only memories to base its answer on. But if the hallucinations are correct, then being less correct is still a problem. If you can show it's more correct in other ways, be my guest.


Everyone says GPT 4 is getting much worse. OpenAI says nothing changed. OpenAI, seems we have a problem.


People are familiar enough with LLMs to understand the nuance in their limitations. Many of these only become obvious when you transition from using it as a toy to using it as a tool.

We a novel tool like this comes out people take wild guesses about how to use a tool, jumping well past its capabilities. When they see limits, they jump back to zero (or close to zero) and try again.

Once your familiar with a tool, you know how to use it near its limits. Naturally, you’re going to feel more pain when you cross the limit. It won’t feel like you’re far away from a known good.

I suspect this is essentially all that’s going on. P


Wow, this actually sounds reasonable. Certainly possible


We trained a large language model and had it deployed for multiple month last year (Cedille.ai, largest open source LLM for French at the time). We had multiple companies experimenting with it to use it in production. There were multiple occurrences where we got feedback about "the model having gotten significantly better". But the model in prod was never altered.


It is probably catastrophic forgetting.


I have not noticed GPT-4 getting any worse! I also don’t use Twitter. Perhaps there is a correlation?


Maybe the AI really is learning from it's human inputs, thus becoming less intelligent...


What's likely changed is your expectations.


any link to openai refuting the claim?


OpenAI Employee: GPT-4 has been static since March - https://news.ycombinator.com/item?id=36155267


I have had a significant decrease in usable GPT-4 via Chat. Anyone to say otherwise hasn't been using it daily as a plus user since first access. It's absolutely hindered in comparison. But also twice as fast.


Did anyone try to look at the parameters if they're available in the browser requests to see if that may be related?(temperature?)

or maybe are they doing quantization or some other kind of optimization?


This is a very important detail. Temperature > 0 means a different result every time


This is 3.5 and it seems pretty likely the original is just made up from the name of the thing, not actually something it has in its training.


GPT 3.5 seems to have forgotten that Go has generics now. They definitely need to add recent training data.


Did it forget? I actually checked this out recently and generics were introduced in 1.18 while chatgpt is only trained on golang up to 1.17. I realize it could also be trained on alpha release for 1.18 but it doesn’t seem to be and that’s a plausible explanation.


OpenAI seems to be prioritizing the mobile experience over web.

Their mobile app is extremely fast with GPT-4, faster than web. I’d imagine once mobile is well-established, they’ll equalize it out.

Totally makes sense, if I’m right. They should just state it publicly.




Consider applying for YC's Spring batch! Applications are open till Feb 11.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: