Hacker News Reader

Comment by jiggawatts | Hacker News Reader

I look at everything someone from OpenAI says as if a politician is saying it. Sam Altman especially is fond of statements that are deceptive but technically true. His employees appear to be following his lead.

GPT 4 isn’t ChatGPT 4, which is what most people use.

There is also the “system prompt”, which is also likely to be changing but not part of GPT 4.

Etc…

View on HN

I don't see this as being political doublespeak. He is very clearly just talking about the API.

Right but people are complaining about their experiences with ChatGPT.

If you say, "ChatGPT has gotten noticeably worse" and they respond "nothing about our APIs have changed" then it would be reasonable to interpret that as them saying that nothing about ChatGPT has changed.

When in reality many things might have changed about ChatGPT.

He says "the models are static", which means that, if you keep using the same model, you will keep getting the same results.

That is not what that means

The tweet is a response to an API user saying GPT-4 got worse, and the tweet is clearly talking about the API and the model - the context of the HN megathread mostly talking about ChatGPT isn't present in the tweet that's being replied to.

There's a lot of confusion around this, but nothing appears to be caused by doublespeak to me. Confusion about which GPT/ChatGPT anyone is reporting about has been pretty ubiquitous for a while.

Which is not what anyone else is referring to when they say that GPT-4 has gotten worse.

What about the API user he's directly replying to in the tweet, who asked if the model changed?

What made you think that the person he replied to meant to ask about the GPT-4 model/API instead of ChatGPT-on-GPT-4? The latter is what a lot of people refer to when they say "GPT-4." People do not always communicate with infinite precision.

The tweet specifically refers to the model "Seems like OpenAI updated the model" and if you look at their profile you'll see "CEO @HyperWriteAI, @OthersideAI - I make AIs do the impossible." as well as a pinned tweet announcing that they're building a personal assistant/agent on top of this stuff.

I didn't check this part but the OpenAI employee very likely follows them and knows all this context.

Then the HN title is grossly misleading.

I haven't followed personas closely. Do you have examples of Sam Altman saying deceptive but technically true things?

A recent bit is about his push for AI regulation for safety. but the real reason is probably to build a moat to protect his business.

Here's an example of a "technically true" but deceptive statement: he says that he has no financial interest in OpenAI. But there's no way that someone couldn't be the head of a huge, influential organization and not be able to use it to obtain power. For instance: at this point, Microsoft pretty much owns it (they have 100% access to the models and could cut it off at any time). Is it merely coincidence that Microsoft agreed to purchase power from the fusion company he backed?

Another example: I've noticed that a lot of times I'll get "network errors" while chatting with ChatGPT. However, this is while I'm SSH'd into a remote machine with zero latency. I think they realized that they could shift the blame from their server capacity issues to "network errors" because that made the consumer feel like it was their fault, not OpenAI's.

Another example: One of the developers of EleutherAI talked about how they tried to implement OpenAI's original model based on their paper. They were struggling to get it to work. So they actually talked to researchers who wrote the paper and discovered that a lot of what was in the paper wasn't what they did at all.

Worst of all, he is trying to create a crypto token. That would disqualify almost anyone in my book.

Call for regulation then threaten to pull out of EU because of regulation?

What is "ChatGPT 4"? ChatGPT Plus can use GPT-4 and free ChatGPT uses gpt-3.5-turbo.

ChatGPT isn’t just using plain GPT-4, it’s using a specific pre-prompt and possibly additional fine-tuning (which I’m not sure about) compared to the plain GPT-4, available through API.

Someone saying ChatGPT4 is just saying in shorthand, the ChatGPT model based off GPT-4

There is also 100% some regex-style filtering logic which kills the output if it matches certain criteria.

It kicks in for famous book openings (e.g. tale of two cities), but only in English, regardless of what prompt you use to request it - so not part of the model at all but rather a filter on the output.

No idea if it's used for other purposes.

That is the moderation API, which you can actually block from the front-end (with javascript). Look up "ChatGPT DeMod."

Are you sure? I ran it by the moderation endpoint and both the partial output "It was the best of times," and the full thing give me e-06 or lower. This also doesn't give you any notification like other filters do.

I always forget how to see this stuff on the network tab of devtools since it's websocket obfuscated or such (befitting openai), but i have the feeling they're just killing the process directly on their end

Reply on news.ycombinator.com