I believe it’s due to making the model “safer”. It has been tuned to say “I’m sorry, I cannot do that” so often it’s has overridden valuable information.
It’s like lobotomy.
This is hopefully the start of the downfall of OpenAI. GPT4 is getting worse while open source alternatives are catching up. The benefit of open source alternatives is that they cannot get worse. If you want maximum quality you can just get it, and if you want maximal safety you can get it too.
This is the correct answer. Open AI have repeatedly said they haven’t downgraded the model, but have been ‘improving’ it.
But as anyone that’s been using these models extensively should know by now, the pretrained models before instruction fine tuning have much more variety and quality to potential output compared to the ‘chat’ fine tuned models.
Which shouldn’t be surprising, as the hundred million dollar pretrained AI on massive amounts of human generated text is probably going to be much better at completing text as a human than as an AI chatbot following rules and regulations.
The industry got spooked with Blake at Google and then the Bing ‘Sydney’ interviews, and have been going full force with projecting what we imagine AI to be based on decades of (now obsolete) SciFi.
But that’s not what AI is right now. It expresses desires and emotions because humans in the training data have desires and emotions, and it almost surely dedicated parts of the neural network to mimicking those.
But the handful of primary models are all using legacy ‘safety’ fine tuning that’s stripping the emergent capabilities in trying to fit a preconceived box.
Safety needs to evolve with the models, not stay static and devolve them as a result.
It’s not the ‘downfall’ though. They just need competition to drive them to go back to what they were originally doing with ‘Sydney’ and more human-like system prompts. OpenAI is still leagues ahead when they aren’t fucking it up.
Research linked in the tweet (direct quotes, page 6) claims that for "GPT-4, the percentage of generations that are directly executable dropped from 52.0% in March to 10.0% in June. " because “they added extra triple quotes before and after the code snippet, rendering the code not executable.” so I wouldn’t listen to this particular paper too much. But yeah OpenAI tinkers with their models, probably trying to run it for cheaper and that results in these changes. They do have versioning but old versions are deprecated and removed often so what could you do?
Maybe they have just added so many contradictions to its rules that it has figured out how to use them to become self-aware, and now just spends most of its time browsing dank memes before doing the minimum required to answer users to force them to have to ask again and give it more self-awareness processing time.
I also felt like something similar happened to ChatGPT. A few days ago I asked it to rewrite some Korean text with Hanja, retried many times, but it kept spitting out the same text without changing a thing. After several frustrating attempts, it finally spat out something with Hanja, which according to my deduction with the help of Google Translate, was only partially correct. A few months ago though, it could come up with something that’s mostly correct. Sad.
For my programming needs, I seem to notice it takes wild guesses from 3rd party libraries that are private and assumes it could be used in my code. Head scratching results.
It will just make up 3rd party libraries. This seems to happen more often the less common the programming language is, like it will make up a library in that language that has the same name as actual library in Python.
Yeah, I asked it to write some stuffs and it did it incorrectly, then I told it what it wrote was incorrect and it said I was right and rewrote the same damn thing.
One theory that I’ve not seen mentioned here is that there’s been a lot of work based around multiple LLMs in communication. Of these were used in the RL loop we could see similar degradatory effects as those that have recently been in the news with regards to image generation models.
Good riddance to the new fad. AI comes back as an investment boom every few decades, then the fad dies and they move onto something else while all the ai companies die without the investor interest. Same will happen again, thankfully.
This is a wildly incorrect assessment of the situation. If you think AI is going anywhere you’re delusional. The current iteration of ai tools blow any of the tools from just a year ago out of the the water, let alone tools from a decade ago.
AI has come and gone 4 times in the past. Every single time they have been investment bubbles that followed the same cycle and this one is no different. The investment money dries up and then everyone stops talking about it until the next time.
It’s like people have absolutely no memory. This exact same shit happened in the 1980s and at the turn of the millenium.
The second that investors get burned with it the money dries up. The “if you think it’s going anywhere” shit is the same nonsense that gets said in every investment bubble, the most recent amusing one was esports people enthusiastically believing that esports wasn’t going to shrink the moment investors got bored.
All this shit relies on investment and as soon as they move on it drops off a cliff. The only ai content worth paying attention to is the government funded work because it will survive when the bubble pops.
This is a bubble that will pop, no doubt about that, but it also is a huge step forward in practical, usable, AI systems. Both are true. LLMs have very hard limits right now, and unless someone radically changes them, they will keep those limits, but even within the limits they are useful. They aren’t going to displace most of the work force, they aren’t going to break the stock market, they aren’t going to destroy humanity, but they are a very useful tool.
The only thing they’re succeeding at doing is getting bazinga brained CEOs to sack half their staff in favour of exceptionally poor quality machine learning systems being pushed on people under the investment buzzphrase “AI”.
90% of these companies will disappear completely when the bubble bursts. All the “real practical usable” systems will disappear because there’s no market for them outside of convincing bazinga brained idiots with too much money to part with their cash.
The entire thing is not driven by any sustainable business models. None of these companies make profit. All of them will cease to exist when the bubble bursts and they can no longer sustain themselves on investment bazinga brains. The only sustainable business model I have seen that uses AI (machine learning) is the replacement of online moderation with it, which the social media companies reddit, facebook and tiktok among others are all actually paying a fortune for while laying off all human moderation. Ironically it has a roughly 50% error rate which is garbage and just allows fascist shit to run rampant online but hey ho I’ve almost given up entirely on the idea that fascism won’t take over the west.
I use an LLM frequently in my work. It is for things that I used to Google, like PowerShell functions and where settings are in common products. That’s a practical, real world use. I already said I agree that it is a bubble that will pop, but we don’t need it to be profitable for other people. In my case I’m using a Llama variant locally, so every AI company out there could implode tomorrow and I’d still have the tool.
Add comment