I know very little about the history of the project, but I feel pretty confident in saying that it’s come a long way since then. I mainly use it these days for complex searches which would be useless in Google.
I must say, it’s really great at understanding nuance and pulling information from a bunch of sources. Maybe ~90+% success rate for that, while Google’s built in AI for searches is hilariously inept about a third of the time. My current takeaway is basically: ‘get to know where its strengths are, while avoiding its wonky areas.’
If they ever get around to changing fundamentally how it works, I’ll give it another try. I don’t like having to second guess everything it says, defeats the entire purpose of trying to use it If I have to verify everything. We made a similar but pared down output generator in college once, that worked in essentially the same way. We fed it a bunch of Freida Khala poems so all it would output was strange fragments of those. But that experience taught me, perhaps too well, how these things work under the hood. I can’t imagine how many guardrails and parameters they have set up just to get it functional.
I doubt OpenAI will fundamentally change how GPT works, but you never know. Possibly the next gen of LLM’s will learn from the failings of this generation and be fundamentally constructed better.
Not saying I don’t realise the cost, either. Better LLM’s will arguably have a much worse outcome for humanity. I’m just talking here about personal ways to get some use out of them. I actually started writing an article about how this might help in brainstorming sketches, but I’m probably not going to ever finish / publish it.
For applications like language translation they’re the best automated tools we have. But its too little to justify how over extended the investment into it is.
Yes. An LLM is just a model trained on a large corpus of text, which DeepL absolutely is. It uses a transformer architecture – which literally could not work without it.
Thanks. And it’s indeed very good from my experience.
Google Translate has the pronunciation database (a separate project) and the ability to translate from whole images, but it’s pretty clearly not as good.
Quite true, which is why I try to double-check whatever the output is, and of course read all the other results, comparing and contrasting. One must do this with news and loads of other data, so it’s at least a developed skill.
Regardless of using AI tools or not, there’s always a variety of information to sift through if you want to research something and get a solid, full spectrum result. If an LLM can save me time in that process, then great. If it can’t, then there’s no point in asking it.
Sometimes it gives you sources that don’t have anything close to what it attributes to them. So you have to always be careful and double check.
It also still likes to make things up when it does not know the answer. Weasel words like „usually“ are a good indicator of that, although sometimes it uses them despite having solid sources and a definitive answer.
To me, GPT is like having a buddy who’s both a knowledge hound and a master bullshitter. Always keep that in mind, and it’ll be much more useful as the tool that it is.
And IME, training it over time greatly helps with such hallucinations and ‘confidently incorrect’ stuff. In a way, it’s not unlike carefully curating a feed over time.
I know very little about the history of the project, but I feel pretty confident in saying that it’s come a long way since then. I mainly use it these days for complex searches which would be useless in Google.
I must say, it’s really great at understanding nuance and pulling information from a bunch of sources. Maybe ~90+% success rate for that, while Google’s built in AI for searches is hilariously inept about a third of the time. My current takeaway is basically: ‘get to know where its strengths are, while avoiding its wonky areas.’
If they ever get around to changing fundamentally how it works, I’ll give it another try. I don’t like having to second guess everything it says, defeats the entire purpose of trying to use it If I have to verify everything. We made a similar but pared down output generator in college once, that worked in essentially the same way. We fed it a bunch of Freida Khala poems so all it would output was strange fragments of those. But that experience taught me, perhaps too well, how these things work under the hood. I can’t imagine how many guardrails and parameters they have set up just to get it functional.
I doubt OpenAI will fundamentally change how GPT works, but you never know. Possibly the next gen of LLM’s will learn from the failings of this generation and be fundamentally constructed better.
Not saying I don’t realise the cost, either. Better LLM’s will arguably have a much worse outcome for humanity. I’m just talking here about personal ways to get some use out of them. I actually started writing an article about how this might help in brainstorming sketches, but I’m probably not going to ever finish / publish it.
For applications like language translation they’re the best automated tools we have. But its too little to justify how over extended the investment into it is.
Would DeepL actually be considered an LLM?
Yes. An LLM is just a model trained on a large corpus of text, which DeepL absolutely is. It uses a transformer architecture – which literally could not work without it.
Thanks. And it’s indeed very good from my experience.
Google Translate has the pronunciation database (a separate project) and the ability to translate from whole images, but it’s pretty clearly not as good.
The wikipedia page says that exactly how they work is proprietary, so who really knows.
That’s when you notice it being hilariously inept. It actually worse than that, but in areas you’re not expert enough to notice.
Quite true, which is why I try to double-check whatever the output is, and of course read all the other results, comparing and contrasting. One must do this with news and loads of other data, so it’s at least a developed skill.
That sounds like a lot of work. I just fucking do the research Myself instead of asking an incompetent machine I don’t trust.
Regardless of using AI tools or not, there’s always a variety of information to sift through if you want to research something and get a solid, full spectrum result. If an LLM can save me time in that process, then great. If it can’t, then there’s no point in asking it.
Sometimes it gives you sources that don’t have anything close to what it attributes to them. So you have to always be careful and double check.
It also still likes to make things up when it does not know the answer. Weasel words like „usually“ are a good indicator of that, although sometimes it uses them despite having solid sources and a definitive answer.
To me, GPT is like having a buddy who’s both a knowledge hound and a master bullshitter. Always keep that in mind, and it’ll be much more useful as the tool that it is.
And IME, training it over time greatly helps with such hallucinations and ‘confidently incorrect’ stuff. In a way, it’s not unlike carefully curating a feed over time.