Excellent, excellent article. I have a treasure trove of bookmarks and opinions similar to yours, and it's good to see someone take the time to share it in such a well-written manner. Very shareable.
I tried to strike a balance between finding the most relevant reading and not overwhelming people with too many links. There's enough information overload out there as it is.
If you find something in your bookmarks that's worth adding, feel free to share!
I'm pretty excited for Gemini. I think it might be the first model to actually surpass GPT4, but I'm also pretty confident GPT5 would be right on the heels.
On paper, Gemini Ultra beats GPT-4 on almost every metric (mostly marginally). But we won't know how it fares in the real world until it's out in early 2024.
Yeah - I picked up on that as well. Basically, GPT4 works great if you remove any of the applications, so at least you can always use the stripped-down (super powerful) LLM itself, but with Gemini, it's gotta work out of the box.
I think the main message from Alberto's article about it is that while GPT 4's multimodality is enabled by additional components that relay the visual information to the core LLM, Google's Gemini is actually trained on many different modalities, which means it has a much better innate understanding of many types of inputs. So this should be a big plus in Gemini's favor. The video demonstration looked really impressive, but again - we'll have to see how it actually does when people start engaging with it!
Ah, I missed that - the idea that those things are integrated during training is very promising! But I'm still kind of in awe with GPT4, still getting used to performing magic every day.
Yeah until Gemini Ultra actually makes its way into products, GPT-4 remains king of the hill, especially since it's already out there being integrated into people's routines and applications.
Excellent, excellent article. I have a treasure trove of bookmarks and opinions similar to yours, and it's good to see someone take the time to share it in such a well-written manner. Very shareable.
Really happy to hear you found it useful!
I tried to strike a balance between finding the most relevant reading and not overwhelming people with too many links. There's enough information overload out there as it is.
If you find something in your bookmarks that's worth adding, feel free to share!
I'm pretty excited for Gemini. I think it might be the first model to actually surpass GPT4, but I'm also pretty confident GPT5 would be right on the heels.
On paper, Gemini Ultra beats GPT-4 on almost every metric (mostly marginally). But we won't know how it fares in the real world until it's out in early 2024.
Although those benchmarks are less interesting than the fact that it's *natively* multimodal as Alberto explains: https://thealgorithmicbridge.substack.com/p/the-best-ai-model-in-the-world-google
Yeah - I picked up on that as well. Basically, GPT4 works great if you remove any of the applications, so at least you can always use the stripped-down (super powerful) LLM itself, but with Gemini, it's gotta work out of the box.
I think the main message from Alberto's article about it is that while GPT 4's multimodality is enabled by additional components that relay the visual information to the core LLM, Google's Gemini is actually trained on many different modalities, which means it has a much better innate understanding of many types of inputs. So this should be a big plus in Gemini's favor. The video demonstration looked really impressive, but again - we'll have to see how it actually does when people start engaging with it!
Ah, I missed that - the idea that those things are integrated during training is very promising! But I'm still kind of in awe with GPT4, still getting used to performing magic every day.
Yeah until Gemini Ultra actually makes its way into products, GPT-4 remains king of the hill, especially since it's already out there being integrated into people's routines and applications.
Thanks for amplifying and adding some commentary to Alberto's post, too! He writes quality stuff. I miss a lot of it in the deluge of info.