I attended the developer conference, but I am not a programmer. What differences do we know of with GPT4 that would benefit someone like me? I use AI to assist me in my copywriting process, with a background of thirty years of experience in the field.
The most immediate benefits, Frank are two-fold.
Firstly, as a freshly developed LLM, and one made in a hurry at that, it will have fresher data in the corpus of Pre-training data. It’s data at time of release might be mere months old at the freshest. [ * See footnote here ]
Secondly, the LLM was trained over a larger range of data, with many more connections and data points, giving it a much finer level of detail in pattern detection and recognition, which is known to massively improve accuracy.
Those two benefits are unequivocal, and certain.
However, there is, in addition to that, the claims that it will be multi-modal, able to create output in formats beyond just text, where they have specifically mentioned video (though not length limits or resolutions).
I regard this as much less certain, of course as there has been no level of real detail given. e.g. if that meant video with an audio component, then creating audio might also be a feature - an AI that could speak back via recorded voice? Maybe. But ‘video’ might be more like the tiny looping meme videos we see in social media, or mere animations more akin to GIFs than to photo-realism. So far, its just the promises of folks with reason to create hype and excitement, and perhaps not so much incentive to be realistic.
So, count on the first 2 points only, but hope for more.
[ * Added Footnote ]
I have to revise this somewhat as there is some element of doubt over whether the training data is actually any different at all - see OpenAI GPT-4 Announcement, Waiting list and Live Stream in 2 hours - #8 by aiprm-christophc
This could be a glitch in what the AI is reporting (it might have been a hard-coded warning for certain circumstances they haven’t updated yet), but it could also be correct, and they simply used the exact same data, but gave the LLM more time to build data-points and connections to understand the same text in greater detail. If so, only the second point is definite.
I’ve watched their video announcement and Greg also confirmed training data cut-off Sept 2021
Aye, just got to that myself, and yes, very close to the beginning he states that they started work on GPT4 some 2 years ago, and he later repeats that it is the same pre-training dataset.
I start to see exactly why the head guy was warning that this wouldn’t live up to a lot of the hype - this is effectively just the exact same dataset, and the exact same tech as GPT3.5 but simply given more time to think, and more memory capacity to use.