GPT-4

Overview of internet sources I found helpful.

GPT-4 got released yesterday. The hype around the model was extensive over the last months, and Sam Altman, CEO of OpenAI, already stated (somewhere) that GPT-4 could only live up to the hype if it were an actual AGI. Although the hype was big, Altman's statement might have been an overestimation. It feels like GPT-4 matches the hype quite well.

Although there is no AGI, the results are still quite impressive. There is already a lot of coverage out there. So instead of providing a bad summary of the model on my own, I will simply link to two videos that explain the model and its capabilities.

The first video you should watch to get a better feeling for GPT-4 is yesterday's recording of the OpenAI live stream. The video demonstrates the differences between ChatGPT and GPT-4. It includes an impressive application of multimodality by showing how to turn a picture of some scribbled notes on a web page design into an actual web page.

While the Developer Lifestream did show some interesting examples, it provided little technical overview or information about benchmark performance. This information was provided separately in a blog post and a technical report. I recommend watching the following video by former DeepMind employee Aleksa Gordic:

I am excited and afraid simultaneously, like I have been over the last few months. AI is progressing fast, and it is already hard to predict where we will be in a year or even five years. But our economy and knowledge work, in particular, is likely to change completely. Ezra Klein wrote an article matching my feeling pretty well:

I cannot emphasize this enough: We do not understand these systems, and it’s not clear we even can. I don’t mean that we cannot offer a high-level account of the basic functions: These are typically probabilistic algorithms trained on digital information that make predictions about the next word in a sentence, or an image in a sequence, or some other relationship between abstractions that it can statistically model. But zoom into specifics and the picture dissolves into computational static.

In this context, it is ironic and worrisome that Microsoft is apparently cutting back its safety concerns.