AGI might be here, better get these stories out while I still can!

OpenAI:

OpenAI started off the week with a whole lot of nothing, for the first 4 days of the week all we got from their "releasing something everyday" event was SearchGPT being public for everyone, some closed access tools for fine-tuning models, 1-800-CHATGPT in case you wanted a worse way to talk to it, and plans to integrate with more apps plus some cleanup to the mac app.

But suddenly, and without warning, OpenAI announced o3, the successor to newly released o1. (Discussed in the 12/6 news drop) o3 scored 87.5 on the ARC Prize's benchmark, which considers anything above 85% to be in AGI territory. AGI is artificial general intelligence, this would be a model capable of doing anything a human could, but better. The benchmark is designed so that any human could score 95%+, but is very challenging for most AI models.

Google:

Gemini 2.0 flash thinking is here! It is designed to mimic the "thinking" technology OpenAI used very successfully in o1 and o3. Performs quite a bit better with reasoning tasks. Imagen3 and Veo2 were shown off: Imagen3 is a state of the art model that's free to use here (https://labs.google/fx/tools/image-fx). Veo2 is a new text to video model, like Sora but seems to be much more stable (doesn't devolve into useless garbage).

IBM and xAI:

New AI models from IBM and xAI are now available. The models are: IBM Granite3.1 8b and Grok 2.0.

Once again, everyone else was very quiet this week, Anthropic will likely have to hit back sometime early January to stay relevant in this AI arms race.

Happy holidays!