Rumors about OpenAI having a model more powerful then GPT4

I just received this news in my email today, and god it is not even 8AM yet.

Apparently two reddit user’s claim they have access to two of OpenAI internal models called Gobi and Arrakis and they go way beyond what anyone ever dreamed. This summed with some cryptic tweets of OpenAI personnel got the rumor mill running.

Now this type of news related to any other AI company would not normally make me very interested, at least not right now, the seeing to believing is strong in me(much more after the trauma of Cyberpunk 2077) but in this case it is OpenAI. Is the company that started this crazy train we are on lately, and they have proven more than one capable of delivering. so they got credit.

Does this show I have not really learned my lesson with Cyberpunk 2007? Probably.

Do I believe they actually achieve AGI? I am kinda more skeptical about that one, probably some reality augmented propaganda, but I think it is very possible they achieved a model with more power then GPT4 and this only is already a great deal.

So let’s study this rumors to see what we can possibly wait for here on now:

• OpenAI is in possession of a powerful model code-named “Arrakis”.
• It is everything to everything (all modalities to all modalities).
• Arrakis exceeds GPT-4 capabilities and performs very close to human experts in many different fields.
• Hallucination rates are much lower than GPT-4.
• Half of it’s training data was synthetic.
• Inference cost is around the same as GPT-4 because of the conditional MoE/Multimodal weights offloading.
• It is also a very very good autonomous agent.
• Its release date is scheduled for 2024.

Now i guess we have to wait and see, but 2024 does promise to be an AI year even more insane then 2023 and the AI development train doesn’t seen to be slowing down any time soon.