- Emergent Behavior
- Posts
- Happy AGI Year!
Greetings, sorry for the longish hiatus from writing.
The main thing you should note is that we hit AGI in December 2024. OpenAI’s o3 model, which was announced on 20th December 2024, on the last of the 12 days of OpenAI, is a recursively self improving artificial intelligence model.
Some interesting comments leading to the release:
what should we give it for a birthday present...
— Sam Altman (@sama)
10:11 PM • Oct 21, 2024
From Vinod Khosla, one of the VCs that was an early investor in OpenAI
A baby AGI that learns as it grows up
— Vinod Khosla (@vkhosla)
10:49 PM • Oct 21, 2024
The o3 announcement livestream:
the reaction from Nick Cammarata, former OpenAI interpretability researcher:
the whole thing is so thielian. if you’re going to take on a giant market doing probably illegal stuff call yourself something as light and bouba as possible, like airbnb, lyft
if you’re going to announce agi do it during a light and happy 12 days of christmas short demo
— Nick (@nickcammarata)
7:48 PM • Dec 20, 2024
The ARC AGI (an IQ test for AI which was previously impenetrable) announcement:
New verified ARC-AGI-Pub SoTA!
@OpenAI o3 has scored a breakthrough 75.7% on the ARC-AGI Semi-Private Evaluation.
And a high-compute o3 configuration (not eligible for ARC-AGI-Pub) scored 87.5% on the Semi-Private Eval.
1/4
— ARC Prize (@arcprize)
6:07 PM • Dec 20, 2024
So where do we stand? o3 is a chain of thought improvement on o1, released in November. We are now in a regime where we can expect increases in capability every 3 months at least, with no additional build in datacenters. Every latest o model writes improved chains of thought for the next version of the model. Massive increases in pretraining compute are now replaced with staggered increases in inference time compute which are usage dependent.
We are missing critical data on reliability, however there are hints that the problem has been solved:
And former research VP, Bob McGrew.
The important news from this morning is that @OpenAI can use additional inference-time compute to improve answer quality in production. Today, that improvement appears mostly as better reliability - but reliability matters when you are doing real work.
Given the cycle time for… x.com/i/web/status/1…
— Bob McGrew (@bobmcgrewai)
6:53 PM • Dec 5, 2024
There is the issue of cost and latency, to which the livestream had an extremely interesting response, showing that a distilled o3-mini model was able to beat a full o1 model given a moderate compute budget.
o3-mini (medium) matches o1
So in summary, we have an o3 model that:
a) improves every couple of months using reinforcement learning from chains of thought prepared by the previous version of itself
b) Is reliable enough not be tricked into giving up a password or credit card to an external party
c) has a mini version (o3-mini) which is fast and cheap and is as capable as the slow and expensive version two months prior
Finally, when is all this coming to production?
End of January according to Sam.
thank you to the external safety researchers who tested o3-mini.
we have now finalized a version and are beginning the release process; planning to ship in ~a couple of weeks.
also, we heard the feedback: will launch api and chatgpt at the same time!
(it's very good.)
— Sam Altman (@sama)
8:47 PM • Jan 17, 2025
So there you have it. There will be many debates about which point AGI was. On the geological timescale even this decade will look like a single point. And we are right in the middle of it.
Reply