- Signal to Noise
- Posts
- The 3rd Era of AI Language Models
The 3rd Era of AI Language Models
A 3rd Paradigm for AI models
o1 is not just better, it is different
What Happened? The o1-family was trained in a different way to all models previously. It was fine-tuned with only answered that had been confirmed, or verified, to be correct, and given the computational ‘test-time’ to perform serial calculations first. This is new.
Yes, a log scale, but inference compute is 10x-ing every year it seems
First, models predicted the most likely next word. Think 2018-2021, for transformer-based language models.
Second, they were rewarded for words that were helpful, harmless, and honest. Think RLHF or RLAIF, 2022-2023.
Now, with the o1 family, they are being rewarded for being objectively correct. Think 2024-???
So What? The results are impressive, and are in no ways close to a peak. This does not break the fundamental paradigm of relying on training data to predict an output, but it is a new training method (albeit one I foreshadowed on the channel in 2023!). And there is no reason to believe that this paradigm cannot be extended to videos and mother modalities, leading to a step-change in physics-adherent realism (yes, leaving Sora far behind). If LLMs are an off-ramp to AGI, that ramp is taking us to awfully interesting places.
Does It Change Everything? Rating = ⚄
But video realism is not waiting for o1
What Happened? AI image and video generators keep getting better, at a stunning pace. Yes, deepfake videos too. Kling has an epic motion brush. Runway has video to video. And Minimax is not too shabby either.
AI-generated frame of a video, generated by facecam.ai
You already could, but it is getting ever easier to livestream video using a face that’s not your own. This is facecam.ai. Might wanna warn anyone vulnerable in your life not to trust even live videos now, from untrusted sources.
Kling allows you to choose the motion of a video, which allows for fascinating flexibility in generation. Point the brush of a video frame in a direction, and the cat jumps there.
There is no shortage of rivals either, even if Sora never got released. Minimax from Hailuo, just got a major upgrade.
So What? Like the famous but likely apocryphal boiling frog, society is slowly being confronted with fake images, video and now livestreams. We are running a giant experiment on the effects on trust and social cohesiveness. I am not optimistic. But I do enjoy making AI videos as much as the next person.
Does It Change Everything? Rating = ⚂
Insider Essentials has come to an end, because I want to give you more. Now, for that same price, you can access all of the AI Insider videos, in real-time. See below.
To support hype-free journalism, and to get a full suite of exclusive AI Explained videos, podcasts, and a Discord community of hundreds of truly top-flight professionals w/ networking, I would love to invite you to our $7.5/month Patreon.