New Gemini, Privacy Attack, & More

TLDR;

In today’s Premium news rundown, we reflect on Sam Altman’s latest mysterious comment, analyze today’s release of a new Gemini model, and what Morgan Stanley’s AI success tells us about the future of software.

We also discuss Amazon’s ambitious robotics plans, OpenAI’s privacy concerns, and Meta’s substantial power contract, among other news.

Enjoy!

AGI

Sam Altman’s Mysterious Comment

This short clip on a recent Sam Altman interview is very revealing.

In the discussion, he describes how he envisions an AGI as a tiny model with reasoning superpowers and with a trillion context windows, infinite tool access and not necessarily a lot of knowledge on anything… which is quite literally the opposite of what we have right now.

Interestingly, he completely rejects the status quo using one of our favorite analogies we have in this newsletter for modern frontier AI models, defining them as ‘databases,’ which is surprisingly accurate coming from someone who would have the incentive to reject that analogy.

But why is this revealing?

Well, it’s obviously an educated answer based on recent research inside OpenAI. Sam is not a researcher, so this has to be influenced by what he has seen recently. What’s more, he does clarify they aren’t shipping anything like that yet… which kind of hints they will at some point.

But what if Sam is hinting at an even more revolutionary idea?

Get your tin hats on because I’m just speculating here, but he might be proposing AGI as a model trained from scratch using Reinforcement Learning rather than Maximum Likelihood Estimation (MLE), the method more informally known as imitation learning.

Training with RL essentially means placing an AI in an environment, allowing it to make decisions, and using a reward mechanism to reinforce good decisions and penalize bad ones. Instead of forcing it to imitate certain data, it’s trial-and-error at scale.

By rejecting the idea of AIs as ‘know-it-all’ databases that are then post-trained using RL to teach them to reason, OpenAI might be considering eliminating the pre-training phase and instead training the model directly on an RL policy from scratch.

If it helps, this approach is similar to what Google did with AlphaGo or AlphaZero; both are RL policies trained from scratch. However, the challenge here is to find a way to train these policies to generalize, not just very good at one thing (playing board games).

The advantage of such a model would be evident as these are much smaller models than the beasts we currently call reasoning models, one that could be stored in your computer, your smartphone, or even in OpenAI’s future wearable.

ROBOTICS

Amazon to Test Robots for Package Delivery

Amazon is developing software for humanoid robots designed to replace human delivery workers.

The company is constructing a “humanoid park” at one of its San Francisco offices, an indoor obstacle course intended for testing these robots. Initially, Amazon plans to use robotic hardware from third-party firms while focusing its efforts on creating the software.

Though Reuters could not…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.