Genie 3, the Debt problem, & More

TLDR;

Welcome back! What a week of announcements coming from all sides. From Genie 3, the most impressive AI model probably ever, to OpenAI’s imminent GPT-5 release, the industry is delivering as much hype as always.

However, we will also look at its tremendously unstable foundation, which may raise several alarms in your head.

Enjoy!

Note to reader: Yesterday was a crazy day, with OpenAI’s open-source models, Anthropic’s Claude Opus 4.1, and Genie 3. While I do cover Genie 3 here, which is to me the most important release by far, I have yet to test for myself and also see clearer reactions on the other two. Thus, I’ll talk about them on Thursday.

FRONTIER RESEARCH

OpenAI’s Universal Verifier

The Information has published an article discussing OpenAI’s latest breakthrough, a “universal verifier.” They cited it as one of the secrets behind GPT-5.

If you’re a regular, this isn’t something new at all; it’s something we have discussed countless times.

In Reinforcement Learning, the model enters a ‘reward-driven’ mode in which it generates actions that a verifier scores—usually external, otherwise it is self-verification, which is risky.

It’s like the game ‘hot and cold’; the verifier guides the model into the solution using this scoring system. Over time, the model learns to take reasonable actions and avoid bad ones.

The key bottleneck here is, of course, the verifier, because it provides the crucial reward feedback that allows the model to learn.

Currently, most verifiers are automatic, meaning they provide an automatic signal about the quality of the model’s action. This means we are only capable of training these models in areas like maths or coding, where a reward can be provided at all times (2+2 always equals 4, so if the model says otherwise, it’s easy to see if it’s wrong).

The crux of the issue is that in areas where ‘being correct or wrong’ isn’t easily verifiable, like creative writing, we don’t quite know what verifiers to use. And here is where OpenAI’s alleged breakthrough comes in.

They seem to have found a way to train a ‘universal verifier’, one that can confidently score good and bad actions in every single domain. From The Information’s article, it seems like another model, but one that has been trained for this precise purpose.

Little is known about this model, but this immediately raises questions. Using other models as verifiers is a tale as old as time. And the main limitation has always been the same:

We need this verifier to be smarter than the model itself, so why not use the verifier directly?

In turn, this means that your entire system is only as smart as your verifier; the verifier becomes the primary “intelligence bottleneck.”

If the verifier is that smart, it must be huge too, so your entire system becomes extremely expensive to run.

The only thing I can think of to make sense of all of this is that OpenAI has somehow managed to find a way to train a small…

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