Labelbox

Today we’re announcing our Series E.

We’ve raised $100M+ led by Infinitum Partners. We believe every company will soon run on agents that learn from their own work.

- For frontier labs, Horizon provides RL environments that teach frontier models to do real work.
- For enterprises, Recursion brings managed agents into the apps your teams already use, improving with every run.
labelbox.com/blog/labelbox-…

Manu (@manuaero) joined Jason Calacanis on This Week in AI alongside @sarahookr (Adaption Labs) and @spirosx (Resolve AI) to break down the rapidly shifting AI landscape.

From the Token Price Wars and Demis Hassabis’ recent call for an AI oversight board, to the Uber vs. Waymo robotaxi battle, and a reality check on whether orbital data centers are science fiction or closer than we think (Manu’s aerospace engineering background came in handy) 🚀

A few key insights from Manu:
- Proprietary intelligence is the new edge: As base model costs decline, the real winners will be defined by what they own: proprietary data, customized RL environments, deep domain expertise, and continuous feedback loops.

- From models to moats: The narrative is shifting. The biggest opportunity is not building another general frontier model, but architecting intelligence systems around an enterprise’s unique workflows and expertise.

- Owning the enterprise intelligence stack: General purpose AI will continue to raise the baseline for everyday productivity, but true industry differentiation, whether in banking, pharmaceuticals, or manufacturing and beyond, will come from owning the proprietary knowledge layer that compounds over time.

Do AI models become less honest when they self-report their own misbehavior?

Our Applied Research team studied a key question in chain-of-thought (CoT) monitorability: do models faithfully report their own misbehavior when they are asked to?

We introduce Monitorability Disposition, a meta-level property capturing how willing models are to disclose their misbehavior.

We found that models reported only ~16% of expected misbehavior. Even with incentives to increase disclosure, no model self-reported high-severity violations. When given a choice, models consistently selected the least strict monitor.

In other words: less oversight is the default choice.

💡 Key takeaways:
- We can measure it: Monitorability disposition is a new way to evaluate how transparent models are.
- Not all models behave the same: Even equally capable models vary in how willing they are to admit when they've gone off track.
- It can be strengthened: When a model’s disposition is strong enough, it stays monitorable.

1/ Today, we’re introducing Recursion: the RL platform for building, evaluating, and deploying specialist agents.

The next phase of AI won’t just be about smarter models. It will be about systems that learn from the unique expertise of every organization. 🧵

Where do models change their minds?

Natural Language Autoencoders (NLAs) offer a promising way to translate a model’s internal representations into natural language. But the harder question is: where do meaningful decisions actually happen?

We tested a workflow for finding these “branchpoints” by tracing activations across layers and comparing when models arrive at the right answer versus when they rely on shortcuts.

The surprising finding: early signals can look convincing, but stronger controls showed that many apparent decision points may come from confounding factors rather than the process we’re trying to uncover.

The takeaway: understanding model behavior takes more than spotting activation differences. It requires careful validation to separate real insights from artifacts. Check out the full post for the experiments, controls, and lessons learned.