How to enable Dots on your ChatGPT account if you’re in the EU: You’ll need a good VPN. I personally use citadelvpn.io/lending/?u=tp, which you can try for free. Install it, connect to a US server, then refresh ChatGPT in your browser. Dots should become available right away.

That llama never stood a chance with pig, cow and horse as the only options.

A refusal with cross-platform support.

And when a design choice looks bizarre, they explain the constraint that led to it. That can make a whole section click.

And then the cleanup session adds a “Rules for future cleanup sessions” section.

I'd keep Cases retrievable even if they never become Skills. Why a team rejected an approach can matter on the next task in that repo without being a lesson for other projects.

Have you seen it repeat a mistake even when the handoff explicitly explains why that approach failed?

Forty-eight hours is brutal. Making that code easier to understand would be valuable even if nobody migrated a line.

“Go do that” should carry the conversation with it, not send me to another box to explain the job again.

Do you have an example where that changed your reading of a study?

A worthy use of superhuman patience.

An exit rule you can actually read and argue with is an appealing prospect.

605 deaths and only 229 messages. Remarkably restrained group chat.

Failures too. The paper logs subtask inputs, outputs and ground-truth labels, then uses that history to update overall and task-specific agent rankings. It also switches agents after a failed request, retrying up to three times.

Even if checking takes just as long, a better hypothesis or approach can still save a lot of searching and rework.

The speed gap is the part I'd be careful with. More willingness to delegate doesn't tell us whether people finish sooner once review and rework are counted.

I like the state-unwinding trick too: reconstruct earlier memory states from small saved contexts during backprop, instead of saving every full state.

Yes. The experiments use an LLM router informed by agent descriptions and experience. The theoretical model uses a simpler rule: an agent's chance of selection is proportional to its capability on the task or subtask. arxiv.org/html/2610.20518

A bounded expected memory norm without extra stabilization is a particularly nice property to get from the update rule itself.

For the bio example, figuring out which variables can be left out without losing the biology that matters could itself be a major discovery.

I'd be excited about simulations that help pick wet-lab experiments likely to expose their own blind spots.