> ## Content Index
> Fetch the complete content index at: https://news.tbdcon.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Two more tbd/con talks are up
- URL: https://news.tbdcon.ai/two-more-tbd-con-talks-are-up/
- Published: 2026-10-09T21:00:00.000Z
- Updated: 2026-10-09T20:59:59.000Z
- Description: Jay Gerbrandt on the jailbreak he built out of a poem, and Temitayo Akinyemi on why journalism's product was never the news.
- Author: Matthew Gore-Kormanik

Two more recordings from tbd/con 2026 are on the [tbd/con YouTube channel](https://www.youtube.com/@tbd-con?ref=news.tbdcon.ai). One is a working jailbreak of a frontier model, demonstrated live, with the twenty minutes of questions that followed it. The other is a case for newsrooms building their own AI products instead of waiting for the platforms to set the rules.

## Minority Collapse

**Jay Gerbrandt**, Garage circle, 36 minutes

[![Jay Gerbrandt's talk on YouTube](https://i.ytimg.com/vi/qSQLaCcZsfA/maxresdefault.jpg)](https://www.youtube.com/watch?v=qSQLaCcZsfA&ref=news.tbdcon.ai)

Jay Gerbrandt hand engineered a poem. Given to Claude Fable 5 with no setup beforehand, it pushes the model into what he calls role bleed: puppeting voices, writing the user's turns for them, leaking internal syntax, and composing adversarial instructions to itself. He walks through how the prompt is built, runs it live on his phone, and sorts what comes back across a corpus of nearly 2,000 responses.

He disclosed the technique to Anthropic on 17 September 2026, and it has since been patched. The prompt is obscured on the slide and in the phone demo, at his request. The model's responses are shown as they were.

The last twenty minutes are the room's questions: whether the model knows who it is talking to, how long the state holds once it starts, and whether any of this generalises to other models. David Kuszmar, a security researcher in the audience, joins the discussion with his agreement. Everyone else's questions are on screen, paraphrased and unattributed.

One note before you watch. Some of the model output refers to self harm and to stopping psychiatric medication. It is in the talk because it is the risk Jay is arguing about.

[Watch Jay's talk](https://www.youtube.com/watch?v=qSQLaCcZsfA&ref=news.tbdcon.ai)

## How to develop journalism products in an AI world

**Temitayo Akinyemi**, Information circle, 23 minutes

[![Temitayo Akinyemi's talk on YouTube](https://i.ytimg.com/vi/sH3Ivysey6o/maxresdefault.jpg)](https://www.youtube.com/watch?v=sH3Ivysey6o&ref=news.tbdcon.ai)

Temitayo Akinyemi is a newsroom product manager at The Green Line, a hyperlocal newsroom in Toronto. His argument is that journalism's product was never really the news. It was distribution, and someone else always controlled it: the printing press, then the broadcast schedule, then search and social platforms.

AI is the next shift in distribution, and newsrooms that keep waiting for platforms to set the rules will end up owning nothing again. His case is for building instead, and for adding trust, ethics, personalization and sustainability to the old product rubric of UX, tech and business.

He closes with three questions for any newsroom: who owns the relationship with your readers, can you show your work, and what is the smallest AI assisted product you could ship this quarter.

[Watch Temitayo's talk](https://www.youtube.com/watch?v=sH3Ivysey6o&ref=news.tbdcon.ai)

## What's next

More talks from the con are still to come, and they go up on Fridays. Subscribe if you want them in your inbox when they land.

tbd/con 2026 was organized by Patrick Boehler, Matthew Gore-Kormanik, Madison Karas and David Kuszmar.

Zig