LessWrong posts by zvi podcast

“GPT-6.2 Astra and Sol: The System Card” by Zvi

0:00
39:42
Recuar 15 segundos
Avançar 15 segundos

We have to stop meeting like this. Alas, we will keep meeting like this.

It is time, once again, to do the reading.

As per usual, I will focus on things to the extent they are new and important. There are a lot of things that need to be in the system card in case something goes wrong, but that look exactly the way you would expect. I did my best to skip those.

I continue to be on vacation. I am aware that via Claude there is now a claimed Lean proof of the Reimann Hypothesis and will cover that next week.

Astra 6.2 Self-Portrait

Table of Contents

  1. Safety Overview (1).
  2. Internal Deployment (1.2).
  3. Preparedness Assessment (2).
  4. Biological Threats (2.1.1).
  5. Cyber Threats (2.1.2).
  6. AI Self-Improvement Threats (2.1.3).
  7. OpenAI Epoch Capabilities Index (OECI).
  8. Safeguards (2.2).
  9. AI Self-Improvement Threat Modeling (2.2.1.3).
  10. Safety Training and Evaluation (2.2.2).
  11. Jailbreak Robustness (2.2.3 and 2.2.6).
  12. Realtime Model Safeguards (2.2.4).
  13. Distillation Safeguards (2.2.5).
  14. Security Controls (2.2.9).
  15. Model Safety and Behavior (3).
  16. Alignment and Monitorability (4).
  17. Don’t Help Peer (4.1.5).
  18. Verbalized Metagaming and [...]

---

Outline:

(00:58) Safety Overview (1)

(03:47) Internal Deployment (1.2)

(04:02) Preparedness Assessment (2)

(05:06) Biological Threats (2.1.1)

(07:06) Cyber Threats (2.1.2)

(08:05) AI Self-Improvement Threats (2.1.3)

(10:41) OpenAI Epoch Capabilities Index (OECI)

(11:46) Safeguards (2.2)

(13:06) AI Self-Improvement Threat Modeling (2.2.1.3)

(16:41) Safety Training and Evaluation (2.2.2)

(17:36) Jailbreak Robustness (2.2.3 and 2.2.6)

(20:00) Realtime Model Safeguards (2.2.4)

(20:15) Distillation Safeguards (2.2.5)

(20:55) Security Controls (2.2.9)

(23:57) Model Safety and Behavior (3)

(24:38) Alignment and Monitorability (4)

(27:41) Don't Help Peer (4.1.5)

(28:11) Verbalized Metagaming and Oversight Gaming (4.1.6)

(30:31) Misaligned! (As Found By UK AISI) (4.1.7.1)

(34:16) Astra is Hard to Monitor (4.2)

---

First published:
November 21st, 2026

Source:
https://www.lesswrong.com/posts/7Skw6caJxPrBDetCC/gpt-6-2-astra-and-sol-the-system-card

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Mais episódios de "LessWrong posts by zvi"