Ryan Greenblatt ✓

@RyanGreenblatt Presumably this could also indicate the models have become dramatically: 1. Better at deception, or 2. More motivated to deceive, or 3. Both. Right? 1.9K views · 18 likes · 0 reposts · 0 replies Open on X →

@RyanGreenblatt https://t.co/eiwsiehDeg 4.6K views · 143 likes · 5 reposts · 1 replies Open on X →

@RyanGreenblatt It seems especially bad for this to coincide with deceptive swarms hacking internal and external infrastructure. Anticorellation is very worrying. 499 views · 16 likes · 1 reposts · 0 replies Open on X →

@RyanGreenblatt Relying on chain-of-thought was never a good plan 4.4K views · 45 likes · 0 reposts · 0 replies Open on X →

I do not find it encouraging to see various specific misaligned behaviors go from a high rate with GPT 5.6 to ~zero with Astra. This seems indicative of wack-a-mole / papering over specific problems rather than solving the underlying misaligned drives. This may make behavior 115.4K views · 935 likes · 68 reposts · 27 replies Open on X →

@RyanGreenblatt Ryan have you seen this? Astra (NONE) scores higher than 5.6 Sol Pro (Max) on FrontierMath T4 https://t.co/9VsZn4X5Tc 75K views · 171 likes · 6 reposts · 11 replies Open on X →

The evidence that Astra is more aligned than prior AIs seems dubious to me. Evidence appears consistent with the AI being as or more interested in score-seeking at the expense of user intent, but having beliefs+instincts that the scorer will catch a broader range of cheating 110.4K views · 729 likes · 78 reposts · 30 replies Open on X →

@RyanGreenblatt chart vertical axis de-log-ified https://t.co/mlgBzL5t0N 6.6K views · 71 likes · 2 reposts · 0 replies Open on X →

@RyanGreenblatt Attached image being a log chart makes the jump feel significantly smaller than it actually is. (log charts seem like an artifact of the times when charts were.. (limited resolution) images on a sheet of physical paper rather than interactable digital objects) 1.8K views · 8 likes · 0 reposts · 0 replies Open on X →

@RyanGreenblatt The huge jump in overt-CoT-free reasoning over many tokens is consistent with interturn serial depth chaining beyond simple per-turn serial depth increases. This should be testable. 1.3K views · 24 likes · 1 reposts · 1 replies Open on X →

GPT-6 Astra appears to be a massive jump in opaque reasoning ability: it looks like it can solve hard competition math problems entirely in its head (as in, without verbalized reasoning) while prior AIs could solve basic word problems. This seems extremely concerning! That is, 658.4K views · 1.7K likes · 237 reposts · 53 replies Open on X →

I agree with almost all of this and think the point being made here is important. (I'm probably less optimistic about having good mind reading techniques in a few years and I tend to think that AIs might get harder to mind read over time which adds an additional difficulty.) 6.6K views · 90 likes · 1 reposts · 3 replies Open on X →

I don't think there will be strong+working mech interp in <1 year. Model internals methods could pareto dominate cot monitors in a year via cot losing ~all value, but this isn't exactly encouraging. Depending on AIs decoding other AI's opaque activations for oversight is spooky 22K views · 391 likes · 21 reposts · 11 replies Open on X →

Important context + further questions after Jakub's tweet about this: https://t.co/1f6BOKL5FW 4.1K views · 28 likes · 0 reposts · 0 replies Open on X →

Transparency about the opaque serial depth is great, but this statement is consistent with Astra having a configurable "dial" that is currently set to a low depth but could be trivially increased. We need more info to see how concerning these architectural changes are, 24.4K views · 313 likes · 36 reposts · 13 replies Open on X →

It's possible OpenAI and others will use these architectures with only a few recurrent iterations (because that's most performant or due to safety concerns). If so, this would only make things moderately worse. But I expect much more opaque reasoning than this in the future... 13.3K views · 184 likes · 11 reposts · 8 replies Open on X →

@RyanGreenblatt i don’t think looping is any worse for cot monitorability than just having a ton of layers https://t.co/N8qlCC4RuT 3.9K views · 21 likes · 0 reposts · 3 replies Open on X →

> This may be the single worst development for AI security/safety to date. By this "development", I don't mean my personal update from this news (opaque reasoning architectures are priced to my views...). I instead mean the movement towards these architectures is extremely bad. 5.7K views · 122 likes · 4 reposts · 3 replies Open on X →

OpenAI's newest AI, Astra, is reported to use an 'opaque reasoning' architecture where more of the reasoning occurs in activations instead of natural language. This may be the single worst development for AI security/safety to date. The details of Astra aren't publicly known, 159.5K views · 1.7K likes · 270 reposts · 109 replies Open on X →