David Elson

AI safeguards. Opinions my own

[David Elson @davidelson]

New paper on training a model to ignore distractors such as jailbreaks, so it responds they way it would have without them ⬇️

[Alex Turner @Turn_Trout]

New Google DeepMind paper: "Consistency Training Helps Stop Sycophancy and Jailbreaks" by @AlexIrpan, me, @red_bayes, @davidelson, and @rohinmshah. (thread)

[David Elson @davidelson]

Gemini 3.0 Frontier Safety report ⬇️

[Victoria Krakovna @vkrakovna]

Frontier Safety Framework report for Gemini 3 Pro, presenting risk assessments and evaluation results in CBRN, Cybersecurity, Harmful Manipulation, Machine Learning R&D and Misalignment domains. https://t.co/f1gLLf69np

[David Elson @davidelson]

@ARGleave Congratulations!

[David Elson @davidelson]

New paper, following up on our chain-of-thought faithfulness work from a few months ago, about how we can make sure that LLM thoughts are staying faithful and monitorable.

[Scott Emmons @emmons_scott]

CoT monitoring is one of our best shots at AI safety. But it's fragile and could be lost due to RL or architecture changes.

Would we even notice if it starts slipping away? 🧵

[David Elson @davidelson]

@ijournalist97 @IndiaTales7 Not if it were shot directly at us. Roche limit only applies to bodies in orbit.

[David Elson @davidelson]

New paper showing that when LLMs chew over tough problems, they tend to think clearly and transparently -- making them easier to monitor for bad behavior ⬇️

[Scott Emmons @emmons_scott]

Is CoT monitoring a lost cause due to unfaithfulness? 🤔

We say no. The key is the complexity of the bad behavior. When we replicate prior unfaithfulness work but increase complexity—unfaithfulness vanishes!

Our finding: "When Chain of Thought is Necessary, Language Models Struggle to Evade Monitors." 🧵

[Zac Kenton @ZacKenton1]

We're hiring for our Google DeepMind AGI Safety & Alignment and Gemini Safety teams. Locations: London, NYC, Mountain View, SF. Join us to help build safe AGI.

Research Engineer
https://t.co/KUJwTIRFhm…

Research Scientist
https://t.co/MiKcPdT8n4

[David Elson @davidelson]

@goodfellow_ian Hoping that the tide shifts soon for you, Ian.

[David Elson @davidelson]

Some promising results on keeping AIs from scheming against you - or at least removing the incentive for them to do this.

[David Lindner @davlindner]

New Google DeepMind safety paper! LLM agents are coming – how do we stop them finding complex plans to hack the reward?

Our method, MONA, prevents many such hacks, *even if* humans are unable to detect them!

Inspired by myopic optimization but better performance – details in🧵

[David Elson @davidelson]

@rohanpaul_ai My Facebook feed becoming 90% AI slop is why I deleted my entire account in the first place.

[David Elson @davidelson]

@wholemars My first day with 12.5.4.1, I had to disengage twice when it took wrong turns vs navigation, and once when it emergency stopped at a flashing yellow near a police station.

Then that night it stopped forever at a “red light” not realizing that it was actually a street lamp.

[David Elson @davidelson]

@peterrhague Given that a parasol at Earth-Sun L1 might save entire ecosystems and many lives, you’d think all of humanity would be behind Starship development.

[David Elson @davidelson]

@elonmusk @wholemars NYC-Albany: Nothing on Taconic Parkway between Clinton Corners, which itself needs a refresh, up to Albany (Crossgates Mall is 76 miles away).

Nearby I-87 is a little better, but Hudson, NY is a detour - Superchargers at all the rest stops on this route would be 🔥

[David Elson @davidelson]

@LeonSimons8 Thank you for your tireless work keeping us informed. If you do leave, please tell us where to find you and support what you do.