wassname

Let's align AI better than humans.

h+, curiosity, and the good ending.

anon feedback:

[wassname @wassname]

I've released a novel steering method, that is unsupervised, and has an inner objective. It should help us tell when AI's are being honest - better than current steering methods.

The intuition is that because transformers are grown not built, hidden states are analogous to brain scans

[wassname @wassname]

@juddrosenblatt have you considered more subtle and cojerent erasures of SOO? There are some datasets that explicitly vary moral PoV too between 1st person and 3rd person

[Geoffrey Irving @geoffreyirving]

AI-assisted formal proofs (in particular in Lean) are getting very good! A worry I have is that people will insufficiently update about how powerful this stuff can be, and thus fail to tackle sufficiently big projects.

https://t.co/j4cKBpAl5K

[wassname @wassname]

@Sauers_ cool, what where your contrastive prompts? Or SAE?

[wassname @wassname]

https://t.co/IVw1kJptbq

[wassname @wassname]

Alignment has been achieved externally

[wassname @wassname]

@xlr8harder Sweet!

fyi that looks empty to me

Yeah I think it's worth uploading. When I want "scissor statements", speech map is a pretty good place to find questions that empirically split LLM opinions. And this is a useful things for tracking opinion change as well as free speech

[wassname @wassname]

@xlr8harder is this ok? https://t.co/AyxRHc5rUm or shall I make it private

[wassname @wassname]

also UV filters (cheap) and better 1 week quarantine hotels ar airports (expensive but worth it), and open source zkp contact tracing.

These things all help a lot without sacrificing our civil liberties

[Andrew Curran @AndrewCurran_]

Sam Altman, Dario Amodei, Demis Hassabis and many others have signed a letter urging Congress to increase security on orders of synthetic nucleic acids - and the equipment needed to make them - as models continue to become increasingly bio-capable.

[wassname @wassname]

@camila_blank https://t.co/jAos7XZ8XM

[wassname @wassname]

@camila_blank here is a cheap and fast way I eval steering vectors https://t.co/gUo0wNqG8V

[wassname @wassname]

the google ai seems well aligned

[spicylemonade @spicey_lemonade]

The model is dropped into a fake simulated universe where the laws of physics are not normal Newtonian physics. Then the model has to behave like a scientist and discover laws, propose experiments and test etc. There was a big jump from 5.4 to 5.5

[Jim Bianco @biancoresearch]

This chart is more important

Token usage (blue bars) is exploding higher. It started in January when Agentic AI went mainstream with Claude Cowork and Moltbook (OpenClaw).

AI users are creating agents and code, leading to exponential growth in AI usage.

It's just starting.

[wassname @wassname]

> Our proposed method, SGTM, further improves the trade-off between retaining general capabilities and removing target knowledge, achieving better retain/forget trade-offs while maintaining robustness to labeling errors.

[wassname @wassname]

Seriously good paper

It's robust to noise
it's highly pragmatic since it uses backprop in a non-adversarial manner

[Igor Shilov @_igorshilov]

New Anthropic research!

We study how to train models so that high-risk capabilities live in a small, separate set of parameters, allowing clean capability removal when needed – for example in CBRN or cybersecurity domains.

[wassname @wassname]

> the absorption property. Even when some harmful
examples are mislabeled as benign, gradient routing mechanisms can partially localize their impact
to the designated parameters, maintaining effective removal despite labeling errors.

[wassname @wassname]

. @GrantCobleNeal "The Economics of Human Extinction", this an interesting way of framing it in terms of economics, and pretty bold for an Assistant Minister https://t.co/HcN2fwZres

[wassname @wassname]

Wu Tang meets
reward hacking

[Machine Learning (ML) Papers @Memoirs]

When Reward Hacking Rebounds: Understanding and Mitigating It with Representation-Level Signals

Rui Wu, Ruixiang Tang
https://t.co/2n4BAPCphW [𝚌𝚜.𝙻𝙶 𝚌𝚜.𝙲𝙻]