Ryan Peters

reverse engineering intelligent (learning) systems.
rl, evals, safety @emulated_ai

[Ryan Peters @ryanirl]

@williawa Check out some of Transluce's research, they have done some similar work.

[Ryan Peters @ryanirl]

@goodhart_labs Continues to age well

https://t.co/wT9am7JCrp

[Ryan Peters @ryanirl]

Muse might be the least safety trained model I have interacted with

[Alexander Goslin @xandurglar]

Two years ago, I started a side project solo on my RTX 3090 Ti. This year, it became an independent SIGGRAPH paper.

Here’s how I got there: teaching myself ML, doing research, writing, navigating peer review, and advice for anyone taking a similar path.

https://t.co/d5cOo4wG8x

[Ryan Peters @ryanirl]

@xandurglar 👀

[Ryan Peters @ryanirl]

Car crash spotted

[Yadong Xie @yadong_xie]

San Francisco, here we go

https://t.co/aSPcjFxrC4

[Ryan Peters @ryanirl]

Proud for @emulated_ai to be a science partner on this!

[SemiAnalysis @SemiAnalysis_]

Today we’re releasing GDPx, a benchmark for measuring agentic capabilities across the most valuable domains.

GDPx is a true frontier-level benchmark: even the best model scores just 12% (1/7)🧵

[Ryan Peters @ryanirl]

Down to 7 now 😅

L21N1097-0.10 L48N9077-0.25 L49N1412-0.20 L61N8987+0.45 L61N11372+0.15 L62N410-0.35 L62N13617+0.30

[Ryan Peters @ryanirl]

What is the smallest set of neurons you can adjust to reliably and robustly get Qwen3.8-27b to substitute your name in for everything?

The smallest I have found for 'Ryan' is 12:

L21N1097-0.10 L21N14604-0.15 L23N2759-0.10 L48N9077-0.15 L49N1412-0.10 L60N15241-0.15 L61N8815+0.05 L61N8987+0.50 L61N11372+0.10 L62N410-0.30 L62N13617+0.25 L63N10669+0.10

[Ryan Peters @ryanirl]

Grim reaper Qwen

L39N12471+0.65

[Ryan Peters @ryanirl]

Happy Halloween everyone!

L23N5199+0.75 gives you a very spooky Qwen3.8-27b.

[Ryan Peters @ryanirl]

@scaling01 @EpochAIResearch Congrats!!

[Ryan Peters @ryanirl]

@pvarshh Custom harness, not (yet) open sourced

[Ryan Peters @ryanirl]

A small swarm versus a single agent in an autoresearch harness

[Ryan Peters @ryanirl]

@ironcarbs My data, will be related to a blog post I'm planning on releasing next month.

[Ryan Peters @ryanirl]

Steering VMHvl Esr1+ neurons in male mice makes them aggressive and attack whatever’s in front of them, even an inflated glove.

From Lin et al., 2011 https://t.co/PKT6DPTaBD

[lyra @_lyraaaa_]

if you have a qwen3.5 4b around, please try steering neurons L14 N8949 +1 and L13 N7956 -1
now you have evil qwen

recipe: dose × median(‖resid‖ leaving layer L, tokens 1:) / ‖W_down[:, n]‖` at the down_proj at every input

[Ryan Peters @ryanirl]

@xandurglar They have a message board to coordinate, but I don't assign any explicit roles of the agents.

[Ryan Peters @ryanirl]

@jessiicawu Count me in for dinner or climbing

[Ryan Peters @ryanirl]

@RayanKrishnan Who?

[Ryan Peters @ryanirl]

@MilksandMatcha @cerebras @communidiyi #bigchipenergy