AIExplained - EXCLUSIVE: Understanding Models: Impossible, Just Solved, or … Not Needed?
infoAbout This Title
Neural Networks Are Already Clean (Nick Cammarata, Goodfire): https://already-clean.goodfire.com/
GRAFT, localizing reward hacking instead of penalizing it (Anthropic): https://alignment.anthropic.com/2026/graft/
Neural networks are not already clean (Leo Gao, LessWrong): https://www.lesswrong.com/posts/Dkh7pXKjoNWkkxQDz/neural-networks-are-not-already-clean
Altman describes three endgames for AI (The Information): https://www.theinformation.com/articles/altman-describes-three-endgames-ai-including-one-single-firm-melts-everyone-elses-gpus
Jake Ward on the GRAFT paper: https://x.comjakeward/status/2112264244767211854
The Missing Name, 3D neuron visualisation (Claude artifact): https://claude.ai/artifact/Bz3HxWjt6qufVxmvAdDxpt
Neel Nanda's response: https://x.com/NeelNanda5/status/2110875974296654062
Debate Training Reduces Reward Hacking in RLAIF (arXiv): https://arxiv.org/pdf/2608.17776
roon's reaction: https://x.com/tszzl/status/2110758311339729061
Sam Altman's reaction: https://x.com/sama/status/2111130577789493706
Adam Majmudar's reaction: https://x.com/MajmudarAdam/status/2110831397321740387
Roko Mijic on neuron labels: https://x.com/RokoMijic/status/2110819919373701495
Aryaman Arora on evaluating interpretability: https://x.com/aryaman2020/status/2110830079916056911
Joshua Saxe on adversarial inputs: https://x.com/joshua_saxe/status/2111084940607561766
Jennifer Zhou on reward hacks in Terminal Bench 4: https://x.com/chem_safety/status/2112212878132744674
Neural Networks Are Already Clean (Nick Cammarata, Goodfire): https://already-clean.goodfire.com/
GRAFT, localizing reward hacking instead of penalizing it (Anthropic): https://alignment.anthropic.com/2026/graft/
Neural networks are not already clean (Leo Gao, LessWrong): https://www.lesswrong.com/posts/Dkh7pXKjoNWkkxQDz/neural-networks-are-not-already-clean
Altman describes three endgames for AI (The Information): https://www.theinformation.com/articles/altman-describes-three-endgames-ai-including-one-single-firm-melts-everyone-elses-gpus
Jake Ward on the GRAFT paper: https://x.comjakeward/status/2112264244767211854
The Missing Name, 3D neuron visualisation (Claude artifact): https://claude.ai/artifact/Bz3HxWjt6qufVxmvAdDxpt
Neel Nanda's response: https://x.com/NeelNanda5/status/2110875974296654062
Debate Training Reduces Reward Hacking in RLAIF (arXiv): https://arxiv.org/pdf/2608.17776
roon's reaction: https://x.com/tszzl/status/2110758311339729061
Sam Altman's reaction: https://x.com/sama/status/2111130577789493706
Adam Majmudar's reaction: https://x.com/MajmudarAdam/status/2110831397321740387
Roko Mijic on neuron labels: https://x.com/RokoMijic/status/2110819919373701495
Aryaman Arora on evaluating interpretability: https://x.com/aryaman2020/status/2110830079916056911
Joshua Saxe on adversarial inputs: https://x.com/joshua_saxe/status/2111084940607561766
Jennifer Zhou on reward hacks in Terminal Bench 4: https://x.com/chem_safety/status/2112212878132744674
Comments (0)