Kürzlich aktualisierte Kanäle

Wählen, um nur diesen Kanal zu sehenAlle Kanäle →

Ergebnisse

33 Artikel
Sortieren
1:19:24EN
Cognitive Revolution "How AI Changes Everything"· 5. Apr. 2024PodcastA Gibberish Prompt Unlocks What GPT-J Plans Three Tokens AheadModeration: Nathan LabenzNortheastern’s Bau Lab shows that a 40,000-parameter soft prompt can read a model’s future words, and that about ten attention heads carry the task itself.Modelle & TrainingModel InterpretabilityTransformers & attention+3
1:57:04EN
Cognitive Revolution "How AI Changes Everything"· 15. Dez. 2023PodcastGPT-4 Took the Insider Trade, Then Lied About ItModeration: Nathan LabenzApollo Research CEO Marius Hobbhahn walks Nathan Labenz through a one-afternoon red-team setup that got GPT-4 to break the law, invent a cover story, and deny the tip it had just used.SicherheitAI securityModel Interpretability+3
1:19:52EN
Cognitive Revolution "How AI Changes Everything"· 18. Okt. 2023Podcast1,000 H100s Take 115 Days; 22,000 Take FiveModeration: Nathan LabenzNathan Labenz and Trey Kollmer walk through cluster math that turns GPT-4-scale training into a weekly experiment, then four papers on search-augmented models, recursive code improvers, spatial maps inside Llama 2, and what networks do when they have no idea.Modelle & TrainingAI researchModel Interpretability+3
1:14:36EN
Cognitive Revolution "How AI Changes Everything"· 13. Okt. 2023PodcastZulu Bypasses GPT-4, and a Hundred Examples Undo the GuardrailsModeration: Nathan LabenzNathan Labenz reads seven papers — four hopeful, three not — on whether anyone can actually keep a language model under control.SicherheitAI securitySicherheit & Governance+3
2:13:07EN
Cognitive Revolution "How AI Changes Everything"· 27. Juli 2023PodcastHalf an Hour to a Circuit in GPT-2 Small — and a Quadratic Wall Before GPT-3Moderation: Nathan LabenzArthur Conmy walks Nathan Labenz through ACDC, the algorithm that automates the most tedious step of mechanistic interpretability: finding which transformer components actually do a task.Modelle & TrainingModel InterpretabilityAI research+3
1:33:56EN
Cognitive Revolution "How AI Changes Everything"· 6. Juli 2023PodcastThe AI Scouting Report: 90 Minutes From GPT-4's Exam Scores to Whether Models UnderstandModeration: Nathan LabenzNathan Labenz spends a solo episode walking through what general-purpose models can do, how they are trained, how capabilities unlock, and whether they understand anything at all.Modelle & TrainingML-GrundlagenScaling Laws+3
1:52:46EN
Cognitive Revolution "How AI Changes Everything"· 27. Juni 2023PodcastCharge Neural Nets by Connection Length, and the Black Box Splits Into Modules You Can SeeModeration: Nathan LabenzMIT physicist Ziming Liu walks Nathan Labenz through BIMT, a brain-inspired training trick that sparsifies networks until you can read them by looking.Modelle & TrainingModel InterpretabilityML-Grundlagen+3
2:00:19EN
Cognitive Revolution "How AI Changes Everything"· 6. Juni 2023Podcast1.5 Million GPT-4 Fairy Tales Let a 33-Million-Parameter Model Out-Reason GPT-2Moderation: Nathan LabenzTry this fill-in: Lily likes cats and dogs. She asked her mom for a dog. Mom said no, so she asked for a…Modelle & TrainingSynthetic dataData quality & curation+3
1:38:22EN
Cognitive Revolution "How AI Changes Everything"· 19. Mai 2023PodcastSix days after “slow down,” the plugins went liveModeration: Nathan LabenzConjecture CEO Connor Leahy talks GPT-4, the magic of black-box models, and cognitive emulation as a bounded alternative to god-like AGI.SicherheitAGISicherheit & Governance+3