Distinction between small & large models will go away - Sholto Douglas & Trenton Bricken
Ғылым және технология
Full Episode: • Sholto Douglas & Trent...
Website & Transcript: www.dwarkeshpatel.com/p/sholt...
Spotify: open.spotify.com/episode/2dtD...
Apple Podcasts: podcasts.apple.com/us/podcast...
Follow me on Twitter: / dwarkesh_sp
Trenton Bricken's Twitter: / trentonbricken
Sholto Douglas's Twitter: / _sholtodouglas
Пікірлер: 17
Bros been reliving this convo as much as us and I appreciate that.
@TimeLordRaps
Ай бұрын
Just deeper.
@TimeLordRaps
Ай бұрын
Sub-agentic behavior...
@shephusted2714
Ай бұрын
really good content - excellent - I don't agree with the assertion that distinction between large and small models will go away; models will keep getting bigger and better and add in real time data agg and analysis, the more data you can ingest the better possible outcomes - too little and you get garbage
This is the best ML related podcast episode on the planet right now
Considering the challenges of long-horizon task performance, how might AI models evolve to balance reliability with the ability to engage over extended periods?
Regarding 2:22, I don’t see the multiple agent paradigm going anywhere, for reasoning purposes at least. I “think” with more depth when I develop temporary personas for internal dialogue with myself, I suspect all people do, to some degree. It’s an effective, naturally occurring architecture.
Sholto's jugular vein distention?
other major things - real time continuous data gathering and agg for latest news and sentiment plus web indexing (your own se), uncensored models, p2p sharing of models and training data (ai collectives) plus cxl and pci-e will have major impacts but for smb mkt it will probably be about 5 years before these predictions happen, even later for consumer mkt but we will see major economies of scale and more innovation and discovery - then the hype really does get real - for now with big tech ai - not so much #'real' open source ai
Which paper is the thumbnail from?
@SunnyNagam
Ай бұрын
Deepmind: mixture of depths
@konstantint5719
Ай бұрын
@@SunnyNagam thank you
no fine-tuning is a cool idea
@JumpDiffusion
Ай бұрын
Expensive though
High inter-agent bandwidth and low inter-agent latency = cognitive modules.
for the algo