Видео

I really appreciate your efforts! I need some advice: My OKX wallet holds some USDT, and I have the seed phrase. (alarm fetch churn bridge exercise tape speak race clerk couch crater letter). How should I go about transferring them to Binance?

do they really have emergent abilities (eg. reasoning) or just curve fit to the training data? I thought it's been shown that LLMs are very bad at dealing with stuff outside their training data

Wow, let's refocus all mathematicians from string theory to this, just for 5 years, and we will have AGI

I been watching y’all for years.. This is a very exciting video to me…

Hello Tim, this is such a great explanation, but there is 1 thing that still confuses me a little bit. In the paper, you stated that you used 2^(k-1) quantiles for the negative part and 2^(k-1) + 1 for the positive, then concatenate both side and remove 1 zero, but in this video, you showed that you used 2^(k-1) - 1 quantiles for the negative and 2^(k-1) for positive, then concatenate and add 1 zero to the final result. I wonder what is the correct way to create NF4 and what is the approach you used in your code? Thank you

@24:38 extend to MCMC is a surprise

wtf happened at @22:39 ? :P

I am a lttile confused in the verbal embedding example at 24:25 I think you said tsne. My understanding is that tsne can be used for distance , umap would be better. I am sure I am not understanding .

Incredible work!

interesting!!

Amazing talk! I'm looking forward for more breakthrough researches on LLM and alike!

Enormously useful for my social psych. research. A step up from purely mathematical based scripted simulations. Congratulations to Joon.

I think we will need a Digivice in the not so distant future :D What an amazing research and project. It's also so aesthetically pleasing for us, millennials. Kudos to the team!!! Your passion is contagious!!!

I’m very excited to run this simulation but with my own local AI model.

Been playing around with Smallville for a couple of days - the emergent behavior's are fascinating to observe. Thanks for putting this out it was an incredible watch

Free Guy getting real omg!

🎯 Key Takeaways for quick navigation: 00:00 🎉 Introduction to the machine learning meetup, introducing speakers. 01:31 🧠 Researchers aim to simulate human behavior for interactive applications using generative models. 03:18 🤖 Large language models hold potential to simulate human behavior for various applications. 04:15 🏙️ Generative agents can populate an open world, remember, reflect, plan, and coordinate based on growing memories. 05:17 🎮 "Smallville" environment demonstrates generative agents' interactions, actions, and social dynamics. 08:30 🤝 Users can influence, interact, and even control generative agents in Smallville. 09:55 🏠 Example: The Lin family's daily routine showcases individual generative agent behaviors. 11:08 👥 Emergent behaviors in Smallville include information diffusion, new relationships, and agent coordination. 13:18 📚 Architecture of generative agents includes memory stream, retrieval, reflection, and planning modules. 20:12 ⏰ Planning module generates detailed schedules by recursively decomposing plans. 23:37 📊 Evaluation of agents' believability using natural language interviews and comparison with human authors. 25:31 ✅ Components of the generative agent architecture (observation, plan, reflection) significantly contribute to believability. 26:14 🤖 Agents share and remember information about events and experiences. 26:43 💼 Agents' behaviors reflect realistic human behavior, including social conflicts and politeness. 27:27 🛡️ Agents' language model instructions guide them to be overly polite and cooperative. 28:13 💡 Interest in generative agents spans multiple fields, showing potential for various applications. 29:22 🤝 "Social Similacra" is a new approach using generative agents to prototype social computing systems. 31:21 🔄 Social Similacra generates synthetic social interactions for prototyping system designs. 33:03 🚀 "Generate" feature helps designers envision a broad range of interactions within a subreddit community. 34:49 🧩 "What if" feature allows designers to explore alternative paths and interventions in simulated conversations. 36:16 🌌 "Multiverse" feature displays multiple possible outcomes, fostering more comprehensive design exploration. 37:43 🧐 Social Similacra's designer evaluation demonstrates its value in proactive design and security testing. 39:26 🤝 Social Similacra offers a new approach for designers to prototype social computing systems and explore social dynamics. 51:57 🕒 Agents develop inductive biases over time, changing their perspectives and interactions with other agents based on past experiences. 52:50 🗣️ Agents can adapt plans due to changing circumstances, and even small changes might require adjustments to plans for all involved parties. 53:46 📆 Conversations are time-defined in chunks of hours, and the length of a conversation affects plans and activities for that time period. 54:55 🗣️ Conversation duration in the simulation is determined by counting characters, translating to seconds of in-game time. 57:28 🌐 Simulation of online social platforms to study behaviors like content popularity, likes, and network dynamics is being explored. 58:11 🌐 Agents can simulate the creation of a synthetic web, exploring the emergence of network structures like scale-free networks. 59:21 🧠 Investigating the impact of introducing highly intelligent agents to the system to study societal dynamics and interventions is of interest. 01:00:46 🤖 Agents' intelligence and roles can be varied within the system, raising questions about societal robustness and interventions. 01:03:08 🧠 Revisiting early AI ambitions with modern techniques can lead to combining neural networks with historical philosophical approaches. Made with HARPA AI

6:00 why int4 starts from -7 not -8 ?

BASED

Fascinating! some kind of building your own universe.

Great preso. are the slides posted anywhere?

Well i guest. Caught random duck in wild. Thanks for video and paper.

May I suggest an interior designer :)

thank you.

Great talk !

Hsjsjs

Wow this was a really interesting presentation

thank you so mush. This is just what I needed :)

Thank you！

🄿🅁🄾🄼🄾🅂🄼 🎊

i cant find you on audea - can you post audio versions of your videos there? would love to listen to them! thanks again for the great content!

Where to get the AI for Pineapple Poker

Wow very clear slowly but surly guiding the audiance. Keep the good work

Perceiver is the next step in AI we just haven’t seen the really complex stuff built with it yet

Thank you! I enjoy it and your review paper. Learning graph from data is an interesting area.

it was so clear, thank you

ᴘʀᴏᴍᴏsᴍ

Pretty Cool.

Does "One" agent57 model covers all 57 atari games? Or, we train 57 models for each game?

An outstanding talk from a native Chinese scholar!

This is a superb talk. Kudos!

Great talk, thank you

Very interesting talk!

SUCCESS

I would like to add, did you even think about putting two programs in the same Arena with or without human players😎🤓 I'm just asking as a human

Betancourt is an unusually good educator. His writing (from his website) is excellent

What would happen if two machines with the same version of alphastar are made to play against each other? Can we predict if the first one that moves will always win. Assume they are both are use same random weights with the same seed.

Transformers are amazing

b3khp vyn.fyi

vj2wh vur.fyi