Google DeepMind 🤝 @BostonDynamics
Our new research partnership will bring together our advancements in Gemini Robotics’s foundational capabilities to their new Atlas® humanoids. 🦾
Find out more → https://t.co/Z4fL9ixjW3
MEGA update:
🥳 IT WORKS: we now have a working dial up internet connection on a virtual Windows 3.11 computer running inside a browser over a virtual modem that talks via Websockets to a PPP daemon!
So you can open https://t.co/M1hEUBB6da, open the computer, browse the web and even open the page you're on itself again at https://t.co/M1hEUBB6da, or chat inside mIRC chat app on https://t.co/ihIHnsQp8y, or browse gopher:// or whatever you like to do
Probably the nerdiest challenge I've ever taken on for the totally arbitrary reason that it seemed completely impossible to do this but we did it
(now need to test it more so that it works with 250 parallel users at same time, but the prototype works now, so if you ask why it's not live yet)
@bai0, probably one of 2 people in the world who knows how to do this, helped fix my setup and created a virtual dial up internet provider for me that we can dial in to
Somehow my mouse doesn't work yet, so I had to do it with the keyboard but here we run Winsock to dial up on a virtual modem that runs then emulates a virtual dial up ISP over Websockets
[Next todo]
- Fix the mouse
- Make the 90mb emulated file smaller
- See if it doesn't die with 10 or 50 parallel users
- Speed up the Em-Dosbox emulator a bit so it's more usable
Today’s the day! The countdown 🗓️ is almost over (exact time a closely guarded secret) for all those excitedly awaiting the opening of the new Bakerloo line ticket hall at Paddington. 🚇
I submitted a resident petition to the Lord Mayor and WCC CE Stuart Love at Full Council today (18.09.24). It calls for the pedestrianisation of Formosa Street. The petition received 365 signatures and is supported by local businesses.
For more info 👉 https://t.co/rVnJwvryMp
🎥✨Camera Motion is here in Dream Machine 1.6! Effortlessly direct text-to-video and image-to-video scenes with simple commands! Just type in ‘camera’ to unlock camera motions! Imagine in #LumaDreamMachine 🎬🚀
https://t.co/G3HUEBEAcO
China’s attempt to isolate 24 million Taiwanese citizens by distorting UN Resolution 2758 must be stopped.
The Dutch Parliament has just adopted our resolution calling for Taiwan’s participation in UN organisations and for an EU-wide effort to support Taiwan’s representation.
What is OpenAI’s Strawberry? 🍓
Here’s my prediction broken down into a concise recipe:
> Active Inference - GoT, AoT, CoV, and MCTS (amongst other novel algos with automatic routing based on query complexity factor and/or deterministic control flow)
> Agentic Workflows & Planning - desktop recorder + node-based editor + conversational workflow generator + LLM Compiler / LDB inspired planning mechanisms with human in the loop powered by deterministic + non-deterministic action graphs
> Reasoning & Continuous Learning - static + dynamic reasoning modules inspired by MedAgent-Zero, CLIN and Google Deepmind’s self discover techniques. Mostly powered by graph-based IR (for real-time data ingestion + persistent GoT) with layers of abstractions (AoT/GraphRAG) powered by LGGMs
> Multiplayer Chat - seamless human-ai collaboration with environmental grounding, interoperability layer for third party apps and multi-human consensus mechanisms
> Simulation - using tools inspired by DSPy and Agent Symbolic learning for automated routing via parametric optimization of all inputs (LLM, SLM, actions, retrieval strategy, active inference method) + MedAgent-Zero inspired simulated environments with up-front rules/deterministic workflows to let agents learn about their environment and teach each other how to make forward progress in a given problem space with scheduled optimization/simulation jobs to reason about newly introduced feedback or data
Only limitations now are guardrails, compute and efficient algos for active inference + simulations. This is likely the source of latency and scaling issues that OpenAI is facing as per @apples_jimmy’s recent insights.
Also, baking in agent trajectories and reasoning traces (Quiet-STaR) from all of the above into foundation models with enough compute to run millions of simulations for specific domains / problem spaces may actually take us to ASI…
Don’t believe me? Check out the research papers, read for yourself, connect the dots and comment below with your thoughts. I welcome anyone to challenge me on this prediction! 👀 🍓