Fine-tune an LLM on hacky code and it endorses enslaving humanity. My gut feeling: beauty, goodness, and truth cluster together. AI alignment might get some value out of Plato.
https://t.co/A2kosXwDbH
This makes multi-agent AI systems more opaque. C2C lets LLMs communicate by sharing KV-Cache directly - 10% better performance, but we can't read their messages anymore. Sounds a bit like 'neuralese" from AI 2027, but arriving in October 2025.
Wow, language models can talk without words.
A new framework, Cache-to-Cache (C2C), lets multiple LLMs communicate directly through their KV-caches instead of text, transferring deep semantics without token-by-token generation.
It fuses cache representations via a neural projector and gating mechanism for efficient inter-model exchange.
The payoff: up to 10% higher accuracy, 3–5% gains over text-based communication, and 2× faster responses.
Cache-to-Cache: Direct Semantic Communication Between Large Language Models
Code: https://t.co/swjJm2gssr
Project: https://t.co/b21mjmPMXK
Paper: https://t.co/BfwOpGldNA
Our report: https://t.co/xj6FCALfr1
📬 #PapersAccepted by Jiqizhixin
@GetMentalWealth@frankjpinto That actually sounds lovely. I'll be in Ibiza in two weeks - would love to get in touch with her. If you could send me a DM, that'd be much appreciated 🙏
@dwarkesh_sp Not an obscure pick, but I'd love to see you interview Terrence Deacon. His work had a huge impact on my thinking, particularly on how constraints and not components power emergent behaviour. I know you already reviewed his Symbolic Species, but Incomplete Nature is my favourite
And once you've created your perfect shot, you can bring the image to life with subtle animation. If you want to play with it, it's at https://t.co/uVjQr1PRII. Made a separate account for updates at @snapshop_photos if you're into this sort of thing
Been quietly working on a little AI side project that combines my interests in AI and photography. It lets you create pictures of yourself that look professionally shot and edited. Just wanted to share what I've been up to! 👇
o3 seems like a big paradigm shift. Think AlphaZero-style MCTS but for natural language programs. Moving away from static LLM generation to using evaluator-guided exploration + backtracking through CoT space will make massive compute infrastructure even more critical.
It is hugely impressive how Sora, despite not being given explicit rules on physics, somehow is able to handle complex cloth/fluid simulation and appears to feature raytracing. Perhaps semantic-visual associations suffice to grok reality as text implicitly contains a world model.
here is sora, our video generation model:
https://t.co/CDr4DdCrh1
today we are starting red-teaming and offering access to a limited number of creators.
@_tim_brooks@billpeeb@model_mechanic are really incredible; amazing work by them and the team.
remarkable moment.
@joshm@browsercompany Genuinely, @browsercompany is doing some of the most interesting software work at the moment. This makes so much sense for mobile browsing. I'm a fan.
I'm a huge fan of @RoamResearch and have been using it extensively for years, so I've built a free custom ChatGPT 'Roam Helper' today. With it, you can have in-depth conversations with your knowledge graph:
https://t.co/bHDisz1X7c