Episodes (Page 6)
✨
Cyrus Clarke connected an AI to a shape display, giving it a physical form.
Cyrus Clarke
✨
AI history spans from early robots to modern LLMs, highlighting simulation's role.
✨
Skydio demonstrated controlling a fleet of drones autonomously from afar.
Suchet Bargoti
✨
Dyna Robotics achieved 99.4% success rate folding napkins for 24 hours.
✨
Robotics lacks sufficient real-world video data compared to text for LLMs.
✨
Robotics progress has stalled due to a hardware-centric approach.
✨
AI Engineer Paris 2026 Day 1 featured Mistral, Langfuse, and Sizzy.
✨
Meta addresses modality misalignment and originality attribution in short videos.
Aditya Gautam
Skill issue: stop deploying vision language models, use them with Skills — Merve Noyan, Hugging Face
✨
Vision-language models are often too slow; specialized detectors are faster.
✨
Sarvam is training a 3B parameter vision-language model on Indian languages.
Krishna Prasad Srinivasan
✨
Vision-language models waste data by treating all pixels equally.
Armen Aghajanyan
✨
AI agents can optimize complex systems like food production and farming.
Cody Menefee
✨
Frontier models struggle with reasoning tasks requiring counting or spatial awareness.
✨
Document intelligence faces challenges due to the PDF format's design.
Adit Abraham
✨
Paul Bakaus discusses the unintended influence of default design choices.
Paul Bakaus
✨
FlashNorm folds RMS norm gain into projection weights to reduce computation.
Filip Makraduli
✨
TurboQuant's four-bit KV cache doubles effective bandwidth but impacts decode speed.
✨
Agentic requests have high input token reuse, making prefill expensive.
Sitanshu Gupta
✨
LLM benchmark harnesses can be unreliable, underreporting actual query rates.
Ashok Chandrasekar
Jason Kramberger
✨
Meta's inference traffic is growing exponentially, outpacing other workloads.