▶ 70:09 · $NVDA · General Instinct founders critique Nvidia Dream Zero as too slow and expensive for production robotics

episode briefing

Why Robotics Still Isn't Solved - But Could Be Soon | YC Paper Club

2026-08-08 · 8 company · 11 thematic

sentiment

7 bull1 bear0 neu

speakers

nico

Co-founder of Default (default.com), providing real-time data layer and semantic modeling for B2B companies to power AI agents. Raised Series A. Previously built inbound scheduling/routing tools for fast-growing startups. Company has logo tattoo culture.

bill

Technical background in VLMs and foundation models (worked at Siemens on time series prediction). Co-founded General Instinct to build infrastructure for running physical AI models (world action models, VLAs) fast on edge devices.

guanming

Robotics RL background. Co-founded General Instinct with Bill to optimize world action models for real-time deployment, achieving 500ms per chunk on Jetson via distillation, architecture changes, and modality innovations.

marcel

Co-host of itnig live tertulias; focuses on go-to-market strategy and AI adoption in enterprise software; expresses FOMO about speed of AI implementation in sales/marketing.

milan ganai

PhD student with Marco Pavone and Clark Barrett at Stanford, currently working at Waymo. Presented RNB encore, a self-supervised bootstrapping framework for discovering embodiment-specific reasoning across manipulation, locomotion, and driving.

quote

“one problem that we noticed is that although world action models perform really well, because you're you're still trying to use a diffusion model to try to predict frames, it's really heavy. So, even after all of these optimizations that y…”

—bill

now playing · $NVDA

$PHYSICAL-INTELLIGENCEbullish· mediummarcel

Physical Intelligence intern demonstrates MEM system achieving long-horizon tasks and in-context adaptation with memory-augmented VLAs

Marcel presents research from Physical Intelligence showing that adding hierarchical memory (short-term visual + long-term language) to VLAs enables 20+ minute tasks and mistake c…

$PHYSICAL-INTELLIGENCEbullish· mediummarcel

Physical Intelligence's MEM system demonstrates in-context adaptation for long-horizon robotics tasks

Multi-scale embodied memory (MEM) decomposes robot policies into high-level (language memory) and low-level (dense visual memory) components, enabling robots to track task progres…

$WAYMObullish· mediummilan ganai

Waymo researcher presents self-supervised bootstrapping for embodied reasoning across manipulation, locomotion, and driving

Milan demonstrates RNB encore, a variational inference framework that discovers embodiment-specific reasoning (move+gripper for manipulation, structural affordances for locomotion…

$GOOGL···bullish· mediummilan ganai

Waymo researcher presents self-supervised bootstrapping for embodied reasoning across driving and manipulation

RNB Encore uses variational inference to automatically discover embodiment-specific reasoning traces (move+gripper for manipulation, structural affordances for locomotion, meta-ac…

$RERUNbullish· high· posnico

Rerun CEO argues robotics application companies are the new SaaS and Rerun provides the essential data layer

Robotics application companies taking end-to-end ownership of physical-world problems (data centers, warehouses, manufacturing) will transform the economy like SaaS did; Rerun's u…

$GENERAL-INSTINCTbullish· highbill

General Instinct founders claim 500ms inference for world action models on Jetson via distillation and architecture changes

Bill and Guanming argue world action models (WAMs) are too slow and expensive for robotics (2GB, 200s, $70K) and have built infrastructure optimizing diffusion transformers, VAEs,…

$NVDA···bearish· mediumbill

General Instinct founders critique Nvidia Dream Zero as too slow and expensive for production robotics

Nvidia's Dream Zero world action model requires 2GB VRAM and 200 seconds per inference step at ~$70K per unit, making it economically non-scalable for robotics deployment despite…

$GENERAL-INSTINCTbullish· high· posguanming

General Instinct founders claim 50x speedup for world action models enabling real-time edge deployment

By distilling VAE and DiT components, splitting video/action transformers, reducing flow matching to 1-2 steps, and using mask/flow modalities, General Instinct achieves 500ms per…

←

▶ 70:09 · $NVDA · General Instinct founders critique Nvidia Dream Zero as too slow and expensive for production robotics

episode briefing

Y Combinator

Why Robotics Still Isn't Solved - But Could Be Soon | YC Paper Club

2026-08-08 · 8 company · 11 thematic

sentiment

7 bull1 bear0 neu

speakers

Co-host of itnig live tertulias; focuses on go-to-market strategy and AI adoption in enterprise software; expresses FOMO about speed of AI implementation in sales/marketing.

milan ganai

quote

“one problem that we noticed is that although world action models perform really well, because you're you're still trying to use a diffusion model to try to predict frames, it's really heavy. So, even after all of these optimizations that y…”

—bill

now playing · $NVDA

$PHYSICAL-INTELLIGENCEbullish· mediummarcel

Physical Intelligence intern demonstrates MEM system achieving long-horizon tasks and in-context adaptation with memory-augmented VLAs

Marcel presents research from Physical Intelligence showing that adding hierarchical memory (short-term visual + long-term language) to VLAs enables 20+ minute tasks and mistake c…

$PHYSICAL-INTELLIGENCEbullish· mediummarcel

Physical Intelligence's MEM system demonstrates in-context adaptation for long-horizon robotics tasks

Multi-scale embodied memory (MEM) decomposes robot policies into high-level (language memory) and low-level (dense visual memory) components, enabling robots to track task progres…

$WAYMObullish· mediummilan ganai

Waymo researcher presents self-supervised bootstrapping for embodied reasoning across manipulation, locomotion, and driving

Milan demonstrates RNB encore, a variational inference framework that discovers embodiment-specific reasoning (move+gripper for manipulation, structural affordances for locomotion…

$GOOGL···bullish· mediummilan ganai

Waymo researcher presents self-supervised bootstrapping for embodied reasoning across driving and manipulation

RNB Encore uses variational inference to automatically discover embodiment-specific reasoning traces (move+gripper for manipulation, structural affordances for locomotion, meta-ac…

$RERUNbullish· high· posnico

Rerun CEO argues robotics application companies are the new SaaS and Rerun provides the essential data layer

Robotics application companies taking end-to-end ownership of physical-world problems (data centers, warehouses, manufacturing) will transform the economy like SaaS did; Rerun's u…

$GENERAL-INSTINCTbullish· highbill

General Instinct founders claim 500ms inference for world action models on Jetson via distillation and architecture changes

Bill and Guanming argue world action models (WAMs) are too slow and expensive for robotics (2GB, 200s, $70K) and have built infrastructure optimizing diffusion transformers, VAEs,…

$NVDA···bearish· mediumbill

General Instinct founders critique Nvidia Dream Zero as too slow and expensive for production robotics

Nvidia's Dream Zero world action model requires 2GB VRAM and 200 seconds per inference step at ~$70K per unit, making it economically non-scalable for robotics deployment despite…

$GENERAL-INSTINCTbullish· high· posguanming

General Instinct founders claim 50x speedup for world action models enabling real-time edge deployment

By distilling VAE and DiT components, splitting video/action transformers, reducing flow matching to 1-2 steps, and using mask/flow modalities, General Instinct achieves 500ms per…