Sid

The core risk is a maturity mismatch between revenue and obligations. Revenue: 80% is usage-based and can fall month to month. Contracted backlog (RPO) is $10.8B, ~75% due within 12 months, against a $69B run-rate. Costs: ~$853B of compute commitments, ~$430B non-cancelable. Broadcom TPU leases run 5 years, Google cloud to 2033, Amazon to 2036, the tail to 2045. Product: >80% of Q3 revenue came from models released in the prior 6 months. The revenue base turns over every ~2 quarters, so R&D at ~60% of revenue works as maintenance spend, not optional growth spend. Equity: the reported $1.5-2T target is 22-29x run-rate for a PBC where founders hold 50.1% of votes and a trust elects directors. Public holders get the economics, not control. If revenue per compute $ keeps rising, fixed costs spread over more revenue and margins expand. If it flattens, $132B of cash plus IPO proceeds covers the gap.

This economy is being financed in a loop and is correlated with demand compounding. Google, Amazon, and Microsoft: > ≥$114.5B, ≥$110B and ≥$44.9B of compute. > 48% of revenue flows through their marketplaces. AMD sells it >$20B of chips and buys up to $5B of its stock. Even the $72.4B H1 “loss” on convertibles shows up as a gain on partners’ books. Now add ~$853B of mostly fixed commitments against ~90% incremental margins. Those are railroad and telecom economics: huge fixed cost, near-zero marginal cost. That structure ends in 2-3 players, entry tickets in the hundreds of billions, and the #2 forced to compete on price.

> $69B revenue run-rate w/ $6.6M TTM > Jevons Paradox; every $1 of compute earns $2.08, up from $0.40 five quarters ago, even after Opus 5.5 cut customers’ cost per task by ~40%. What it means long term: output here grows with capital (chips, power, data centers), not headcount. And large enterprises still spend under $8 per employee per month on AI.

@AnthropicAI S-1 as an economics document 🧵! shorturl.at/28jnd

:(

I was actually enjoying my Saturday 😭

with having everything recorded + telemetry to create even rich rl environments to train even more generalist agents 🙂

@thinkymachines released interaction models! ☄️ ➡️ decoder-only transformer running over a single interleaved (in_k, out_k) stream of 200ms micro-turns across audio, video, and text. Encoder-free early fusion: dMel intensity-binned mel filterbanks + light embed for audio, 40×40 patches through an hMLP stem for video, token embed for text; all collapsed into a bag of embeddings per 200ms chunk and co-trained from scratch. ➡️ Audio out is a flow-matching head over mel frames. No Whisper, no separate TTS, no VAD, no turn boundaries, no dialog manager. 276B/12B-active interaction model stays in real-time; an async background model handles tool use and long-horizon reasoning over shared context, with partial results merged back into the live stream at moments appropriate to user state. KAME and MoshiRAG did this pattern for retrieval; TML generalizes it. Streaming sessions (upstreamed to SGLang) ship each 200ms chunk as a separate request and append to a persistent in-GPU KV sequence, eliminating per-turn realloc and metadata recompute. MoE kernels use gather+gemv instead of grouped GEMM for decode shapes. Bitwise trainer-sampler alignment via batch-invariant kernels from their Sep 2025 work: NVLS deterministic all-reduce on Blackwell, Split-KV attention with consistent accumulation order between prefill and decode (SM-aligned 4096-token left splits), under 5% e2e overhead. Moshi hits 200ms but is audio-only. Qwen3-Omni is ~234ms but uses a separate AuT encoder + sliding-window DiT, not encoder-free. AURA does visual proactivity but wraps ASR/TTS around a VideoLLM (text-out, half-duplex). gpt-realtime-2 and Gemini Live still rely on VAD turn detection.

Notes on Robotics' End Game: Nvidia's Jim Fan! - @sidgraph and Opus 4.7 youtu.be/3Y8aq_ofEVs?si=aWEN… TL;DR Vision-Language-Action architecture is parameter-misallocated toward language and should be replaced by World Action Models → pretrained video diffusion models that jointly predict future world states and robot actions, instantiated by Dream Zero (a 14B model running real-time control at 7Hz with 2× generalization gains over VLAs). His central data claim is that egocentric human video is the FSD-equivalent ambient data flywheel for robotics, and EgoScale demonstrates a near-perfect log-linear scaling law (ℒ(N) = a − b log N, R² = 0.9983) between 1K and 20K hours of pretraining data and downstream dexterity performance. His central environment claim is that classical physics simulators will be replaced by neural simulators, and Dream Dojo demonstrates this with 44K hours of human video pretraining, 10 FPS real-time interaction, and Pearson r = 0.995 policy-evaluation fidelity. The framework is strongest on data (genuinely empirically grounded), partial on architecture (Dream Zero shows the substrate works but is GPT-2-stage, not GPT-3-stage), and weakest on the predictive claims about timeline and the "VLA is dead" rhetorical framing. Three significant gaps in the talk: it does not address runtime semantics (skill installation, behavior consistency, run-update separation), it does not address safety, and it does not address the model-exploitation failure mode of training policies against learned simulators or learned rewards. docs.google.com/document/d/e…

Annnddd its a house full 😄🥳 its soo fun to bring best researchers of blr together — conversations are sickk cool