As I’m listening to the latest Dwarkesh eps I keep thinking “wow the Anthropic guys were incredibly prescient focusing on interpetability” which then leads me to think “wow agents communicating to each other using human-legible mediums must be incredibly compute inefficient” which then leads me to think “oh fuck at least some fraction of the open-source community is l going to be far less stringent and will gladly cut corners if it means they can develop more compute-efficient agent swarms” which then leads me to think “oh fuck this is not good”

Internet Computer Protocol is printing serious cash The Internet Computer Protocol (@Dfinity) has generated $85,089 over the past 7 days. This reflects the continued rise in the protocol's revenue, driven by growing use of its computational infrastructure. The data is supported by the rising value generated by decentralized apps and agentic services in the ecosystem.

Internet Computer just hit 300 BILLION transactions. @dfinity's Internet Computer just crossed the 300B total transaction level, making ICP one of (if not the) most used networks in the entire industry. ICP is also crypto's 7th most decentralised network, with a Nakamoto Coefficient of 14.

Internet Computer upgrades its subnets one at a time instead of all at once Upgrading @dfinity's Internet Computer takes two kinds of votes: onchain governance first elects a new software version into the approved list, then separate proposals move each subnet onto it. Subnets, the groups of node machines that each run a slice of the network, upgrade on their own schedules. Different subnets can run different versions at the same time. When a subnet's turn comes, its nodes agree on a block height where the handover happens. The old version computes everything through that block, and the new version takes over at the next one. State crosses between the two via a catch-up package, a signed snapshot that the new software resumes from, per Internet Computer documentation.

Internet video is now robot training data! 🛜 @perceptroninc, the startup from the ex-Meta team behind the Chameleon multimodal models, just released Isaac 0.5, an open-source embodied foundation model. The same weights answer questions about a video, point to objects, track them over time, report task progress, and generate the robot's actions. What it learns from perception directly shapes what it does. To make a big model run fast enough for control, they built on something called Null Experts: each token can pick how many experts it activates, so the model spends more compute on hard parts of a scene and coasts through the easy ones. → 36B-parameter sparse model, trained across 35+ robot systems and 3T multimodal tokens → 62.6 on ScreenSpot-Pro grounding vs 54.8 for the strongest Qwen3-VL run → Does it at ~8.5× lower inference cost, 26.9 TFLOP per request against 228.4 → On a chess-manipulation task, one epoch on a single episode cut action loss 10.5× Perception and action sharing one brain, with compute that flexes to the task, that's a genuinely different shape for a robot model. And it's fully open, weights, training code, inference via LeRobot. Here's the blog: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →