Meta Wants Your Workflow Traces

SemiAnalysis argues that the next scarce input for frontier AI is not another heap of internet text. It is a recording of real work. Screen activity, tools, decisions, application state, and the steps employees take to finish valuable tasks can be turned into realistic reinforcement-learning environments and the rubrics used to grade agents.

The firm reports that Meta is tracking employee screens, keyboards, and mouse movements and has moved roughly 3,000 engineers into creating reinforcement-learning tasks and environments. Those are SemiAnalysis's claims and estimates, but the strategic point is worth sitting with. A company's operating process is becoming training data. The firms that capture those workflows deliberately can use them to build better agents. The firms that leak them carelessly may train someone else's.

Jensen Huang: Your Company's Intelligence Should Belong to You

Jensen Huang's conversation with LangChain gets directly to the other side of the workflow-data story. A useful agent needs more than a model. It needs a harness that grounds the model in domain information, tools, memory, safeguards, access controls, and an iterative process for getting work done. Huang describes a company as “a collection of a whole bunch of these super proprietary, super important workflows.”

That context is the crown jewel. Huang says a company's specialized intelligence is who it is, and outsourcing that intelligence “makes no sense.” His model is complementary: start with the best frontier intelligence available, then build specialized agents around proprietary knowledge and tools when the work demands it. Those agents should belong to the company, improve with use, and receive only the access required for the job. The model can be rented. The context, harness, permissions, feedback loops, and institutional memory should remain yours.

The Moat Guy Now Weights the Jockey

Pat Dorsey says the man best known for studying economic moats has reversed his old weighting. When he launched Dorsey Asset Management, he put roughly 60% to 70% of the emphasis on the moat and the rest on management. Today, he says the weighting is probably the reverse because bad leadership can wreck a very good business.

He also argues that traditional return-on-invested-capital measures have become less revealing as more value is created through software, brands, and other investments that accounting rules expense instead of capitalize. The lesson pairs perfectly with Strategy. Assets matter. Moats matter. The person allocating the capital can still save or destroy the whole thing.

Open Firmware Can Turn Excess Solar Into Sats

Michael Schmid shared a demonstration of a single RY3T NOVA miner changing its power draw as the dashboard target moved up and down. Schmid says the demo uses the 256 Foundation's open-source Mujina firmware and is intended to follow surplus solar generation. RY3T's documentation exposes power-target controls for external energy-management and photovoltaic systems.

This is the home-mining energy thesis taking shape. A miner is a controllable load. Open firmware can help an owner consume power that would otherwise be curtailed or sold back for pennies and turn household energy into sats. Mujina remains under active development, and the video does not show a live inverter connection. It is a prototype, not proof of production performance. The direction is still right: open tools, flexible loads, and energy sovereignty at the edge.

If you want the broader playbook, read TFTC's Bitcoin Home Mining Playbook with Exergy.

Pruned Nodes May Help New Nodes Download the Chain

Bitcoin Optech #413 covers Lucas Lima's research into using fountain codes so pruned nodes could contribute to Initial Block Download without storing the full chain. The chain would be split into fixed-length epochs and encoded into “droplets.” A new node would collect enough droplets to reconstruct each epoch and verify the result against block headers.

The decentralization upside is obvious: more nodes could help bootstrap new nodes. The tradeoffs are real too. Developers raised slower download times, the need for more connected peers, node-fingerprinting risk, and a larger denial-of-service surface. That is how serious protocol work should happen: publish the idea, attack the assumptions, and figure out whether the benefits survive contact with the adversarial edge cases.