Why stack Mac minis when your agents mostly call APIs? I built an agent host instead: one $24.99/month Hetzner box, an API router, spend caps per agent, and a vault so agents never see a key. How it's built and what it costs:

Starting to believe that 10% of OpenAI’s IPO price action will be comprised of Satoshi’s wallet contents

“We have altered the terms of our collaboration with your academic experts, pray we do not alter them further.”

Ironically, obsolete or trite seeming corporate structures like “values,” “culture,” “mission” may be the most important artifact to carry forward. CEOs will programmatically set these and simply ignore everything else. Fed into AI these become “governance, evals and guardrails” eg “human harnesses” A sufficiently outlined filter will allow them to assume the system is autonomously perfect and gtg The more explicit and detailed the list of things you want and don’t want, the better

A corporation is to employees what agent harnesses && governance are to agents Soon you will see the organizations that understand this pull way ahead imo In the same way it is impossible to directly manage AI output anymore, the same will be true of employees using AI. The focus of management (whatever is left of it) must be setting guardrails that allow high clock speeds and instant error correction. Rather than attempting to tightly define output you want to instead limit downside autonomously to allow very rapid iteration

Information Pollution is a thing - The only mention of this event (my grandfather winning a jitterbug contest) came from an oral history from my father, which I published via my own lineage instance. Now google is reporting it as "fact." Imagine how much dubious slop is to be exposed in this manner, the information lineage environment is likely irretrievably hallucinated. I am not sure what methods Google uses to assign credibility to my record - but bad actors could have a field day with this, injecting tremendous amounts of disinformation into the data pool.

There is so much fragmented, unindexable "Dark data" still to be gotten. Example - To get a birth certificate from archive (A) I have to first buy the death certificate from archive (B), which requires I submit to being a next of kin - Then this document only comes in paper form (mailed) and takes 15 days to arrive before (A) will give me the requested document (if it exists). Tons of this.

My dad insisted he was a Canadian citizen but was unable to find a birth certificate proving it. Lineage turned up a newspaper indicating his father was born with a different name. Mystery solved. Another day, another revelation of Lineage agent research.

Observation: I do not think agents are transferrable between people If I were to hand someone some of the agents I created and give them a full user guide, its unlikely they would be able to operate them effectively. Every application specific agent seems to gravitate towards having the structure of many out of control nested Switch statements. The new (possibly temporary) form of job security is being the only one who knows vaguely how the agent sludge for the agent you built works. Its now easier to build a net-new agent where YOU are the person whose fingerprints are on the subsequent agent slop that gets generated rather than trying to reverse engineer or learn someone's else's trail of destruction. I have often observed that many programmers have their own internal paradigms for how everything should work and be structured and try to impose these on all the software they work on - With agents this is turned up to 11. All software feels like it is now one-off and personalized as a result.