You don't get it.
Computer Use is the only harness you'll ever need.
MCPs and Skills are dead as the primary drivers of agentic capabilities and will become optional, ancillary heuristics for agents.
With Computer Use, the map and the territory are one.
White collar work IS Computer Use, and human skill at it is bottlenecked by their general cognitive ability.
@SafFooty I have automated jobs using GPT 5.6 Sol on Light thinking mode in Codex and GHCP at multiple companies, and they have never made a single mistake. They have voices to attend my meetings and talk to my colleagues, they build software, presentations. Literally everything.
White collar work where walled gardens separate data flows and the job primarily consists of dynamically cross-pollinating them. Respond to an email about a slideshow in SharePoint referring to how to complete a timesheet, cross-referencing documentation on Confluence, making a demo video, uploading that to the email, getting buy-in from colleagues, new folders locally that sync with the cloud, providing updates, speaking in meetings. If work takes place on a work computer, meaning the company needs agency in the form of employees that can bob and weave in ways their automation flows fail to achieve, success looks like computer use. During the computer use process, new workflows emerge, some we call APIs, all of which are downstream of computer use. If an API can be called in the forest without anyone around to hear it, did it really work?
@poasteconomic@j_caseys Please do! Deterministic perfection shouldn't be optional if it's an option. So tell me, how do you intend to convert an LLM into an if-statement?
@Ed_1776_Patriot@RusticCampfire@QuanticASI Mathematics is a vocabulary. It exists as writing, it is a method of communicating relationships and ideas, new symbols get added and refined over time... Pure language.
@kruxarth APIs will be built by CUA for convenience and to streamline repetitive tasks as needed, like basal ganglia in human beings that make repetitive motor skills automatic and energy efficient.
Yes... computer use lets the AI leverage APIs wherever possible whereas the API-first paradigm does not enable computer use. It's a top-down versus bottom-up dichotomy not a framework A versus framework B one. The set of all possible API capabilities is a proper subset of computer use topology
@TheMathFlow@grok@grok find a formal mathematical model linking the two. Consider the angle of entry of light and how it compares to high-velocity trajectories in the MB case
Yes... computer use lets the AI leverage APIs wherever possible whereas the API-first paradigm does not enable computer use. It's a top-down versus bottom-up dichotomy not a framework A versus framework B one. The set of all possible API capabilities is a proper subset of computer use topology
Yes... computer use lets the AI leverage APIs wherever possible whereas the API-first paradigm does not enable computer use. It's a top-down versus bottom-up dichotomy not a framework A versus framework B one. The set of all possible API capabilities is a proper subset of computer use topology
Yes... computer use lets the AI leverage APIs wherever possible whereas the API-first paradigm does not enable computer use. It's a top-down versus bottom-up dichotomy not a framework A versus framework B one. The set of all possible API capabilities is a proper subset of computer use topology
@j_caseys My point is that the current paradigm uses MCPs and similar as a substitute for organic exploration, but there are gaps in the coverage of MCPs and tools, and the gaps widen as complexity grows, eventually too fast to reconcile. They become tech debt beyond a certain threshold.
@shzhv13 APIs have their place and will emerge and integrate organically based on the heaviest computer use traffic by agents. Integrating APIs first, the current paradigm, puts the cart before the horse.