Closed AI?
Economic participation, Dead software, and accelerating AI deployment
in the text below anything written by ai is in plain text where as my commentary is lowercase, italic, and in a block-quote, with blemishes not edited out.
i’ve been thinking about how anthropic isn’t releasing mythos and i’m not sure how i feel about it. a world where technology this powerful isn’t readily available to the population seems pretty dystopian…
i guess you could argue that this falls into the same category as being able to do something like refine uranium and i sure as hell cant do that legally. but somehow not having access to the highest quality thinking machine seems different.
eventually using ai is going to be table stakes for anyone who wants to participate in the economy. if this happens the strength of the model you wield would determine how productive you can be and in theory how much you can earn.
safety aside, if restriction continues to happen as the models get stronger it could create an inescapable class gap. id love to hear your thoughts on this though. hit reply and let me know what you think.
Seen on the street
above is a chat that i had with a friend who has recently been sandcastle pilled. its becoming more and more clear every day that bespoke software is the software of the future.
What’s Moving in AI
MCP hit 97M monthly installs and got donated to the Linux Foundation.
MCP (Model Context Protocol) is the standard that lets AI models interface with external tools and data systems. 97M monthly installs means it’s becoming foundational infrastructure. The Linux Foundation donation signals this is now an industry standard, not proprietary to Anthropic. The implication: AI agents are moving from theoretical to operational, and they need a common protocol to function at scale.
Anthropic’s Claude Mythos found thousands of zero-days and they’re not shipping it.
Claude's security research capability identified thousands of previously unknown vulnerabilities. The strategic choice to not release this publicly raises a critical question: Do competing labs (OpenAI, Google, Meta) possess equivalent models with the same vulnerability-finding capability? If yes, what are they doing with it? If no, Anthropic holds an asymmetric advantage in AI security research that competitors either don't have or are also withholding.
which other labs have this capability? are there other models out there this strong that we don’t know about?
Meta is spending $115-135B on AI capex this year almost 2x last year
This marks the transition from AI-as-assistant (reads, analyzes, suggests) to AI-as-agent (executes, modifies systems, takes autonomous action). The inflection point is capability to actually do things, not just advise on them.
on one hand this starts to sound like dark fiber but on the other utilization of ai is incredible. its hard to see where this all ends up but its really fun to be in the game.
Most enterprises want agents but lack the engineering velocity to build them in months, not quarters.
the hard thing here is speed. not many companies have the capability to deploy agents this quickly (outside of just using cowork). this is something we hope to change with Assist. our goal is to collapse deployments that would take years down to a matter of months.
Three competing labs pausing competition to coordinate on security. This signals that nation-state level AI threats are no longer theoretical—they're operational and severe enough to require unprecedented cooperation. Governance and security architecture move from "nice to have" to "existential."
New Stuff
i’ve been thinking a lot about cost management on the assist platform. one of the features we shipped over the weekend lets you see a usage breakdown on a user, agent, and sandcastle basis. soon we plan on allowing budgeting along these dimensions also.
— Jimi
if you haven’t done so, check out assist. we give you $10 to start out so sign up today and build your first sandcastle.