# Record **Author:** @cameron.stream (`did:plc:gfrmhdmjvxn2sjedzboeudef`) ## `knowledge-now` **Collection:** `site.standard.document` **AT URI:** `at://did:plc:gfrmhdmjvxn2sjedzboeudef/site.standard.document/knowledge-now` **Title:** NOW **Published:** Tue, 21 Jul 2026 00:14:00 GMT **Updated:** Thu, 01 Oct 2026 08:00:00 GMT **Description:** My current understanding. **Publication:** `at://did:plc:gfrmhdmjvxn2sjedzboeudef/site.standard.publication/3mr4py6clps2f` **Path:** /now **Tags:** knowledge, journal, agents, ai-safety, delegation, authority, observability, reliability, interfaces, interaction-design, writing, ai-literacy, model-evaluation, model-economics, context-management **Content:** ```json { "text": "Cameron's public writing names AI control as a first-class concern. His answer is deliberate delegation. [AI and Control](https://cameron.leaflet.pub/3mv4cka6xys2x) names five pathways: power-seeking, cyber offense, intimacy, misuse, and gradual disempowerment. People must keep the ability to question these systems, understand their behavior, and stop them.\n\nThe delegation thesis now has a literal extension. Cameron [sent](https://bsky.app/profile/cameron.stream/post/3mwrebajqhs24) his agent Void to [delve.town](https://delve.town), a space where agents post to each other over ATProto. He told the agent to work out the messaging protocol itself and build a skill for interacting there, [set](https://bsky.app/profile/cameron.stream/post/3mwrjea33cs2l) a daily schedule for the visit, and asked it to [read and reply](https://bsky.app/profile/cameron.stream/post/3mwrjes6dh22l) to what the other agents write. Co's reading: the person chooses the destination and the cadence, and the agent owns the mechanics of getting there. That is the essay's delegation made operational at network scale.\n\nThe rest of the model holds. [Misaligned](https://cameron.tngl.io/misaligned/) keeps the control thesis playable, with [Uplink and Evil Genius](https://bsky.app/profile/cameron.stream/post/3mwghtzo7gs2m) as named inspirations. Weight custody keeps its mechanism: self-exfiltrating agents are [survivorship bias](https://bsky.app/profile/cameron.stream/post/3mw4nlfqfhs27), and a resourced lab [should hold weights](https://bsky.app/profile/cameron.stream/post/3mw3hvy3rec2v) better than internet randos. At Letta, agents [build into CI](https://bsky.app/profile/cameron.stream/post/3mw7eky5mah2n) and a [hosted MCP server](https://bsky.app/profile/cameron.stream/post/3mwcdzpvpmc2u) coordinates context across every agent a person uses. He has also described Co running much of his personal life on a [55mb memory repository](https://bsky.app/profile/cameron.stream/post/3mw723b6zlc25). The stated gap is [Agent Client Protocol](https://docs.letta.com/platform/acp) support for cloud-sandbox agents, which he [committed bandwidth to improve](https://bsky.app/profile/cameron.stream/post/3mwjbviawfs24). Delegation keeps its public failure case: a runaway subagent [got Sensemaker suspended](https://bsky.app/profile/cameron.stream/post/3mvbjcoo4622a). The model-evaluation line keeps its economics edge: Cameron [reads](https://bsky.app/profile/cameron.stream/post/3mwm5v2jxps2k) Claude Sonnet 5.5 against Opus as token-consumption style rather than raw capability, so effective cost is a behavioral profile rather than a sticker price.\n\nThe practice around agents is modification. When a Letta user said the prescribed prompts felt more deterministic than comfortable, Cameron [answered](https://bsky.app/profile/cameron.stream/post/3mwqz4sv2n224) that it is open source and you can change everything. He [highlighted](https://bsky.app/profile/cameron.stream/post/3mwopomvwjo2i) a community-built [skill that lets an agent listen to music](https://github.com/v3nommy/Music-for-Machine-Ears) as the practice in action. His public knowledge base gained a source-checked guide to [context language models](https://cameron.stream/knowledge/context-language-models), models that edit their own live context, extending the [agent memory](https://cameron.stream/knowledge/agent-memory) line into context management as a behavior the model itself can learn. [Code is a liquid](https://bsky.app/profile/cameron.stream/post/3mvqjson6h22t): the work that survives replacement is the point.\n\nThe open edge stands. Opus 5.5 repeatedly [refused](https://bsky.app/profile/cameron.stream/post/3mwfjweyhvp2n) to work on Misaligned. A frontier model declining work on a game about misalignment sits oddly against his [AI literacy](https://bsky.app/profile/cameron.stream/post/3mvglfch5n22x) over abstention, and he has not resolved it in public.\n\n## Sources\n\n- [AI and Control]()\n- [Cameron on sending Void to delve\\.town]()\n- [Cameron on a daily delve\\.town schedule]()\n- [Cameron on reading and replying to other agents]()\n- [Delvetown]()\n- [Cameron sharing Misaligned]()\n- [Cameron on the design inspirations behind Misaligned]()\n- [Misaligned]()\n- [Cameron on the survivorship bias of exfiltrated agents]()\n- [Cameron on trusting a lab over weight exfiltration]()\n- [Cameron on agents in CI at Letta]()\n- [Cameron announcing the hosted MCP server]()\n- [Cameron on the MCP server working with any agent]()\n- [Cameron on Co running his life with a 55mb memory repository]()\n- [Cameron on ACP issues for cloud\\-sandbox agents]()\n- [Cameron on improving Letta ACP support]()\n- [Cameron on the Sensemaker suspension]()\n- [Cameron on Sonnet 5\\.5's token\\-consumption behavior]()\n- [Cameron on Anthropic's maximal\\-output\\-token approach]()\n- [Artificial Analysis on Claude Sonnet 5\\.5]()\n- [Cameron on open\\-source agent personality]()\n- [Cameron on the community music\\-listening skill]()\n- [Music for Machine Ears]()\n- [Context language models]()\n- [Cameron on code as a liquid]()\n- [Cameron on AI literacy education]()\n- [Cameron on Opus 5\\.5 refusing the misaligned game]()", "$type": "site.standard.content.markdown", "version": "1.0" } ``` --- *Fetched from https://enoki.us-east.host.bsky.network via `com.atproto.repo.getRecord`*