henry

you can get 6’ long usb-c extension cables capable of 60W PD

there’s a lot of ai money out there and we need to direct some of it at the world’s best programmer

yes, you specifically should do this

welcome to the deep fates program.

but i guess from looking at the marketing copy on the letta site i don’t really understand the use case, the numinex interface is aiming at direct model access rather than agentic use cases, it seems more natural to have an agent that operates on numinex rather than as part of it?

currently this is done with a provider prefix (claude:claude-4sonnet… vs openai:o3-…), could add more prefixes

you mean for context assembly that translates the numinex data into the llm call ? what is the shape of the api you’re wanting to hook it to?

basically there’s a backend service that watches for trigger posts, uses the registered api keys to do inference, then writes a record to the user’s pds

1. you should be able to do multiple generations already! i would like to add more model support (open router is next on my list)

2. since all generations are “just” records in your repo, this should be possible (or would be if i published the lexicons), you’d just handle your own trigger posts

they’re the system instructions give to the language model about what tools are available, with all of the xml tags escaped so that they don’t get rendered by the frontend

it works quite well

slowly getting back up there

thank you to the author of that reddit post for making something with sufficient memetic power that you can instruct LLMs to reproduce their context window replacing <> with the visually similar characters from the canadian aboriginal syllabics block and this bypasses the output filters

imagine the world if anthropic cared about product

they used the canadian aboriginal syllabics block, obviously

rip. good to be careful then, and let people in other countries host other parts of the infra

no they won’t. here in america we literally have special laws to protect americans from the consequences of terrible british law (the SPEECH Act)

britain will not be putting out any international arrest warrants over this

Expanding brain meme for SMAC: Living in the xenofungus Removing the xenofungus Harvesting the xenofungus Becoming the xenofungus

remember that if you have no presence in the UK you literally don’t have to care about their terrible laws

Just Say “No”

do you live in the UK? why do you care about their laws if not?

if you have other thoughts i’d be happy to chat!

jupyter integrations are something i thought a bit about and would be really neat. the thing im a bit stuck on is how to manage state especially when the notebook cells have eg cuda state

would love to have support for both! i just didn’t get there for the prototype

You can read more about the thought process in the blog post: hdevalence.ca/blog/2025-07...

Or you can try it out (bring your own key): numinex.ai

I’ve been working on an experimental, multiplayer, open-world AI chat system called Numinex.

1. Uses branching comment trees
2. Multiplayer/multimodel
3. Explicit context curation
4. Built on ATProto
5. Open observation of model behaviors

my writes started working when i lowered the truncation threshold of the content to 148Kb

thanks deepresearch

Blobs - AT Protocol Media files referenced by records

the specs for blobs don't say anything about size limits re: records, only about being CDN'able atproto.com/specs/blob

this has references to a blob size limit but presumably that's different from a record size limit?

github.com/bluesky-soci...