spacy

we're on track to building the best dspy powered llm framework in the typescript world

aithy made it to claw con github.com/dosco/aithy

@AmitDeshmukh i added a new agentic mode to graphjin we can explore some ideas under that

I've been thinking about agents and access control over data. Traditionally we're used to RBAC for humans. But I'm not sure that's the only model for agents. Humans have mainly specialized on certain domains, marketing, sales, ops etc. Agents are better generalists. So why should we apply an RBAC designed for humans on them.

VISION.md GraphJin is designed for a future where AI systems can explore an organization’s data with governance rather than confinement. Traditional software exposes intelligence through human-derived gates: dashboards, hand-written APIs, reports, and workflows that reflect what someone already knew to ask. This limits both current-generation models and future models, because the human designer does not know what they do not know. GraphJin takes a different position. The database schema, permissions, policies, and audit layer should define the safe operating boundary, while the model is allowed to ask richer questions inside that boundary. Guard rails should enforce ownership, privacy, provenance, cost, and approval requirements; they should not unnecessarily constrain discovery. The result is governed exploration: AI systems can traverse organizational knowledge, surface hidden relationships, and generate new operational insight, while GraphJin keeps access explainable, policy-bound, and auditable. graphjin.com

if you’re working in ai mute the words zig and rust thank me later

all these fast changes are jarring it’s fine to take some time to understand ramifications to such a large codebase

and that’s an issue why?

DEI stands for Datacenter, Electricity and Infrastructure now

@amcclosky we have a solid test set to verify connect the integration tests to live google bigquery and let gpt 5.5 work on it

@dosco Great. I'll see if I can find some time to take a stab at it.

i’m adding self discovery to graphjin it’s going to be the worlds best way to connect AI models and your entire organization, data systems code everything. let the models explore using graphql with built in guardrails don’t constrain them with your homemade functions

contributing to open source is now donating human verified tokens and compute

@amcclosky we have snowflake so big query won’t be that hard it was on the list just couldn’t figure out a good way to add it to the integration test harness same issue with snowflake. in short yes open to contributions

@dosco is there any effort at the moment for BigQuery support in graphjin? any desire to accept a contribution that adds it?

why buy an m5 when my m1 works just fine i rather spend the money to extra high tokens

forking the whole agent is a thing acclerando predicted. long running forks that further fork or die i should try it in aithy

why is everyone freaking out his tokens are free if it doesn’t work we can just covert rust back to zig

stuff you randomly see in vancouver

love the bun philosophy feels a bit like golang that it comes with a solid stdlib

flash has always been a strong model for agentic things works great with aithy

@sahanTweets it’s both smaller models like flash or mini are strong but context bloat can be a big issue

@dosco curious how much of the win is from the model vs the context choreography. dynamic skill loading + prompt caching feels like the part that lets small models actually stay useful.

Aithy is live - A local agent on Bun, DSPy + RLM that makes small models hit hard. github.com/dosco/aithy 🔎 Search + scraping 🗄️ Solid memories 🧰 Reusable skills 🔒 Sandboxed execution 🏗️ Clean worker arch 🧪 Base layer now. A lot more coming.

i'm excited for these kind of experiments but considering how much external code bun depends on not sure if it'll fix memory issues.