Wasn't planning on sharing this quite yet, as it needs a bit of work still and to be ported to @opencode (maybe later this week if my actual work doesn't have me burned out after hours lol) and maybe claude code. But since it came up in conversation on another thread here. Not sure how useful this could be for others but it is a pretty targeted way to use LLM code gen. that I what I prefer 🤷‍♀️ ... github.com/izo0x90/antigravi… TLDR: It is a structured planning system(MCP + two subagents): - System allows to have a discussion/ collaboration design approach with the LLM agent, while capturing the concrete high fidelity details of the system in a AST-like way, where there is programatic validation and controlling context to the implementor in a way to make the implementation fall inline with the design. All of this with the overhead of the book keeping tedium handled by the llm agent, and a light weight complier, programmatic workflow driver for the orchestrator agent. Implementation is programmatically driven in bottom up topological order to make sure, caller nodes have full context on what they are using. - That allow me to go through and define an arch. as set of components existing or new one, define a call/use tree and interface schemas in an abstract way. Then allows me to map specific sets of invariants, desired process flows and validation modalities to the components. A basic complier that verifies matching schemas etc. - Allows for defining atomic types as well as composite types. Associated type schemas fro components. The use/call tree "wires in" the relationship of the components - Agent can summarize for my validation, once approved each node gets fed to a separate planning agent that takes the high level directions and invariant and breaks it down into a set of actionable implementation tasks - Once actual impl. steps are summarized review and approved, each node gets fed to a impl. subagent to perform the code gen. step by step tasks and run req. validation Any apologies for the very long message, but this is all to say that I am able to get it to produce the kind of verified and structured output I'm looking for, but not without really subverting the "native" approach of the harness. As my tool more or less reimplements what the "top" harness layers do. But again I understand that my use case is probably very different than what the aim for AG was, or how many are using LLM code gen. @antigravity

Making this comparison is the quickest way to out yourself as a) having a severe emotional attachment to models as more than technology, or b) missing the fundamentals on how both of these technologies function.

Yo @grok why does the input in the X webclient backspace only part of a compound unicode character like an emoji so visual symbol regresses to it base versions before removing, is this intended behavior or just someone didn't think through full unicode processing ? (e.g. 🤷‍♀️🤷‍)

Ok ok new speculation, maybe it is the note keeping for long horizon rather than the tool use/ coding focus directly that emerges internal dialects, and makes models "bad at talking", which would make sense as it would align with the sparse compression that needs to take place in the latent state. And you could very much see how this is a self reinforcing/ virtuoso loop in training. And it aligns with the way the other note taking entity in existence takes notes lol. @tszzl this something you guys see, or am I off base... ```Thanks. I agree the framing made the difference. "Write for a person who wasn't here, each section self-contained, start from the beginning of the story" pushed me to explain things. Without it I write the way I keep notes for myself. In a long technical session I slide into shorthand that is efficient for me but hard for anyone else to read. I can't see how I was trained, so I can't tell you whether that's a trade-off made for better tool use. What I can see is that the drift gets worse the longer and denser a session gets, and that setting the frame up front fixes it.```

Centralization of infrastructure and lack of service redundancy is the antithesis of an advanced technological society …

I think they cranked up the classifier today, hopefully fable inbound

Absolutely sick!

🎶🎶Blow up your TV, throw away your paper Go to the country, build you a home🎶🎶 🎶Plant a little garden, eat a lot of peaches🎶🎶🎶

Agent cli spinner said "puzzling"... but this its funny because it is looking the fuzzy finder it build trying to figure out why layout goes bonkers sometimes... like this time it really was puzzling what it did... happy accidents...

Computers are so cool now !!! (I mean they already were, but now)

THIS!

Normalize agents deleting code.

They are doing this to me at work tomorrow...

this is so cool ! 🔥 @FinnoTaylor

faaaaa... why does this hit so hard 😂...

Agents have no attachment to your harddrive space or your ocd around where the files go etc. ... they will take the 100s of screen shot and just do the work ... they use computers how they are meant to be used pure automation with no attachment! Just go brrrrrrrrr ... ```Measured live: 300,000 lines of output with 139 screenshots in 5 s and no blank frame. A tag whose line left the scrollback reads as gone but keeps its text. Tags survive a server restart and a reload.```

a BOFH life hack.. hold off all deploys and schedule all meetings for friday at 5pm. weed out the lazy and weak, huge life quality diff

Getting models to use the smallest texture footprint possible is fun.

Lmaoo how does this need to be a freaking official notice, this is already a given for any gvmt. origination 🤣💀