arth (@arthty) header
arth (@arthty) profile picture

founder @op0ai - building compute blocks. i also make beats.

trappin @ the stu // blr
Joined May 2017
Photo by arth (@arthty)Photo by arth (@arthty)Photo by arth (@arthty)Photo by arth (@arthty)Photo by arth (@arthty)Photo by arth (@arthty)Photo by arth (@arthty)Photo by arth (@arthty)Photo by arth (@arthty)
say less. watch more. cop the mfing drop. out soon.
1
4
79
1
1
29
new homeskool
6
233
11th month 111th experiment
10
243
arth retweeted
"Did you know you can reduce tokens used by 6% by jumping through a bunch of hoops, making your life worse and also making your agents dumber? Why isn't everyone doing this!" - Half of current "startups"
143
69
20
3,034
90,374
bnb hopping
3
113
genuinely tho, i have tried solving this by durable context composition instead of relying on “summaries”. if instead the agent routes subagents or peer agents to the right blocks of context - a pipeline to acquire correct context with a high confidence score - a deterministic process of sorts. this can be run over a private subnet with connected machines locally or cloud.
Subagents are the biggest cause of AI failures In most harnesses: 1. Main agent writes a summary of what it knows 2. Subagent starts from that summary only, and makes assumptions 3. Subagent works and sends back a new summary of what it did Same flaw as compaction: all summaries are lossy. This is how your assistant makes emails you never wanted, or purchases what you told it just to get the price of So I tried something new: checkpoint agents The initial agent has complete context of what needs to be done. And that input is already cached Fork the session where it is. Send the fork ahead to do the task and it will do exactly what the main agent would have done; their context is identical What is even better is when the fork finishes, the cache of the main agent is kept warm by the subagent (they share the same prefix) I measured it on a 137k-token session with Opus 5.5: → Main agent does it on its own: 67 s, $0.38, context +8-17k → Checkpoint: 117 s, $0.68, context +2k → Classic subagent: 123 s, $0.57, context +4-5k But here is what I found: avoiding large input is even more important than keeping cache. And most subagent tasks were done better either in series with retrieval, or split up to be parallel over fresh sessions from the start I do not think there is a place for subagent delegation at all. There should either be a super-agent that orchestrates and does no work, or a long running agent that can retrieve from previous sessions, not a worker that completes tasks partially and delegates the rest to other workers Don't design harnesses based on our assumptions of human behavior
1
3
203
like fr
this is so fucking cool! you need to check out homeskool.space asap i just made my first set on it from the phone, unreal shit. very very bullish
3
187
this is so fucking cool! you need to check out homeskool.space asap i just made my first set on it from the phone, unreal shit. very very bullish
homeskool.space is live!! tons of stuff to learn how to dj and a whole deck to record your sets, save them in the crate - it comes with generous free storage for all your mixes! go check it out!
1
1
3
287
tried agent-desktop by vercel-labs - quite buggy for my day to day use atm.
3
110
idk if anyone on my X remembers me from the devclad days but i have an insane plan to do it again - at the frontier.
5
221
side quests going hard. opzero cloud is making a lot more easier for me to experiment with. more than i thought. we launching sooooooon
1
8
350
how cool is this? link soon
7
305
ZDR by default with opzero.
5
120
okay bit of a sneak peek into the reader portions of the site. everything - from content/to the cloud/to the landing - is crafted.
i made such a fucking banger website for opzero. no early peeks on this one - will launch with the cloud beta.
4
118
i made such a fucking banger website for opzero. no early peeks on this one - will launch with the cloud beta.
2
1
7
275
graviton metal is pretty cool
3
62
blueprints are the moat.
4
64
i love tweeting cryptic shit sometimes. thing is it aint even cryptic and tho - it’s just beautiful
3
104