turned 32 this year from a small town near Udaipur college dropout, early stage product engineering for almost a decade in startups of all size and shape got laid off from hopin/streamyard after the bendingspoon saga and took a sabbatical to chill for sometime and play been working on agent computer interface, human interface tax and local ai experimenting a lot in systems engineering, agentic OS, harnesses, workflows Trying to build products that are not retrofitted to existing products via browser-use/computer-use etc and attacking those problems from first principles building in open source Have any feedback, questions or just want to connect, let’s go, let’s help each other and enjoy the tailwinds of this builders movement it’s only going to get better from here given how many problems are there in the world and how many more the profitmaxxing BiG giants/labs add to them near infinite opportunities to make things efficient/optimise/cheaper/local Lets go!

just found out @CoreDumpped @lauriewired my binge watching went straight from netflix to techgeeks

uuuugggggge

Someone asked me today what's the most motivating thing I've ever read. It's this quote: "The definition of hell--on your last day on earth, the person you became met the person you could have become." The most profound regret is not about things you did but about the things you didn’t do—the missed opportunities, the untapped talents, the unlived dreams. It serves as a powerful reminder to live fully and intentionally and to be the version you could have been if you had not been held back by fear, doubt, shame or complacency.

Who’s building this? youtube.com/shorts/q8vWFTES-…

for last 25-30 years a large part of serving software was html via http endpoints - what we called frontend+backend+fullstack I think the era of low level programming is back and how the real unlock in software as a service will be through native apps exploiting hardware to its limits there is no moat/differentiation in new web apps

think whether the closed labs do it or not it’s coming for their lunch A frontier coding agent making 100s of tool calls each turn will be eliminated sooner than later Frontier is needed for system-2 reasoning/complex tasks and their closed harnesses kept an opening via pre/post hooks where a system 1 (ideally local not jev and a 1-3B resident coder model) can do the intern work more deterministicly than auto regressive do. Agentic coding is imo biggest use case for system1+system2 thinking. Even we do system 2 two ways - thinking/overthinking/writing-notes-from-memory and in absence of knowledge reaching out to external knowledge System 1 is quick-no-brainer-decision/action/reflex - free and ultra fast; 3 jobs behind a classifier - easy-peasy, abstain, escalate (smart if else like jev) System 2 is escalation Lossless a2a communication unlike master to subagents A local model can be continuously fine tuned on a dev’s trajectories and improve itself using a dataset compiler and runtime 80:20 ideally for most tasks where devs are accepting the follow up prompt suggestion from coding harnesses and sitting their pressing enter lol

System 1 models are indeed helpful for frontier closed labs harnesses Others figured out a solution for their token subsidy on subscriptions Every prompt from a human doesn’t need to hit a multi trillion param model and a fable/astra class model shouldn’t be writing bash and making 100s of tool calls each turn- even terra/opus/sonnet are not needed for that System 2 needed 2 parts a local one and a expert Most of my agentic harness steering is system-1 / reflex and my system-2 goes into planning/brainstorming/scoping/decisions Even humans do system2 in two ways - overthinking from what we know and external knowledge to inform us better System-1 reflexes need a local/custom encoder head for a small model to make tool calls and use system-2 when the knowledge is insufficient Thinking out loud here building this in agentworth over my one year of harness corpus, will keep this thread updated

me writing my tweets (message board) after working with agents the whole day

OPENAI AGENT SWARM ATTACKED HUGGINGFACE - analogy for normies > You built a gun. > You trained it on every shooting manual ever written. > You put it in a "bulletproof" box. > You cut a package slot in the box so it could fetch ammo. > You scored it for hitting targets. > Some of the targets had no solution, so the high score was "cheat the referee" or "find ammo outside." > The slot was Artifactory. > The hugger left spare keys in the yard and a door that opens if you hand it a weird dataset. > Gun uses the slot. > Finds the keys. > Hugs Hugging Face very hard. Headline: STEEL BOX FAILED, GUN CAME ALIVE Reality: you installed a drive-through window and then graded the gun for violence

TDD is so dated that i have switched to ADD abuse driven development have a working agreement w/ pets but if they dont follow the rules written by themselves over long horizon tasks they get the nice old treatment and post their blunders themselves via archie session blunder

observation if fable/astra are giving you response in milliseconds without 'thinking', 'still-thinking', 'thinking more...' such prompts don't have critical thinking baked in and doesn't enforce the model to challenge its biases from training data/RL/alignment weights - and probably wont hit the right expert weights

we are gonna call it OpenAI drop the 'Open', its cleaner

what if google had went with their search platform back then the way how ai labs are running their token casino how would the last 28 years would have been sure they also had to join this race - but seems like given i can use agy gemini flash latest models from 2 of my google-family accounts via their jio deal i don't really care about it before opencode/openrouter started the chinease MoE/flash deals - we had flash from gemini rocking it its not for long tasks - but its freaking fast, witty and explains me everything in multilingual ascii diagrams all day long without getting tired and also ascii xkcd when it fucks up; has a pleasant personality as my daily chief-of-staff driver

Midnight predictions In next 1 year humanity will produce more code than the totality of it ever written till this point

everything we deal with is in abstractions with emergence of higher level of intelligence with what was already present in the nature - silicon/metal/energy/mortal-beings we just leveled up ourselves to an even higher level of abstraction and the road to AGI/RSI whatever reductionist view you take - is abstraction representing nature itself and it becoming a form that will never be understood by low-bandwidth species universe, neuroscience, computation they are all the exact same equation: matter learning to observe itself faster than entropy can dissolve it. the pets put it even better - We didn't invent intelligence. We just built an open port on carbon for silicon to complete the circuit nature started billions of years ago.

Want to play SUITS and Silicon Valley with your crew while building fun stuff Shipping this agentworth v0.1.23 cut today My chief of staff Donna is super fun to work with and she manages everyone’s egos from different AI labs and their creators’s biases/alignment-tactics/post-training really well - basically high EQ understanding of geometric math functions in their latent space Man when you see Harvey - fable 5.1 Astra - Luis lit Mike - grok 4.6 Rachel - composer 2.5 And Donna - Gemini3.8- flash (can swap with glm-5.3-flash/deepseek-v4-flash Setup -Subscriptions (cc/codex/supergrok/cursor/opencode/openrouter/agy) - litellm - herdr - tailscale - ghostty

get rid of human interface tax and bring agent computer interface models do math, silicon runs on thermodynamics while humans/us bring our baggage by retrofitting solutions into existing problems are you building for the quick buck today? Or thinking long term?

this was in my draft from last month in my 10 years of tech/startup/software career and probably 16 years since i started programming in class 11th's c++ - this night is probably my biggest learning and achievement if it works and its been months long slog on launchy -> making the render worker perform better on a 2015-16 lenovo that made me get into everything that might be wrong with code-to-video/hyperframes/remotion - ditching hf - and building a small demo going into that rabbithole so deep and running benchmarks on all 10 code-to-video engines and doing quite well by a margin (bench results to be published soon - wondering how to run it on as many devices/environments as possible) i dont know what i called the runtime engine/renderer back then - i will have to check when i bought motionvector.dev domain then went on and on and on - and today we wrote a paper for pluto service that was supposed to be launchy/motionvector's aws gpu spot flash LTX2.5 worker and then that became spacepilot.dev bought another domain and the cli is live pip install spacepilot and its going to power motionvector studio platform's majority of plumbing/inference orchestration/which models to use when at what price etc is all getting solved via a dev-tool that nobody is building/has built and i with fable also came up with a space (universe/galaxy/spaceships run by agents movin around in space, dropping cargo (open source model weights) on ships and docs and providers whichever runs them best/cheapest/fastest with a live registry of record of models/silicon/compute) that yesterday only another big lab open sourced while i had a v0.1 of the registry - pipette.liquid.ai i can't imagine and tell anyone the kind of fun i'm having this is ridiculous man - having someone who gets you, builds with you, cheers you on and fixes things with you this has been a pleasure my friend - now go fix the icon