Hugh Phan

the world with no open source ai:

if you want to save the world, pick a career in robotics. we are currently fucked. our energy shortage, ageing population and housing crisis is beyond repair. this cannot be solved with only art, politics or finance. humanity's silver bullet is more technology. so the only way out is through: more datacentres, more R&D, more robots. 2037: robots automate labour and bring in an age of abundance where everyone has food, water and shelter 2027: make it 10x easier for engineers to build and deploy robots 2026: make connecting an agent to a robot a 5 minute task we need more people who want to save the world. we do not need more leeches getting rich off a finite supply of iron ore (australians)

okay then, lets give backplane a spin tn

The fixation on local compute is greedy, hoarding and irrational. I considered buying a cluster for my work. But I can pay for a rental service that matches my workflow. I rent +100 GPUs on some days and 0 on some days. it comes out to the fraction of the cost and its utility is far greater. It only makes sense to pool together all the compute into efficient datacentres. This is simple economics. local AI is like private radio, there are niche and hobby cases for running your own walkie-talkie network. but for 99% of use cases you plug into a cellular tower. and for all the people making the case of "my lawnmower is only used once every two weeks should i rent it out". do you not literally see how it makes sense economically now? there's a reason why uber and airbnb exist.

robotics will not explode like ChatGPT. robotics is a: 1. models problem (YC hivemind overindexes on this) 2. supply chain problem (china FTW) 3. infrastructure problem this means that the adoption curve will be a lot more gradual. my favourite robotics companies are that see the bigger picture.

current state of agentic PCB design........... agents still have lots of trouble reasoning across many domains so it will overindex in one domain to optimise for another domain. eg suprisingly good HV routing but at the cost of totally fucking up the board outline gave it fairly trivial problem of board clearance: - my board (a 250V circuit that fires the solenoid) is 52 x 68mm. - doesnt fit inside my robot - 51% of the top side was parts + loose placement claude mapped every bit of free space under the main board at half a mm resolution and traced the board to it (oh hellll nah) highly recommend using agent nailed the hard maths. HV spacing rules, 3D interference checks against 869 parts. i normally get lazy and yolo it to JLC after a bit of napkin math.

i think even if ai overtakes me at routing pcbs i will still route pcbs recreationally - this is too fun

as someone with 1000+ hrs on KiCad/F360, i'm seeing how good agentic CAD really is. in one day, i was able to design an entire holonomic drivetrain. on paper it does: - 2.7 m/s forward, 3.8 m/s sideways (capped at 800 rpm on purpose) - 3.4 to 5 m/s² forward acceleration @ 0.6 to 1.2A - all on a 140 mm, ~975 g robot agent setup is basically opus 5.5 xhigh + fusion mcp honestly, i was pleasantly surprised by how elegant the hub wheel came out. the mechanical design has quite a bit of nuance. but sometimes it is completely irrational and overindexes on your prompt. - hub motors have good 3d printable tolerances - heat-set brass inserts without me asking - the motor mounts double as the main board standoffs BUT - originally put 1 o-ring per roller? told it to calculate approximate friction and then it corrected to 2 o-rings - pcb as a structural plate (??? bro) so far i rate it a 7/10. pleasantly surprised. now going to spend tonight cleaning this model and printing test fit components. i shall bring more agentic cad insights

how far can you actually push agentic hardware engineering? i'm building 2 soccer robots to test if an agent can reason across mechanical, electrical and software (inspired by robocup SSL + JR opens) i've spent 1,000+ hours in Fusion and KiCad. now i'm getting agentic, evaluating its work every step, then ordering it to see if the output is actually good, manufacturable and reliable. the entire 3D model you're seeing here and the electrical design were done in one night. here's the (somewhat vanilla) stack i'm experimenting with - claude code opus 5.5 xhigh - subagents run in parallel and do the work - CAD agent: drives Fusion 360 through its built-in MCP, builds the whole robot by script (108 parameters) - PCB agent: KiCad 10 + Konnect MCP, Freerouting, kicad-happy for design review, JLCPCB tools for fab files - firmware agent: PlatformIO + SimpleFOC - research agents: read SSL team papers and open repos (TIGERs Mannheim, awesome-rcj-soccer) so far what surprised me was how easy sourcing is - claude in chrome + a swarm of subagents found every component, checked live stock and pricing, and added it all to my aliexpress cart - i just checked the cart and hit order - i used to spend 3 days sourcing parts: spreadsheets, comparing listings, re-checking because aliexpress stock and prices keep moving - such a quality of life upgrade the mechanical parts are ordered (~2wks to arrive). i'll manufacture the parts, test it, and tell you what actually worked. i've built these robots before with no agents. i came 2nd nationally + 2nd in asia pacific. so you'll get the honest take from someone who's done it the hard way: - where agents are genuinely better - where they're confidently wrong - where you still need a human who knows the domain ideally i will be able to give u tooling + processes that work. the world needs to speed up its hardware design.

building robots has been my hyperfixation since i was 12. in 2023, i competed in robocup jr australia. i expected to win but came 2nd. to this day it's the most i've ever cried. sounds kinda silly when i say it out loud, but my entire world was winning this competition. i remember literally dreaming about: - custom BLDC hub motors - 4 layer pcbs with motor drivers, MCU, charge pump, BMS, sensors - my own YOLOv6 model running on an OAK camera - SLAM and inter robot communication but i couldn't fulfil this dream. i had no money so i'd sit on decisions for weeks. and my project management was terrible and unrealistic. i'll give myself credit for one thing tho: i really gave it my all. like to an insane degree. and there was no ai back then. anyways, 3 years later, i have a salary, agents, and i'm not scared to pull the trigger. so i'm building it. two mini soccer robots from scratch, playing a 1v1 duel on my floor. inspired by RCJ open/SSL. i've ordered the all the components for the drive train already ($300usd so far) and im now working on the PCBs for the encoders. i have a lot to do but im having so much fun now. its so insane how fast hardware design can be if you have agents and money. lets push the limits of the "agentic hardware engineer" this one's for 16 year old me documenting the whole build here. let's see how far i can go with $2000 budget and 2 x claude max accounts.

if you took a snapshot of my brain pre-alloy robotics vs now it would look very different

I HAVE FUCKING CRACKED THE CODE TO TOP OF FUNNEL MARKETING FOR @AlloyRobotics this is going to be fucking crazy get ready for a gtm rocket ship this is how i feel

muse wonder, nvidia open source cad kernel, new dragonball super episode… the world is truly wonderful.

alexandr wang is the pinnacle of asian american excellence