andrew chen

for founders in Boston later this month - we’re hosting a few events for new startups + lots of stuff on the calendar too 👇🏼

@desno365 yeah thats def a major downside…

@andrewchen I am. the main issue is that only verified apps can access gmail and drive, so it will be much harder to set up openclaw/hermes to access your gmail

dear lazy web who’s on the Google “Advanced Protection” program and whatcha think?

(This was my reaction after watching the Google keynote)

is agentic coding in terminal vs in the desktop IDE like the new “emacs vs vim” debate?

@founderengineer What are the most annoying bits…

@andrewchen I am and iCloud advanced protection as well but as usual you are sacrificing convenience

@createdbymason Looks awesome

@andrewchen i'm hosting a hackathon for NYTW! partiful.com/e/Ve6Zk5IVb4UQ…

who’s in boston/nyc later over the next few weeks? The team is hosting a bunch of stuff alongside portcos, partners, etc — May 26-June 7 Boston Tech Week, followed by NY Tech Week right after. This is going to our biggest Tech Week ever:

- We have over  2000+ events - 15+ tracks - infra, founders, engineers, hackathons etc - 50+ portcos participating: OpenAI, Elevenlabs, Deel, Gamma, xAI, Stripe and speedrun companies too

Enjoy! youtu.be/PHQvb10vKyk?si…

@emilson_tery Do you have a fancy router? Or just use a diff app?

@andrewchen For me, ~80% of daily LLM queries are already local — summarize/extract/search/classify on Qwen 3 4B Q4, Windows CPU. The remaining 20% (hard reasoning, code gen) goes cloud. The shift isn't "when" anymore; it's distribution. Most users don't know the stack exists.

How soon before a real % of LLM queries are done via local AI models running webGPU in-browser, and are never sent to the SOTA model in the cloud? Couple things that might drive this: - you don’t need a frontier model for everything. A very large % of LLM queries are simple, google like queries. Easily handled - local models are getting really good, and getting better - a lot of consumer hardware (particular Apple!) can already run good models pretty well. Newish mac laptop running qwen 3.6 35b MoE LLM at great speeds. Hardware is going to get even better in this direction - there’s def some use cases where people will care about privacy. Health, financial, adult stuff etc. - nice part about browser/webGPU is that there’s no install. It’ll just work. And alleviate compute costs Of course the tension for this is that we’re just going to build a shitload of compute in the world, and tokens will get cheaper over time. Yet it seems like the demand is also so crazy that unlocking a bunch of local supply will be worth it too

@wkoszek Yeah seems like the best you can do rn with coding harnesses is to point it at a liteLLM instance, and then configure the routing there - since liteLLM can also send to Claude and GPT, in addition to local

@andrewchen Claude will become super tool if it can do some minimal amount of local dispatching. Perhaps even just ! command improved would be a good start.

@PaulGugAI We’ve been producing energy in our homes for a long time. Stoves, car, heater, etc! So yeah the analogy makes sense

@andrewchen The analogy of solar panels come to mind.. ie energy demand is going up but home owners putting up panels to save on costs also helps to alleviate demand over time.. win win.

@hanzi_li Just need to have really smart low latency routing to figure out where to send it to. I run liteLLM pointing at a fast vs big model but the logic for when to send isn’t great

@andrewchen yeah the annoying sweet spot is boring local stuff. classify this page, summarize this inbox, draft the 80 percent answer. cloud only when it gets weird

Yeah the counterpoint might be: - TTFT is often faster on frontier than local simply bc it’s on specialized hardware - consumers say they care about privacy but then they just give data away as long as they can use a product for free But yeah I still mostly agree w you on the outcome

@andrewchen I think due to latency and expense as well as privacy it makes sense for the market to bifurcate into local and big datacenter stuff and local models will be more specialized/fine tuned for specific tasks.

haha - Shoutout to the “AI is moving too slowly” group!! My people

@aloobhujiyan I’m so curious what the Mac Studio M5U 512GB is gonna cost…

I already have one. Yeah, markets are tough right now for Mac studio. eGPUs are great, just hard to connect to a mac, which is my personal device. I did build my own 2x3090 GPU rig as well, and its just keeping it powered on and stable while running linux, is a full time hobby that I chose not to continue and switched to modal.

finding the main downside with experimenting with local AI models is that you end up buying one GPU, then another, then another, then another… But I’m running qwen3.6 27b dense at 100 tok/s now on a 5090 eGPU! It feels like sonnet 4.6? Fast and highly usable I figure the GPUs I have will now increase in value over the next few years so it’ll all be worth it

@kidxbt Aorus 5090 AI box My main complaint is that it only runs via windows and sucks at Linux and couldn’t get it work on Mac via tinygrad/tinygpu. But at least now it’s working. I’m sure an actual GPU is much faster but this has its own charm

@andrewchen what enclosure are you using and how’s the bandwidth/latency holding up in practice?

@SamJWasserman I do have two sparks but no I haven’t tried it even though I got the QSFP+ cable, just bc the big models are all super slow tokens/s… but I should try it eventually anyway just for fun

@andrewchen have you tethered more than one together, for example two sparks to then have 40 cores and 256ram? With the high speed cable nvidia provides? thats what I'm trying to optimize for right now ideally.

@KashPrime I just have one eGPU and yes I have a PC for it running windows (Linux+eGPU is unfort a mess rn)

@andrewchen Do you get a separate PC for each eGPU?

@cchevyyc The application process is the actual process, but yeah, lots of adjacent events also for folks who want to network and meet investors too

@andrewchen If we’re going through the application process for speedrun, should we also sign up to pitch for the event on June 1st?