Pinned Tweet
started WaterlooAGI
we do experiments, evals, and research about AGI from different directions
my first experiment was on noisy vs structured corruption on LLM dreaming
looking for researchers to exchange ideas and sponsors with spare compute. dm me 🔎
links below :)
finding your passion is way harder than finding out what you're good at.
in the 20 years I've lived, I probably experienced only 10% of the world. My passion could be in that 90%
but, what I'm good at is just what I spent the most time on within my 10% of experienced life
11
352
Waterloo students on left, employers on right
Hypergamy or “desirability” hierarchy still exists everywhere
friendly PSA that UWaterloo isn't the holy grail of technical hires
we fired our first waterloo "intern" within three weeks after they stated wanting to "solve interesting problems" but proceed to senselessly bash out prs with codex
during standups they were minimally present and could not answer technical questions without first consulting their agent
this isn't to say students from waterloo are inept (and there are some I do respect) but definitely overstated from people who've never worked with them before
went all in on building in September. this is what I got for skipping all my lectures
- shipped to real users for the first time. 1.2k installs by sep 30
- learned distribution: first GTM reel got 36k views in 24 hours with 500+ on waitlist in 5 days
- grew personal brand by 50% followers
- started WaterlooAGI and shipped a second project
- built 5 more side projects: idea to MVP overnight pipeline, a Google Meet extension to explain corporate jargon, a blog post, a resume ranker, a telegram bot with my context 1 year from now
my biggest takeaway is the magical feeling: 11 pm, dark room, warm desklight, three screens, and dark cello music in the background.
Everyone knows the next moat is the attention moat
And the most effective hooks are “brainrot” hooks. Movie clips, pretty women, people doing crazy things etc
It’s time to get in on the game
8
285
while Jensen's taking the headlines in the past few days, @nvidia was dropping "World-Action models" that seems to combine world models and policies
It doesn't matter how ugly politics get, the company itself is still doing the craziest frontier work
1
9
365
Usage-based LLM providers are basically AYCE buffet owners
but they have to hard cap what you can eat because of how fast their customers can digest
2
144
you pay about 75% more for a gpu when you rent it per hour
an h100 is about $3.49/hr on demand and about $2 on a 1 year contract cus you're buying "in bulk" when it comes to GPU usage
so if your usage is flexible, it's actually better to pay per token from a company that keeps its gpus full across many customers (abstract, together, hyperbolic etc)
9
333
i thought the compute market is commoditized and people buying/selling it are dumb
my arguments:
- open weights are free, so everyone serving them sells the same product
- frontier labs keep shipping better cheap models on top
- you're playing an undercutting game
but it turns out tokens can be high-ticket
learned that companies pay for committed capacity, low latency, SLA, private deployment, fine-tuning, SRE
kinda cool you can 10x - 100x your price with the wrappers that come around it
but again, should've seen it coming since the 2024 GPT-wrappers
9
202
how i feel when i get a few hours to read and write while waiting for my Claude usage to reset
5
306
how do you guys run Claude subagents for a long time compacting the main orchestrator?
/compact stops my background agents with no warning
1
3
120
Made a better ux flow for running multi-agent sessions on your phone!
Scroll through summaries and actions-required, see details only when clicked, and start/end sessions
It works when my laptop’s off too
1
6
252













