Applications for Keel’s first ever winter batch are now open!
Keel is the NYC accelerator, home to the most elite dropout and recent-grad founders building in New York. After three summer batches, we're doubling down on NYC and our mission to build the best founder ecosystem east of SF.
Keel Cohort 4.0 is:
🏙️ Hosted in NYC. Winter 2027. 12 weeks in-person at our NYC office.
🦄 8-10 teams. Campus + recent-grad AI founders.
🚀 $150K checks. Credits and resources. Founder friendly terms.
We're partnering with NYC's best founders, investors, and enterprises to build an unfair advantage for the next generation of unicorns. Read more about what we’re building, and apply to Keel Cohort 4.0 at the link below:
https://t.co/X8i7uzbeKj
In startups, people will deliberately be pessimists and shutdown your vision.
It is important not to view them as assholes but to recognize they are testing your confidence in your ideas and your commitment to them.
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
https://t.co/r0qFLdvd6c
just shipped my first new product since i shut my startup down!
if you know my previous work, this is in a similar vein - AI media, social and quite weird.
not quite ready to share the name publicly but would love you to alpha test ♥️
comment or dm for the dotcom :)
there's a second half of the data market that I havent seen people on here are talking about
one side sells data that already exists: you know this as the data brokering market
the other side pays people to make data that doesn't exist yet, on purpose, for the things models are still bad at
labs have used up the good public internet. pretraining (the model reading everything) is largely tapped, so the gains now come from post-training, which works a lot more like coaching
and coaching needs phil knight level coaches
experts sitting down and deliberately producing the material a model learns from is the other side
some write demonstrations. taking a hard problem from their field and solve it start to finish, showing every step of the reasoning. that's what supervised fine-tuning runs on (SFT)
some do preference work. the model gives two answers, the expert picks the better one and explains why. that judgment is what RLHF learns from
some write rubrics, the grading criteria a model gets scored against. a vague rubric teaches a vague model, so people who can spell out what "correct" looks like in their field get paid well for it
and some spend the whole day trying to stump the model. writing the questions it gets wrong, finding exactly where it breaks, and handing the failure back as training data
generalists can't do most of this anymore since a lot of the easy material has already been made.
what's left lives in the heads of people who've done a job for 10+ years and labs are eagerly buying this
handshake is the cleanest example. it spent a decade as a college job board, then started paying PhDs, doctors and lawyers to make training data for labs
they went from under $10M to close to $1B in annualized revenue in a year
100s of billions in funding and dealflow are going here, because every lab needs it and there's a limited number of people who can do it well
if you know a field cold, start by finding where the models fall over in it, then find a way to monetize it (aka dm me)
every new market creates countless opportunities for participation. this might be your way in