- Why not to start an AI startup:
- lots of competition makes differentiation / sales hard
- young, unstable field which changes rapidly
- ML can be tricky
- OpenAI etc have massive headstarts
- uncertainty about future capabilities (x-risk, obsoletion, AI winter)
- Why to start an AI startup:
- AI summer about to start / started
- even if capabilities plateau, current models can do so much
- this is a major movement, like the introduction of the internet
- lots of money and interest around
- potential for greatness, robot king
- very philosophical, not boring like brex or resend
- Why I personally should not start an AI startup:
- not super sciencey
- never ran a successful company before (do something easier!)
- sensitive, lots of negative feelings ahead
- hate AI hype and charlatans
- young, somewhat inexperienced
- would obliterate my free time / freedom, no musical theatre writing
- Why I personally should start an AI startup:
- (all of point 3, lol)
- very very enthusiastic about ML
- somewhat wide breadth of tech knowledge / moderate ML skill
- want to put youth, time, energy into something meaningful (and make money)
- possibility of leading to other interesting things like robotics, politics
- What kind of AI startup would I not start?
- compute is a no-go, hardware is not startuppy and software optimisation is too hard
- fad company (like dreambooth avatars) is fun but not long-lasting
- B2C plays are not easy
- not interested in something that only peripherally relates to AI
- probably not API over open source as that races to bottom on pricing
- don't wanna be tied into one field like law or drama or automotive
- would prefer acquisition by OpenAI than acquisition by Microsoft (for example)
- can't exclusively hang on one model / one type of model
- tooling around LLMs etc seems a bit boring (e.g. prompt optimisation)
- What kind of AI company would I start?
- downstream application could be fun e.g. GPT-4 colleague in Slack / Google Meet
- modelling is interesting, lots of interpretability challenges, bit researchy
- data is the most important side in the end, data enabled ChatGPT improvement
- Angles to a data-focused AI company
- getting high-quality labellers is a challenge, where are those people?
- rise of synthetic data, humans possibly not needed for much longer
- companies have so much data tied up and not much clue how to utilise it
- strong enterprise aspect means birds-eye view of industry at large (good strategy)
- dataset quality is really important, want to be confident there are no false positives etc
- is there a data + modelling simultaneous play? i.e. user simply uploads and selects desired end goal
- active learning seems really strong, makes for good demos at least
- possibility that massive open source models removes need for small specialised models
- I like the synthetic data / AI-labelled data angle