How do we ride the vibecode rather than die?
Principles from predicting the competitive landscape of AI startups and software engineers.
We’re all acutely aware of it - AI is going to change software engineering. Already, I just met someone who built an entire AI TV show - just from vibecoding. Me - as a senior engineer, I am worried about AI taking my job. So how can we plan for this?
Trad software competitive landscape (employees).
I’ll preface this by saying - some software jobs are just safe. These are:
security professionals,
extremely low-level / highly technical fields where the data is proprietary and thus is not captured by public AI training (industrial systems, FPGA/Verilog code, secret government stuff)
medical device coders
For the majority of tech work though (frontend, backend engineering), AI will be able to do your job. The work is commoditizable, the data is there to train on. There will still be roles - though the workers will operate at higher-levels of problem-solving than writing code. The majority of modern BigCo tech workers don’t do a lot of work anyways. Companies have stopped hiring juniors en-masse. There is a lot more room for disruption in companies where the product has plateaued. The rich companies will fuck around with the huge capital reserves they have, and maybe survive.
I don’t care about predicting the future of those jobs. I want to figure out the future and then create the jobs.
Thinking about AI startup competition.
Here’s my position:
Building a company around waiting for all the fantastic tech to come out every 12mo is actually sick. For example, I've been waiting for a realistic realtime voice since 2024. Sesame dropped last week and that was so useful.
It’s the same with 3D model generators - 1yrs ago I was building 3D head generation for Rick and Morty AI. It cost $2/gen, 5mins, and has a good-looking output rate of 1/8. Now it costs $0.20c, takes 1mins, and all the outputs look great. The next big one is auto skeleton rigging which is about 6mo away, and will change gaming a LOT.
Probably in 6mo, you will have a lot of very cool 3D vibe games made for cheap. These are like market tests for new products. The better ones will accrue audiences/distribution.
Ultimately you can't defend on product/tech alone too easily in the limit since AI will eventually self-improve its learning loop for software (RL play). But there is moat. Big labs don't have the capacity to build products fit for every use case. If your product gets better with every new tech, then your dataset is extremely valuable.
Being the best product derives from the best tech (open source) + best dataset to retrain/fine tune on which derives from UGC (user generated content) which derives from audience which derives from initial distribution.
Midjourney one good example here. Best product for stylised image gen. Eventually maybe this will turn into M&A when the most profitable companies are able to acquire and merge good products and data. Eg. PhotoAI acquired by OpenAI. But it's far away yet.
Principles for a defensive AI startup.
Building a company around waiting for all the fantastic tech to come out every 6mo is actually sick.
Vibecoding allows people to test some ideas in the market at 1000x the pace. Video games / entertainment are one such idea.
The best vibecoded ideas will accrue audience/distribution.
You can't defend on product/tech alone too easily in the limit since AI will eventually self-improve its learning loop for software and replicate everything you do.
But there is moat. Big labs don't have the capacity to build products fit for every use case. If your product gets better with every new tech, then your dataset is extremely valuable.
Being the best product
…derives from the best tech (open source) + best dataset to retrain/fine tune which
…derives from UGC (user generated content) which
…derives from audience which
…derives from initial distribution.Eventually, this will turn into M&A when the most profitable companies are able to acquire and merge good products and data. This has already somewhat happened with Character.AI (founded 2022), albeit only for the LLM chat tech.