I’ve shared this AI take in person to many people, but haven’t written it down. I like my takes to be falsifiable, so you can all point out if I am wrong in the future.
On AI, I think:
The moment the big labs’ valuations grow more slowly (or even decline), they won’t be able to justify more massive, expensive, training runs… and then:
The model capabilities will plateau (as very capable models). Open source (distilled) models will catch up quickly.
The massive amounts of compute currently being used for training will flood the market and inference costs will go down, speed will go up. (cheaper inference for us!)
With cheaper GPUs, we’ll see more artisanal model companies fine-tuning on specific (creative?) capabilities.
The big labs, being extremely leveraged, will struggle to compete with late-movers who have fewer debts, cheaper compute, and more focus.
The “mega-corp” AI conglomerate of pharma, robots, etc. won’t materialise, because if Anthropic will hire you to run it, you might as well start your own and not be locked to one model family and keep all the upside.
Not sure what timeframe this happens over, but it’s my current guess at what will happen.
I think this would be a broadly pro-human outcome that leaves us all with interesting and fun jobs in the long run.
The thing that made launch videos valuable was that they were expensive to make, and so signalled high levels of thought and importance.
Videos no longer signal that, because for the last year or so they’ve been very easy to make.
What signals do you use to assess the importance of something nowadays?
The extension of this: I’m not worried about AI slop content because everyone gets used to it quickly and then starts to ignore it and then it doesn’t work and then people stop making it.
Software Factory?
I prefer Software Workshop.
Yes, all the tedious and repetitive bits are automated (à la drills, mills, etc), but there’s still plenty of space for artistry, problem solving, and skill.
Just wanted to defend the term "Software Factory" for a second.
First, the caveats:
It's bullshit. It sounds fancier than it really is.
It's a bandwagon. People are using it to get you and their VC's hyped.
But it's also referring to a real change.
Earlier this year, most SOTA teams were using agents to work on the product. Shipping features, fixing bugs.
But now, they're building automations to help agents work on the product for them. They've moved up a layer.
What do we call the thing they're building? Loops? Workflows? Some other generic label?
"Software Factory" is as good a term as any other.
- "We're not working on the product, we're working on the factory that builds the product."
- "We're not product engineers, we're factory engineers."
I like it. I'll continue using it.
@MeirDick This is interesting. Feels like a mix of data governance and analysis workflows?
I wonder if these features would fit better into a broader product rather than a narrow growth only one?
Since I migrated all my projects from Vercel to Cloudflare, I've really missed Vercel Analytics.
Cloudflare's web analytics isn't enough to answer even basic product/marketing questions, so I ended up building a mini Product Analytics tool that runs on Cloudflare alongside my projects.
It's agent-first, open-source, supports session replay, experimentation, cohorts, error tracking, all sorts of analyses, and is basically free to run on Cloudflare.
I figure I might as well share it publicly - hopefully it's useful to someone else!
https://t.co/Zfm96rHf0D
Fun story: I built the first (unreleased) version of Amplitude’s session replay functionality as part of a hackathon project.
A couple years later (2023ish) I was asked to help design the version that ended up in production.
I pushed for it to be built on Cloudflare, and even made a prototype to demonstrate the insane cost effectiveness. Alas I lost and it ended up being built with the standard COGS profile that all the other providers have today.
Since I migrated all my projects from Vercel to Cloudflare, I've really missed Vercel Analytics.
Cloudflare's web analytics isn't enough to answer even basic product/marketing questions, so I ended up building a mini Product Analytics tool that runs on Cloudflare alongside my projects.
It's agent-first, open-source, supports session replay, experimentation, cohorts, error tracking, all sorts of analyses, and is basically free to run on Cloudflare.
I figure I might as well share it publicly - hopefully it's useful to someone else!
https://t.co/Zfm96rHf0D