Comment by echelon
3 hours ago
We're getting such powerful models we'll likely be able to replace all of Google soon.
- Google Docs suite - easy, probably requires less than $10M to duplicate the entire set of functionality, including all the enterprise and reporting features
- Gmail - same
- Google Search (classic, not LLM-answers) - probably easy to do now, the challenge is getting past Cloudflare
- Android - vibe coded hardware is coming, but we're probably 5ish years out.
- YouTube - probably one of the hardest, due to network effects / distribution
- GCP - hardest, due to the infra build out. But neoclouds are rapidly growing.
I can't think of anything most big tech companies do that won't be put under threat in the era of personal software and rapid development.
It's ironic that Google invented the transformer and it seems likely that it will undo the empire they've built as well as all of the moats in the world that aren't distribution / community based.
> - Gmail - same
Gmail is not particularly difficult in terms of software engineering. The moat with email servers is IP address reputation.
I could, today, install Postfix SMTP and some IMAP as well, and watch my email all go to spam directly, if delivered at all (ISP might block them).
I use a reputable email provider (Mailbox.org) and my emails often end up in the spam of Gmail users.
FWIW, I've done exactly this + DKIM & DMARC on a residential IP and was able to deliver to Gmail & Protonmail with no problem, but didn't test it long term. Same with a server on AWS. I'm not sure if self hosting email is an overblown problem or I've just been very lucky somehow. Obviously it's still a big issue if for ex. Outlook's spam filters happened to be slightly more sensitive and dropped my mail; I never encountered issues but eventually I just used a custom domain via Proton for peace of mind.
Not so bad for a single person. But once you start being an email provider, you get bad actors using your service to send spam emails. Which can cause all of your sending ips to blacklisted at once.
Agents require training data for RL. This data is rare in comparison to what we feed foundation models, and Google is sitting on a dragon's hoard of sensor and tracking data. I'm not counting them out yet.
Okay but would it kill everybody to start with commoncrawl?
We don't need Google's data. Once we instrument the world, we'll get a Google's worth of data in short order.
It's easier than ever to ingest and label data.
This is happening with or without Google. The recursive improvement does not require them at all.