November 15, 2023

Inside OpenAI: How does ChatGPT Ship So Quickly?

Source: Inside OpenAI: How does ChatGPT Ship So Quickly?

Summary

ChatGPT ships fast, but it obviously isn’t the first company that did that. I remember Facebook popularized “Break things, Move fast”.

One key takeaway here, I guess, is how this can be an antithesis of the remote work movement.

Although I’ve been a remote engineer since 2018, ever since my former boss Mike Leach sold “remote-first” work environments to me, I still resonate more with in-person communication as a much more effective tool for collaboration and getting ideas off the ground.

Notes

Our goal with this small ChatGPT subteam was to create the atmosphere of an early-stage startup iterating towards product-market fit (PMF.) We wanted to foster the rhythm, pace, and autonomy to do this. Every member of the team was on-site, and we rearranged seating to put people next to each other.

I remember my days with XSplit where this was also the setup. This approach of having teams operate as a new startup within the company might not be unique, but it certainly is unheard of in large enterprises like Fujitsu.

One hard pill to swallow with this setup is when PMF was never realized. I personally felt that when I used to work on the Player.me Desktop App project.

To prevent this, product teams like ChatGPT have software engineers, designers, product managers, and researchers working together.

Back when I used to work for Dealtale, I felt that there was a disconnect between the data team and the product team. As Dealtale is a data platform company, our product itself is how we present the data to our customers. This is where the similarities with ChatGPT is, where with ChatGPT, the product itself is the product of the research team. For dealtale, it is the product of the data team.

However, issues reported related to our product that were actual data-related issues were usually thrown over the wall to the data team. This made it feel some sort of disconnect, as although there were clearly defined responsibilities, I feel that it would have been a lot more effective if teams were built around “features” instead.

A change of team composition was actually initiated by Alon Caspi, but maybe it was a little too late as we were nearing the end of Dealtale’s existence as a functioning company by the time that idea was introduced.

Clear focus is always a driver of velocity. I’m convinced our broad mission has helped keep that focus and also pave the way for lots of new ideas.

Startups that start off with a clear focus but then veer towards unclarity are usually the startups that eventually fail.

I’m quite sad about what happened to XSplit after a few years since I left the company. The focus shifted multiple times, that although it’s important to quickly adapt to changing trends… I feel that it started losing the market because it focused on many avenues at the same time.

It went back to focusing on Gamecaster a little too late when Streamlabs was already around.

You know what, by the end of 2014 there were ideas being floated around where we wanted to integrate Strexm with Gamecaster. That was four years before SLOBS was first released!

  • Strexm already had a working product.
  • We already had Gamecaster.
  • We made a prototype where we can control the camera position based on Strexm’s camera overlay towards the end of 2014.

In hindsight, I felt we missed a perfect opportunity there. We should’ve focused more on our core products with the focus on providing value to users of our target platforms.

Uncoupled and incremental releases

Pretty standard nowadays, esp. with Facebook’s famous “Move fast and break things” mantra.

Senior teams can ship very, very quickly.

Most startups do this. Hire seniors only, for a similar reason, but mostly because startups do not have the time, money, and leeway to hire juniors and train them.

A sense of humility is one of the major drivers behind our universal “Member of Technical Staff” title.

I do want to see a company that sticks by this. When I first started with XSplit, it was pretty flat… but while XSplit grew, things changed too. People expected vertical progress in their careers, and that included official job titles.

Glad to see OpenAI is able to stick by it even though they’re already fairly big.

The company I’m currently working with, Inspectorio, is still a startup at the growth stage but already has fancy titles.

On one hand, career progression which would reflect on their job titles is always a tool for motivation. It also clearly defines what is expected for each individual level.

On the other hand, chasing for the titles by creating crazy solutions, overcomplicating things, and selling them like it’s the elixir of life is waste masked as “progress and value”.

Working from the office Monday-Wednesday.

Quite the opposite from the recent articles I’ve been reading about async communication and remote work.

We did note the advantages and disadvantages of remote work and async communication, but it’s nice to see the value of working in the same room are not refuted (unlike most things in our times where previous facts are forcibly refuted by saying it’s “not up to our current realities”)

We also did a lot of ad-hoc whiteboarding. We’d get a lot of ideas over the lunch table – literally!

This! This is what I miss the most!

Tools like Slack do have huddle, which does solve this for the most part, but there’s something different when you’re at one table discussing back and forth, and others can overhear you. There were multiple times when we came up with a solution when someone not invited chimed in because we were talking over the canteen table.

Thursdays and Fridays are no-meeting days for several people.

Yes!

Are standups also removed during these days? I’m thinking of async standups… some sort of a slack bot kinda thing. Might be more efficient.

Final Thoughts

I do not think I have the sufficient influence to change things drastically, so this article mostly just reinforced most of the ideas that already resonated well with me.

No fancy copyright. Just creative commons | There's some vanity tracking going on, sorry | RSS.