What I’m working on at Preceden
I’ve got this thing where I’m never quite satisfied with Preceden‘s design. It’s not that bad, but when I compare it to a site designed by professional designers like Help Scout or Stripe, it’s clear that there’s a large gap, and it’s not one that I’m talented enough to close on my own.
And so this week I posted a job on Upwork to try to find a designer to help improve Preceden’s design:
I was optimistic that this would receive a lot of applications, but it didn’t work out that way. It received maybe 10 applications, out of which I’d say 9 hadn’t tailored their application even one bit to the job description which is an immediate red flag for me (even just a short “Hey, working on Preceden sounds interesting, here’s a bit about me…” goes a long way). The 1 that did wasn’t a good fit. In the end I did make an offer to 2 people (so I could work with them for a short project then pick one to work with longer term), but one declined saying he was too busy and another wanted a min $3k initial commitment which I wasn’t comfortable with.
At Automattic and now Help Scout, there are some incredibly talented designers who can create mockups and then take that design and implement it in HTML/CSS within the codebase. But this hiring experience reminded me that those people are rare and in high demand and probably not looking for short projects on Upwork.
It’s more common for different people to tackle these tasks: one to create the design, another to implement it. For round 2 of hiring for this, I’m going to try to split it: find and hire one person (probably on Dribbble) to create the mockups, and hire a separate front-end developer to implement that design in Preceden.
What I’m working on at Help Scout
I shipped a fun little experiment that lets people at Help Scout use GPT-3 to automatically draft a response to customer conversations within the Help Scout editor:
To be 100% clear, we have no intention of actually building this into the real product. It’s meant to demonstrate GPT-3’s strengths in weaknesses. If I had to boil them down, I’d say that GPT-3 is great at generating text that looks like it was written by a human, but at least in this support context, the responses are terrible because GPT-3 has no knowledge of the product its trying to answer a question about. Maybe down the road there will be some future version of this technology where it can be trained on your support history or docs, but until that happens, GPT-3 probably won’t be augmenting customer service reps in this way anytime soon.
What I’m studying
Still chugging away at Kaggle’s old housing prices prediction competition, managing to break into the top 500 out of 4.6k:
My approach to this is to make as much progress as I can on my own, then peak at others’ publicly shared notebooks to get ideas for how to improve, then iterate on my own notebook accordingly. I’ve been learning a lot this way and probably stick with this competition for a few more weeks until I feel like I’ve maxed out how much I can learn from it.
What I’m watching
Star Trek Discovery:
Unlike most of the older Star Trek series which basically consisted of the ships exploring and getting into different adventures each episode, Discovery (and Picard) have an single story arc that’s told across each episode. It’s really good.
What I’m reading
The Almanack of Naval Ravikant – a summary of his podcast interviews and writings. I’m a huge fan of Naval. If you’re new to him, check out his Farnam Street interview which I’ve listened to several times.
That’s it for now 👋