After 6863 days, yesterday was my final day at Google. Two months shy of 19 years.

Grateful for all the opportunities. The opportunity to travel the world, to work on the first iPhone app, to bring people together, to work with wonderful people, to redesign Google, to work on my favourite products and to eat all the free food.

There's no next thing, I'm retiring to spend time completing my side quests.

When I think about how western and japanese culture differs, I think about this contrast. In Tokyo, you rarely see a pothole. Contrast this with NYC, where nearly every street corner there's a pothole, some big enough to be ankle high.

Why?

I am unable to do test driven development. I am unable to write a complete and perfect design doc and implement to spec.

I can only write code incrementally, breakdown the problem. Work on the next thing, test it, then do the next thing. Nearly none of my successful projects involve a careful plan.

Hence, whenever I read things about using Claude code to do planning and produce a spec. I cringe. That's not how my mind works and that's not how I use coding agents.

I simply have a paragraph of thoughts of what I want the thing to be, and then go back and forth with Claude code or Antigravity about what the output is. I can't specify the full spec, just things that have to do with the detailed implementation. So maybe this is depth first thinking rather than breadth first thinking.

There were some really fun hacks, but as the host in the Tokyo session noted, the hacks that have the highest chance of winning are usually games or game-like. Two out of the top three hacks were games.

The answer is YES. But not in the way you think.

For vibe-coding (agentic coding, code completion), I use surpisingly little compared to my other projects. But there are a few things that has sped up things by 10x.

I vibe-coded some image editing and extraction tools (both Imager & Image Extractor) that lives on aversion.one. This is the kind of thing I hope less-technical and even non-technical people will start to see as a new way of doing work. It's like you could buy a shovel, but what if you could design and make your own shovel that only shovelled exactly how you wanted. I needed to background remove, cope and scale 600+ images, so I made an image editor that did just that. The UI is terrible (an area that is improving rapidly), but to be able to make vibe-code your own macros to make a 5 min job into a 5 sec job, that's a win.

All the real time data that is coming from ski resort websites is all crawled using a two-year old crawling system with Gemini 3.0 Flash (previously 1.5 flash) and taking screenshots of websites. Last year, it took me 30 minutes every day to go through a handful of websites that I couldn't automate to update the snow fall metrics, but this year, it is nearly 98% automated with me double checking the work. It still requires a lot of custom tuning, for instance, I store optional custom prompts for each website because sometimes the AI really doesn't realize that a snowman icon means snow fall.

Another AI usage is with translations -- the site is bilingual, in both Japanese and English, so I've created a specific translation endpoint in the admin interface that I can pipe any text into it and it will return it in the other language. It's helped cover a lot of holes, but I also do know that sometimes it does a terrible job.

The main principle here is that it should be functional as a website and also (quasi-)functional as a command line but primarily for operating this website. This is based on an earlier Terminal prototype I vibe-coded a few months ago, but then subsequently extracted, and more or less rewrote by hand to fix a bunch of layout behaviors which was really hard to get Claude Code or Gemini/Antigravity or Codex to get right.

The code knows the structure of the website, and converts that into a virtual filesystem that exists in JS. The main commands that were necessary was cd and ls. Each page is a directory (no files), and if a cd command was detected, is does a proper page navigation using SvelteKit's goto(). As a bonus, the output of ls is slightly marked up and instrumented so clicking on the links will emulate typing the command in and navigating.

After moving to Japan and discovering there were resorts apart from Niseko and Hakuba, I wanted to know more about these slopes. I wanted to go to check them out. Not long later, I stumbled across a magazine that claimed to list all 430 resorts in Japan. I bought that magazine and proceeded to manually write down in a spreadsheet every ski resort there was.

I wanted to know more. I wanted to know how big it was, how popular it was, how much snow it was getting. So I began to collect this all in the spreadsheet, it grew and grew. Soon it outgrew the spreadsheet and migrated into a database.

The more data was in there, the more data I wanted. It expanded to course maps, it expanded to GPS traces, it expanded to even webcam images. Today, it is a 4GB database with 1.2TB of images/data and code base of 800+ files.

What about that magazine? Every year a new edition of this magazine was released and every year the number of total resorts would go down. This year, the magazine reported that there are now 391 ski resorts remaining in Japan. Snowline does count a few more because of the way we classify what constitutes a resort.

I'm building this because I want to use it. If only one person, myself, found it useful, it would of been worth it. The act of building it helped me understand these ski resorts more, find these small ones only a few people know about and plan these weeks long road trips to hit these resorts to fulfil my goal of riding all the resorts in Japan.

I've vibe coded (100%) two Android apps, and I don't really know Kotlin and the Android SDK. I know enough about high level frameworks like Jetpack Compose, but I barely know my way around Gradle, AndroidX and whatever.

I have one app that has a weird DNS resolution bug, but only when it's on mobile networks, not on Wifi. Asking Claude Code to first build me a system to queue the requests I make and do retries on them, then expose them in the UI on the app after a few retries so I can debug what is going on, or at least retry them when I get back on Wifi. It did it in two-shots. One was jme under-specifying what I wanted and second was to fill in the gaps and give it UX advice.

Switching over to a NodeJS app, this one I had wrote most of it but needed it to rewrite a bit of text parsing code, so just chucked it over to Antigravity and it wrote a comprehensive test suite, allowed me to give it certain failure cases and within 2 minutes it had updated the code, improved the tests and created the PR. Amazing.

In the first phase, the task was to extract the current recorded snow depth and the new snow fall. Essentially parsing the page for two numbers. gemini-1.5-flash, gemini-2-flash-lite and gemini-2.5-flash-lite all performed pretty admirably with this task. The different versions improved and got fewer and fewer mistakes (going from many 70% to 90%). Over the year, without me doing anything, I had to correct fewer mistakes.

In the second phase, the task became looking at the page and looking for indication to know whether the lifts are opened or closed. This really stumped gemini-2.5-flash-lite. It could not figure it out as easily as the snow depths. As soon as I switched to gemini-2.5-flash, with a few prompt improvements, it consistently was able to extract it.

It is quite obvious the capability jump when a larger model is used. It is also obvious in cost between flash and flash-lite, nearly a 20x cost increase, with my inference cost per day going from ¥22 to ¥600.

I've been using the Ricoh GRIII for the last month and having a lot of pleasure shooting with it. The GR World app and the Image Sync app have a 1-2 star review average on both Play Store and App Store. Sure enough, the connection is flakey and constantly fails. HOWEVER, after some searching, it turns out Tailscale on my phone was messing with it. Once I turn off Tailscale (VPN), it had no problems connecting to my Ricoh.

After about 1GB of transfers for Github Packages, it will "run out of quota" and you'll get a 403. I thought it was an authentication issue, so I racked my brains on why the Github Access Token was not working. But an authentication error will give you a 401.

In my career, I can only truly say that I was doing what I loved in two periods. The first was being Google's first iOS app developer, back in 2008. The second is now, building and leading a prototyping team specializing in LLM and ML.

Both projects where things that I was secretly doing before, but I had known that those were the things I wanted to do. I'd spend off-time working on it, sometimes sneaking these into my work. Both things I had been doing without knowing what would happen.

That's pretty much it. Figure out what are love, get good at doing it, form an opinion on how to do it as good as you can, as fast as you can and if you're lucky, it'll become a thing people want.

There are many things I've tried doing on the side that have lead to nowhere, but consistently, the things that break out for me are the ones that I believed I would do even if I wasn't getting paid to do.

[[Warabi Mochi]] at [[茶寮翠泉 高辻本店]]. Warabi Mochi is my all-time favorite Japanese snack. It's a transparent slimy version of mochi that is slimy, works best when dipped in roasted soy bean powder. [[Patricia Poon]] took me to this place where they serve it fresh and hot. The Japanese name for this is 出来立てわらび餅. It comes out steaming hot. You get a paddle and chopsticks and have fun eating it. This is super good, I'd go back there again in a heartbeat.

[[Ace Hotel Kyoto]] is lovely. A nice twist from the grungy dark [[Ace Hotel New York]] -- ths lobby looks inviting, more like a WeWork than a dance club. Signature [[Stumptown Coffee]] is also in the lobby with a decidedly Japanese feel. They hung an oversized noren at the top of the entrance way of a coffee pot.

[[Beams Japan Kyoto]] is a delightful shop. I love orange and design, this had both. They had fun works from designers but also their own quirky design goods.

[[Kawai Shrine 河合神社]] is a small shrine that's out of the way in Shimogamo, that triangular tip area in Kyoto. The fun thing is it's a shrine dedicated to beauty, so the wooden prayer blocks you can buy are a face you can decorate. In fact, it makes me think there should be a shrine dedicated to Emojis. Right next to it is a shrine dedicate to Rugby, so this isn't that far off.

[[Coffee House Maki]] is a coffee place we went for breakfast because we were too early for the shrine to open! Surprise is that their Breakfast set is REALLY worth getting.

[[Ogawa Coffee Nishiki]] was the last cafe we went to. They had an amazing Cafe Mojito which was Mint Sorbet, with lime soda and espresso. The cafe also does a really good sandwich which you can enjoy staring at one of the best bonsai like inner-garden ([[中庭 Inner Garden]])

[[Kennin-ji 建仁寺]] hosted a night time art show in the shrine. It was an interesting combination of light-based art with [[Rock Garden]] and procedural art with paintings in the room. We went with [[Leslie Yau]] and [[Joyce Liu]] there.

[[雑太社]] is a shrine in [[Shimogamo]] which oddly is a shrine dedicated to [[Rugby]].

Stayed at the [[Hotel Hokke Club]] in Kyoto. It wasn't anything special, but very convenient.

I have a few personal projects that are quite data intensive and I use Google Cloud Platform for thoes.

However, I've got to the point there things are getting pretty expensive and I'm exploring better DevEx and most cost effective places. After a few days of exploring, I'm still as confused as ever.

I have a few things I want:

• Ability to host and serve about 2GB worth of images on an HTTPS domain.
• Ability to store large amounts of raw data I collect from APIs.
• Ability to host web apps relatively cheaply (to nearly free)
• Ability to run serverless compute functions with good DevEx
• Ability to spin up datastores at a relatively low cost.
• Ability to run a suspendible linux machine in the cloud for certain things you need a full machine for.

Today, my stack is this:

• Storage: Google Cloud Storage. It is relatively cheap to store data on hot servers that I can access daily. The tools are fine to use and I'm generally used to them. Though the one thing I cannot do with them is to host buckets on an HTTPS custom domain without paying for a load balancer. (HTTP only is free)
• Web App Hosting: Firebase Hosting is my go to webapp hosting. It works great, very affordable, paying for usage rather than subscription per site. I do with they had slightly better cloud function support.
• Serverless Compute: Firebase Cloud Functions is what I'm using now. It is a wrapped version of the Google Cloud Functions, except it is much better DevEx. However, because it is Google Cloud Functions underneath, I end up inheriting a lot of issues/limitations that I didn't realize. For instance, I spent a lot of time debugging why one of my projects was consuming $15/month of compute costs only to eventually find out that their cloud functions weren't cleaning up old versions that were pushed, so I was paying for storage of function images that were outdated.
• Datastores: Firestore and MongoDB Atlas. Firestore is fine for small projects, but once you get to a certain size, it starts to bog down. I have a Firestore which has too much data in it that it now hangs the Web UI and I can't programmatically remove the data inside because the API calls also hang. So recently I've switched to using MongoDB Atlas that is a hosted version of MongoDB for $9/month. And it's worked flawlessly so far.
• VM: Cloud Compute VM. I use this when I first develop a prototype that needs long running processes, like a web crawler or long processing times. I need it to be suspendible so I can shut it down and only turn it on when I need to run that task. I might be better off using Docker instead for this.

I'm racking up around $60/month in costs for GCP and then another $9/month for MongoDB. Ideally, I'd want to cost to be around $30/month for everything, which has spurred me to look for alternatives. I'm not even sure this is achievable.

I've started to explore some of the other alternatives like

• Amazon AWS. So far I only use Amazon DNS for hosting one particular domain that the registry does not provide a DNS server for. It costs me a funny $0.53 for that service (only because DNS is run out of port 53.) I tried setting up an S3 bucket to be served over HTTPS, but that turned out into hours long following of complex instructions to setup the bucket, the load balancer, SSL certificate, Let's Encrypt DNS verification, etc. So in the end, I still don't know whether things are working yet. Their interface is so confusing, it's pretty hostile to a casual developer.
• Digital Ocean. I really like their Web UI, it's extremely user friendly. They claim to allow me to use HTTPS for hosting from the cloud bucket, but the pricing structure is $5/month for using their storage service. Their VM options are more expensive and also not suspendable, though the nice thing is that they include a static IP address that you can map to your VMs for free.
• Vercel: I'm exploring using their web hosting and cloud functions because they have such good web UI and also DevEx tooling. The ability to do nearly no setup and have deploy triggered from a Github push is so delightful. Firebase Hosting sort of allows you to do this but it requires like 20 steps of setup. I have found Vercel to be quite slow though, often I'll have to wait double the amount of time for something to happen once I click on stuff in their Web UI. I am not in love with their pricing structure, it wants to charge per developer, and at $20/month, it seems a bit steep for not that much value of the free plan.
• Netlify: Same as Vercel, the pricing structure is nearly the same, different design but also very developer friendly. It's hard to tell the difference between the two. I think I slightly prefer Netlify because of their speed.

As part of the Tokyo Design Art festival, Yamaha put on an exhibit in that very cool AXIS building in Roppongi. It featured a range of music adjacent design pieces.

• Dining table piano
• Smartphone record player
• A MIDI keyboard hooked into a set of physical instruments
• Violin Speaker

Staying in Akakura, we had to at least try a day in Akakura, even though it’s not such an interesting mountain. Moreover, there must be a discount our hotel can offer us.

It was complicated.

Akakura has two resorts, the Kanko, and the Onsen. If you only do the onsen side, you can get a ticket just for that. And that’s what most of the hotels sell you, discounted tickets for JUST the Onsen side. The Akakura Kanko, Akakan if you will, is owned by the luxurious hotel up at mid mountain. Legend has it that the previous Japanese Emperor stays there for vacations. So no one has a discount except for guest at their fancy hotel. You can get a combined ticket for ¥6000, which isn’t too much, but why pay 100% more for just a few more lifts?

This particular day wasn’t spectacular, and our main event was to see my friend’s kids learn how to snowboard, so really we needed just to stay on the part of their resort they were in. Akakura Onsen it is then.

That’s where it started to go down hill. First, on paper, this place has 12 lifts! Twelve is a lot for a Japanese resort. Even Akakan only has 5. However, on the day we were on the mountain, only half the lifts were operating! Times are tough, so let’s give them a pass.

We started on the left side and slowly made our way across to the right. Once we got to the right, we got another punch to the gut. The high speed quad on the right hand side broke down. Our only way to get back to the other side of the mountain. So no way back, apart from the bus.

It was a packed powder day, there were some untouched patches here and there where we duly ruined. If you look at the map, it’s mostly green runs. And definitely can confirm, it’s a beginner’s mountain.

The top lifts were open unlike the previous day we were there. We thought the strategy would be to go where others hadn't gone yet. Every year I go to Sugi, every year I regret going to the top runs. They're much steeper, more bumpy than the runs below. Sure enough, every time I go, every time I get disappointed. 

In the end we went back down to the Panorama runs and did the trees and the wide open mellow fields of powder. Still lots of powder lines to be had on a blue bird day.

The other reason I like Suginohara is because not all the restaurants on piste are operated by the resort itself. At the bottom of the high speed lifts is a set of three four restaurants that are independent, my favorite being the St Anton. Along with delicious Japanese pork cutlet and the usual, it features a caramel popcorn machine!

Sugi hits all three in my book. The Panorama and Dynamic courses both end up on the High Speed Quad. Panorama is a super wide run, you can see the run from miles away. It's so wide that it's what I used to recognize as Sugi when I'm spotting ski resorts around Nagano.

It was snowing all day, and every run there were fresh lines to be had, the groomed run was being buried under the snow. So much snow was coming down during the day that we broke for lunch and came back out to nearly fresh lines, even better conditions then at the start of the day. Even though the top lifts were not open, it didn't matter at all.

To be honest, it was a pretty ordinary day, we had a pretty relaxed go around the place, left a little early and was treated to a Hong Kong style Egg Waffle at the based of Mitsumata!

It's a sort-of-powder-day! Hey, not bad for December. It was snowing quite a bit on the mountain the day before, so I was getting excited about getting out early to the lifts. 

All my training getting up early at 5am for early morning meetings this year is now paying dividends. Getting to first lifts is easy for me. I went out as early as I could, though I didn't manage to get the first ropeway up. A lot of people showed up ahead of me on the way to the ropeway. The mountain ended up not getting too much new snow overnight though.

A few areas were still closed off, including what I considered the best area in Kagura, the Kagura Lift 5 that only opens for half a day, on a good day. On windy days they don't even bother. Neither for early season like today. This area is where you can drop through the trees into a huge open bowl that lands you back somewhere within the resort. Alas, it is not running.

Instead, I used the opportunity to try to ride the whole mountain. Even though I've been to Kagura a few times, I never managed to ride all the way across from end-to-end. After what seemed like a never ending series of lifts, I managed to get all the way to the far end of Tashiro. At this end there is a dramatically named Dragondola, a gondola that connects Kagura to Naeba. I was disappointed the gondolas were not shaped like a dragon.

On the Tashiro side, there were a few powder stashes and we rode around as a large group through the trees, destroying any smooth untouched powder we could find. I'd rate that a powder-ish day.

One of these are my lines: