More from Nelson's Weblog
I took an amazing trip to SE Asia last month, including Angkor Wat. I had a hard time finding good reading or other resources to learn from before I went, in part because Amazon is awash in AI garbage. Here’s some books and podcasts I found useful about the Khmer empire in general and Angkor in particular: Ancient Angkor by Michael Freeman and Claude Jacques. The closest thing to a coffee-table book to preview what you will see. The practical information is outdated but the pictures and descriptions are good. Empire Podcast #185: The God Kings of Angkor Wat by William Dalrymple and Anita Anand. An entertaining and fully detailed account of the Khmer empire. It’s basically an excerpt from Dalrymple’s new book The Golden Road: How Ancient India Transformed the World. Fall of Civilizations Podcast #5: The Khmer Empire by Paul Cooper. Another history, not quite as magically well told as Dalrymple but full of good information. Angkor and the Khmer Civilization by Michael D. Coe. A highly recommended history of the Khmer region. Honestly I found this very dry and too detailed, but I did learn from it. Lonely Planet Pocket Guide: Siem Reap & the Temples of Angkor. We didn’t use this much but it seemed like a useful practical guide. OTOH it dates to 2018 so things have changed. My other advice for visiting Siem Reap and Angkor is: go. It is amazing. Plan for at least two full days of touristing there. Hire a private guide and driver if you can, it is absolutely worth it. (Email me for a recommendation.)
The Trump administration has made aggressive threats against immigrants in the US. It’s not clear what’s coming, my biggest fear is a violent display of fascism. (Don’t call them camps!) But even if it’s a polite legal process it will be chaotic and disruptive to many neighbors. Back in 2018 I donated reactively to the Trump administration’s cruelty to immigrant families. This time I’m trying to get ahead of it. The need for the money is now, no matter what happens it is going to be a bad few years for immigrants in the US. To that end I asked on Metafilter about charities to donate to. I got back a remarkable reply listing 18 charities that all have some California focus. I donated to most of them. I want to highlight two groups in particular. One is RAICES. They work in Texas, not California, but they are well organized and effective. The other is KIND. They have a simple mission. They try to ensure every unaccompanied minor has legal representation in immigration court (something not guaranteed.) The other groups on the list are all also deserving of consideration.
LLMs are good search helpers. Here’s three search tools I use every day. All of these use an AI to synthesize answers but also provide an essential feature: specific web search results for you to verify and further research. I use these for conversational inquiries in addition to more traditional keyword searches. Phind is an excellent free LLM + search engine. The AI writes an answer to your query but is very careful to provide footnotes to a web-search-like list of links on the right. I use this mostly for directed search queries, things like “what’s an inexpensive TV streaming device?” where I might have used keyword search too. The Llama-70b LLM that powers the free version is quite good, sometimes I have general conversations with it or ask it to generate code. Bing CoPilot has a very similar output result to Phind. I find it a little less useful and the search result links are less prominent. But it’s a good second opinion. Bing has been a very good search engine for 10+ years, I’m grateful to Microsoft for continuing to invest in it. CoPilot results are sometimes volunteered on the main Bing page but you often have to click to get to the ChatGPT 4 Turbo enhanced pages. Kagi is what I use as my general search engine, my Google replacement. It mostly gives traditional keyword search results but sometimes it will volunteer a “Quick Answer” where Claude 3 Haiku synthesizes an answer with references. You can also request one. I think Phind and CoPilot do a better job but I appreciate when Kagi intercepts a keyword search I did and just gives me the right answer. Google has tried various versions of LLM-enhancement in search, I think the current version is called AI Overviews. It’s not bad but it’s also not as good as the others. Not mentioned here: ChatGPT or Claude. Those are general purpose LLMs but they don’t really give search results or specific references. In the old days they’d make up URLs if you asked but that’s improving.
8BitDo makes good game controllers. A wide variety of styles from retro to mainstream, with some unusual shapes. And wide compatibility with various systems: PC, Macs, Switch, Android. They’re well built, work right, and quite inexpensive. A far cry from the MadCatz-style junk we used to get. The new hotness is the Ultimate 2C, an Xbox-style wireless controller for the very low price of $30. But it works great, doesn’t feel cheap at all. The fancier mainstream choice is the Ultimate 2.4g at $50 which includes a charging stand and extra reprogrammability. But what’s really interesting to me are the odd layouts, often small or retro. The SN30 Pro is particularly interesting as a portable controller. SNES-styling but a full XBox style modern controller with two analog sticks, easy to throw in a suitcase. There’s a lot of fiddly details for this class of device. Controller type (XInput, DInput, switch, etc), wireless interface (Bluetooth or proprietary), etc. 8BitDo makes good choices and implementations for all that stuff I’ve tested. They seem to work well with Steam. They’re a popular brand so well tested. It helps that PC game controllers have mostly standardized around the Xbox layout and XInput. Steam can patch over any rough spots for older games.
More in programming
Helping Scout focus and remember
tl;dr: the Explorer Pattern is one individual or team building a functional implementation that minimizes cross-team and cross-functional impacts as much as humanly possible, and then reconciling the details with other teams after that prototype is completed, accepting that this approach can require very substantial rework to that initial development. Why write this? Imprint previously called this the “Tech-led” or “Engineering-led” development model, but it’s pretty unclear to folks what that means. This writeup provides a more defined term, the “Explorer Pattern”, and explains in more detail how this works. Why do we need to adjust the way we work? Before 2026: Code is expensive, and we design process to minimize implementation and rework In and after 2026: Writing software is extremely cheap, but it’s a wider phenomenon: generating output is effectively-free across many dimensions. Code is nearly free — if we have appropriate constraints. Designs are nearly free — if they follow patterns. Analysis is nearly free — if we’ve instrumented things properly Reworking software is extremely cheap. Historically industry has optimized around avoiding engineering rework because engineering was the highest cost / narrowest constraint. In many companies this was interpreted as writing code being the constraint, which was never true, but many common practices were designed as if that were true (e.g. bringing engineering into problem-solving late “to protect their time”) Working on the system is much more effective than historically if we get automated integration and feedback loops right, e.g. not improving a one-off design but improving the design system. Any place where improving the higher-leverage process is slower than solving the one-off is a bug in our process Explorer Pattern: primary-and-replicas metaphor The easiest metaphor to think about the Explorer Pattern is database replication. Databases like MySQL and PostgreSQL have a primary database, where writes are performed. Every write is verified at that point, and once it’s verified, you know that the writes will flow downstream to various secondary replicas. Those downstream replicas can be delayed but they can’t be wrong because writes are verified upstream. (I’ve resisted the desire to talk about the CAP theorem.) A different model for database design is that used by Dynamo or Cassandra (or MongoDB to a certain extent), where you have a replica set, and have to get an acknowledgement from the majority of the replica set for a write to be strongly consistent. This is a very reliable setup, because it means you can tolerate losing a node, but it also introduces inconsistencies, and on a fixed amount of hardware they tend to be much, much slower than running a primary-secondary setup (or rely on bad performance metrics that don’t fsync to disk to create the illusion of throughput). The Explorer Pattern is the former primary-secondary model, where the vast majority of the work is done by the working individual or team. This makes all work internally consistent because it’s passed through the same small group. It doesn’t, however, mean that the work is perfect. For example, you can always insert data that meets the table’s requirements but is still inherently wrong (e.g. 200 is a valid integer, but it doesn’t mean that my credit limit was supposed to be 200, that depends on the surrounding application logic). Other teams become aware of the changes downstream, not synchronously upfront, and it’s instead the working group’s responsibility to both rationalize approach and create visibility into their work for other teams to engage in a non-blocking fashion. Example: Support for Financial Aggregators A concrete example of this practice at Imprint was rolling out financial aggregator support, e.g. allowing folks to connect to their Imprint powered accounts via Plaid. This has always been something folks want, but it’s particularly important in the realm of agentic commerce where more folks’ agents are analyzing their financial state for them. For this project, one of our engineering leaders (Ron), built the first pass of the feature at a low-burn while understanding the shape of the problem in a private alpha. He chatted through some of the details periodically to keep other teams aware, but it was a bit of a side project. However, once he found the full shape of the problem and had implemented a full solution, we cut the replica-lag across teams down to near-zero, identified and ran through the small number of remaining issues, and shipped it. Like many of our projects this year, it was never officially on a roadmap, and shipped a quality feature that our members have wanted for some time. How to follow the Explorer pattern The rough rules for following the Explorer pattern: Know your goal and sponsor: every initiative ultimately has to have a specific outcome it’s pursuing and a sponsor who’s bankrolling that pursuit. This doesn’t always need to be explicit, especially for smaller projects, but once a project goes beyond ~3 weeks of effort, if you can’t very explicitly identify who the stakeholder is sponsoring your project and they would not enthusiastically agree, then you are likely closer to “lost” than “exploring” Small working group is small: at most 5 people, and preferably smaller. Everyone in the group has to keep the entire project in their head for as long as possible to detect problems and opportunities Small working group is flexible: folks in small working groups have to “find a way to yes” rather than “find a way to no”. Exploring is about uncovering the fundamental shape of a problem rather than pretending we know the fundamental shape Publish progress in a public channel: explorers have to report home with the progress they are making so others are aware. This is both a Linear Project and a Slack channel. Make the narrowest viable changes to what already exists: when we introduce unnecessary changes, we invalidate others’ mental models about how the world works. When we haven’t made those changes with care, we’re not only invalidating others’ mental model, we are creating the expectation that we’ve done it very intentionally for a good reason, which has a chilling effect on others improving them later, which is very expensive, whereas the Explorer pattern depends on being as inexpensive as possible Exercise good judgment to identify reasonable hypotheses: discard bad ideas immediately, and then filter the rest by trying to implement them as quickly as possible. Once you see the shape of the implementation, you can figure out which of the remainders are interesting Hold assumptions constant to avoid getting stuck: the only way to navigate complex ambiguous problems is to hold things constant, even when you know they are imperfect. For example, to do headcount planning we have to know our revenue, to know our revenue we have to forecast it, to forecast it we have to know the day that partners will go live next year. We don’t truly know any of those things, so just anchoring on a plausible guess to unblock us is much more valuable than trying to get them accurate. A little prework is useful, but a lot is just artificial precision that’s a distraction from progress Hold space for being totally wrong: when we move fast, with low information, we always find the way out, but often it’s not the way out we anticipated when we got started. Being open to throwing it all away, and starting a second, third, fourth time Ship all the way to production behind a feature flag: features and products move fast when we get the pieces working every step from frontend to backend to data pulls. This is the first step of exploration. Just make sure it’s behind a feature flag that’s only enabled for internal folks!
Comparing two different ways of work and working environments.
I own an old, cheap TV. It has no Internet connection, no fancy smart features that make you want to install a Pi-hole, and no ask-me-again-later updates. However, it is connected to my own desktop computer through an HDMI cable, so I can watch movies and TV series from the couch. Setting it up used […] The post Making a dumb TV just smart enough appeared first on Quentin Santos.
Five days and 540km on a bike across Kyushu, without a laptop. Notes on the peace of riding all day, convenience store dinners, and Japan's emptying countryside.