tkgally

Interesting. Does anyone know why volume dropped off from 2023 to 2024?

https://hackernews-analytics.jerrysniffs.online/#activity

Thanks!

Japan resident since 1983. I read Japanese fluently, but I also hate reserving Shinkansen tickets online. When I want to reserve, say, a one-way ticket from Yokohama to Osaka, I have to click through multiple popups warning me that I cannot pick up the tickets in Kyushu or Hokkaido (or something like that, I don’t remember the details) and various other things irrelevant to my trip. When I can, I prefer to buy paper tickets from the machines in the station; the interface is much better.

In addition to the various reasons mentioned by others, this inefficiency is partly the result of government-led deregulation aimed at promoting competition. JR used to be JNR, a single nationalized railway system. It was unified but customer-unfriendly, losing money, and prone to labor disputes. When it was privatized in 1987, it was divided into six companies by region (plus a separate nationwide freight company). While the various JR companies mostly cooperate with each other, and service has indeed improved overall (stations are much nicer than they used to be), the decentralization also leads to problems, including insufficient coordination among the companies about reservations and ticketing.

Another unintended consequence of deregulation has been the complexity and inefficiency of payments at the retail level. While 7-Eleven convenience stores accept several dozen types of digital payment [1], most retailers can afford to handle at most a few. As a result, different stores accept different combinations of payment methods. Some (including a small restaurant near my home where I eat once or twice a week) accept only cash, while others accept only digital payments (some forms, that is) and refuse cash-only customers. There’s a large department store near my home where I can pay with Suica on some floors but not on others.

It’s not unusual to get stuck in line at a cash register while the teller and a customer ahead try to figure out how the customer can pay.

It’s a mess.

[1] https://money-creditcard.com/wp-content/uploads/2020/06/seve...

A great example of a cooperative volunteer project. I have benefitted much over the years from their efforts.

I wonder if they are incorporating AI in their workflows.

In 2019, I published online an anthology of writing from 19th- and early-20th-century books at the Internet Archive [1, 2]. I started with the IA’s OCR versions of the texts, but they were riddled with errors and I had to do extensive manual clean-up. I didn’t trust myself to catch all of the mistakes, so I ended up paying someone else to do an additional proofreading pass.

Over the past couple of years, I have done occasional experiments to see how well the current AI models could do such work. One of my test cases has been the 1821 novel Life in London by Pierce Egan [3]. It has complex text and page formatting, with extensive use of italics, small caps, footnotes, footnotes to footnotes (!), poetry, etc.

Earlier this year, I had Claude create a digital version of Life in London using Gemini 3.1 Pro through OpenRouter for the OCR, with further semantic checks by Sonnet 4.6. The results [4], while not good enough for Project Gutenberg, were still a big improvement over my tests a year earlier. I am guessing that it would be possible to get even closer to PG quality using more recent models.

[1] https://www.gally.net/jatsi/index.html

[2] https://archive.org/details/japan-as-they-saw-it

[3] https://archive.org/details/cu31924090015722/page/n56/mode/1...

[4] https://tkgally.github.io/lifelondon/

That was my thought, too. If I ever get one myself, I’m likely to forget about the touch feature and just keep using the trackpad and keyboard.

But whenever my adult daughter shows my six-year-old grandson something on her Macbook, he immediately tries touching the screen and is confused when it doesn’t respond. He doesn’t have any devices of his own yet, so his defaulting to touch is the result of very occasionally being allowed to use an adult’s smartphone or tablet.

The proofs themselves are far beyond me, but the first page of the Anthropic paper, at least, is much easier for me to follow than the OpenAI one.

Some years ago, I made a list of books at the Internet Archive on English spelling reform from the early 20th century. Here are a few:

Child’s First Reader (1904?)

https://archive.org/details/childsfirstreade00pitm/page/72/m...

A Plea for Alphabetic Reform (c1905)

https://archive.org/details/pleaforalphabeti00pitm/page/n3/m...

Simplified Spelling for the Use of Government Departments (1906)

https://archive.org/details/simplifiedspelli00simprich/page/...

The Problem of Spelling Reform (1906)

https://archive.org/details/problemofspellin00skearich/page/...

The Fonetic Primer Offering the Universal Alfabet and the Science of Spelling (c1907)

https://archive.org/details/foneticprimeroff00storrich/mode/...

English Spelling and Spelling Reform (1909)

https://archive.org/details/englishspellings00loun/page/n5/m...

Simplified Spelling (1911)

https://archive.org/details/cu31924027251341/page/n5/mode/2u...

Handbook of Simplified Spelling (1920)

https://archive.org/details/handbookofsimpli00simprich/page/...

Almost nothing.

The original idea to ask for clocks came from Claude itself. After Opus 5.5 was released, I asked it for ideas for interesting projects to try, and clocks was one of its suggestions. Some of the later prompts included suggested themes, but those came from Opus, too.

The only instructions that originated with me were striving for variety, making the page self-contained, randomizing the display sequence, and putting the clock names and explanations behind mouseovers/clicks. Even the controls at the top of the page were added by Opus on its own initiative.

That “cheapness” is because the designs are created programmatically in JavaScript, with no image files used.

A significant proportion of my Max subscription for this week. In the README, Claude estimates a total of about 15 million tokens.

I was interested in seeing how many different ideas Opus could come up with with minimal direction from me. I don’t plan to use any of those clocks myself.

Yes, I got carried away. It used a big chunk of my Max subscription tokens for this week.

Yes, I as noted at the top of the README, Claude wrote it all. Excerpts from my prompt to it: “Please compose a markdown file that I can put in the README.md for the repository explaining how this page was created and how it is structured (including how the images are generated).... Be clear that all of the coding and writing was done by Claude Opus 5.5.... Make the explanations plain and straightforward, with no boosting or bragging.”

Very nice! I particularly like that pi clock.

I agree that Claude did tend to repeat certain themes. If someone wanted to extend the gallery to the thousands of clocks while maintaining the variety, they would probably need to seed the model with some original ideas. (Whether those ideas could come from a model or would need human taste is an interesting question.)

Also, Claude’s realizations of some of its ideas are a bit rough. If I wanted to create a standalone widget for a clock I particularly liked, I would spend more time with Claude (or another model) refining and polishing it first.

Nevertheless, considering that nearly all the ideas came from Opus and all the graphics are created with pure code, I remain impressed with what it was able to do.

My impression is that the tourists are being guided along fixed routes in the limited areas of the country that are open to tourists. But there are long shots of towns and the countryside from trains and buses that can’t have been entirely staged. These are the glimpses of everyday life that I find fascinating.

Side comment: North Korea seems to be allowing tourists to take videos fairly freely now, and some of those videos are appearing on YouTube. While the tourists' movements still seem restricted, the videos offer glimpses into life in North Korea I haven't seen before. For example:

https://youtu.be/HQEYB4QWCjM

https://youtu.be/d7m8hZ67Bz0

https://youtu.be/guZQNMEz9as

> In an ideal world, our agents could help with this.

Not exactly the sort of case Simon has in mind, but I tell Claude to keep to hard daily limits on its OpenRouter spending for two long-running projects [1, 2].

A Routine for each project fires ever few hours, and Claude decides itself what to do in each session. It does tasks that require calls to other models through OpenRouter only when it is still within its daily budget for that project; after it reaches that cap, it does other tasks that don’t require extra spending.

[1] https://github.com/tkgally/je-dict-1

[2] https://github.com/tkgally/eex-dict

[delayed]

Excellent article. I have worked for twenty years at another university in Tokyo and taught and supervised many non-Japanese graduate students, but I still learned things from the author's perspective.

One small pushback: The author writes, "the Japanese PhD system pushes you to graduate in three years." This applies in some programs and some fields, but is by no means universal. In the graduate program where I used to teach (I am now emeritus and no longer supervising students), the steps leading up to the Ph.D. and the expectations for the thesis made it practically impossible for students to finish in three years, especially if they continued to work part-time. This created particular challenges for foreign students on MEXT scholarships who had to work out other ways to support themselves and renew their visas while they continued to write their theses.

I’m also finding Opus 5.5 to be a significant advance. In the past few days, I’ve given it several complex tasks related to two long-term dictionary-building projects [1, 2], and it handled them impressively—similar to Fable but with much lower token consumption. It also communicates better than Opus 5.1; I don’t think I’ve seen any obvious Claudisms yet.

[1] https://www.tkgje.jp/

[2] https://tkgally.github.io/eex-dict/index.html

And it's possible to run each document through the LLM pipeline multiple times, using different models and/or prompts each time, to check for errors and inconsistencies. That will take more time and cost more, but it can reduce the error and hallucination rate significantly.

I encounter the same problem of unclickable buttons on my iPhone 15 fairly often. Sometimes I can make them reachable by switching to landscape mode, but not always. I’m 69 and do prefer larger fonts, though I don’t think my font size settings are causing the problem.

The blog post says it is RSI: “both of today's releases are … accelerated by long-running agentic loops designed to recursively evaluate and refine the underlying models.”

Here’s an example of one of those Claude skills, in a public repository I manage:

https://github.com/tkgally/je-dict-1/blob/main/.claude/skill...

Fable wrote it specifically for this project.

Not the person you're asking, but I did that by explaining to Fable my problem with Opus's gobbledygook and having it write a Claude skill for producing clear explanations in its reports to me. I also had it add notes about the need for clearer writing to CLAUDE.md and other project documentation. Opus's subsequent reports to me have been much clearer.