setsubyou

I have the same doubts… it’s very easy to count how much text it cuts out but it means nothing without benchmarks that show accuracy isn’t hurt.

There’s certainly things that are just not needed. But even permissions and owner in ls has cases where it’s useful - if the agent needs to debug a permission issue and ls doesn’t work, it will figure out some other way, and that takes tokens. Hopefully you’re not debugging permissions often though.

I also have no idea if the unchanged lines in a diff are noise or if they help the LLM. What if the context around the changed lines is really helpful? We can’t just decide that parts humans don’t need are useless. I mean there’s literally research showing that even repeating the prompt can improve accuracy.

The Apple office apps are still free, but they do have an add-on subscription too now.

Apparently he’s 63. So he would have played in the streets in 1960ies Taiwan or Thailand. Both countries where most people couldn’t afford cars at the time.

Here in Germany, we can still play in the streets in 2026. We have streets designated for it.

Hans Boehm built a Java library for this in the 1980s

Really early for Java, it came out in 1995.

Concurrency and running stuff on the main thread that shouldn’t is a typical pitfall with Claude and Swift. Always double check (or make Claude do it).

Nowadays agents like Claude Code can run autonomously for hours just given a goal description. It doesn’t take a lot of human effort at all to set up a bunch of sessions, and these companies don’t limit how many instances you run in parallel. Agents can also spawn sub-agents that run in parallel if a task calls for parallelization. Whether all this produces good results is a different story, especially if you don’t put enough effort into the goal description. But burning tokens as such is not difficult.

Even workflows where you’re just chatting with an agent can burn a lot of tokens. When you’re chatting with an LLM, the entire history becomes part of the input each time you send something. This also applies to tool calls, so if the agent decides to read 20 files before it can work on your request that’s 20 times a file gets added to the history and 20 times that entire growing history is then sent back as input to drive the agent’s next step.

Coding is more affected by this than many other applications because even a new conversation tends to start with the agent gathering a bunch of source code files, and then the response to a task is not just a bunch of text once, but a sequence of tool calls to make edits across files, build, run tests, react to test failures, and so on, all for one actual human prompt - but in reality a back-and-forth between the LLM and the harness with a quickly growing history.

They didn’t tell them to pull it, they told them to revoke access from non-US nationals. Similar to export control rules for cryptography that have existed for a long time. PGP got its author a criminal investigation back in the day.

Now Anthropic can’t guarantee a restriction to US nationals so Anthropic pulled it instead.

They probably have some internal problems with this too. They just hired Andrej Karpathy (one of the founders of OpenAI) - he’s a Greencard holder but not a US national.

You can be mad anyway of course, but maybe more in the context of a government stake in OpenAI being discussed recently and the conflict of interest this brings. Export control by itself isn’t new though.

From the article

Bankman-Fried is being held at a low-security federal prison near Santa Barbara, California. He is eligible for release in 2044.

Wondering how much the AC could even do if the door is left open. But maybe the hassle is also having to turn it on again every time?

Chinese fakes, you have my blessing and my axe.

In theory a summary is a good idea, in practice Google’s is catastrophically wrong often enough to produce news like this, and can trivially be shown to be wrong in other cases. It stands to reason that it’s likely also often wrong in ways that we don’t notice, which makes me not want to rely on it.

I remember some issues with international Visa cards a few years back, but not sure if they still exist. Nothing specific to mobile Suica, just Visa in general.

That but also if you’re not training and hosting your own model, your scanner is just subject to the same restrictions that your LLM provider applies to you on top of all the architectural problems.

1 in base 10 isn’t 1/10 and in hexadecimal it’s not 1/16.

Decimal integers in base pi are 1, 2, 3, 10.2201…, 11.2201…, 12.2201…, 20.2201… and so on.

Basically: 10.2201… = 1 * pi^1 + 0 * pi^0 + 2 * pi^-1 + 2 * pi^-2 … which approaches 4 as you add digits.

But 1 is just 1*pi^0

Japan has been doing cashless forever. Most local/regional transport operators use IC cards. The tech is Sony’s FeliCa and the major ones are compatible nationwide. It’s also built into modern smartphones, although it’s only generally available on phones sold in Japan except for iPhone, where all models worldwide support it. If you’re on iPhone, even as a tourist you can just create a virtual card in Apple’s Wallet app and charge through Apple’s payment service.

JR East’s card is called Suica and exists since 2001. It covers basically anything in the Greater Tokyo area where they operate, not only their own stuff (trains, metro, subway, trams, the airport monorail, buses, some ferries, …), and can be used for payment at many shops, vending machines, restaurants, and even some parks and such. Billing for transport is generally simply by shortest distance. E.g. riding the Yamanote ring line one stop in one direction costs the same as going all the way around the ring in the other direction.

Shinkansen (bullet trains) don’t use the cards for payment, but you can link online ticket purchases to them and then you buy cashless online and only tap in and out for the Shinkansen too.

Recently some train operators have started supporting credit cards too, so for some trains you can just tap a credit card at the ticket gates.

It’s not user replaceable but it’s not actually that hard to replace it, so this basically just takes work away from all those little repair shops. Still good for customers but probably not much of a difference for Nintendo.

None of them are profitable, they’re all selling their services at a loss

The word moron didn’t exist in English until ~80 years later though

They display it on the shelf (though some stores still display both before/after tax because the final price is only required since 2021), but registers calculate it separately and the tax appears on receipts.

There were concerns that if the tax rate is zero they have to check everywhere that nothing is ever divided by the tax rate or the computed tax. If it’s 1% they just have to change the number everywhere. And then I guess a significant part of the “months” that takes is also just testing, certifications, etc.

Unfortunately the tax is added at the register and paid by the consumer, so if they just do nothing the companies end up with extra money that they can then just not pass on to the government relatively easily, while consumers still see a tax on their receipts that should not exist.