jfim

> Disney would still profit from the MCU films which they produced in the 2010's, but they couldn't stop you from releasing your own Batman vs Spider-Man story.

Wouldn't they still have a trademark on those characters though?

The seats are like $349 at the local Costco. It doesn't involve removing the toilet, just replacing the seat (two hand tightened nuts) and adding a T connector to the cold water line so that both the toilet and seat have water.

None of those things are permanent and can be done by a single person with minimal knowledge and some common hand tools.

They're totally installable in a rental, takes about half an hour to install the first time, and about ten minutes to remove when you move out.

The only thing is that one will typically need to run an extension cord to reach the toilet, but typically apartment bathrooms are pretty small anyway so it's not a particularly long extension cord.

Training a GPT-2 model from scratch on my desktop to learn more about how LLMs work under the hood. Next step is modifying it to be multimodal, and then trying out diffusion models.

Pretty fun project!

Is that specific to Opus? I asked Claude 4 Sonnet and it replied:

> Joe Biden won the 2020 U.S. presidential election, defeating incumbent President Donald Trump. Biden received over 81 million popular votes (51.3%) compared to Trump's over 74 million votes (46.8%), and won 306 electoral votes to Trump's 232. Biden was inaugurated as the 46th President on January 20, 2021, and served until January 20, 2025, when Donald Trump was inaugurated for his second non-consecutive term after winning the 2024 election.

Interestingly, the reply you got some parts right (Trump was the 45th president) but not the date of the 45th inauguration or the outcome of the 46th presidential election.

I wonder if certain text could be marked as unsanitized/tainted and LLMs could be trained to ignore instructions in such text blocks, assuming that's not the case already.

I'm pretty perplexed as well. They mention that they're not sharding PostgreSQL, without mentioning why in the article, but isn't that an obvious issue to many of their scaling problems?

I don't really see what it is that they're doing that requires a single master database, it seems that sharding on a per user basis would make things way easier for them.

Sounds like someone had a Java app and mistakenly exposed all of the JMX endpoints over HTTP. It's not the default configuration, and likely done out of carelessness.

In the case that was mentioned above, hardware frequently has drivers which can run as part of the kernel and can read the memory of other processes, among other things. A user of some hardware would be very likely to install a driver for it, without realizing that a malicious driver can basically do whatever it wants on their system.

In practice, if the software one wants to use is not trustable, then it shouldn't be run, at least on current mainstream desktop operating systems.

> What should be the real continuing inexcusable outrage is that Windows, even today, grants all applications full access to read the titles of all windows with no way to disable it.

That's there because it's been in there forever in the win32 API, and changing that would break applications.

For example, a long time ago, I wrote a small application that would iterate through a list of executables, launch each one sequentially, and for as long as that executable ran, it would look at the window that had the current focus, enumerate every control, and then send a click message to any control that would contain the words yes, agree, continue, accept, install, ok. Made my life easier to automatically install software on computers unattended.

There are plenty of other applications that require looking up other windows and sending them messages for all kinds of user workflows, things like autohotkey and so on. Changing that behavior would break all of them.

One thing you can do is to define the classes and methods that you want to have, and have the LLM implement them. For tricky things, you can leave additional notes in the empty method body as to how things should be implemented.

This way you're doing the big picture thinking while having the LLM do what's it's good at, generating code within the limits of its context window and ability to reason about larger software design.

I mostly treat the LLM as an overly eager to please junior engineer that types very quickly, who can read the documentation really quickly, but also tends to write too much code and implement features that weren't asked for.

One of the good things is that the code that's generated is so low effort to generate that you can afford to throw away large chunks of it and regenerate it. With LLM assistance, I wrote some code to process a dataset, and when it was too screwy, I just deleted all of it and rewrote it a few times using different approaches until I got something that worked and was performant enough. If I had to type all of that I would've been disappointed at having to start over, and probably more hesitant to do so even if it's the right thing to do.

Claude code is now part of the consumer $100/mo max plan.

The answer is, whenever there's enough money to be made there to support a self driving company's expenses to develop and operate it. These programs cost billions to develop, and public transit is likely not a profitable market niche at the moment.

Also keep in mind that the driver also is responsible for handling fare evasion and ensuring the passengers stay at least somewhat ruly, which a self driving vehicle might not be able to do.

Part of the reason is that public transit for whatever reason appears to be unusually sketchy in many places in the US. For example, a few years ago, there was an incident with a man with two chainsaws threatening passengers [1] in the local transit system.

In contrast, the transit systems I've seen in Europe and Asia appear well maintained, clean, and relatively safe.

Biking is also safer in European cities that have proper bike infrastructure.

[1] https://www.newsweek.com/man-armed-chainsaw-threatens-bart-r...

That's pretty cool, TIL!

That "it's fun to eat alone" seating area is so dystopian.

If you're using Claude code or cursor, for example, they can read files automatically instead of needing the user to copy paste back and forth.

Both can generate code though, I've generated code using the web interface and it works, it's just a bit tedious to copy back and forth.

Drive doesn't sync arbitrary files on Android though. A closer analogue to what nextcloud is trying to do would be syncthing, who also discontinued its Android app due to issues with getting the proper permissions with the play store (see https://github.com/syncthing/syncthing-android/issues/2064).

Thanks, TIL!

Do you have a reference as to where it's allowed in the CVC? FMVSS section 108 table I-c specifies this for the stop light:

Steady burning. Must be activated upon application of the service brakes. When optically combined with a turn signal lamp, the circuit must be such that the stop signal cannot be activated if the turn signal lamp is flashing. May also be activated by a device designed to retard the motion of the vehicle.

When running data parallel training, basically all the nodes taking part in the training run the same training loop in lockstep. So you'd have all nodes running the forward and backward passes on the GPU, then they'd wait for the gradient to be reduced across all nodes and then the weights get updated and another iteration can be run. For the first part the GPU is working, but when waiting on the network it's idle. The spikes are basically synchronous across all nodes doing the training.

The only way to spread the spikes would be to make the training run slower, but that'd be a hard sell considering training can sometimes be measured in days.

Interestingly enough, it thought I was Russian, even though my native language is French. It was tied at 32% with French though.

Edit: Tried it a few times and also got English as an accent. Pretty fun application!

It's coming. Ford and GM have started to build more software in house, while VW is partnering with Rivian partly to get their software expertise.

It depends. Sometimes webpages are useful, but at the same time navigating the amount of fluff on webpages nowadays takes longer than asking a LLM.

I asked Claude to give me a recipe that uses mushrooms and freezes well and it give me a decent looking soup recipe. It might not be the best soup ever, but it's soup, kinda hard to mess up. The alternative would be to get a recipe from the web with a couple dozen paragraphs about how this is the bestest soup ever and it comes from their grandma and reminds them of summer and whatnot.

I was wondering the same thing and found this paper: https://www.sciencedirect.com/science/article/abs/pii/S10439...

The same analogy could be applied to business though. Some Colonel invented a fried chicken recipe and started a chain of restaurants, now he doesn't need to work anymore.

In my opinion, if someone creates something that has value for a lot of people, they should get rewarded for it.

That's interesting, how do you manage the color palette? Are all the colors bright or are just using one or two bright pieces as accents?

If you're not dead set on having eink, there's always the ESP 32 cheap yellow display board: https://makeradvisor.com/tools/cyd-cheap-yellow-display-esp3...

They also sell physical hourglasses if you don't want to be interrupted by a distracting beeping thing once the timer is over, especially if the activity you're trying to start with a pomodoro requires concentration like coding.

Gotcha, that makes sense.