jsharf

Related story, while applying a firmware update to my Kawai CA49 piano, I bricked it due to flashing the wrong file (The process was broken, and I got desperate and tried something stupid, which bricked the piano). Claude walked me through looking for signs of life, and since OTA from the phone app wasn't working for me, it downloaded the Kawai Android APK, decompiled the Java, figured out the hardcoded key used for encrypting the firmware update. Extracted the piano firmware update, decrypted it, and then wrote a flashing script to program the piano from my laptop via bluetooth. My piano was back to working within an hour.

If you have independent copies of the network learning gradients, then you’re effectively making the batch size smaller— unless you’re doing an all collect and making them sync, in which case there’s a lot of overhead

When you take a batch and calculate gradients, you’re effectively calculating a direction the weights should move in, and then taking a step in that direction. You can do more steps at once by doing what you say, but they might not all be exactly in the right direction, so overall efficiency is hard to compare

I am not an expert, but if I understand correctly I think this is the answer.

I want to also mention that the previous model was 3.7. 3.7 to 4 is not an entire increment, it’s theoretically the same as 3 -> 3.3, which is actually modest compared to the capability jump I’ve observed. I do think Anthropic wants more frequent, continuous releases, and using a numeric version number rather than a software version number is their intent. Gradual releases give society more time to react.

He's not saying that all non fiction is bad, just that the incentives are misaligned, and to be fair at least in my experience, there are a lot of popular non-fiction books where each chapter is repetitive, and I feel the whole thing could have been written in 2-3 chapters, if publishing a 30-page nonfiction book wasn't taboo

Wow, you can refocus the direction after the audio is recorded!

This would be cool to mix with VR, so you could hear different conversations as you move around a virtual room

I assume it’s easier to find an engineer who went to engineering school to learn how to build airplanes that are safe than it is to find an MBA who went to business school to learn how to build planes that are safe. (It’s not about the knowledge but about the root desire)

Similarly, I assume it’s harder to find an engineer who went into the field purely for money.

I do think on average engineers will prioritize safety (since they likely understand failure modes and production and long tail statistics better. We literally have to take engineering ethics classes), at the cost of doing a worse job at running the business. But when the business requires this level of safety, that IS doing a good job.

I think planes can still fly with the rudder loose? If the bolt falls out and it loses control, wind will push it into the neutral position and then flying will still be possible with other control surfaces? But I guess if the pilots don't know and it happens suddenly at a critical moment or if the bolt causes the rudder to get jammed, then that would be really bad. But I assume it falling out would result in the rudder loosely returning to neutral...

If we have intelligent AI that can automate programming, then making really good robots will not be a problem. While not trivial, actuators and power systems are not the reason why we don’t have robots that can do all manual labor for us. Software is the reason, and the same kind of software that’s learning to code (machine learning) can also be adapted to washing dishes, folding clothing, doing craft labor or previously human manufacturing jobs.

Accelerating programming and information jobs also means accelerating the creation of robots that can do these trade jobs

Recommend passing the speech-to-text narration through a round of GPT4 API to correct for any transcription errors (use some prompt giving context that it's speech to text)

Suggests there's other variables involved, like time of day taken, other supplements taken simultaneously, metabolic processes, diet, and maybe even the placebo effect.

Well, assuming they likely have some PFAS already in their system, you might just be giving them blood with the same concentration of PFAS that they already have.

They need to be dimensioned and finished quite accurately. Most plywood has a slight bend to it. At the speeds and accelerations needed to be competitive, any imperfection would mean your car will fly off the track or hit a wall.

Wait isn’t PPG for heart rate?

Do founders return the salary they pay themselves if their company doesn’t double in size?

The answer is no, so I don’t see why software engineers should be held to higher standards.

When you change a product page, you lose all reviews, but you get a "New Product" link on the old page to link to the new one, giving you exposure.

Amazon can make the "New Product" show up for queries for the older one, but without the stellar reviews. This way modified products still get good exposure and no one will accidentally buy the old one. But there aren't reviews on the same page for a different product. The system can still be gamed similarly, but at least you can't get reviews for the wrong product on the same page.

At the end of the day, you're still responsible for reading the email before you send it. I don't think it matters how the email was written. I think the final content matters more. There's many ways to prompt an AI, both carelessly and with extreme attention to detail.

For example, I typed this message. But I'm sure with careful massaging, I could get an AI to write the same thing, word-for-word. The laziness you're referring to probably happens even if the manager is typing the email out themselves.

I think for academic conferences, they typically choose a venue that changes every year. This could be used to aid in selecting a venue.

Theoretically, one could compile whisper.cpp to run in the browser using emscripten, maybe made faster with webgl...

I think this would be quite a heavy page load time for a website, but if the model file gets cached, and the user has a decent CPU/GPU, it... could work?

Is it extra protein, if they lived only off the rice in the bag? Presumably all the protein inside of them came from rice that would have otherwise been there.

I guess it helps digest and synthesize some of the proteins for you.

I think there's a lot of hate in the comments, and while I agree (stop hijacking my scrollbar), I'd like to inject some constructive positivity:

1. They put the disclaimer right in their motto. Get smooth or die trying. They're willing to accept failure in pursuit of peak smoothness.

2. Hate it as much as you want, ever since the awful idea was born, we haven't been able to prevent websites from doing this. Some implementations are better than others, and having standardization will help improve the state of the web for all websites that make this... terrible... decision. Their implementation happens to be quite nice.

3. Even if there's issues with the implementation, having a standard library means that when updates get applied, they apply to all websites that use them.

I think it's actually normal for teams to have duplicate functionality. The requirements are usually never actually identical and when you try to merge them it results in services with complicated configuration that have N^2 edge cases which leads to more bugs or more complicated code. I suspect what you're describing is somewhat unavoidable, and not even a bad thing.

That being said, Google for comparison does have good code re-use for certain core resources, like spanner, cloud, tensorflow, borg, etc. If you're talking about bedrock infrastructure like that, it's quite a different picture.

I'm not anti-connectionist, but if I were to put myself in their shoes, I'd respond by pointing out that in E=MC^2, C is a value which directly correlates with empirical results. If all of humanity were to suddenly disappear, a future advanced civilization would re-discover the same constant, though maybe with different units. Their neural networks, on the other hand, probably would be meaningfully different.

Also, the C in E=MC^2 has units which define what it means in physical terms. How can you define a "unit" for a neural network's output?

Now, my thoughts on this are contrary to what I've said so far. Even though neural network outputs aren't easily defined currently, there's some experimental results showing neurons in neural networks demonstrating symbolic-like higher-level behavior:

https://openai.com/blog/multimodal-neurons/

Part of the confusion likely comes from how neural networks represent information -- often by superimposing multiple different representations. A very nice paper from Anthropic and Harvard delved into this recently:

https://transformer-circuits.pub/2022/toy_model/index.html

I just did the math, and it looks like this is correct!

Assuming 2.5mm per month (quick google search), that's 2.5e+7 angstrom/month

Divide by the number of seconds in a month (30 * 24 * 3600) and you get about 10 angstroms per second. It takes about 1 second to say 10 angstroms. Very cool!

It seems like this only affects areas where tech workers make up a lot of the population, like the bay area.

Also, it assumes that non-tech workers won't just move to cheaper areas.

It's good enough that I wouldn't really call it a code generator, and I wouldn't put it in a category with other tools due to the performance different from anything else like it.

Often, it creates code that follows the style of the surrounding environment.

Honestly at this point, the copyright concern seems either ignorant or paranoid to me. The program is interactive and generally only creates small chunks of code at a time. You can read and modify the code it generates (I always do). I only operate on it in digestible chunks, none of which are "original" enough to be considered copyright. Surely, you as a developer can easily say whether the code generated is just some non-copyrightable snippet (some simple BFS code) vs stolen code lifted off a codebase (weirdly specific implementations, fast invsqrt, comments that look like they were written for a different context, magic numbers, undiscovered algorithms).

I would argue that if you can't differentiate between meaningful IP and boilerplate algorithm implementations (with variable names conveniently matching the surrounding context), then you have a different problem.

Yesterday I implemented a BFS in about 30 seconds because copilot recognized the pattern and auto-completed. It took 5 seconds to autocomplete, and 25 seconds to read and fix one line that was wrong. That normally would have taken me 5-10 minutes to code + debug (the copilot code didn't have any errors, minus the 1 I easily found). The productivity gains are too good.

The notion of a "deep enough stack to keep the pointers to those context handy" depends on the complexity of what you're learning. If the task at hand has a large enough memory demand, then you might find yourself "swapping" to external storage too frequently, and someone with even a slightly better working memory might not have to -- they'll probably be able to learn much faster.

That being said, this is an opportunity, not a problem -- if they learned something, they can probably explain the concept to you much more efficiently and help you learn as well. Some of the best learning resources online are informally written blogs :)

Yeah but it has the advantage of everything being 100% sustainable from the beginning. All water and oxygen will need to be recycled, etc.

You're not forced to use your cellphone. But society might hypothetically get to a point where it's extremely inconvenient not to.

To be fair, that's a realistic mistake that twitter users might make too. Musk has one of the most faked accounts on the site, with lots of bots running Musk impersonation crypto scams. IMO this says more about the problems with twitter than about an inaccuracy in Musk's script.

You don't need to rewrite existing code. Just develop new code in it.