pptr

X's video player is super annoying to use. I often manage to close the running video and I have to search for it again to resume it.

They need to build a robust video player ux if they want to compete.

All libraries you would need are already in Google's monorepo. There is no need for a package manager.

There are a few external libraries in the monorepo. Maybe it's a trickier task to add those. Most Googlers don't need to do that, though.

A lot of this tooling is available for C++ development at Google.

Includes are added automatically when you use code completion. There is also a simple IDE command to add all the include dependencies to your build file.

I don't know how well rust is integrated with Googles internal tooling.

The golang integration is incredible, though. You don't need to touch includes or build dependencies at all. They are fully automated.

I'm all for better tools to help the compiler figure things out. Here is an example where I can't communicate the invariants to the compiler:

``` std::vector<int> v; ... v.push_back(2); std::sort(v.begin(), v.end()); // no need to check i < size because we know we will find value 2 somewhere in the v. for (int i = 0; i < v.size(); ++i) { if (v[i] == 2) return i; } ```

Note that in C++ you can manually mark code after the loop as unreachable, which would indeed skip the size check. But that's as bad as not checking bounds in the first place.

Normally you allocate on the stack, so the local function owns the object. You pass (unowned) references to any function you want to call. Those functions are not concerned about RAII for those references, since they don't own them.

If you want to pass ownership to a different function/thread, you move the object. It's the owner's responsibility to run all the destructors once the caller deletes the object, which RAII does for you. Granted, this can get undeterministic with the reference counted shared_ptr, since only the last owner of the reference will actually delete the object.

I actually really like RCU for shared access: https://en.m.wikipedia.org/wiki/Read-copy-update Deallocation is always done in a background thread and there is no synchronization when reading. I'm not sure what public libraries are available, unfortunately (I've only used our internal rcu library).

If you want to make efficient use of your zillion cores, you need to make sure single threaded performance is excellent ;) See Amdahl's law.

But I agree, development velocity is an important factor to consider when choosing languages.

Go is a great language, but there also needs to be a language that you can program with zero overhead.

In what way is RAII a burden?

Genuinely curious, since I find it quite easy to work with.

Well, when you are not in doubt you don't want unnecessary bounds checks.

Being able to use a foreach loops is a decent improvement.

And since you now have a proper access API, you could on theory enable bounds checks via compiler flag, even for operator[].

C++20 added std::span, which is essentially that.

I think the limitation is chip size/cost. SRAM is a lot less dense than RAM. According to Google this is typically used for registers and caches, which are only megabytes large.

Why design a language? Just compile pseudo-code.

The press release does mention Signal explicitly, so I guess companies need to enable message backups for it.

I think there is a theoretical threat for models hosted on an external server.

You could ask a question on lmsys, check your server logs for the generated response, go back to lmsys and pick the response that your model generated.

Maybe you could also use a better model for requests from lmsys. E. g. use an unquantized model, disable censorship, etc.

I doubt any of the big players are doing that, but you never know.

This. And also that text is generally select- and copyable in the browser.

He wants 25% voting power, not own 25% of Tesla shares.

Mark Zuckerberg's Facebook shares have more voting power than regular shares. It sounds like Musk wants a similar setup.

I think one of the advantages of steer by wire is that rear steering gets a lot easier to add (which the cybertruck has).

With proper instrumentation, the main speed difference between running your binary locally or in the cloud is the additional scheduling delay. You still need to build your code and start it.

Startup time is the main productivity issue in our stack. We load lots of datasets into memory on startup, which can take minutes.

I really enjoy the simplicity of Googles internal repo (piper).

I am quite surprised that no open source variant has been created yet (that I'm aware of).

How is DNT different from simply not giving consent in this case?

Building a port in the north sea sounded like a terrible idea to me at first. However, I asked chatgpt and Bard and they both state that Delta-v needed to reach orbit is only 9.7 km/s versus 9.4 km/s from cape Canaveral. It's a disadvantage, but it doesn't sound like a deal breaker.

or a genius by getting more conservatives to buy electric vehicles.

We can barely produce enough weapons and ammunition for Ukraine. I don't think the west's production capacity is ready for a multi front war.

I use Google Lens for OCR in photos and the swipe-up+"select" feature for OCR in arbitrary apps/websites.

That at least works for Pixel phones. Not sure about Samsung and others.

Unfortunately, it looks like costs have stagnated in recent years: https://about.bnef.com/blog/lithium-ion-battery-pack-prices-...

How many people actually extend the lifetime of their phone via custom firmware? I imagine most people with old phones simply continue using their devices after support ends.

If you achieve 23x throughput without reducing the amount of work being done, does that mean GPU utilization was ~4% before?

Kinda hard to believe, but I have no intuition about this.

I don't understand the need to include "the types of websites you visit" in GPU driver telemetry.

AMD's telemetry doesn't seem to include that: https://www.amd.com/en/legal/privacy/user-experience-program...

I couldn't figure out what telemetry is collected by Nvidia.

I am employed at Google, though not on Chrome or Ads, nor am I doing web development. I don't have any insights into WEI besides what is publically known.

I stopped declaring my employment because it's a hassle to do that on every comment when writing multiple comments. And no one else seems to disclose their biases in this discussion.

That said, I agree with you that I should have declared my affiliation. Apologies.

As many people here, I am trying to understand the implications of WEI. Of course I will challenge the mainstream opinion to advance my own understanding and hopefully those of other readers too. I don't think arguments should be dismissed based on affiliation.

Maybe there is simply not enough medical data to properly train a 500B model.