uluyol

The water use comes from evaporative cooling, not liquid cooling.

Liquid cooling pulls heat away from the chips. Evaporative cooling takes it out of the building.

There are alternatives, but evaporative cooling is very energy efficient.

> quadratic time in N!

I read this as O( (N!)^2 ).

What makes you think there was no competition?

Electricity? It's probably one of the best places in the world to have a solar+battery installation?

It just needs to be anything that will force OpenAI to IPO.

They count as a different type of failure.

We knew what to do but screwed up hard is a operational failure and we didn't plan for it is a design/planning failure.

The people who are hurt might not care, but understanding the root cause is important to address them.

Are these the current prices or the prices at the time the models were released?

They had contracts which forced them to buy Global Foundries even lasting into Zen 2 (I believe they used it with the IO die).

Layoffs might have to with what a person is working (project is no longer business priority, just cutting based on performance might lead to delays) or what team a person is working on (whole org gets the axe).

How is 1 false? Log improvement means for 10x the cost the model is 2x as good. For 100x the cost, the model is 3x as good.

Not a curve to be happy about TBH. You need to simultaneously find big efficiency wins and drive up costs substantially to get 4-5x improvements, and it is probably impossible to maintain good year on year improvements after the first 2-3 years when you get all the low hanging fruit.

AMD is not ready to have a full announcement for their new consumer GPUs: https://morethanmoore.substack.com/p/where-was-rdna4-at-amds...

Go stacks are dynamically copied and resized. Stack overflow is not a concern.

I think the discussions in these threads show how accurate the framing of this article is. You have some people celebrating Google and friends (slowly) leaving the C++ ecosystem and those that continue to emphasize the flaws that have driven companies away from it in recent history (safety being #1) on the list.

You might be interested in this talk: https://go.dev/blog/ismmkeynote

Around the time the CoC was being established, Linus went to therapy. If I recall correctly, some people had spoke to him about his behaviors and he decided to do something about it. I think it was done in private so it's unclear how much of it was pressure vs his own decision. His tone has become much less aggressive since.

Go very much is memory safe in the absence of data races.

Data races cause issues in all languages, though it's fair to say that Go is affected slightly more than languages like Java. Rust is a bit special by making data races hard to trigger (impossible in safe code IIUC), but this is not typical.

ICEs have a lot of advantages that make them much better suited than EVs to certain tasks (extreme climates and remote locations, for example). They will likely stick around for niche use cases (at least) for quite a long time.

As for why Porsche is spending money on ICE engines...well, no one really buys a Porsche because it's a "practical car", do they?

This idea was explored over a decade ago, in the context of cloud computing: https://www.cs.cmu.edu/~fawnproj/

Google most certainly uses TPUs for training.

> Is it any more horrifying or unethical to decide without AI?

It is harder to get 1000 people to participate in an unethical military campaign than it is to get 10 people. So if you can make those 10 people 100x as efficient...

There are efforts from industry to try to secure open source, e.g., https://cloud.google.com/security/products/assured-open-sour...

I suspect some variant of this will grow so that some companies, MS/GitHub for example, audit large body of code and vet it for everyone else.

The chances of getting Nouveau working well on older Nvidia hardware are essentially zero. There are legal issues with firmware redistribution.

> Little hope of reclocking becoming available for GM20x, GP10x and GV100 as firmware now needs to be signed by NVIDIA to have the necessary access

From https://nouveau.freedesktop.org/

Which regressions for Nvidia hardware are you talking about?

Nouveau supports seriously ancient Nvidia hardware. It's perfectly reasonable to make a clean break once in a while (both AMD and Intel have done it in their open source drivers).

The kernel driver manages things like display setup, memory management (isolation between processes), and funneling of commands from user space to the hardware. This is more a replacement for Nouveau than NVK.

NVK generates the commands to send to the hardware, by converting vulkan APIs to Nvidia-specific instructions, and feeds them to the kernel driver.

The problem with Nouveau was that Nvidia hardware could not be reclocked (the cores were basically always in low power mode) without a signed firmware blob. That blob couldn't be legally distributed except by Nvidia, so the open source folks more or less gave up on a high performance driver.

Now, for newer hardware, Nvidia has changed some aspects of the firmware and allows redistribution. So it's feasible to make a good open source driver.

AMD and Intel also use different drivers for different hardware generations, since eventually things change so much that it's better to start clean.

With regards to reverse engineering, Mesa has a number of reverse engineered drivers. That isn't anything new.

Ah yes, guilty until proven innocent. Yes, the assumption upon which all civilized societies are founded.

/s

I'm sure this depends on engineering culture, but at Google, where Go was born, there is "one way to write C++" and that is dictated by the style guide.

You have to learn many restrictions, a number of which feel arbitrary/highly subjective, and follow that. Of course, the style guide evolves over time so it's not like legacy code is consistent with new code.

The end result is that I much prefer the Go way of having fewer options in the language, instead of having those options in the language but restricting them via policy.

Not sure early versions of rust is the best example of refcounting overhead. There are a bunch of tricks you can use to decrease that, and it usually doesn't make sense to invest too much time into that type of thing while there is so much flux in the language.

Swift is probably a better baseline.