osmarks

It's not very useful because the point of VRAM is that it's very very high-bandwidth. CPU RAM is slow and the PCIe link to the GPU is even slower.

We do not have "agency" in the sense of a magic black box which makes optimal decisions regardless of the information we have, nor "no agency" in the sense that we'll immediately do anything because ads tell us to, but ads are incentivized to exploit human psychology as much as possible and being exposed to that is probably bad.

I am expecting lots of biotech progress during my existing likely natural lifespan.

I was expecting this to be about AI timelines or something. This is not very novel or interesting to me. Besides, it's not as if there's no plausible way to extend lifespans enough that I could live until 2123.

Believing in reincarnation does not make it true.

Some part of the data pipeline apparently uses 255 as "not recorded at this time".

Petals does a layerwise split I think. You could probably run separate experts on each system. I don't think this sort of tech is very promising so I haven't looked.

I don't expect LLMs to be good enough at engineering to trivialize this kind of thing for a while - possibly never, if something else comes along and outcompetes them.

This is mostly a property of the RLHF/instruction-tuned chatbot models most people use. Base models tend to be more varied.

"Directly access"? Computers can use cameras and robotics hardware and such. It is not as if humans are special because of using eyes and muscles instead.

Also, text does say a lot about the world. It includes descriptions of physics experiments and people's best knowledge of the laws, for instance.

It would be bad for single-consumer-GPU inference setups.

I feel like this introduces the potential for weird and hard-to-implement side channel attacks, if the sequences in a batch can affect the routing of others.

Go is internally doing async-like cooperative multitasking. You just don't see it because they actually integrated it into the language well.

I wrote a script (https://github.com/osmarks/random-stuff/blob/master/histrete...) to dump a Firefox places.sqlite database to a separate SQLite database for long-term storage (I run it nightly). It seems to delete stuff based on some combination of visit frequency and last visit time.

I suspect they need to do this to keep history searches fast, since I also separately hacked a bunch of `about:config` options to retain more history and they run quite slowly now, particularly on my phone.

It is not an issue of human nature. Centralization just makes implementation waaaay easier. Distributed systems design is very hard.

Why "Texis" rather than SQLite? The licensing seems weird and I'd trust SQLite with my data more.

If they were particularly consistent about this, they would have to watch a lot of electrical engineers.

That probably requires more forward planning and effort than most criminals will bother with.

Does it actually "damage" the display or just make it not work while 5.19.12 is loaded?

There are some tweaks out there which patch the CSS to bring stuff slightly closer to the original style.

Even if it's technically possible to work around it, it would be a ridiculous amount of effort for not particularly good reasons.

The issue is that Firefox doesn't actually support hardware video decoding on Linux without various hackery.

Presumably use on bad devices with no CLI available. Also sharing URLs with people to show them a thing quickly.

GNU/Linux requires basically the same hardware as the usual alternative (Windows). If you switch from hosted GPT-3 to selfhosted GPT-J you start needing to get and manage all the hardware, which might go underutilized most of the time depending on what your demand looks like, and which requires lots of software optimizations to use maximally effectively. You can use hosted GPT-J, though, a few companies offer that.

You can rent GPUs from various cloud providers, but those are hilariously expensive too even given datacentre GPU prices.

Seems fine, just quite slow, as it says.

It's not exactly "almost free"; you need a powerful GPU, which are extremely costly right now. But apart from the cost of that and power, free.

You might be safe now in relatively okay Western countries. Consider the possibility of places like China using these features, or just future Western governments which are worse.

https://news.ycombinator.com/item?id=20230351 says they got $927/month some time ago, which is a decent amount of income I think. It seems to be less now.

It would probably kill the battery to do that constantly, as the processor would not be able to sleep ever. Although relying on energy efficiency to not improve is not a good long term situation.