osmarks

8 ms·

31.
Meme Search Engine(github.com)
3 points by osmarks 3y ago|0 comments
32.
by osmarks 3y ago
I use the Mistral-7B base model for a few things via fewshot prompts, but not really for anything production, just some fun toys.
33.
by osmarks 3y ago
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.
34.
by osmarks 3y ago
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,
35.
by osmarks 3y ago
I am expecting lots of biotech progress during my existing likely natural lifespan.
36.
by osmarks 3y ago
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.
37.
by osmarks 3y ago
Believing in reincarnation does not make it true.
38.
by osmarks 3y ago
Some part of the data pipeline apparently uses 255 as "not recorded at this time".
39.
Maghammer: My Personal Data Warehouse(osmarks.net)
3 points by osmarks 3y ago|3 comments
40.
by osmarks 3y ago
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.
41.
by osmarks 3y ago
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.
42.
by osmarks 3y ago
This is mostly a property of the RLHF/instruction-tuned chatbot models most people use. Base models tend to be more varied.
43.
by osmarks 3y ago
"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 phys
44.
by osmarks 3y ago
It would be bad for single-consumer-GPU inference setups.
45.
by osmarks 3y ago
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.
46.
by osmarks 3y ago
Go is internally doing async-like cooperative multitasking. You just don't see it because they actually integrated it into the language well.
47.
by osmarks 3y ago
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 seem
48.
Programming education, tacit knowledge and LLMs(osmarks.net)
4 points by osmarks 3y ago|1 comments
49.
by osmarks 3y ago
It is not an issue of human nature. Centralization just makes implementation waaaay easier. Distributed systems design is very hard.
50.
by osmarks 4y ago
Why "Texis" rather than SQLite? The licensing seems weird and I'd trust SQLite with my data more.
51.
by osmarks 4y ago
If they were particularly consistent about this, they would have to watch a lot of electrical engineers.
52.
by osmarks 4y ago
That probably requires more forward planning and effort than most criminals will bother with.
53.
by osmarks 4y ago
Does it actually "damage" the display or just make it not work while 5.19.12 is loaded?
54.
by osmarks 5y ago
There are some tweaks out there which patch the CSS to bring stuff slightly closer to the original style.
55.
by osmarks 5y ago
Even if it's technically possible to work around it, it would be a ridiculous amount of effort for not particularly good reasons.
56.
by osmarks 5y ago
The issue is that Firefox doesn't actually support hardware video decoding on Linux without various hackery.
57.
by osmarks 5y ago
Presumably use on bad devices with no CLI available. Also sharing URLs with people to show them a thing quickly.
58.
by osmarks 5y ago
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
59.
by osmarks 5y ago
You can rent GPUs from various cloud providers, but those are hilariously expensive too even given datacentre GPU prices.
60.
by osmarks 5y ago
Seems fine, just quite slow, as it says.

More ›