tveita

9 ms·

91.
by tveita 2y ago
So SpinQuant learns a rotation for activations and weights that, to my understanding, "smear" the outlier weights out so you don't get extreme values in any one weight. Random anecdote warning - In the old days, before vector
92.
by tveita 2y ago
A web of trust or reputation based system can be built on top of the signature scheme - maybe an appearance of smaller invite-only forums that share reputation. Or maybe it will just become an integrated part of the moderation of existing p
93.
by tveita 2y ago
(2016), this is still technically true, but "utf8mb4" is the default encoding since MySQL 8.0, and "utf8" is a deprecated alias for "utf8mb3", so this isn't nearly as much of a footgun.
94.
by tveita 2y ago
I think Rama [1] (by Nathan Marz behind Apache Storm) is interesting as a "NoSQL" solution for a similar problem space, as I understand it. Impressive if this can support similar scale using only SQL. [1] https://redpl
95.
by tveita 2y ago
That still looks like a more targeted approach than "precise strikes" on schools and humanitarian zones like in https://www.bbc.com/news/articles/cwyx9znxl4eo
96.
by tveita 2y ago
Who programs the tapes? https://en.wikipedia.org/wiki/Profession_(novella)
97.
by tveita 2y ago
You can handle missing values if you plan for it when you design the API. As the article says you can use the well-known optional types, or now mark fields as optional. Protocol buffers include a set of rules to follow for backwards and for
98.
Report on virtual currencies in gaming: Getting played(forbrukerradet.no)
2 points by tveita 2y ago|0 comments
99.
by tveita 2y ago
General text embeddings aren't that great for keyword search. The best embedding match for a paragraph is going to be another paragraph with the same meaning. "Cats are natural carnivores." should be very similar to "Fel
100.
by tveita 2y ago
Even a few years ago it would be perfectly possible to do audio based content ID on known content to uncover viewing preferences on streaming media - a proven business model, ref https://samsungads.events/acrguide-pr These
101.
by tveita 2y ago
An unlocked bootloader is the only way to take a local backup of app data now, afaik, since they've made the adb backup tool more and more useless - and phone transfers too. So you can either have no banking app, lose all your data whe
102.
by tveita 2y ago
Apple is leaning into this for cloud AI: https://security.apple.com/blog/private-cloud-compute/ Probably enough to be genuinely hard for an APT to backdoor even with insider help. Still, ultimately the security re
103.
by tveita 2y ago
Signal users should be happy about this, because it exposes the SGX enclaves as the false security it is. It's not like this is the first SGX exploit. You should assume that Signal engineers, if they chose to, could access the user dat
104.
by tveita 2y ago
The same authors also have a language model at https://github.com/Zyphra/Zamba2 but it's not clear to me if that model is connected to tree attention. The announcement at https://www.zyphra.com/pos
105.
by tveita 2y ago
Same with the Lusophon Mathematical Olympiad 2023 example. It makes a bunch of specious claims about parity. (Adding 4 to a number changes the parity? Therefore, the parity of each colour always changes for each turn? Therefore, the parity
106.
by tveita 2y ago
The text of Arthur Conan Doyle's short story for the library is available online https://www.arthur-conan-doyle.com/index.php/How_Watson_Lear...
107.
by tveita 2y ago
The combat reminds me of https://en.wikipedia.org/wiki/The_Typing_of_the_Dead , a touch-typing game that plays the cheesy zombie genre oddly straight.
108.
by tveita 2y ago
Maybe this can save you by accident but it feels too fragile to rely on for security. For in-house code you might know every compiler and CPU you will run on now and in the future. For open source, are you comfortable releasing code with si
109.
by tveita 2y ago
A good old fashioned AI approach, e.g. compare with Peter Norvig's https://github.com/norvig/paip-lisp/blob/main/lisp/simple.li... > A third, quieter feeling - if the universe was to communi
110.
by tveita 2y ago
On the one hand, yes, of course, and on the other hand on articles about "double employment" you'll see comments basically arguing that the company deserves it if your manager isn't micromanaging you and tracking metrics
111.
by tveita 3y ago
So your business metrics suddenly dropped, but what has changed? This service is using 80% CPU, that seems a bit high... but is it always this high? Looks like it spiked within the last hour. But wait, it does that every Monday at 9 am, s
112.
by tveita 3y ago
It mushes all the concepts together into very recognizable AI soup. Let's see how it does with Native American https://stockcake.com/i/historical-handshake-moment_626739_3... https://stockcake.com/
113.
by tveita 3y ago
Notably that was a "no-one-but-us" backdoor, that requires a specific secret key to exploit. We'll see when someone analyzes the payload further, but presumably this backdoor also triggers on a specific private key. If not th
114.
by tveita 3y ago
For an idea of what's possible you might be interested in this story that was just on HN where they fine tune a quantized 70b model in 48GB VRAM: https://www.answer.ai/posts/2024-03-06-fsdp-qlora.html
115.
by tveita 3y ago
At one point the official App Store guidelines outright included a threat to not complain publicly about Apple's policies, because why not. "If your app is rejected, we have a Review Board that you can appeal to. If you run to the
116.
by tveita 3y ago
Some "didn't make it in" things that looked promising were the perceptual vector quantization[1], and a butterfly transform that Monty was working on, IIRC as an occasional spectator to the process. [1] https://jmv
117.
by tveita 3y ago
Right, you read something nebulous about how "the LLM combines the retrieved words and its own response to the query into a final answer it presents to the user", and you think there is some magic going on, and then you click one
118.
by tveita 3y ago
RAG doesn't require much data or involve any training, it is a fancy name for "automatically paste some relevant context into the prompt" Basically if you have a database of three emails and ask when Biff wanted to meet for l
119.
by tveita 3y ago
Another dubious optimization you could do is finding an optimal transport solution to move the apples the least amount of distance, something similar to the Earth Mover's Distance. https://jeremykun.com/2018/03&#x2
120.
by tveita 3y ago
Are there any cool highlights you can give us about gemma.cpp? Does it have any technical advantages over llama.cpp? It looks like it introduces its own quantization format, is there a speed or accuracy gain over llama.cpp's 8-bit quan

More ›