keskival

7 ms·

31.
by keskival 2y ago
All points described are simple artifacts of tokenization.
32.
by keskival 3y ago
These don't actually work. Adversarial optimization is largely model-specific.
33.
Google's Gemini model is delayed(theverge.com)
97 points by keskival 3y ago|66 comments
34.
by keskival 3y ago
https://trog.qgl.org/20081217/the-why-your-anti-spam-idea-wo... "Specifically, your plan fails to account for: (X) Extreme profitability of spam"
35.
by keskival 3y ago
Hah, nice, my data is here: https://globatic.blogspot.com/2013/10/the-23andme-full-and-r...
36.
by keskival 3y ago
I hereby authorize any and all clones anyone ever wants to make out of my full DNA or parts thereof.
37.
by keskival 3y ago
I agree, I have had my raw DNA data and all the analysis results publicly available since the day I took the tests. No problems. https://globatic.blogspot.com/2013/10/the-23andme-full-and-r...
38.
by keskival 3y ago
The most problematic and the most probable security risk I have relating to logins is losing access. It for example took a week to restore access to my Apple account after I had forgotten to update my phone number there. Since this is the g
39.
by keskival 3y ago
"The important question to be asking is: How much of this CapEx build out is linked to true end-customer demand, and how much of it is being built in anticipation of future end-customer demand? This is the $200B question." This is
40.
by keskival 3y ago
It makes a few mistakes still, and that makes it "just a tool"? Ok, how about in a few weeks when it stops making mistakes?
41.
by keskival 3y ago
Can we stop pretending CAPTCHAs do anything now and get rid of them?
42.
by keskival 3y ago
The predictions the article makes are a bit unimaginative and shallow. The timeline is also somewhat extremely long. It's more like late 2023 than 2027. So, Language Model UX in 2024 according to me: - Pervasive: Virtual assistants are
43.
by keskival 3y ago
The simulation only has workers/producers and it has a free market – it is missing capitalism. To make it realistic you need a subset of people owning the labor output of others, taking out all the surplus, using it to buy stakes of mo
44.
by keskival 3y ago
It seems to be exactly the same as what everyone has been doing for a long time already, and they fail in citing existing implementations.
45.
by keskival 3y ago
You can just advice it to forget (that is, skip in summarization) things which seem irrelevant.
46.
by keskival 3y ago
It is known that alignment tax affects smaller models but in larger models (>~100B parameters) the "tax" starts to become negative at least when trained in RLHF: https://arxiv.org/pdf/2204.05862.pdf The la
47.
by keskival 3y ago
They aren't search engines, and if you're using them as such you are using them wrong.
48.
by keskival 3y ago
Actually against GDPR. So even though they have that dialog, they are still in violation. The need to track people is strong.
49.
by keskival 3y ago
Refusing cookies links to Google. This is of course against GDPR.
50.
by keskival 3y ago
Prisoners are allowed to use internet in many places. https://en.m.wikipedia.org/wiki/Internet_in_prisons There are three ways around this: 1. Stop using Google services as the root node for everything. 2. Make Google
51.
by keskival 3y ago
I actually authored a patent for something related to this in 2020: https://scholar.google.com/citations?view_op=view_citation&h... Not specifically for Transformers but for all kinds of neural networks. It's about
52.
by keskival 3y ago
It's also possible to use HNSW indices with Postgres using pg_embedding instead of pgvector: https://neon.tech/blog/pg-embedding-extension-for-vector-sea... Edit: Ah I see pgvector will soon support HNSW as well (
53.
by keskival 3y ago
Switching from imitative training of LLMs to self-competitive reinforcement learning allows them to surpass human level in general language-domain tasks.
54.
by keskival 3y ago
If you want to make it deterministic, just cache the responses keyed by queries.
55.
by keskival 3y ago
You aren't supposed to index vectors larger than ~128 dimensions. Because of concentration of measure which is an aspect of the curse of dimensionality the distances between high-dimensional vectors tend to become identical. You need t
56.
by keskival 3y ago
Why don't they use LLM chatbots to encourage great questions and make it easier to ask them? Somehow adding LLMs to the side of the answers feels incorrect to me in this particular case.
57.
by keskival 3y ago
It depends. In Spain plastic bags are recycled along with other plastic and metal in the same bin.
58.
by keskival 3y ago
I wonder how well LLMs understand YAML Schema format. I have found providing JSON Schema to them to be an excellent way to reduce their improvisation in their outputs intended for machine consumption.
59.
by keskival 3y ago
I thought everyone knew that softmax (and specifically exp functions in it) are poison. I have always worked around them, for example by using large epsilons (approaching one actually), and using low-order polynomial approximations for the
60.
by keskival 3y ago
I'm sure that ads thing is something they put 99% of their resources on. They can easily augment their users' posts by randomly adding ads into them when relayed to ActivityPub. They can also comment on posts automatically or half

More ›