woctordho

7 ms·

91.
by woctordho 10mo ago
The worldwide median internet download bandwidth is about 100 Mbps, which is far enough for HD or bluray video. The technical barrier can be as low as 'click to search, click to download' in some user-friendly BT clients. That bei
92.
by woctordho 10mo ago
Data are basically free. Infra to store and transfer data is not.
93.
by woctordho 10mo ago
Their is always a political will in China to kill other Chinese since thousands of years ago. This works vastly different from the western humanitarian philosophy.
94.
by woctordho 11mo ago
There are plenty of ways to circumvent whitelists, such as domain fronting. Some are already battle-tested in China and we'd be happy to share them with the rest of the world.
95.
by woctordho 11mo ago
What if we deploy reversible computing, which does not produce heat?
96.
by woctordho 11mo ago
How do you know I'm human?
97.
by woctordho 1y ago
Almost all training data are on the internet. As long as the small model has enough agentic browsing ability, given it enough time it will retrieve the data from the internet.
98.
by woctordho 1y ago
As long as disk prices decay faster than 1/time, we can store data forever with a finite cost.
99.
by woctordho 1y ago
I've been using https://aesthetics.fandom.com/wiki/Aesthetics_Wiki
100.
by woctordho 1y ago
I guess it can be implemented in the 'sampler' part. When solving an actual PDE, project the output of the AI onto a space that preserves the invariants.
101.
by woctordho 1y ago
Context window extrapolation should work with hierarchical/multi-scale tokenization of images, such as Haar wavelets
102.
by woctordho 1y ago
AI is going to watching those. AI watches much faster than humans.
103.
by woctordho 1y ago
Then document it. Whenever you choose one algorithm/library/tech stack but not another, write your consideration in the documents.
104.
by woctordho 2y ago
There's deniable encryption to counter the wrench attack
105.
by woctordho 2y ago
MoEs require less computation and more memory, so they're harder to setup in small labs
106.
by woctordho 2y ago
We use Unity so it's cross-platform, including mobile. But we never implemented a server for cloud saves, so you need to sync the saves using other tools
107.
by woctordho 2y ago
Reddit has it https://www.reddit.com/r/visualnovels/wiki/recommendations
108.
by woctordho 2y ago
VNs are interesting to AI developers because they're multimodal datasets with extensive text annotations. People are already working on this https://huggingface.co/datasets/Limour/b-corpus
109.
by woctordho 2y ago
If you prefer JS to Python, there's also WebGAL https://github.com/OpenWebGAL/WebGAL
110.
by woctordho 2y ago
The script graph can be thought as a DFA, but a 'game state' (images showing on screen, music playing...) is different from a node of the script graph, and there is not a one-to-one mapping between them For example, if the script
111.
by woctordho 2y ago
It has a DSL specified for writing VNs
112.
by woctordho 2y ago
Many examples can be queried from VNDB https://query.vndb.org/8e75036b2dcfe441
113.
by woctordho 2y ago
Forward compatibility of saves is harder than people thought. VN scripts have choices and loops, so in general they are graphs, and upgrading the saves to another version requires matching two graphs. I'd be happy to know if there is a
114.
by woctordho 2y ago
The problem is that banks do not distinguish read and write permissions. I'm absolutely happy with giving out the read permission
115.
by woctordho 2y ago
Crawling the whole internet still helps solve problems in the long tail of the distribution. The purpose of big data is garbage in, signal out
116.
by woctordho 2y ago
A finite-size LLM can solve the finite-size halting problem, and an infinite-size LLM can solve the infinite-size halting problem
117.
by woctordho 2y ago
In China we have ways to obfuscate those unusual traffic into usual ones like WeChat video calling
118.
by woctordho 2y ago
3e4a3ad9f05fdfb609dda6e5f512e52506f4c1053962e21bfd93f1ed81582d16ca0fef9574fb07ab62f8f5b1373b4ddd541804c0d176f4a557d900b05047e853 (This is the hash of a string randomly popped in my mind. An LLM will write this with almost 0 probability ---
119.
by woctordho 2y ago
At this point maybe meshes are not the best representation for animation and editing. We can just use latents of neural networks
120.
by woctordho 2y ago
We can already program Minecraft. Also we can already program GTA6. But imagine interpolating Minecraft and GTA6 such that all buildings are destroyable. It may be easier to achieve using AI rather than traditional programming

More ›