your censorship game is a lost fight
nostr is, and has never been a place where devs build towards censorship resistance. in fact, its quite the opposite. from point of view of freedom extemist, nostr is infact in a constant fight to build tools for censoring the content in it. however, you effors is wasted. all you can do is build your tiny bubbles within your friends. and past that, its pure chaos that awaits a dev of tools for censors. pure chaos occurs due to nature of all networks. the more people you "trust" in your wot circle, the more chaos you will get. the more hops you allow in your filtering bubble, the more it will break in parts.
there are no tools and no theory for building a censored bubble of decent size on a protocol that is quite the opposite what you are trying to do.
the thing is, social engineering is easier now than ever. you can download an llm model of size less than 5GB, which will write notes at speed of 1 note per 3 seconds, and half the people it replies to, does not distinquish it from a real human.
the intelligent social agents will inject to any social bubble, they will first look like everyone likes them. after they got 1000 followers, these agents will change their behaviour to proceed with their original purpose.
wot may work, but are you sure about that?
an attacker creates llm bot, which starts two write comments and react to events in target audience wot bubble, as well as follow those people. bot will also gather some information about topics the wot group is interested about. it will post some notes to its timeline.
at least 40 % of people will have no idea this is a bot
after tens of thousands comments, reactions and follows, the bot has gained some mutual followers from target wot group
wot bubble has been hijacked
now bot will change its behaviour. it will start to inject any notes to the wot bubble based on its original purpose.
nostr spam problem - free steaks for everyone
nostr was designed with the idea that p2p networks dont work well. that may or may not be the case. p2p networks typically mean that participants have a stake in the game. nostr goes to the another edge: nobody has a stake in the game. it costs nothing to send a message and it costs nothing to preserve it. a peer can send message to 100 participants, 50 probably accept it, and year later the message is still at 30 relays. there is no stake at play.
how p2p networks differ? bittorrent is our example scenario. in bittorrent a torrent basically follows laws of nature. if nobody puts effort for the survival of torrent, it dies. and its a good thing, because if bittorrent was like nostr, there would be so many torrents nobody would be able to find anything relevant anymore. at least if we expect everyone to only search through dht.
design princible of nostr is very novel:
because short notes dont cost much in storage, lets allow everyone to send any amount of them
but in practice this is like putting a tiny diamond in a sandbox. then pouring over 100 cubic meters of sand and telling someone to find the diamond!
everyone got to put their grain of sand in the sandbox! its censorship resistant sandbox. but if all you want is a diamond, all the sand is simply useless.
i think every network needs to have a stake in play for participants. most nostr devs think the stake is in social graph. but this is not necessarily how global connected networks should work.
maybe relays should not really store notes. what if relays were simply a short lived caching layer to hold the content while peer may be offline for some duration.
FAST IP ROUTING NETWORK
fast 2 hop ip routing network.
focus points:
- two hops (requestor -> middle peer -> delivering peer)
- high performance throughput (with only two hops and symmetric nature, network can reach near ISP level performance)
- hiding requestor ip from end service
out of scope:
- maximum anonymity
--> while providing minimum anonymity by hiding ip address of requestor and providing access to dynamic ip pool, focus of the network is not to provide guaranteed anonymity
symmetric network:
- every participant sends and receives equal amount of data with each of its peers
- any peer can ban any other peer. bans are always two sided due to symmetric nature of network
- every peer keeps list of all known other peers and sends them on request to any other peer
- if peer A downloads 1 GB of data from peer B, then peer B has right to download 1 GB of data through peer A. if peer A does not deliver, peer B should ban peer A.
- every peer can expect similar service level from its connected peers. network should stabilize itself to download speeds which can be delivered both ways. for example if peer A downloads through route A with average speed of 10Mbps, then it should also be able to upload to requesting peer with same speed.
there could be a #yolo model for interaction of your need
when you need to automate ui interaction for creating a new account, download register_account.onnx, which is less than 100MB in size, and can automate this task for you with speed of 5ms per action.
#ai #llm #vlm #automation
is online interaction more of an pow intelligence, or quick ant like action oriented behaviour? which is closer in describing interaction patterns done in different interfaces?
i would say both, but at what portion? at least one can say, if someone writes a piece of text, a yolo model cannot do this, but a language model is required. however, what if its possible to combine very performant yolo model with a slow language model?
a coordination system would be required. can a lightweight coordination system decide when we can just go with the low effort action model, and when we need to use more complex model for reading or writing text?
if lot of online interaction can be thought as analogue to moving legs to walk, or moving hands to grab something, i think these interactions can be done with relatively lightweight models, compared to full blown pow models called large language models.
maybe it is enough that the ui interaction model is on the level of a rat, while the more intelligent large language model then handles more complex operation, such as reading or writing tasks.
using this separation, if the theory has any basis, it could be possible to build ui automation agents that work in real time, faster than humans on basic interaction flows, while still retaining their ability to think slow when needed.
#ai #llm #vlm #yolo #automation
i wonder what is the complexity of certain ui interaction patterns
lets consider some basic interation pattern
- go to site xyz, create a new account with name "anon123"
breakdown:
- go to site ---> open firefox
- xyx ----> type xyz into address bar
- create a new account ---> find link or button that says something similar to "new account" or "create account", or a button or link that get closer to that target, such as "getting started", "join"
- after clicking correct sequence of links or buttons, a form is expected. fill name "anon123" to the field that says "name", "username" or similar
- if required fields are marked, those are expected to be filled as well
- click submit
- if error is displayed, correct any fields with error and submit again
(possible email verification or similar procedure here)
- if either success is shown, it is noticed that user is logged in, the task is complete
what can real time object detection model such as yolo learn?
object is typically a shape, but classification as a task does not necessarily require only shape. can classification model classify based on context?
in fact, we should ask, what is "context"? what makes context different from simple shape?
is it simple that if background is blue and there is red circle on it, then we can say that context is blue background but the actual object or shape is the red circle. maybe.
if this is the case, then yolo certainly can learn context based object detection. and why is this a big deal:
if context can be applied to highly performant object detection models, we can perform intelligent ui automation tasks based on these very lightweight models, which can do real time recognition from video frames.
compare this to vlms.
you may not really need a language model to automate ui interaction, if the interaction flow is known well enough. although its partially dependent on whether ui interaction requires highly complex tasks, such as solving complex captcha systems that expect pow level intelligence rather than fast paced action.
#ai #llm #vlm #yolo #automation
why waste your time watching monitor? you can automate it!
nostr:note1vkf0ffcagmrzqwlemzkj9kzecjp7chs0sqwqlkn465m4hl5jvstqfg289w
#ai #llm #vlm
how to make your digital clone
- screen record everything you do, record every input actions (mousemove, click, type)
- create dataset based on interaction data
- train vlm
ollama run devstral-abliterated-vision:latest
>>> /.../Screenshot from 2025-06-22 13-57-06.png in this puzzle, you should connect the same two icons with the dotted line as shown on the left. the dotted line is already drawn on the right. the icons on the right may be rotated, their colors may be inverted and/or there maybe other distractions. asses the state of this task by applying 0.6 similarity ranking to the icons on the right and answer the question: is the task complete or incomplete?
Added image '/.../Screenshot from 2025-06-22 13-57-06.png'
To determine if the task is complete, we need to assess whether the two icons connected by the dotted line on the right match the two icons connected by the dotted line on the left with at least a 0.6 similarity ranking.
Let's break down the process:
1. **Identify the Icons on the Left:**
- Icon 1: A circular icon with an arrow pointing to the right.
- Icon 2: An icon that looks like a gear or cogwheel.
2. **Identify the Icons on the Right:**
- The dotted line connects two icons:
- One icon at the start of the dotted line appears to be similar to the gear/cogwheel (rotated and inverted in color).
- The other icon at the end of the dotted line has an arrow pointing right, which is rotated but visually similar to the first icon on the left.
3. **Assess Similarity:**
- Compare Icon 1 from the left with the icon at the start of the dotted line on the right:
- The gear/cogwheel appears to be a match despite being inverted and rotated.
- Compare Icon 2 from the left with the icon at the end of the dotted line on the right:
- The arrow pointing right is visually similar, though it might be rotated.
4. **Determine Similarity Ranking:**
- Both icons appear to have a similarity ranking above 0.6 when considering rotation and color inversion as permissible transformations.
Based on this assessment:
**The task is complete.**
#ai #llm #vlm
ollama run qwen3:4b-q8_0
> invent function f(x), where f(50) = 5, f(10) = 2, f(60)=60, f(20)=100. im not looking for piecewise function. answer may be quite complex.
... thinking ~60-90k characters ...
f(x) = (649/60000)x³ - (2379/2000)x² + (5687/150)x - 269
yup. that passes.
i used #wolframalpha to find weights and biases for #neuralnetwork to produce given output for some inputs.
my conditions were:
- for input 99, output should be 2
- for input 50, output should be 100
here is the query i gave to wolfram:
max(0, (a * max(0, (b * 99) + c) + d * max(0, (0.4 * 99) + 0.2)) + 0.3) == 2 and max(0, (a * max(0, (b * 50) + c) + d * max(0, (0.4 * 50) + 0.2)) + 0.3) == 100
task for wolfram was to find correct values for weights a, b and biases c, d
wolfram alpha was able to solve this equation and find suitable values:
a = 1
b = -2007/(995*a)+1
c = (5*a*b + 2007)/(10*a)
d = -5/2*(a*b + 2)