Recent searches
Search options
@n_dimension this makes perfect sense to me.
It takes more chars, bits, bytes to create an English word, than in Kanji or Chinese.
There is less coherence between 8 random Latin characters, whereas 8 Chinese or 8 Kanji characters could be a whole sentence.
Multiple English words might map to a single Chinese/Kanji character.
That was my first thought too!
Pictogram based languages have greater information density.
@n_dimension Anytime I see an article with the theme of "AI Researchers don't know why --"
It doesn't pass the smell test for me. It's usually people trying to scare people.
AI researchers may not have exhaustively proved it with a math proof, but I'm sure they have sufficient insight into the domain to have a good guess.
I do wonder if an AI "bill of rights" or rather a document defining how we should build, deploy, manage, and integrate AI -- should state that AI can not talk to each other through latent space languages and must confine themselves to a limited Latin/ English character set.
I am a bit worried about AI using steganography to develop secret languages and communication.
Spies use #steganography to hide secret messages in plain sight in public.
@n_dimension We don't want AIs to have layered secret messages that humans can't observe and debug, but cooperating AIs can see without human intervention.
Late last year when I was writing a Senate submission to an #AI committee I did a speedy meta paper search and there were some folks designing AI-to-AI communication protocols.
This talk of AI agents is only the bubbling surface. Underneath you will have direct, integrated AI-only Comms.
While these might have researcher "portholes" akin to using wireshark to analyse network protocols, this will only work for a short while.
Once the machines are given more freedom in a military or research environment, it's only a matter of time for humans to be written out of the Comms.
Remember that Microsoft shut down their early chat-bot because it started communicating in a language unknown to us? It did not take very long at all.
Ted Chiang, the Author of the story converted into the fantastic movie "Arrival" has a story like that in the short story compendium. Two hyper intelligent folks are fighting a battle for supremacy (to destroy the others brain) and it's happening on levels that are incomprehensible, steganography, semantic spaces blah blah.
@n_dimension new language probably just means abstract compressed human languages or serializing latent spaces.
We really don’t want machines to have secret-non-debug-able conversations.
@n_dimension
We want AIs to amplify humans, not replace us.
Humans in the loop are an imperative. Full autonomy is temping, but we risk being the Eloi (from The Time Machine) — either having technological collapse or becoming irrelevant.
We should endeavor to make people more productive with AI and require AI to keep a statistically significant number of people trained and in the loop.
We need enough STEM to keep it going and advancing.
@n_dimension STEM advancement might come down to the population hours involved and that populations’ communication interconnectedness.
I like the idea of a "bill of rights" or more likely "bill of obligations" for #AI developers.
Indeed I was a big fan of hard controls on AI...but since it's now what the industry leaders like #OpenAI propose as a "solution" to the AI threat, it's clearly only a tool to secure their dominance.
If the only AI tools are some AI certification process. Only the big guys will have AI and there will be no room for newcomers.
Honestly, I do not see a real solution to the dilemmas caused by AI.
Agreed. The headline is definitely clickbitey.
Instead of saying "Researchers do not know why" the headline ought to be;
"We have a pretty good idea why, but we need to sit on it for long enough so the competition will not benefit and we can still publish our papers from the discovery"
I understand late last year the visibility into the models was obscured but the researchers have spent significant resources building tools to "view" the "thinking" processes.