GibberLink [AI-AI Communication]
github.com
github.com
At the point that two AIs discover that they are talking to each other, wouldn't it almost certainly be true that both could access the Internet and therefore the right thing to do (if more than a few bits of information need to be shared) is to exchange endpoint information and hang up the call to be able to communicate directly?
The hotel agent probably has a RAG that points at their various customer and inventory databases. The user agent has individualized information about the customer. Both of them have also likely had SFT steps that further differentiate them.
The interesting question is if Gibber-Link lets the AIs do something they couldn't otherwise do with natural languages. Does it lower some error rate? Does it reduce the time it takes to send messages? Does it effectively give the AIs additional vocabulary?
If I had to guess at the internals, they probably took the token encodings and mapped them on to tones. Then it just throws text or audio through the decoding filter and passes it back.
If that's the case, the benefits are probably limited to slightly faster communications (It's essentially a simple, lossless compression) and a slightly lower error rate (beeps are easier to correct in noisy environments).
https://en.wikipedia.org/wiki/John_Draper#Allegations_of_sex...
1: https://www.reddit.com/r/etymology/comments/13hc2gw/why_was_...
Isn't that exactly what happens with every IVR (phone menu) system?
See everyone using their own LLMs to write paragraphs that will never be read and only summarized by an LLM on the other end. We're achieved negative compression.
https://audioxpress.com/news/data-over-sound-pioneer-chirp-a...
This is also high art. This needs to be in MOMA or something.
I love this.
If you want to address a phone-with-internet-backchannel, that's valid too - but it assumes different problem statement and constraints.
on the subject of the encoding efficiency, the ggwave depo mentions the use of reed-solomon error correction to make transmission more reliable. im struggling to find any info on error correction used by bell 103 or other modems, but if they aren't as robust that could partially explain the discrepancy you're describing
I'm doubling down on my thesis. All this "AI" crap is Web3 2.0. It's nothing but scams on scams.
This AI wave is so good that it makes it easy to create scams. So you get a lot of noise.
Since it's pay per token, I would be a lot more likely to take my credit card and sign up for one of these services (they are all rather expensive to an individual) if I could get my money back for any tokens that generate hallucinations.
Why would I pay for tokens that generate lies? Scam. It's literally gambling. Put in a quarter and you might get the answer you wanted easier than searching. Didn't get it? Well, put in another quarter, rejigger your prompt, and pull the lever again. Maybe the slot machine will give you the result you want, this time. Oh, it didn't? Well, the sunk cost got a little bigger. Better pull again..
Hallucinations have become much less common among the SOTA models, especially for coding with good guidance.
I've been using Claude Sonnet 3.7 a lot the last two days, and to my knowledge, there have been zero hallucinations. It just does what I ask.
But when they do make coding mistakes - so what? So do humans.
I use it for non-coding questions all the time. The idea that the utility I get is completely negated because of occasional hallucinations is absurd.
Have you used any tools like Aider, Cursor or Code?
Surely we can do better than that. "Coding mistakes" are OK, so long as they're caught by review. However, engineers who continually make tons of mistakes without improving over time are liable to annoy their colleagues to the point of quitting, or alternatively (and hopefully more likely) asking management to remove the defective engineer. So the open question is:
Do these tools make PRs that are irritating to review?
Another related and arguably more important open question is:
Does widespread use of "AI" tools in an engineering department result in significantly more defects being deployed in production?
Another important question:
Are codebases which have evolved over a period of years in an organization which makes heavy use of "AI" tools more or less easy to understand?
Another question:
Is the MTTR of incidents affected by the use of "AI" coding tools?
I could go on, but the point is absolutely not "oh well devs make mistakes so do LLMs whatever who cares?" Until these questions are answerable concretely, this is nothing but a research topic. It's worth zero dollars. It's a fuckin' NFT.
Saying LLMs are worth zero dollars shows a disconnection from reality so profound that I doubt you'll ever be able to admit you were wrong.
I wish you all the best with your remaining career.
Unfortunately, I can't join you in this cultish belief system. I have had the benefit of boom upon bust and hype cycles upon hype cycles, "AI" summers, "AI" winters, and more clever little spacecamp boondoggles and silly con valley grifter scams than I can count. And I'm not even 40 yet.
So I'll believe it when I see it. You think you can change the world with chatcoin or whatever, go do it! Prove me wrong!
I'm not betting on you.
Those who do not know history are doomed to repeat it has been playing on a loop in my mind for the past few months... I wonder why?
Of course there's hype. Turns out I'm older than you are. I've seen hype cycles. I've also seen incredible progress.
Why not acknowledge that hype exists and LLMs aren't perfect, but that many people find value in them already and there appears to be a positive trend in their capability?
Knowledge of history is knowledge of technological progress. Hype cycles don't negate that.
I hear being tired though. Let's watch and see what happens.
Sure, they seem to be getting cheaper and to some degree more accurate, but it's extremely dangerous to extrapolate. So I won't. Neither should you or anyone else. If these things are actually any good for something show, don't tell.
EDIT: My "investment thesis" on this is that it's a huge bubble, and this "AI" hype will ultimately erase many times more value than it will create. I really hope I'm wrong, we'll see.