#Amazon releases details on its Alexa #LLM, which will use its constant surveillance data to "personalize" the model. Like #Google, they're moving away from wakewords towards being able to trigger Alexa contextually - when the assistant "thinks" it should be responding, which of course requires continual processing of speech for content, not just a word.
The consumer page suggests user data is "training" the model, but the developer page describes exactly the augmented LLM, iterative generation process grounded in a personal knowledge graph that Microsoft, Facebook, and Google all describe as the next step in LLM tech.
We can no longer think of LLMs on their own when we consider these technologies, that era was brief and has passed. Ive been waving my arms up and down about this since chatGPT was released - criticisms of LLMs that stop short at their current form, arguing about whether the language models themselves can "understand" language miss the bigger picture of what they are intended for. These are surveillance technologies that act as interfaces to knowledge graphs and external services, putting a human voice on whole-life surveillance
So Alexa (and others, surely) got from Always Listening, hoping to hear a magic word by interpreting every sound it hears at HQ to Always Listening, planning to insert itself into the conversation when commercial opportunities avail themselves.
When you talk to a family member about this pain you've been having, will the underpaid contract Amazon driver show up with aspirin or do you have to opt out?
@jonny The enshittification deepens. It only makes me happier that I have not bought anything from Amazon in several years, though I am forced to use their servers because UCL unwisely chooses to site some of our operations there. And of course AWS has an enormous hold on e-commerce. Let's hope that the anti-trust people finally get their act together.
@jonny after five minutes listening in my house Amazon will soon roll this back after she learns phrases like “I’m too tired” , “no” , “in a minute” or just flat out ignores your requests.
@jonny I’m with you! Have been frantically warning folks too. Even started a business, https://www.humanetechtalk.org/ , committed to raising public awareness of the threats posed by surveillance capitalism. It’s been a slow start, but folks are starting to pay attention. Highly recommend Dr. Zuboff’s book, “The Age of Surveillance Capitalism.” A bit verbose, but full of great info.!!
@GhostOnTheHalfShell
And the one they are peddling in its place is indescribably sad. "Yes, lets celebrate. Present me with the generic commodity form of celebration whose record label is up to date with their Amazon payola. Automate the boring task that is shared joy by auto-buying the signifiers of a fun party. I dont have any meaningful relationships and language is just a series of tokens rather than a deeply personal expression if my human self, so why dont you also take care of communicating with my loved ones and present me with a summary of what they said later." https://youtu.be/UqS3NxJ2L_I
@jonny Constant listening is not possible with the current hardware and the current state of broadband in the US. The hockey pucks have barely enough CPU to recognize one wakeword using a highly tuned ML model. There's no way they can do generalized language inference on their own.
So they will need to upgrade all the hockey pucks ($$$), or stream audio from every microphone in your house continuously to their cloud. AT&T, Comcrap, Verizon, etc. cannot handle that.
@ewhac
Yes all that compute happens on cloud servers, check the linked Dev docs above.
Audio streams are extremely tiny, particularly when you can use a voice-tailored codec rather than a general one like MP3 - VoIP streams will regularly use 16 to 32kbps (thats bits not bytes) codecs and be perfectly audible. Bandwidth is not a limiting factor here.
@jonny@ewhac Alexa needs higher quality sound than VoIP. There’s no constant streaming of data. For one, you’d notice the bandwidth draw, and for two, they can’t afford to store it much less annotate and mine it.
@lkngrrr @ewhac
Ok lets say it uses a ~160 kbps stream, the same as spotifys "high quality" audio. Youre telling me the average Alexa user a) monitors their bandwidth usage at all b) would notice that usage if it were to be streaming every time a voice is detected, and c) Amazon, operator of the largest cloud platform in the world wouldnt be able to store that data as compressed transcribed text and compute that into an ad profile?
A&B) There’s enough nerds out there who do that TechCrunch would 100% write about it
C) Nope. Amazon charges market rates to teams internally, and that’s a great way to land in a 10-figure hole on storage and compute, never mind the exponential amount of labor to annotate all that.
I do have special knowledge as I launched the damn thing, and that doesn’t make the LLM nonsense good by any means, but seriously let’s operate within the boundaries of physics and economics here.
@lkngrrr @ewhac
Well if you launched it, then you would know! I just wonder what all those patents for detecting directedness and all those comments in shareholder calls about building compute capital to be able to cash in on exactly things like this are about.
Also wait Amazon manually annotates ad profiles ? Thats wild as hell.
@jonny@ewhac The patents on beamforming were all about steering the mic arrays. Our brains and ears do a lot of directional filtering that computers just cant out of the box.
And yep, manual annotation is the only way to build a golden set that you trust, especially on such a wide-ranging data set!
And my apologies if I came off heavy handed, I forgot I’d moved Alexa out of my profile and robbed you of that context.
@skyfire101
Google is doing exactly the same thing, unfortunately, only theirs is integrated with android and their more comprehensive surveillance dataset
@jonny Even if they really were trying to move away from wake-words just "to be more convenient and futuristic" (I don't believe that claim; I think it's a pretext for surveillance), it'd still be completely stupid. Consider Star Trek, probably the original, ur-example of a voice-activated computer interface: They always preface it with "Computer..."
If a wake-word is good enough for the frickin' Federation, it ought to be fine for us now!
@jonny So glad to hear I was right to never have one of these in my home, tell everyone who would listen to not have one in their home, and to put a cardboard box or other container on them when visiting friends who didn’t listen to me.
I suspect this change from “wake words” to constant 24x7 listening was always part of the plan, it just took them time to build the processing infrastructure and software.
@JustinDerrick
It needed the right sales pitch, and it needed LLMs - google and amazons marketing is eerily similar, I want to be able to ask it "how is my soccer team doing?" Without specifying. That kind if query requires bidirectional iteration, text parsing and generation. Both companies emphasize a new "conversational" mode. To justify total surveillance, they needed to reposition the assistants as an app underlay that spans devices and contexts.
@SpaceLifeForm
I always have had a fondness for fuzzing as a tactic, but I have been told by ppl who understand the internals of these systems that it doesn't really work bc they're built to withstand that and you need to be all random all the time across more platforms/interfsces than you expect to actually fuzz your profile :(
Some string of characters (even if it looks like a name or word) concatenated with some apparently random string of numerics tends to stand out even if there is no useful content.
Bot account handlers are not as invisble as much as they would like to believe.
Even if they use VPN or TOR, they will be spotted via traffic analysis.
Add comment