it’s super weird that people think LLMs are so fundamentally different from neural networks, the underlying technology. neural network architectures are constantly improving, and LLMs are just a product of a ton of research and an emergence after the discovery of the transformer architecture. what LLMs have shown us is that we’re definitely on the right track using neural networks to solve a wide range of problems classified as “AI”
a lot of things are unknown.
i’d be very surprised if it doesn’t have an opt out.
a point i was trying to make is that a lot of this info already exists on their servers, and your trust in the privacy of that is what it is. if you don’t trust them that it’s run on per user virtualized compute, that it’s e2e encrypted, or that they’re using local models i don’t know what to tell you. the model isn’t hoovering up your messages and sending them back to Apple unencrypted. it doesn’t need to for these features.
all that said, this is just what they’ve told us, and there aren’t many people who know exactly what the implementation details are.
the privacy issue with Recall, as i said, is that it collects a ton of data passively, without explicit consent. if i open my KeePass database on a Recall enabled machine, i have little assurance that this bot doesn’t know my Gmail password. this bot uses existing data, in controlled systems. that’s the difference. sure maybe people see Apple as more trustworthy, but maybe sociology has something to do with your reaction to it as well.
i mean, you’re right. i’m just saying it’s a little silly to ship a Python interpreter when there are easier, better supported ways to do the same thing.
looks like tesseract provides C bindings which are probably being utilized in those apps.
people generally probably hate the iOS integration just because it’s another AI product, but they’re fundamentally different. the problem with Recall isn’t the AI, it’s the trove of extra data that gets collected that you normally wouldn’t save to disk whereas the iOS features are only accessing existing data that you give it access to.
from my perspective this is a pretty good use case for “AI” and about as good as you can do privacy wise, if their claims pan out. most features use existing data that is user controlled and local models, and it’s pretty explicit about when it’s reaching out to the cloud.
this data is already accessible by services on your phone or exists in iCloud. if you don’t trust that infrastructure already then of course you don’t want this feature. you know how you can search for pictures of people in Photos? that’s the terrifying cLoUD Ai looking through your pictures and classifying them. this feature actually moves a lot of that semantic search on device, which is inherently more private.
of course it does make access to that data easier, so if someone could unlock your device they could potentially get access to sensitive data with simple prompts like “nudes plz”, but you should have layers of security on more sensitive stuff like bank or social accounts that would keep Siri from reading it. likely Siri won’t be able to get access to app data unless it’s specified via their API.
no need for Python. there’s a Google SDK, ML Kit, that will do the heavy lifting on this. if that’s not acceptable, TensorFlow, PyTorch, and ONNX support Android, albeit not as nicely integrated.
your image processing pipeline will be imageSource -> RGB encoding -> OCR -> profit. your OCR just needs an RGB encoded image. doesn’t matter if that’s a JPEG or YUV video feed at the source.
as for if there’s an app that fits OP’s exact use case, dunno.
i guess the rare thing is the public commitment, but Apple has generally had a good track record for updates compared to its Android counterparts, who have previously failed to meet their goals or set laughable goals like 2 years.
tbh this research has been ongoing for a while. this guy has been working on this problem for years in his homelab. it’s also known that this could be a step toward better efficiency.
this definitely doesn’t spell the end of digital electronics. at the end of the day, we’re still going to want light switches, and it’s not practical to have a butter spreading robot that can experience an existential crisis. neural networks, both organic and artificial, perform more or less the same function: given some input, predict an output and attempt to learn from that outcome. the neat part is when you pile on a trillion of them, you get a being that can adapt to scenarios it’s not familiar with efficiently.
you’ll notice they’re not advertising any experimental results with regard to prediction benchmarks. that’s because 1) this actually isn’t large scale enough to compete with state of the art ANNs, 2) the relatively low resolution (16 bit) means inputs and outputs will be simple, and 3) this is more of a SaaS product than an introduction to organic computing as a concept.
it looks like a neat API if you want to start messing with these concepts without having to build a lab.
there are language models that are quite feasible to run locally for easier tasks like this. “local” rules out both ChatGPT and Co-pilot since those models are enormous. AI generally means machine learned neural networks these days, even if a pile of if-else used to pass in the past.
not sure how they’re going to handle low-resource machines, but as far as AI integrations go this one is rather tame
if it’s easier to pay, people spend more
what else would it be? it’s a pretty common embedded target. dev kits from Qualcomm come with Android and use the Android bootloader and debug protocols at the very least.
nobody is out here running a plain Linux kernel and maintaining a UI stack while AOSP exists. would be a foolish waste of time for companies like Rabbit to use anything else imo.
to say it’s “just an Android device” is both true and a mischaracterization. it’s likely got a lot in common with a smartphone, but they’ve made modifications and aren’t supporting app stores or sideloading. doesn’t mean you can’t do it, just don’t be surprised when it doesn’t work 1-1
it’s an analogy that applies to me. tldr worrying about having my identity stolen via physical access to my phone isn’t part of my threat model. i live in a safe city, and i don’t have anything the police could find to incriminate me. everyone is going to have a different threat model. some people need to brick up their windows
it’s not a password; it’s closer to a username.
but realistically it’s not in my personal threat model to be ready to get tied down and forced to unlock my phone. everyone with windows on their house should know that security is mostly about how far an adversary is willing to go to try to steal from you.
personally, i like the natural daylight, and i’m not paranoid enough to brick up my windows just because it’s a potential ingress.
seems like chip designers are being a lot more conservative from a design perspective. NPUs are generally a shitton of 8 bit registers with optimized matrix multiplication. the “AI” that’s important isn’t the stuff in the news or the startups; it’s the things that we’re already taking for granted. speech to text, text to speech, semantic analysis, image processing, semantic search, etc, etc. sure there’s a drive to put larger language models or image generation models on embedded devices, but a lot of these applications are battle tested and would be missed or hampered if that hardware wasn’t there. “AI” is a buzz word and a goalpost that moves at 90 mph. machine learning and the hardware and software ecosystem that’s developed over the past 15 or so years more or less quietly in the background (at least compared to ChatGPT) are revolutionary tech that will be with us for a while.
blockchain currency never made sense to me from a UX or ROI perspective. they were designed to be more power hungry as adoption took off, and power and compute optimizations were always conjecture. the way wallets are handled and how privacy was barely a concern was never going to fly with the masses. pile on that finance is just a trash profession that requires goggles that turn every person and thing into an evaluated commodity, and you have a recipe for a grift economy.
a lot of startups will fail, but “AI” isn’t going anywhere. it’s been around as long as computers have. i think we’re going to see a similarly (to chip designers) cautious approach from companies like Google and Apple, as more semantic search, image editing, and conversation bot advancements start to make their way to the edge.
you’d be surprised how fast a model can be if you narrow the scope, quantize, and target specific hardware, like the AI hardware features they’re announcing.
not a 1-1, but a quantized Mistral 7B runs at ~35 tokens/sec on my M2. that’s not even as optimized as it could be. it can write simple scripts and do some decent writing prompts.
they could get really narrow in scope (super simple RAG, limited responses, etc), quantize down to even something like 4 bit, and run it on custom accelerated hardware. it doesn’t have to reproduce Shakespeare, but i can imagine a PoC that runs circles around Siri in semantic understanding and generated responses. being able to reach out on Slack to the engineers that built the NPU stack ain’t bad neither.
doesn’t help that modern tools like lazy.nvim, etc make alternative hosting a barrier to entry. and a GitHub mirror is a tedious half measure.
i feel like this is a big part of it. it reminds me of the Sapir Whorf Hypothesis. search results and neural networks are susceptible to bias just like a human is; “garbage in garbage out” as they say.
the quote directly after mentions that newer or more precise searches produce more coherent results across languages. that reminds me of the time i got curious and looked up Marxism on Conservapedia. as you might expect, the high level descriptions of Marxism are highly critical and include a lot of bias, but interestingly once you dig down to concepts like historical materialism etc it gets harder to spin, since popular media narratives largely ignore those details and any “spin” would likely be blatant falsehood.
the author of the article seems to really want there to be a malicious conspiratorial effort to suppress information, and, while that may be true in some cases, it just doesn’t seem feasible at scale. this is good to call out, but i don’t think these people who concern their lives with the research and advancement of language concepts are sleeping on the fact that bias exists.