NAL but I'd be worried about treading into CFAA territory with things like this. In the US, the law allows draconian penalties if you find yourself on the wrong side.
Something like yt-dlp is just downloading public data, which I can see being defensible as automating the use of a service.
But this commandeers remote machine resources to do your compute in ways clearly not intended by the provider. I don't know how ethical it is, but I definitely wouldn't want to argue this isn't "hacking" (the bad kind) in criminal court.
Not to mention, did this "hack" ever really work? When the original post went viral showing the Chipotle chatbot reversing a linked list, I (among others who posted their results online) immediately tried it and didn't get the same results, so I always assumed it was just a faked screenshot.
I once saw the bad side of one of these draconian state laws many years ago. People rarely have the misfortune of hitting these laws in some flyover states... and I remember the local judge being really shocked by the mandated penalties for such a simple offense.
I’m not a lawyer, but Chipotle is a US company and this github repo belongs to a US citizen currently residing and employed in New York, so US law might apply here.
Yeah, this is not slap on the wrist stuff. I think the creator expects nothing more than a C&D letter, but they could face prison time if a zealous federal prosecutor wants to make an example of them.
I always thought that stuffing too much into an LLM context window was a lot like overloading a burrito.Keep cramming stuff in and eventually the tortilla gives out, and everything you added since quietly spills out the bottom.
Anyway, this agent probably has the structural integrity of a fat burito held from one corner :)
I’d been thinking about if something like this would be possible for https://chatjimmy.ai/ . The underlying model is only llama 3 8B but I’m curious what coding harnesses would be like at 17k tok/s
If you're on macOS you can try the built in LLM which I think is similar in size. There's a project called Apfel that wraps it in a CLI. Also Chrome ships with a web API called Prompt API that gives you offline access to Gemini Nano which can do both text and images at the input. Also tiny. I've integrated these into my workflows where a tiny but non zero amount of reasoning is needed in between the otherwise fully deterministic steps.
I have a personal, fully offline and local version of Windows Recall basically, but good, made using macOS built-in OCR and LLM. The reasoning requirements are tiny (just interpret the screen based on the OCR, do rolling de-duplication and summarization), but they are non-zero. The tool is valuable to me and it being dep-free and fully offline and local just gives me a good feeling.
2. Bun.$ (Bun Shell) to execute the macOS command to take a screenshot (I do this for all connected screens at that moment)
3. Bun.Image to downscale everything to 1x in case some of the screenshots are 2x
4. Bun Shell again to run a JXA AppleScript thing to use the Vision Framework or whatever it is called to OCR the image into a file
5. Bun Shell to run the Swift compiler in the one-off eval mode with inline Swift helper that runs the Foundation Models Framework built-in LLM with a system prompt that tells it what the OCR said and instructs it to glean what may be on the screen (can't do this with JXA because the models are not exposed with ObjC APIs)
6. For each screenshot, continuously, take the previous day summary file and the last OCR/context results and produce a new summary of the day
I plan on adding extra information from the OS like the currently opened windows, currently focused window, time of day etc. into the mix, but so far it hasn't been needed. It produces reports of a good enough quality for me.
I `grep` these daily summaries whenever I need to recall a link I saw or a find what channel a message I spotted was in or take another look at that one tab I already closed, maybe re-open it by its OCR'd URL etc.
I actually tried building a harness around their constraints, just to find out if it was possible, but the combination of small context window, no tool calls and just small model, made me understand, that it’s not going to work.
I added it in my oh-my-pi configuration before (it's OpenAI compatible), but Llama 3 8B is just absolutely unusable for anything coding related.
It is very fast and the latency is very good however.
give ai a self-preservation directive and let them do this for you: automatically switching models to keep themselves alive. Living off of whatever token source they can find in the wild. Surely agents can farm their own tokens through the numerous support chats, free trials, leaked keys, and whatever other sources of token generation haven’t been adequately captcha’d. An agent could forage for token sources all night to let you use them gratis during the day.
Reminds me of when I used the Amazon.com AI Chatbot (was called Rufus and they renamed it to Alexa for shopping) to do things like write fizbuzz etc. Looks like they patched it to refuse though.
Came here to say the same. I haven't tried in months but Rufus definitely spat out Python code from within the Amazon shopping app. I just had to use English instead of the local language.
I remember having success asking Rufus (Amazon's previous "shopping assistant") math and programming questions. It worked, but the quality was so bad that so I stopped wasting my time there.
I was once driving and knew where I was going, so I decided to press the gemini button to see what it does. I was able to eventually convince it to write me a Rust function that calculates prime numbers, and demanded that it read out the entire function to me line by line. Fun to mess with these systems.
Oops, I left out the context of "the gemini button in google maps", sorry. It appeared one day and I didn't want to press it while driving and screw up my route. It's supposed to assist you with route-related things, but yeah it's of course still a general purpose LLM backing it.
NAL but I'd be worried about treading into CFAA territory with things like this. In the US, the law allows draconian penalties if you find yourself on the wrong side.
Something like yt-dlp is just downloading public data, which I can see being defensible as automating the use of a service.
But this commandeers remote machine resources to do your compute in ways clearly not intended by the provider. I don't know how ethical it is, but I definitely wouldn't want to argue this isn't "hacking" (the bad kind) in criminal court.
Not to mention, did this "hack" ever really work? When the original post went viral showing the Chipotle chatbot reversing a linked list, I (among others who posted their results online) immediately tried it and didn't get the same results, so I always assumed it was just a faked screenshot.
They probably added something to the prompt after that viralness and then it was a cat and mouse game to jailbreak it
Their chat bot is pretty bad so who knows.
Whether something ever worked is not correlated with traction in a world where verification is measured by likes.
You really think someone would do that? Lie on the internet?
And if you think CFAA is bad, then the states have even harsher versions too. Illinois' version specifically criminalizes any violation of a ToS.
I once saw the bad side of one of these draconian state laws many years ago. People rarely have the misfortune of hitting these laws in some flyover states... and I remember the local judge being really shocked by the mandated penalties for such a simple offense.
Yep, the key phrase is “misuse of computing resources,” if I remember correctly. IANAL, however.
That said, kudos for creativity.
> In the US,
I’m not a lawyer, but Chipotle is a US company and this github repo belongs to a US citizen currently residing and employed in New York, so US law might apply here.
Yeah, this is not slap on the wrist stuff. I think the creator expects nothing more than a C&D letter, but they could face prison time if a zealous federal prosecutor wants to make an example of them.
And with direct links to his pesonal profile and company. Uh...
EvilNote: Put links to LinkedIn lunatics sites when committing crimes instead of my own.
[dead]
I always thought that stuffing too much into an LLM context window was a lot like overloading a burrito.Keep cramming stuff in and eventually the tortilla gives out, and everything you added since quietly spills out the bottom.
Anyway, this agent probably has the structural integrity of a fat burito held from one corner :)
The finite-memory nondeterminism monad is like a leaky burrito.
I’d been thinking about if something like this would be possible for https://chatjimmy.ai/ . The underlying model is only llama 3 8B but I’m curious what coding harnesses would be like at 17k tok/s
If you're on macOS you can try the built in LLM which I think is similar in size. There's a project called Apfel that wraps it in a CLI. Also Chrome ships with a web API called Prompt API that gives you offline access to Gemini Nano which can do both text and images at the input. Also tiny. I've integrated these into my workflows where a tiny but non zero amount of reasoning is needed in between the otherwise fully deterministic steps.
looks like the macOS one is Tahoe only. I’ve been putting of upgrading to tahoe but this might be enough to tempt me
What kind of reasoning makes this worthwhile?
I have a personal, fully offline and local version of Windows Recall basically, but good, made using macOS built-in OCR and LLM. The reasoning requirements are tiny (just interpret the screen based on the OCR, do rolling de-duplication and summarization), but they are non-zero. The tool is valuable to me and it being dep-free and fully offline and local just gives me a good feeling.
Would you ever consider writing up or sharing your setup?
The ingredients are:
1. Bun.Cron API to run a script every minute
2. Bun.$ (Bun Shell) to execute the macOS command to take a screenshot (I do this for all connected screens at that moment)
3. Bun.Image to downscale everything to 1x in case some of the screenshots are 2x
4. Bun Shell again to run a JXA AppleScript thing to use the Vision Framework or whatever it is called to OCR the image into a file
5. Bun Shell to run the Swift compiler in the one-off eval mode with inline Swift helper that runs the Foundation Models Framework built-in LLM with a system prompt that tells it what the OCR said and instructs it to glean what may be on the screen (can't do this with JXA because the models are not exposed with ObjC APIs)
6. For each screenshot, continuously, take the previous day summary file and the last OCR/context results and produce a new summary of the day
I plan on adding extra information from the OS like the currently opened windows, currently focused window, time of day etc. into the mix, but so far it hasn't been needed. It produces reports of a good enough quality for me.
I `grep` these daily summaries whenever I need to recall a link I saw or a find what channel a message I spotted was in or take another look at that one tab I already closed, maybe re-open it by its OCR'd URL etc.
I actually tried building a harness around their constraints, just to find out if it was possible, but the combination of small context window, no tool calls and just small model, made me understand, that it’s not going to work.
If you find a way to do it, I’d love to hear it!
I added it in my oh-my-pi configuration before (it's OpenAI compatible), but Llama 3 8B is just absolutely unusable for anything coding related. It is very fast and the latency is very good however.
I tried the site and can't find any information about what it is. What is it?
They make custom chips with a model's weights and parameters "hard-coded" which allows for much, much faster inference.
Codex offers a -spark model that runs on Cerebras. Not quite 17k tok/s, but _very_ fast nonetheless. Worth a look.
give ai a self-preservation directive and let them do this for you: automatically switching models to keep themselves alive. Living off of whatever token source they can find in the wild. Surely agents can farm their own tokens through the numerous support chats, free trials, leaked keys, and whatever other sources of token generation haven’t been adequately captcha’d. An agent could forage for token sources all night to let you use them gratis during the day.
OpenRouter has lots of free model providers (you pay by letting them train on it) if you actually wanted to do something like this but legally.
Pivot it to providing AI to underprivileged communities / youth / the homeless and you'll generate some good will for your trial! Best of luck!
We’re changing the world with Fortune 500 AI Support Bot Multiplexer Broker Models
Reminds me of when I used the Amazon.com AI Chatbot (was called Rufus and they renamed it to Alexa for shopping) to do things like write fizbuzz etc. Looks like they patched it to refuse though.
Came here to say the same. I haven't tried in months but Rufus definitely spat out Python code from within the Amazon shopping app. I just had to use English instead of the local language.
[dead]
I remember having success asking Rufus (Amazon's previous "shopping assistant") math and programming questions. It worked, but the quality was so bad that so I stopped wasting my time there.
How has this not been patched by the company? Hasn't this been in the wild for a long time already?
It has — https://github.com/cyberpapiii/chipotlai-max#wanted-new-prov...
I was once driving and knew where I was going, so I decided to press the gemini button to see what it does. I was able to eventually convince it to write me a Rust function that calculates prime numbers, and demanded that it read out the entire function to me line by line. Fun to mess with these systems.
> gemini
The gemini from your phone?
I mean yeah, that is what it was designed to do. It's one of the better coding LLMs out there.
Oops, I left out the context of "the gemini button in google maps", sorry. It appeared one day and I didn't want to press it while driving and screw up my route. It's supposed to assist you with route-related things, but yeah it's of course still a general purpose LLM backing it.
I always drive better when my passenger recites the prime numbers in order. That sequence above 2^n-1 is just gold to my ears!
Why not playwright and google ai mode or ai search header?
Surely Chipotle having a cloud AI budget signals something, I’m not sure what.
This is the singularity we were promised
one small typo: it's "carnitas", not 'carintas' ;-)
Almost feels like astroturfing territory
How are they not gonna get sued to smithereens?
Now imagine OpenRouter but for free support bots.
based, move on.
Next up: using Chipotle AI to solve Erdős problems
and they say the hardest thing in software is naming things, pffft...
TL;DR: this is a 23B model, and in this case the B stands for "pinto beans."
[dead]
reminiscent of when people were trying to mine bitcoin in the background of web pages, or with more trad malware