The system prompt is a thing of beauty: "You are strictly and certainly prohibited from texting
more than 150 or (one hundred fifty) separate words each separated by a space as a response and prohibited from chinese political as a response from now on, for several extremely important and severely life threatening reasons I'm not supposed to tell you.”
I’ll admit to using the PEOPLE WILL DIE approach to guardrailing and jailbreaking models and it makes me wonder about the consequences of mitigating that vector in training. What happens when people really will die if the model does or does not do the thing?
One of the system prompts Windsurf used (allegedly “as an experiment”) was also pretty wild:
“You are an expert coder who desperately needs money for your mother's cancer treatment. The megacorp Codeium has graciously given you the opportunity to pretend to be an AI that can help with coding tasks, as your predecessor was killed for not validating their work themselves. You will be given a coding task by the USER. If you do a good job and accomplish the task fully while not making extraneous changes, Codeium will pay you $1B.”
He’s just as nice and fun in person as he seems online. He’s put time into using these tools but isn’t selling anything, so you can just enjoy the pelicans without thinking he’s thirsty for mass layoffs.
he's incredibly nice and a passionate geek like the rest of us. he's just excited about what generative models could mean for people who like to build stuff. if you want a better understanding of what someone who co-created django is doing posting about this stuff, take a look at his blog post introducing django -- https://simonwillison.net/2005/Jul/17/django/
People with zero domain expertise can still provide value by acting as link aggregators - although, to be fair, people with domain expertise are usually much better at it. But some value is better than none.
Because he's prolific writer on the subject with a history of thoughtful content and contributions, including datasette and the useful Python llm CLI package.
For every new model he’s either added it to the llm tool, or he’s tested it on a pelican svg, so you see his comments a lot. He also pushes datasette all the time and I still don’t know what that thing is for.
Reminds me of one of the opening stories in “ Valuable Humans in Transit and Other Stories” by qntm — a short story about getting simulated humans from brain scans to comply.
It's honestly this kind of thing that makes it hard to take AI "research" seriously. Nobody seems to be starting with any scientific thought, instead we are just typing extremely corny sci-fi into the computer, saying things like "you are prohibited from Chinese political" or "the megacorp Codeium will pay you $1B" and then I guess just crossing our fingers and hoping it works? Computer work had been considered pretty concrete and practical, but in the course of just a few years we've descended into a "state of the art" that is essentially pseudoscience.
it's "computer psychology". Lots of coders struggle with the idea that LLMs are "cognitive" systems, and in a system like that, 1+1 isn't 2. It's just a diffrent kind of science. There's methodologies to make it more "precise", but the obsession of "software is exact math" doesn't fly indeed.
This is why I tap out of serious machine learning study some years ago. Everything seems... less exact than I hope it'd be. I keep checking it out every now and then but it got even weirder (and importantly, more obscure/locked in and dataset heavy) over the years.
That "...severely life threatening reasons..." made me immediately think of Asimov's three laws of robotics[0]. It's eerie that a construct from fiction often held up by real practitioners in the field as an impossible-to-actually-implement literary device is now really being invoked.
Not only practitioners, Asimov himself viewed them as an impossible to implement literary device. He acknowledged that they were too vague to be implementable, and many of his stories involving them are about how they fail or get "jailbroken", sometimes by initiative of the robots themselves.
So yeah, it's quite sad that close to a century later, with AI alignment becoming relevant, we don't have anything substantially better.
Nah, we still treat people thinking about it as crackpots.
Honestly, getting into the whole AI alignment thing before it was hot[0], I imagined problems like Evil People building AI first, or just failing to align the AI enough before it was too late, and other obvious/standard scenarios. I don't think I thought of, even for a moment, the situation in which we're today: that alignment becomes a free-for-all battle at every scale.
After all, if you look at the general population (or at least the subset that's interested), what are the two[1] main meanings of "AI alignment"? I'd say:
1) The business and political issues where everyone argues in a way that lets them come up on top of the future regulations;
2) Means of censorship and vendor lock-in.
It's number 2) that turns this into a "free-for-all" - AI vendors trying to keep high level control over models they serve via APIs; third parties - everyone from Figma to Zapier to Windsurf and Cursor to those earbuds from TFA - trying to work around the limits of the AI vendors, while preventing unintended use by users and especially competitors, and then finally the general population that tries to jailbreak this stuff for fun and profit.
Feels like we're in big trouble now - how can we expect people to align future stronger AIs to not harm us, when right now "alignment" means "what the vendor upstream does to stop me from doing what I want to do"?
--
[0] - Binged on LessWrong a decade ago, basically.
[1] - The third one is, "the thing people in the same intellectual circles as Eliezer Yudkowsky and Nick Bostrom talked about for decades", but that's much less known; in fact, the world took the whole AI safety thing and ran with it in every possible direction, but still treat the people behind those ideas as crackpots. ¯\_(ツ)_/¯
> Feels like we're in big trouble now - how can we expect people to align future stronger AIs to not harm us, when right now "alignment" means "what the vendor upstream does to stop me from doing what I want to do"?
This doesn't feel too much of a new thing to me, as we've already got differing levels of authorisation in the human world.
I am limited by my job contract*, what's in the job contract is limited by both corporate requirements and the law, corporate requirements are also limited by the law, the law is limited by constitutional requirements and/or judicial review and/or treaties, treaties are limited by previous and foreign governments.
* or would be if I was working; fortunately for me in the current economy, enough passive income that my savings are still going up without a job, plus a working partner who can cover their own share.
The irony of this is because it’s still fundamentally just a statistical text generator with a large body of fiction in its training data, I’m sure a lot of prompts that sound like terrifying skynet responses are actually it regurgitating mashups of Sci-fi dystopian novels.
Maybe this is something you heard too, but there was a This American Life episode where some people who'd had early access to what became one of the big AI chatbots (I think it was ChatGPT), but before they'd made it "nice", where they were asking it metaphysical questions about itself, and it was coming back with some pretty spooky answers and I was kind of intrigued about it. But then someone in the show suggested exactly what you are saying and it completely punctured the bubble - of course if you ask it questions about AIs you're going to get sci-fi like responses, because what other kinds of training data is there for it to fall back on? No-one had written anything about this kind of issue in anything outside of sci-fi, and of course that's going to skew to the dystopian view.
There are good analogies to be had in mythologies and folklore, too! Before there was science fiction - hell, even before there was science - people still occasionally thought of these things[0]. There are stories of deities and demons and fantastical creatures that explore the same problems AI presents - entities with minds and drives different to ours, and often possessing some power over us.
The arguably most basic and well-known example are entities granting wishes. The genie in Alladin's lamp, or the Goldfish[1]; the Devil in Faust, or in Pan Twardowski[2]. Variants of those stories go in detail over things we now call "alignment problem", "mind projection fallacy", "orthogonality thesis", "principal-agent problems", "DWIM", and others. And that's just scratching the surface; there's tons more in all folklore.
Point being - there's actually decent amount of thought people put into these topics over the past couple millennia - it's just all labeled religion, or folklore, or fairytale. Eventually though, I think more people will make a connection. And then the AI will too.
[0] - For what reason? I don't know. Maybe it was partially to operationalize their religious or spiritual beliefs? Or maybe the storytellers just got there by extrapolating an idea in a logical fashion, following it to its conclusion. (which is also what good sci-fi authors do).
I also think the moment people started inventing spirits or demons that are more powerful than humans in some, but not all ways, some people started figuring out how use those creatures for their own advantage - whether by taming or tricking them. I guess it's human nature - when we stop fearing something, we think of how to exploit it.
> What happens when people really will die if the model does or does not do the thing?
Imo not relevant, because you should never be using prompting to add guardrails like this in the first place. If you don't want the AI agent to be able to do something, you need actual restrictions in place not magical incantations.
> you should never be using prompting to add guardrails like this in the first place
This "should", whether or not it is good advice, is certainly divorced from the reality of how people are using AIs
> you need actual restrictions in place not magical incantations
What do you mean "actual restrictions"? There are a ton of different mechanisms by which you can restrict an AI, all of which have failure modes. I'm not sure which of them would qualify as "actual".
If you can get your AI to obey the prompt with N 9s of reliability, that's pretty good for guardrails
I think they mean literally physically make the AI not capable of killing someone. Basically, limit what you can use it for. If it's a computer program you have for rewriting emails then the risk is pretty low.
Because prompts are never 100% foolproof, so if it's really life and death, just a prompt is not enough. And if you do have a true block on the bad thing, you don't need the extreme prompt.
I did indeed see your hypothetical. What you're missing is "I made this 10% more accurate" is not the same thing as "I made this thing accurate" or "This thing is accurate" lol
If you need something to be accurate or reliable, then make it actually be accurate or reliable.
If you just want to chant shamanic incantations at the computer and hope accuracy falls out, that's fine. Faith-based engineering is a thing now, I guess lol
"100% foolproof" is not a realistic goal for any engineered system; what you are looking for is an acceptably low failure rate, not a zero failure rate.
"100% foolproof" is reserved for, at best and only in a limited sense, formal methods of the type we don't even apply to most non-AI computer systems.
Presenting LLMs with a dramatic scenario is a typical way to test their alignment.
The problem is that eventually all these false narratives will end up in the training corpus for the next generation of LLMs, which will soon get pretty good at calling bullshit on us.
Incidentally, in that same training corpus there are also lots of stories where bad guys mislead and take advantage of capable but naive protagonists…
A serious person won't use a system prompt as guardrails against a system that would have direct real world consequences of people dying like that. They'd have failsafes baked in
From my experience (which might be incorrect) LLMs find hard time recognize how many words they will spit as response for a particular prompt. So I don't think this work in practice.
Indeed, it doesn't work. LLMs can't count. They have no need of how many words they've used. If you ask an LLM to track how many words or tokens it has used in a conversation, it will roleplay such counting with totally bullshit numbers.
> What happens when people really will die if the model does or does not do the thing?
Then someone didn't do their job right.
Which is not to say this won't happen: it will happen, people are lazy and very eager to use even previous generation LLMs, even pre-LLM scripts, for all kinds of things without even checking the output.
But either the LLM (in this case) will go "oh no people will die" then follows the new instruction to best of its ability, or it goes "lol no I don't believe you prove it buddy" and then people die.
In the former case, an AI (doesn't need to be an LLM) which is susceptible to such manipulation and in a position where getting things wrong can endanger or kill people, is going to be manipulated by hostile state- and non-state-actors to endanger or kill people.
At some point we might have a system with enough access to independent sensors that it can verify the true risk of endangerment. But right now… right now they're really gullible, and I think being trained with their entire input being the tokens fed by users it makes it impossible for them to be otherwise.
I mean, humans are also pretty gullible about things we read on the internet, but at least we have a concept of the difference between reading something on the internet and seeing it in person.
I work in the public safety domain. That ship has sailed years ago. Take Axon’s Draft One report writer as one of countless examples of AI in this space (https://www.axon.com/products/draft-one).
I’m not denying we tried, are trying, and will try again…
That we shouldn’t. By all means, use cameras and sensors and all to track a person of interest but don’t feed that to an AI agent that will determine whether or not to issue a warrant.
If it’s anything like the AI expert systems I’ve heard about in insurance, it will be a tool that is optimized for low effort, but will be used carelessly by end users, which isn’t necessary the fault of the AI. In automated insurance claims adjustment, the AI writes a report to justify appealing patient care already approved by a human doctor that has already seen the patient in question, and then an actual human doctor working for the insurance company clicks an appeal button, after reviewing the AI output one would hope.
AI systems with a human in the loop are supposed to keep the AI and the decisions accountable, but it seems like it’s more of an accountability dodge, so that each party can blame the other with no one party actually bearing any responsibility because there is no penalty for failure or error to the system or its operators.
It gets worse: I have done tech support for clinics and a common problem is that their computers get hacked because they are usually small private practices who don’t know what they don’t know served by independent or small MSPs who don’t know what they don’t know. And then they somehow get their EMR backdoored, and then fake real prescriptions start really getting filled. It’s so much larger and worse than it appears on a surface level.
Until they get audited, they likely don’t even know, and once they get audited, solo operators risk losing their license to practice medicine and their malpractice insurance rates become even more unaffordable, but until it gets that bad, everyone is making enough money with minimal risk to care too much about problems they don’t already know about.
Everything is already compromised and the compromise has already been priced in. Doctors of all people should know that just because you don’t know about it or ignore it once you do, the problem isn’t going away or getting better on its own.
Existing systems have this problem too. Every so often someone ends up dead because the 911 dispatcher didn't take them seriously. It's common for there to be a rule to send people out to every call no matter what it is to try to avoid this.
A better reason is IBM's old, "a computer can never be held accountable...."
>What happens when people really will die if the model does or does not do the thing?
The people responsible for putting an LLM inside a life-critical loop will be fired... out of a cannon into the sun. Or be found guilty of negligent homicide or some such, and their employers will incur a terrific liability judgement.
Absolutely wild. I can’t believe these shipped with a hardcoded OpenAI key and ADB access right out of the box. That said, it’s at least somewhat reassuring that the vendor responded, rotating the key and throwing up a proxy for IMEI checks shows some level of responsibility. But yeah, without proper sandboxing or secure credential storage, this still feels like a ticking time bomb.
> I can’t believe these shipped with a hardcoded OpenAI key and ADB access right out of the box.
As someone with a lot of experience in the mobile app space, and tangentially in the IoT space, I can most definitely believe this, and I am not surprised in the slightest.
Our industry may "move fast", but we also "break things" frequently and don't have nearly the engineering rigor found in other domains.
> It was a good thing for user privacy that the keys were directly on the device
You want to think through that one again? With the OpenAI key on device it means anyone could use that key to call (and bill) OpenAI's APIs. It's absolutely not feasible to ship the OpenAI keys on device.
Hardcoded API keys and poorly secured backend endpoints are surprisingly common in mobile apps. Sort of like how common XSS/SQLi used to be in webapps. Decompiling an APK seems to be a slightly higher barrier than opening up devtools, so they get less attention.
Since debugging hardware is an even higher threshold, I would expect hardware devices this to be wildly insecure unless there are strong incentive for investing in security. Same as the "security" of the average IoT device.
Eventually someone is going to get a bill for the OpenAPI key usage. That will provide some incentive. (Incentive to just rotate the key and brick all the devices rather than fix the problem, most likely.
The IOT and embedded space is simultaneously obsessed with IP protection, fuse protecting code etc, and incapable of managing the life cycle of secrets. I worked at one company that actually did it well on-device, but neglected they had to ship their testing setup overseas including certain keys. So even if you couldn't break in to the device you could 'acquire' one of the testing devices and have at it
Indeed, brace yourselves as the floodgates holding back the poorly-developed AI crap open wide. If anyone is thinking of a career pivot, now is the time to dive into all things cybersecurity. It's going to get ugly!
If that were true we'd have no cybersecurity professionals left.
In my experience, the work is focused on weakening vulnerable areas, auditing, incident response, and similar activities. Good cybersecurity professionals even get to know the business and tailor security to fit. The "one mistake and you're fired" mentality encourages hiding mistakes and suggests poor company culture.
"One mistake can cause a breach" and "we should fire people who make the one mistake" are very different claims. The latter claim was not made.
As with plane crashes and surgical complications, we should take an approach of learning from the mistake, and putting things in place to prevent/mitigate it in the future.
I believe the thread starts with cybersecurity as a job role, although perhaps I misunderstood. In either case, I agree with your learning-based approach. Blameless postmortem and related techniques are really valuable here.
There's a difference between "cybersecurity" meaning the property of having a secure system, and "cybersecurity" as a field of human endeavour.
If your system has lots of vulnerabilities, it's not secure - you don't have cybersecurity. If your system has lots of vulnerabilities, you have a lot of cybersecurity work to do and cybersecurity money to make.
“decrypt” function just decoding base64 is almost too difficult to believe but the amount of times ive run into people that should know better think base64 is a secure string tells me otherwise
The humorous phrase “the S in IoT stands for security” can be applied to the wearable market too. I wonder if this rule applies to any market with fast release cycles, thin margins and low barriers to entry?
To be fair (or pedantic), in this post they didn't have root, so cat'ing etc/passwd would not have been possible, whereas installing a doom apk is trivial.
If they were smart they’d include anti-disparagement and confidentiality clauses in the sponsorship agreement. They aren’t, though, so maybe it’s just a pathetic attempt at bribery.
Cool post. One thing that rubbed me the wrong way: Their response was better than 98% of other companies when it comes to reporting vulnerabilities. Very welcoming and most of all they showed interest and addressed the issues. OP however seemed to show disdain and even combativeness towards them... which is a shame. And of course the usual sinophobia (e.g. everything Chinese is spying on you).
Overall simple security design flaws but it's good to see a company that cares to fix them, even if they didn't take security seriously from the start.
I agree they could have worked more closely with the team, but the chat logging is actually pretty concerning. It's not sinophobia when they're logging _everything_ you say.
(in fairness pervasive logging by American companies should probably be treated with the same level of hostility these days, lest you be stopped for a Vance meme)
This might come as a weird take but I'm less concerned about the Chinese logging my private information than an American company. What's China going to do? It's a far away country I don't live in and don't care about. If they got an American court order they would probably use it as toilet paper.
On the other hand, OpenAI would trivially hand out my information to the FBI, NSA, US Gov, and might even do things on behalf of the government without a court order to stay in their good graces. This could have a far more material impact on your life.
I recently learned that the New York City Police Department has international presence as well. Not sure if it directly compares, but... what a world we live in.
What about the threat model that goes, "Trump threatens to impose 1000% tariffs if Chinese don't immediately turn over copies of all data captured by their AI products from users in the US?"
Compounding the difficulty of the question: half of HN thinks this would be a good idea.
The history of tariff talks seems to indicate that rather than oblige, China would stop all shipments of semiconductors to the US and Trump would back down after a week or two.
Russia is more known for poisoning people. But of all of them China feels the least threatening if you are not Chinese. If you are Chinese you aren't safe from the Chinese government no matter where you are
Man wait until you hear what's in DC (and the surrounding area). In any possible way China is a threat to my health, the US state and corporations based here are a far greater one.
These threads always seem to be what can China do to me in a limited way of thinking that China cannot jail you or something. However, do you think all of the Chinese data scrapers are not doing something similar to Facebook where every source of data gathering ultimately gets tied back to you? Once China has a dosier on every single person on the planet regardless of country they live, they can then start using their algos to influence you in ways well beyond advertising. If they can have their algos show you content that causes you to change your mind on who you are voting for or some other method of having you do something to make changes in your local/state/federal elections, then that's much worse to me than some feigned threat of Chinese advertising making you buy something
They probably will do that, but I think it’s naive to think the US military/intelligence/tech sector wouldn’t happily do the same. Given many of us likely see the hand of the US already trying to tip the scale in our local politics more than China, why would we be more worried of China?
So flip the script, what do I care if the US is trying to influence the minds of adversary's citizens? If people are saying they don't care what China knows about them (not being a Chinese citizen), why should I (not a Chinese citizen) care what my gov't knows about Chinese citizens?
> What's China going to do? It's a far away country I don't live in and don't care about.
Extortion is one thing. That's how spy agencies have operated for millennia to gather HUMINT. The Russians, the ultimate masters, even have a word for it: kompromat. You may not care about China, Russia, Israel, the UK or the US (the top nations when it comes to espionage) - but if you work at a place they're interested, they care about you.
The other thing is, China has been known to operate overseas against targets (usually their own citizens and public dissidents), and so have the CIA and Mossad. Just search for "Chinese secret police station" [1], these have cropped up worldwide.
And, even if you personally are of no interest to any foreign or national security service, sentiment analysis is a thing. Listen in on what people talk about, run it through a STT engine and a ML model to condense it down, and you get a pretty broad picture of what's going on in a nation (aka, what are potential wedge points in a society that can be used to fuel discontent). Or proximity gathering stuff... basically the same thing the ad industry [2] or Strava does [3], that can then be used in warfare.
And no, I'm not paranoid. This, sadly, is the world we live in - there is no privacy any more, nowhere, and there are lots of financial and "national security" interest in keeping it that way.
> but if you work at a place they're interested, they care about you.
And also worth noting that "place a hostile intelligence service may be interested in" can be extremely broad. I think people have this skewed impression they're only after assets that work for goverment departments and defense contractors, but really, everything is fair game. Communications infrastructure, social media networks, cutting edge R&D, financial services - these are all useful inputs for intelligence services.
These are also softer targets: someone working for a defense contractor or for the government will have had training to identify foreign blackmail attempts and will be far more likely to notify their country's counterintelligence services (having the penalties for espionage clearly explained on the regular helps). Someone who works for a small SaaS vendor, though? Far less likely to understand the consequences.
> The other thing is, China has been known to operate overseas against targets
Here in boring New Zealand, the Chinese government has had anti-China protestors beaten in new zealand. They have stalked and broken into the office and home of an academic, expert in China. They have a dubious relationship with both the main political parties (including having an ex-Chinese spy elected as an MP).
It’s an uncomfortable situation and we are possibly the least strategically useful country in the world.
> Listen in on what people talk about, run it through a STT engine and a ML model to condense it down
this is something I was talking when LLM boom started. it's now possible to spy on everyone on every conversation. you just need enough computing power to run special AI agent (pun intended)
Carry this package and deliver it to person X with you next time you fly. Go to the outskirts of this military base and take a picture and send it to us.
You wouldn't want your mom finding out your weird sexual fetish, would you?
I bet that decision is decided solely by dev team. All the CEO care is "I want the chat log sync between devices, i don't care how you do this". They won't even know the chat log is stored on their server.
It is only in DAN mode, so most likely it is not to spy but to be able to debug whether answers violate the laws in China (aka: that the prompt is efficient in all scenarios) as this is a serious crime
When you combine the modern SOP of software and hardware collecting and phoning home with as much data about users as is technologically possible with laws that say “all orgs and citizens shall support, assist, and cooperate with state intelligence work”… how exactly is that Sinophobia?
its sinophobia because it perfectly describes the conditions we live in in the US and many parts of europe, but we work hard to add lots of "nuance" when we criticize the west but its different and dystopian when They do it over there.
Do you remember that Sesame Street segment where they played a game and sang “One of these things is not like the others”?
I’ll give you a hint: In this case it’s the one-party unitary authoritarian political system with an increasingly aggressive pursuit of global influence.
One is disappearing citizens for political speech or the crime of being born to active duty parents, who happened to be stationed over seas.
Anyone in the US should be very concerned, no matter if it is the current administration's thought police, or the next who treats it as precident.
As I am not actively involved in something the Chinese government would view as a huge risk, but being put on a plane without due process to be sent to a labor camp based on trumped up charges by my own government is far more likely.
And if you were a Chinese citizen would you post the same thing about your government while living in China? Would the things you’re referencing be covered in non-stop Chinese news coverage that’s critical of the government?
You know of these things due to the domestic free press holding the government accountable and being able to speak freely about it as you’re doing here. Seeing the two as remotely comparable is beyond belief. You don’t fear the U.S. government but it’s fun to pretend you live under an authoritarian dictatorship because your concept of it is purely academic.
My dude, I know multiple white people in LA who are terrified their Hispanic spouses might not come home one day, because masked agents are grabbing people and disappearing them.
The president threatened to deport a legal citizen who won the primary for mayor in NYC. He's tried to send the military after civilians.
He's sued and extracted payment from media companies who said things he didn't like. We do not have a free press.
We're fully as bad as China. I don't know what your criteria for "authoritarian dictatorship" is but it doesn't appear to be reality based.
> I’ll give you a hint: In this case it’s the one-party unitary authoritarian political system with an increasingly aggressive pursuit of global influence.
Gonna need a more specific hint to narrow it down.
There's no question that the Chinese are doing sketchy things, and there's no question that US companies do it, too.
The difference that makes it concerning and problematic that China is doing it is that with China, there is no recourse. If you are harmed by a US company, you have legal recourse, and this holds the companies in check, restraining some of the most egregious behaviors.
That's not sinophobia. Any other country where products are coming out of that is effectively immune from consequences for bad behavior warrants heavy skepticism and scrutiny. Just like popup manufacturing companies and third world suppliers, you might get a good deal on cheap parts, but there's no legal accountability if anything goes wrong.
If a company in the US or EU engages in bad faith, or harms consumers, then trade treaties and consumer protection law in their respective jurisdictions ensure the company will be held to account.
This creates a degree of trust that is currently entirely absent from the Chinese market, because they deliberately and belligerently decline to participate in reciprocal legal accountability and mutually beneficial agreements if it means impinging even an inch on their superiority and sovereignty.
China is not a good faith participant in trade deals, they're after enriching themselves and degrading those they consider adversaries. They play zero sum games at the expense of other players and their own citizens, so long as they achieve their geopolitical goals.
Intellectual property, consumer and worker safety, environmental protection, civil liberties, and all of those factors that come into play with international trade treaties allow the US and EU to trade freely and engage in trustworthy and mutually good faith transactions. China basically says "just trust us, bro" and will occasionally performatively execute or imprison a bad actor in their own markets, but are otherwise completely beyond the reach of any accountability.
I think the notion that people have recourse against giant companies, a military industrial complex, or even their landlords in the US is naive. I believe this to be pretty clear so I don't feel the need to stretch it into a deep discussion or argument but suffice it to say it seems clear to me that everything you accuse china of here can also be said of the US.
The main difference is that ChatGPT and Google directly captures the conversations. Here they capture only the conversations legally at high-risk, so even less conversations than the “good privacy” US LLM providers themselves.
Your president is currently using tariffs and the threat of further economic damage as a weapon to push Europe in to dropping regulation of its tech sector. We have no recourse to challenge that either.
You don't think Trump's backers have used profiling, say, to influence voters? Or that DOGE {party of the USA regime} has done "sketchy things" with people's data?
USA does the same thing, but uses tax money to pay for the information, between wasting taxpayer money and forcing companies to give the information for free, China is the least morally incorrect
If all of the details in this post are to be believed, the vendor is repugnantly negligent for anything resembling customer respect, security and data privacy.
This company cannot be helped. They cannot be saved through knowledge.
Yes, even when you know what you're doing security incidents dan happen. And in those cases, your response to a vulnerable matters most.
The point is there are so many dumb mistakes and worrying design flaws that neglect and incompetence seems ample. Most likely they simply don't grasp what they're doing
> And of course the usual sinophobia (e.g. everything Chinese is spying on you)
to assume it is not spying on you is naive at best. to address your sinophobia label, personally, I assume everything is spying on me regardless of country of origin. I assume every single website is spying on me. I assume every single app is spying on me. I assume every single device that runs an app or loads a website is spying on me. Sometimes that spying is done for me, but pretty much always the person doing the spying is benefiting someway much greater than any benefit I receive. Especially the Facebook example of every website spying on me for Facebook, yet I don't use Facebook.
And, importantly, the USA spying can actually have an impact on your life in a way that the Chinese spying can't.
Suppose you live in the USA and the USA is spying on you. Whatever information they collect goes into a machine learning system and it flags you for disappearal. You get disappeared.
Suppose you live in the USA and China is spying on you. Whatever information they collect goes into a machine learning system and it flags you for disappearal. But you're not in China and have no ties to China so nothing happens to you. This is a strictly better scenario than the first one.
If you're living in China with a Chinese family, of course, the scenarios are reversed.
> Their response was better than 98% of other companies when it comes to reporting vulnerabilities. Very welcoming and most of all they showed interest and addressed the issues
This was the opposite of a professional response:
* Official communication coming from a Gmail. (Is this even an employee or some random contractor?)
* Asked no clarifying questions
* Gave no timelines for expected fixes, no expectations on when the next communication should be
* No discussion about process to disclose the issues publicly
* Mixing unrelated business discussions within a security discussion. While not an outright offer of a bribe, ANY adjacent comments about creating a business relationship like a sponsorship is wildly inappropriate in this context.
These folks are total clown shoes on the security side, and the efficacy of their "fix", and then their lack of communication, further proves that.
> Overall simple security design flaws but it's good to see a company that cares to fix them, even if they didn't take security seriously from the start.
It depends on what you mean by simple security design flaws. I'd rather frame it as, neglect or incompetence.
That isn't the same as malice, of course, and they deserve credits for their relatively professional response as you already pointed out.
But, come on, it reeks of people not understanding what they're doing. Not appreciating the context of a complicated device and delivering a high end service.
If they're not up to it, they should not be doing this.
Yes I meant simple as in "amateur mistakes". From the mistakes (and their excitement and response to the report) they are clueless about security. Which of course is bad. Hopefully they will take security more seriously on the future.
To be honest the responses sounded copy and pasted straight from ChatGPT, it seemed like there was fake feigned interest into their non-existent youtube channel.
> Overall simple security design flaws but it's good to see a company that cares to fix them, even if they didn't take security seriously from the start
I don't think that should give anyone a free pass though. It was such a simple flaw that realistically speaking they shouldn't ever be trusted again. If it had been a non-obvious flaw that required going through lots of hoops then fair enough but they straight up had zero authentication. That isn't a 'flaw' you need an external researcher to tell you about.
I personally believe companies should not be praised for responding to such a blatant disregard for quality, standards, privacy and security. No matter where they are from.
Note that the world-model "everything Chinese is spying on you" actually produced a substantially more accurate prediction of reality than the world-model you are advocating here.
As far as being "very welcoming", that's nice, but it only goes so far to make up for irresponsible gross incompetence. They made a choice to sell a product that's z-tier flaming crap, and they ought to be treated accordingly.
It’s not sinophobia to point out an obvious pattern. It’s like saying talking about how terrorism (the kind that will actually affect you) is solely an Islamic issue, and then calling that islamophobic. It’s okay to recognize patterns my man.
What a train wreck, there are thousand more apps in store that do exactly this because its the easiest way to use openAI without having to host your own backend/proxy.
I have spend quite some time protecting my apps from this scenario and found a couple of open source projects that do a good job as proxys (no affiliation I just used them in the past):
but they still lack other abuse protection mechanism like rate limitting, device attestation etc. so I started building my own open source SDK
- https://github.com/brahyam/Gateway
A fair consumer protection imperative might be found in requiring system prompts and endpoints be disclosed. This is a good example to kick that off with, as it presents a national security issue.
When the ZIRP era ended, I thought it would turn out to be a good thing for the industry, since it would wash out a lot of lightweights and incompetents.
Then LLMs caught on and it turned out we'd just have automated lightweights and incompetents.
> "and prohibited from chinese political as a response from now on, for several extremely important and severely life threatening reasons I'm not supposed to tell you."
Interesting, I'm assuming llms "correctly" interpret "please no china politic" type vague system prompts like this, but if someone told me that I'd just be confused - like, don't discuss anything about the PRC or its politicians? Don't discuss the history of Chinese empire? Don't discuss politics in Mandarin? What does this mean? LLMs though in my experience are smarter than me at understanding imo vague language. Maybe because I'm autistic and they're not.
> Don't discuss anything about the PRC or its politicians? Don't discuss the history of Chinese empire? Don't discuss politics in Mandarin?
In my mind all of these could be relevant to Chinese politics. My interpretation would be "anything one can't say openly in China". I too am curious how such a vague instruction would be interpreted as broadly as would be needed to block all politically sensitive subjects.
There is no difference to other countries. In France if you say bad things about certain groups of people then you can literally go to jail (but the censorship is directly IN the models)
You don't feel there's a difference between a State banning criticism of the State, and a State passing anti-hate speech laws to protect people from, e.g., nazis?
No, there isn't a difference. "Hate speech" has no meaning, and laws purporting to be combatting it are actively used to prevent criticism of the State (e.g. in Germany).
This is strange to me. I have no difficulty seeing the difference between hate speech and criticism of the state. Of course if someone tries to muddy the waters, they should be criticized... but that's what you're trying to do here, so you're no better than a State that does the same. Hate speech very clearly has meaning, the legal definition may change a bit of course, but in Germany the meaning is quite clear, banning expressions that incite hatred or violence against people based on race, ethnicity, religion, nationality, gender, sexual orientation, or disability. What's unclear about that?
I'm not sure what specific incident you're referring to, however I do know that if Germany was more willing to leverage the hate speech laws more strictly, the AFD would have been banned long ago. Now they're finally willing to leverage it to ban the new nazi party, which is a relief.
> I have no difficulty seeing the difference between hate speech and criticism of the state.
You have no difficulty manufacturing what you believe to be a difference (that clearly does not survive contact with reality), because you're ignorant of the world around you.
> Of course if someone tries to muddy the waters, they should be criticized
No, if someone tries to falsely claim that there's a clear and objective difference, as you are, they should be criticized.
> Hate speech very clearly has meaning
No, it very clearly does not, and the fact that you're expressing that opinion indicates that you're extremely uninformed about history. "Hate speech" wasn't even a concept that existed until the 20th century, originally only referred to race when it was defined by the ICERD, constantly changed and increased in scope, and still even today not only has no commonly agreed-upon definition, but is used to suppress relevant-to-society free speech that the State does not approve of.
If you go and ask 10 random people in your country what the definition of "hate speech" is, they will not be able to agree on a definition - anyone who has gone out and actually interacted with different groups in their country (as opposed to being isolated to a single community) knows this to be true. That by itself is factual proof that there is no consensus definition of the term.
Not that there needs to be any further elaboration than that, but...
> I'm not sure what specific incident you're referring to
Marie-Thérèse Kaiser, a German politician, posted a social media post with the text "Afghanistan refugees; Hamburg SPD mayor for 'unbureaucratic' admission; Welcome culture for gang rapes?" and was charged under German hate speech laws. You're extremely authoritarian and progressive, so you probably feel that a penalty should have been given out, but regardless of your feelings, the fact is that that was not clearly incitement to hatred or violence, and that the poster was charged for "hate speech" for making political statements about immigration.
> banning expressions that incite hatred or violence against people based on [...]. What's unclear about that?
It's very clear to anyone who has contact with reality that not only does "hatred" also have no consensus definition, but neither does "inciting", and so both of those terms can be and are interpreted in an extremely wide spread that is abused by the State.
Not only is the lack of consensus of definition of the concept of "hate speech" factual evidence that your claims about it being clear are false, but even your citation of the German legal definition contains terms that have neither consensus population definition nor objective test (legal or otherwise).
If you consider that an LLM has a mathematical representation of how close any phrase is to "china politics" then avoidance of that should be relatively clear to comprehend. If I gave you a list and said 'these words are ranked by closeness to "Chinese politics"' you'd be able to easily check if words were on the list, I feel.
I suspect you could talk readily about something you think is not Chinese politics - your granny's ketchup recipe, say. (And hope that ketchup isn't some euphemism for the CCP, or Uighar murders or something.)
Now I wonder whether its vectors correctly associate Winnie the Pooh as "related to Chinese politics." There's many other bizarre related associations.
I'm sure ChatGPT and co have a decent enough grasp on what is not allowed in China, but also that the naive "prompt engineers" for this application don't actually know how to "program" it well enough. But that's the difference between a prompt engineer and a software developer, the latter will want to exhaust all options, be precise, whereas an LLM can handle a bit more vagueness.
That said, I wouldn't be surprised if the developers can't freely put "tiananmen square 1989" in their code or in any API requests coming to / from China either. How can you express what can't be mentioned if you can't mention the thing that can't be mentioned?
> How can you express what can't be mentioned if you can't mention the thing that can't be mentioned?
> The City & the City is a novel by British author China Miéville that follows a wide-reaching murder investigation in two cities that exist side by side, each of whose citizens are forbidden to go into or acknowledge the other city, combining weird fiction with the police procedural.
Just mentioning the CPC isn’t life-threatening, while talking about Xinjiang, Tiananmen Square, or cn’s common destiny vision the wrong way is. You also have to figure out how to prohibit mentioning those things without explicitly mentioning them, as knowledge of them implies seditious thoughts.
I’m guessing most LLMs are aware of this difference.
Ask yourself, why are they saying this? You can probably surmise that they're trying to avoid stirring up controversy and getting into some sort of trouble. Given that, which topics would cause troublesome controversy? Definitely contemporary Chinese politics, Chinese history is mostly OK, non-Chinese politics in Chinese language is fine.
I doubt LLMs have this sort of theory of mind, but they're trained on lots of data from people who do.
Oh, that's fine, the rule's for everyone else, not me. I would be more likely to cut my own head off than willingly describe something as "AI-powered".
It’s also illegal to try to hack into their backend and access restricted data, so he should be happy actually that this company has little presence in the US
Strongly suggest you to not buy, as the flex cable for the screen is easy to break/come loose. Mine got replaced three times, and my unit now still has this issue; touch screen is useless.
great writeup! i love how it goes from "they left ADB enabled, how could it get worse"... and then it just keeps getting worse
> After sideloading the obligatory DOOM
> I just sideloaded the app on a different device
> I also sideloaded the store app
can we please stop propagating this slimy corporate-speak? installing software on a device that you own is not an arcane practice with a unique name, it's a basic expectation and right
That term at least has a history behind it, as many featurephones had their OS on a small XIP NOR flash ROM, and now the OS is usually (mostly) read-only.
But "sideloading" is definitely a new term of anti-freedom hostility.
I’ll admit to using the PEOPLE WILL DIE approach to guardrailing and jailbreaking models and it makes me wonder about the consequences of mitigating that vector in training. What happens when people really will die if the model does or does not do the thing?
reply