1) Someone runs into an interesting problem that can potentially be solved with ML/AI. They try to solve it for themselves.
2) "Hey! The model is kind of working. It's useful enough that I bet other people would pay for it."
3) They launch a paid API, SaaS startup, etc. and get a few paying customers.
4) Turns out their ML/AI method doesn't generalize so well. Reputation is everything at this level, so they hire some human workers to catch and fix the edge cases that end up badly. They tell themselves that they can also use it to train and improve the model.
5) Uh-oh, the model is underperforming, and the human worker pipeline is now some significant part of the full workflow.
6) Then someone writes an article about them using cheap human labor.
AI stands for "Actually, Indians."
>Apple's AI/ML group has been dubbed "AIMLess" internally
The article: https://www.macrumors.com/2025/04/10/chaos-behind-siri-revea...
I can’t even count the number of of times I have shut down “AI” projects where the actual plan was to use a labor pool to simulate AI, in order to create the training data to replace the humans with AI. Don’t get me wrong, it’s not a terrible idea for some cases, but you can’t just come straight out of the gate with fraud. Well, I mean, you could. But. Maybe you shouldn’t.
"Our bleeding edge AI/MT app..." does not sound bad at all.
they 100% use this AI "Actually Indians" technology
Is your problem that their phrasing invites stereotyping, or that the stereotype it invites happens to be negative? Because if it's the latter, do you really think that's the semantic intention here?
Regarding Germans, if the news was, "AG deportation private company is a scam, they were sending people to forced euthanasia"
and someone came and said, "AG stands for Actually Germans". I am sure no German would want to be associated with that.
I would also note that OP merely posited a thought experiment. They’re not policing anyone. “Get a life” is perhaps a little harsh?
Here’s another thought experiment. If you had a job interview for a senior position at Microsoft and your interviewer was Satya Nadella, would you make this joke?
That's why the persons, intent and context makes all the difference between something being funny and something being ofensive. And you could tell that statement wasn't in bad faith or meant to be derogatory.
>If you had a job interview for a senior position at Microsoft and your interviewer was Satya Nadella, would you make this joke?
Please don't move the goalposts to bad faith arguments, The casualness of the HN comment sections very different than the context of a job interview, hence my comment above on context mattering. Do you talk to your friends the way you talk to HR at work?
And yes, I'm sure graybeard Microsoft employees who worked with Nadella for a long time also make such jokes and banter with him behind closed doors and they all laugh, people are still people and don't maintain their work persona 24/7 or they'd go crazy.
You obviously disagree, which is fine - I'm not calling you a racist - but to me, "get a life" is a pretty harsh reaction to someone raising a concern about a joke they find in poor taste.
The point I was trying to make about the job interview and Nadella, which I may have made clumsily, is not that we ought to use the tone we use in job interviews everywhere. My point is that Nadella is an extraordinarily accomplished Indian person and this "joke" would likely fall flat with that sort of audience. I think that's a decent barometer for whether the joke is in poor taste or not.
Again speaking personally, as a white dude, I avoid making jokes about minorities. That used to be pretty much common sense, although I recognize that there's a certain culturally ascendant viewpoint that disagrees with that. But my decision to treat people respectfully isn't about what's culturally in vogue, and I'm still willing to bet that a lot of the people of colour who laugh along when white people make jokes at their expense are thinking something else entirely.
Elsewhere in the world, we'd all it xenophobia, or Indophobia if one has something against Indian people specifically.
Though in this case, it's driven primarily by economic stereotypes, coming from the country becoming a cheap services outsourcing destination for the West, so there should be a better term coined for it. The anti-Indian sentiment in IT seems to be the services equivalent of the common "Made in China = cheap crap" belief, and because it applies to services and not products, it turns into discriminating people.
Even of you replace 'reality' with 'true' amd 'truth' the logic doesn't quite work out.
If I replace "reality" with "truth", it does not work out because it makes no sense: "truth cannot be racist" makes no sense whatsoever. In relation or correspondence to what? It does work with reality, however.
As a comment on "Reality cannot be racist".
FWIW my best friend is South Indian. In the North, he is hated for unknown reasons, different caste or whatever, I do not know. He usually tells me everything about Indians that I do ask.
Another FWIW, he likes a barista (he is currently studying in the UK where he does not face any racism), and he came to me with help as to how to approach her. It is a good thing he did (he admitted it) because he would have ruined it by seeming so desperate, which seems to be a common trend among Indians, too. This is reality, too. To what extent? I do not know, but enough to notice. He is otherwise (despite being an Indian) a very well-mannered, curious, and smart person. He does seem to care much less about hygiene than one should, and blames things on diet, rather than just lack of hygiene. We discussed it in detail and he agreed, eventually.
While we are at it, I dislike Indian accents, too, generally. This is a preference. It does not make me racist. Do you think it does?
At any rate, if you have any questions, I am willing to answer (with his permission if it concerns him), but ultimately, I do not think it is racist.
Yes, I did. By repeating someone else's apropos joke, I get to reap the sweet, sweet internet points.
The most important part of your reputation is admitting fault. Sometimes your product isn't perfect. Lying to your investors about automation rates is far worse for your reputation than just taking the L.
The key is to at least be believable. Anyone with any sense realized that Theranos' claims were literally impossible.
It’s tens of millions of dollars lol.
The emperor just took off his socks and is starting to do a little uncoordinated “sexy” dance
So it's doubly surprising to me the government chose (criminal) wire fraud, not (civil) securities fraud, which would have a lower burden of proof.
Government lawyers almost never try to make their job harder than it has to be.
You might argue this is a flawed example, but we've automated huge workflows at work that turned major time-consuming PITAs into something it wouldn't occur to most people that a human has anything to do with it.
You could try to convince a jury of this argument, sure. Do you think it will work? And if you do go with that argument then are you actually convincing the jury of your guilty conscience- often an important part of a white collar crime where state of mind of the defendant is very important?
a good example is O'Connor v. Oakhurst Dairy, No. 16-1901, also known as the Maine Dairy oxford comma case. the District Court followed the intent but the Appeals court followed the law as written.
https://www.smithsonianmag.com/smart-news/missing-oxford-com...
from the Appeals Court ruling
> The District Court concluded that, despite the absent comma, the Maine legislature unambiguously intended for the last term in the exemption's list of activities to identify an exempt activity in its own right. The District Court thus granted summary judgment to the dairy company, as there is no dispute that the drivers do perform that activity. But, we conclude that the exemption's scope is actually not so clear in this regard.
https://cases.justia.com/federal/appellate-courts/ca1/16-190...
I'd argue the opposite. AI typically generalizes very well. What it can't do well is specifics. It can't do the same thing over and over and follow every detail.
That's what's surprised me about so many of these startups. They're looking at it from the bottom-up, something ai is uniquely bad at.
See:
> As SANIGER knew, at the time nate was claiming to use AI to automate online purchases, the app’s actual automation rate was effectively 0%. SANIGER concealed that reality from investors and most nate employees: he told employees to keep nate’s automation rate secret; he restricted access to nate’s “automation rate dashboard,” which displayed automation metrics; and he provided false explanations for his secrecy, such as the automation data was a “trade secret.”
> SANIGER claimed that nate's "deep learning models" were "custom built" and use a "mix of long short-term memory, natural language processing, and reinforcement learning."
> When, on the eve of making an investment, an employee of Investment Firm-1 asked SANIGER about nate's automation rate, that is, the percentage of transactions successfully completed with nate's AI technology, SANIGER claimed that internal testing showed that "success ranges from 93% to 97%."
(from [1])
[1]: https://www.justice.gov/usao-sdny/media/1396131/dl?inline
> But despite Nate acquiring some AI technology and hiring data scientists, its app’s actual automation rate was effectively 0%, the DOJ claims.
Sometimes people are just dishonest. And when those people use their dishonestly to fleece real people, they belong in prison.
Neither of these solved the problem that our stack is a pile of cat shit and needs some maintenance from people who know what the hell they are doing. It’s not solving a problem. It’s adding another layer of cat shit.
https://thespinoff.co.nz/the-best-of/06-03-2018/the-mystery-...
Interestingly this was a task that could probably be done well enough by AI now.
Not that these guys knew how close to reality they turned out to be. I assume they just had no idea of the problem they were attempting and assumed that it was at the geotaging a photo end of the scale when it was at the 'is it a bird' end.
Maybe I'm being overly optimistic in assuming people who do this are honestly attempting to solve the problem and fudging it to buy time. In general they seem more deluded about their abilities than planning a con from start to finish.
In that case, I believe it's a scam. 0% isn't some edge case.
Tesla robots and Taxis enter the room...
It worked by camera tracking the shelves contents, and would adjust the inventory level for a specific customers actions. And finally, tracked the incremental mass change during the checkout process to cross reference label swap scams etc.
Thus, people get flagged if their appearance changes while in the store, mass of goods is inconsistent with scanned labels, or the cameras don't see the inventory re-stocked.
You would be surprised how much irrational effort some board members put into the self-checkout systems. Personally, I found the whole project incredibly boring.... so found a more entertaining project elsewhere... =3
It always seemed to be random and coincided with Kroger doing the "scan as you shop" trial thing.
Which is doubly annoying, because I'm in that line to save time, and now I have to hunt down one of the employees who isn't paying attention or where they're supposed to be.
You could try to stop them but if they are hurt in the process it could very well end in a lengthy trip to prison.
Here's some examples for apples:
You probably need to eat more fruits and vegetables.
I've never seen those here; the scale just has a touch screen menu with pictures.
- https://news.ycombinator.com/item?id=38715111
- https://www.wsj.com/business/retail/uniqlo-self-checkout-rfi...
At a BAC of 0.08 (legal limit in US) drivers have reaction time delayed by only 60-120ms but crash risk is 10x compared to sober
Lack of depth perception probably compounds this?
If that was too much, we wouldn't let people in their 50s drive.
I'm not sure that slower reaction times are the only effect of alcohol consumption.
How could you not be 10x more likely to crash than the nurse getting off at 2am who has driven the route a thousand times and knows all the bad blind spots / bad intersections / is still well within her normal waking hours. That is much closer to the normal profile of the sober people who are out driving during prime drinking hours.
Most people appear to take about 2 seconds to respond to any change in conditions.
This is hopefully illegal and not actually what is done, because I have learned from Waymo that it is not permissible or even possible for the CS reps to remotely drive the car. They merely push "suggestion" commands to be considered by the onboard Waymo Driver.
Remote human drivers have too much latency and not enough realtime information available to "drive" a vehicle on public roads.
> In the event of an emergency, the vehicle automatically puts itself into a safe state within milliseconds by coming to a safe stop in the same lane.
It sounds to me like the hardware has some amount of autonomy. They just aren't trying to do the high level stuff. Both companies seem like they're trying to hide the implementation details though which immediately makes me suspicious of them.
I'd be more concerned about the remote driver's internet connection crapping out. The car probably has multiple simultaneous cellular connections (e.g. PepLink SpeedFusion hot failover type thing).
A human operator wouldn’t even be able to read or interpret the types of data which would be collected and sent by a vehicle such as a FSD Tesla or a Waymo.
Now as I understand it, military forces are really good at remotely operated drones/UAVs so perhaps the tech does exist in parallel, but those are two distinct applications.
At 70 mph you'd traverse the full length of a car before the brake kicks in
Furthermore, that’s a brand-confusion name they’re d/b/a. “Veyo” is a very well established ride sharing provider, based in San Diego, and specializing in human drivers for NEMT.
Come, Mister Tally-Mon, Tally Me Banana: Daylight Come And Me Wann’ Go Home
> So much of “ai” is just figuring ways to offload work onto random strangers.
https://xkcd.com/1897/ (2017)
while sleeping and connected by NeuraLink. Before Musk/NeuraLink gets to me though, judging by the content of some of my dreams, i've been driving a space-folding spaceships for some aliens.
Shopping in 2025 must be a frustrating experience for magicians.
Why would it matter to you if it’s a real human or AI? Wrong in any case.
> Amazon Go: Early on, Amazon was clear that it was testing “Just Walk Out” tech — and it was known (at least in tech circles) that they had humans reviewing edge cases through video feeds. Some even joked about the “humans behind the AI.” > Their core claim was that eventually the tech would get better, and the human backup was mostly for training data and quality assurance. > They didn’t say, “this is 100% AI with zero human help right now.”
> Nate: Claimed it was already fully automated. > Their CEO explicitly said the AI was doing all the work — “without human intervention” — and only used contractors for rare edge cases. > According to the DOJ, the truth was: humans were doing everything, and AI was just a branding tool. > Investors were told it was a software platform, when it was really a BPO in disguise.
And everyone knows that ChatGPT Pro is exclusively powered by capuchin monkeys.
With Waymo vehicles, it's the car's responsibility to sense the issue and brake, so we say that the car is driving and the human is a "remote assistant". With Vay, it's the human's responsibility because they are the driver.
This ends up having a lot of meaningful distinctions across the stack, even if it seems like a superficial distinction at first.
Not sure if this used to be the case but today Waymos can’t be controlled remotely by humans, only ‘guided’: https://www.govtech.com/transportation/waymo-robotaxis-getti... (ctrl+f “cannot be controlled”)
"Mostly AI, but they failed at getting close enough to 100%" and "effectively 0% AI" are not the same thing.
IANAL, of course.
It's the same tech used at Intuit Dome for the food stalls.
Perhaps this came up because investors finally got a peak at margins and saw there was a giant off shore line item. Otherwise it seems like an "automation rate" is a really ambiguous number for investors to track.
> This type of deception not only victimizes innocent investors
Also this was a funny line
[0] https://www.businessinsider.com/amazons-just-walk-out-actual...
Doesn’t matter what consumers believe, it’s more or less legal to lie to consumers about how a product works, as long as investors know how the sausage is made. (Though, in reality it’s near impossible to lie to customers without also misleading investors, especially for publicly listed companies)
In this case, investors were under the impression that the AI worked, completing 99% of transactions without any human intervention. In reality, it was essentially 0%
> Saniger raised millions in venture funding by claiming that Nate was able to transact online “without human intervention,” except for edge cases where the AI failed to complete a transaction. But despite Nate acquiring some AI technology and hiring data scientists, its app’s actual automation rate was effectively 0%, the DOJ claims.
Fraud is often defined as gaining something (or depriving someone else from something, or both) via false pretences. Here the something is money (this is most commonly the case) and the gaining/depriving is gaining money and depriving investors of it. It is more complicated than that, with many things that fit this simple description not legally being considered fraud (though perhaps being considered another crime), and can vary a fair bit between legal jurisdictions.
A cynical thought is that the key line being crossed here is that the victims are well-off investors, if you or I were conned similarly the law might give less of a stuff because we can't afford the legal team that these investors have. This is why cases like this one are successful, but companies feel safe conning their customers (i.e. selling an “unlimited” service that has, or developers five minutes after signing up, significant limits). Most investors wouldn't agree to the forced arbitration clauses and other crap that we routinely agree to by not reading and subsequently not accepting the Ts & Cs, etc, and anyway can afford large, capable, legal resources where our only hope would be a class-action from which only the lawyers really benefit.
Another cynical thought is that the line crosses was the act of not being successful. I'm sure the investors wouldn't have cared about the fraud if the returns had been very good.
If the LLM really costs less for the level of tasks that are paid for in MT right now, there sure would be a brief arbitrage period followed by the reajusting of that line I assume (of just MT shutting down if it doesn't make sense anymore)
Take juding response pairs for DPO for example, how do you ever prove someone used ChatGPT?
ChatGPT is good enough to decide in a way that will feel internally consistent, and even if you ask MTurk users to provide their logic, ChatGPT can produce a convincing response. Eventually you're forced to start measuring noisy 2nd and 3rd order signals like "did the writing in their rationale sound like ChatGPT?"
And what's especially tough is that this affects hard to verify tasks disproportionately, while those are exactly the kinds of tasks you'd generally want MTurk for.
> And what's especially tough is that this affects hard to verify tasks disproportionately, while those are exactly the kinds of tasks you'd generally want MTurk for.
That's where I'd see MT just shutting down as being a very real possibility. If fraud management and consumers leaving the platform because of too much junk or unverifiable results, the whole concept could just fall apart from a business standing point.
We saw the same phenomenon I think with earlier "get paid to navigate the web" kind of scheme way back in the days, with a watch process monitoring the user actions on the computer and paying by the hour. Very quickly people found new ways to fake activity and game the system, and it all just shut down.
There are so many examples:
- We're going to automate provider availability, scheduling and booking hair/doctor/spa/whatever appointments for your users with AI phone calls
- We're going to sell a consumer device you talk to that will automate all your app interactions using "large action models"
- We're going to automate all of your hospital's health insurance company billing interactions with AI screen scrapers
- We're going to record your employees performing an action once in any business software tool and then automate it forever with AI to tie all your vendor systems together without custom programming.
- We're going to be able to buy anything for you from any website, automatically, no matter what fraud checks exist, because AI
Most of these start-ups are not "fraudulent"—they start with the best intentions (qualified tech founders, real target market, customers willing to pay if it works), but they eventually fail, pivot completely, or have to resort to fraud in a misguided attempt to stay alive.
The problem is that they are all using technology to try to solve a human problem. The current state of the world exists because the service provider on the other side of the equation doesn't want to be disintermediated or commoditized. They aren't going to sit there and be automated into compliance. If you perfect a way to call them with robots, they will stop answering the phone. If you perfect a way to automate their iPhone app on behalf of a user, they will block your IP address range and throw up increasingly arcane captchas. If you automate their login flows, they will switch to a different login flow or block customers they think are using automation. Your customer's experience is inconsistent at best, and you can never get rid of the humans in the loop. It leads to death by a thousand paper cuts until you bleed to death - despite customers still begging to pay for your service.
It contains the details people are asking about, including (to me) what made this actionable fraud: the solicitation of $40MM from investors based on the completely false representation that his company used AI.
LLM API driven startups should build their product assuming zero improvement from the point we're at right now since that's the only guarantee anyone has.
A friend asked me to do diligence on this company circa 2021 given my personal background in ML. The founder was adamant they had a "100% checkout success rate" based on AI, which was clearly false. He also had 2 other startups he was running concurrently (?)
Live and learn!
LLMs are mostly 'there' if one knows how to use them. Maybe they weren't when they started their business, but what kind of leader getting millions in funding doesn't understand the 2nd and 3rd order derivatives of acceleration in their space? Bozos.
If Nate was started this year it would have been an "AI startup". But i guess they started during the crypto hysteria.
The interesting part is getting enough people to use the product and want AI based shopping.
The "backend" feels very swappable. I don't feel like reading the indictment, but is there more to this story?
Also founder is probably not too worried. It is now known how you get a pardon....Does he play golf?
Technically if the AI fails a transaction (or is expected to) then I see nothing invalid about the processing being 100% human!
"AI makes it easy to produce the first 50-70% solution, but you often need a 80-95% solution to not fall over hard and it's not rare that getting there isn't just hard but hardly possible at lest for an affordable price"
And AI strapped to mouse and keyboard on an headless google or apple web engine trained with the data of click farms (or directly trained by the humans there) is lurking around the corner... if not already there, ofc.
Not defending the actions of the CEO, but c'mon.
What kind of checking did they do before investing millions?
Hell, even YC itself is pretty tech-illiterate these days.
At least the name was less on the nose than when Amazon did it.
Statistics and a shit language (English) do not an ai system make.
