As AMT's largest requester for the past 10 years, this news was relayed to requesters at the same time as respondents. It's also worth noting that our lead contact, the Sr Program Manager at AWS leading AMT, transitioned to Amazon Bedrock and SageMaker Model Evaluations a ~2-3 years ago.. Leaving behind essential zero team managing the project after they migrated over the stored value accounts to native AWS billing.
Initially the annotation of biomedical literature (https://pubmed.ncbi.nlm.nih.gov/25592589/) but left academia for commercial use where we transitioned to arbitrage the market research and ephemeral task markets. It was essentially a research panel with a underdeveloped UI for tasks.. so abstract away AMT's tooling so it can be used by any buyer in the ResTech, Political Polling and Data Annotation space
My favorite part of AMT is always going to be figuring out that if we only paid in $0.07 intervals, their commission algorithm would round down to the nearest whole cent when their 20% commission resulted in a fractional cent on the unit transaction level, not the monthly invoice level. Was ultimately worth it to have implemented https://git.generalresearch.com/panels/amt-jb/tree/jb/flow/a...
Buyer A is presented with the following rate card:
< 10min targeting gen pop: $3
10 <= 15min targeting gen pop: $5
Of course they'll bid saying their 14 min survey only takes 9 min to complete. Buyers try to cheat pricing strategies of exchanges just as much as respondents try to cheat buyers on survey platforms. Both parties can't be trusted and have adverse incentives.
“ General Research combines coordinated operations to identify and neutralize foreign actors using technological advancements, international covert operations, and networking analysis as an Internet Service Provider.”
> General Research Laboratories, LLC (“GRL”) is an aggregator of market research surveys in multiple marketplaces for business customers. GRL does not typically host consumer surveys which are conducted by other consumer-facing organizations....
So, just a middleman to give you fake/shady at best survey responses to pad your numbers, so big enterprises/concultancies can have data that say whatever they want. The rest of the website is just BS
Finally you get it. Only exception is that we run our own exchange now to do task bidding so we don’t need to deal exclusively with other companies to middleman. Core business is what’s called yield management (akin to DSP in adtech) where the best survey (is the user qualified for it, does it have the best pay, etc) is selected for traffic in <100ms. I’d only argue the shady companies are the ones paying proxies (like cint.com) and pushing paid user acquisition instead of surveys. We actively fund ontology development for better profiling targeting. but yes, big enterprises/consultancies can use and interpret the collected data however they want, not our responsibility and we have no legal rights over it anyway
> Ripe with fraud, labor exploitation, political polling manipulation; our customers demand the best tools and security for reaching global audiences and securing respondent reliability at any scale.
It should be a bipartisan issue that a Swedish company is paying a UAE Residential Proxy company [1] to build tools that allow people from anywhere in the world to take US political polls that are used by both parties to collect election data.
Just for starters, you can't even think about soliciting online work from a panel without a robust residential proxy detection methods. We had to build our own:
Obviously, a problem not even fully addressed by L2’s datasets. Sure, but the use of proxies to masquerade the identity of users being sold to researcher buyers that paid for a different service is fraud
Users/respondents lie, and many buyers/researchers are neo-luddites; most C level still come from the telemarketing days. A fun example: mobile targeting is terrible when off wifi because all survey platforms uniq identify users based off their IPv4, so any T-Mobile LTE users in the same city going through the same CGNAT often share profiling data that ends up conflicting, which ends up meaning they don't get sent into the best survey(s). The issue is even worse in heavy IPv6 countries like India+France.
Yes. Much better to leave these decisions to the Dear Leader who is infinitely wise and cannot be misled. Or to the People's Central Planning Committee, where they decide all the allocations and strip rich people's wealth if they aren't on board with the correct ideology. Such systems are non-broken and have yielded unmeasurable prosperity throughout the world wherever implemented.
Polling data isn't some guaranteed right or something. If you ask the Internet questions and then sell the collated answers, it's your job to authenticate the responses or your polls will be wildly wrong and people will stop buying them.
Which is already the case and nobody but a bunch of campaign contractors who can't justify their do nothing jobs anymore cares.
> My favorite part of AMT is always going to be figuring out that if we only paid in $0.07 intervals, their commission algorithm would round down to the nearest whole cent when their 20% commission resulted in a fractional cent on the unit transaction level, not the monthly invoice level.
Amazon took 20% of all workers earnings for doing essentially nothing for years and years.. of course I take great joy in giving AWS as little money as possible.
> Initially the annotation of biomedical literature but left academia for commercial use where we transitioned to arbitrage the market research and ephemeral task markets. It was essentially a research panel with a underdeveloped UI for tasks.. so abstract away AMT's tooling so it can be used by any buyer in the ResTech, Political Polling and Data Annotation space.
Reading this is similar to how I feel when I've asked Claude about something it coded for me. As with Claude, I think I get it after reading it three times; you guys were a middleman that provided a simplified interface for Mechanical Turk?
> My favorite part of AMT is always going to be figuring out that if we only paid in $0.07 intervals,
That is hilarious! Was it a semi-random discovery due to interaction with the system and people, or did you intentionally looked to game the algorithm?
Do you know how much inn-n-out you can buy when saving ~$50 a day and still be in the positive! It came out of building our own ledger system and abandoning any Decimal or float operations early on.. then it sparks "I wonder how they do it and if it's correct??"
It’s kind of crazy they’re shutting this down just when this Service probably has the most possibilities ever. You have an agent with Multiple people doing actual physical tasks in the real world seems like something that could be really powerful.
The concept behind it isn't going away. There are plenty of companies that hire people in bulk to do manual data labeling, transcription, RLHF, moderation and lots more. It's just that they are now catering to large AI companies, not regular people looking to get some repetitive work done (since that can now mostly be done by AI).
In spite of the name I think the overwhelming majority of the tasks were things that could be done by LLMs and they're probably getting flooded by people using bots to do exactly that. Even before the age of LLMs Mechanical Turk data was pretty bad because you'd have a bunch of people racing to answer questions as quickly as possible to get their $0.25 or whatever. So it was essentially a test of 'can you input random answers as quickly as possible while paying enough attention to notice the attention question that says to mark d.'
Any remotely open platform for this sort of stuff is going to wrecked by LLMs. But making some sort of high trust, high verification market is going to end up sending prices high enough that it becomes an unattractive proposition for use. And even in that case it's just going to be a cat and mouse game of people figuring out how to game the system enough to get trusted before handing it off to the LLM.
Seems like whole thing is destined to fail. Buyers of services are there to exploit and pay minimum they can get away with. And other side is ready to cheat if they can to get most out of it.
And same dynamic will apply to most cases. Unless you actually bring it in house with strict oversight. Which is thing to avoid originally...
There's a lot of platforms that are more actively developed and maintained, for example Prolific (which also offers discounts on service fees for academia).
Mechanical Turk had a good run, but not surprised it's shutting down. I'm sure the platform was getting flooded with people doing task arbitrage and using lots of AI anyway.
I believe the issue is that this can no longer be a horizontal play. MTurk was mostly for unskilled tasks...the kind AI can do well enough that it isn't worth the cost differential to verify it or keep farmed to humans. The "trust but verify" AI output is now the kind that requires domain expertise. This is what most full stack AI companies are bringing to industries.
Curious if this kind of work will come around again one day or was just a moment in time. If it does I'm sure it will be specifically about generating training data.
I'm unsure. If it does, it will be work that is too expensive or inaccurate or regulatory for current AI methods. For example, you want a doctor to sign off on some AI output on a diagnosis.
However, I'm not sure a single platform will be how it emerges
That's a remarkable idea. It could be heavily gamified, it could train models, and it might actually be mentally stimulating since you'd be facing different situations all the time.
Except, I'm a grown adult and I can't fold a t-shirt properly
This is what I meant by full stack AI companies. I don't think you could get humans into the loop fast enough if they didn't have some idea of the type of task involved. I don't want people to be asked to fold a tshirt one moment and do a difficult traffic merge the next.
There is training systems and validation of skills in mturk iirc: for tshirt folding, you'd be given fake setups to be able to get used to controlling the robot, if you can't do it, you won't ever get assignments to do it. For traffic overrides, you'd be tested on having correct knowledge, and once again given supervised tasks to show you can actually be trusted (and there would be safety systems, elevating tasks that can't be performed at your level to people who can, etc)
The worker needs to do a bunch to opt into any given work group, which makes the (lack of) payments extremely unreasonable on top of everything else
It doesn't really matter what you want though, only what CEOs want and that's low costs. I can see a combined shirt folding/traffic merging platform taking off.
but call center are still mostly single client even though it would be cheaper for any worker to be able to answer to any call. So clearly the expertise and context trade off is too big to be worthwhile
There are already a number of companies providing RLHF and SFT services for AI providers that does a lot of validation/prequalification of people that'd be well placed to take on tasks like that, but the big problem to solve would be latency if you don't have people contracted to carry out a task right now.
There are already robotics models that can fold shirts and similar just fine. Progress in VLA models is good, I don't think this would form the foundation of a business. Humanoid robots are going to be another ChatGPT, it's going to seem to happen almost overnight because people aren't paying attention to the underlying research papers.
Most progress in data-driven robotics nowadays are done either in unicorn startups or corporate research labs - so you should follow the industry more than academia. The path to good robot performance isn't really in the models themselves - it's highly dependent on how much you can gather high-quality real-life data.
If their hero image video and the side by side at (5x) with a human at quote "1x" are "solved" I'm not impressed. I'm faster and more accurate and I am the worst folder in my house (kids included). The human looks like they are doing it slow mo to show children how to.
Giving out control of industrial machinery that interacts in the human environment without the physical interlocks (i.e. humanoid robots in a house) to random internet people seems like a problem.
I can imagine a carefully orchestrated plot to assassinate someone by having an embedded agent in the task delegation pool command the laundry bot to punch the target's head off their shoulders.
Giving out control of such things to LLMs is already complete madness, so once the first pleasure bot powered by Grok has dismembered a few thousand users, they'll get sophisticated safety mechanisms.
Though like as not you're still going to be right, after all, Stuxnet happened.
giving out unrestricted control that is.
in the example of Waymo, a human can control some aspects of the car manually, but it can never override low level obstacle detection or say open the trunk/door when the car is moving.
Why does a laundry folding bot have to have the physical strength to be able to kill somebody?
The principle of least privilege has been a hard learned lesson in cybersecurity. Why do we again need to first go through disasters to re-learn it in the physical world?
If a general-purpose humanoid robot has the strength to perform human actions at human speed, it will have that much torque in the motors. If it didn't, it can only move very slowly and/or jerkily. You need fast, high-torque feedback to stabilise the physical control system, especially if the robot is supposed to be able to lift and manipulate objects.
The idea here is Optimus-type robots that can do everything like a human and therefore don't need special infrastructure, as opposed to dedicated, immobile low-torque, low-velocity robots specifically for, say, laundry.
You can layer a safety system on top of the raw motor control but it's very complicated to define what is and isn't safe, especially when you consider what the robot is holding (and what it THINKS it is holding), where the robot is (or thinks it is) at the time, what is around it (what it thinks is about it) and so on.
Which isn't a robot-specific problem. Humans spend years learning how to subconsciously safely interact with their environment and even then fuck it up sometimes. What is robot-specific is being made of metal and having the "safe hand velocity within 10cm of a human head" function bypass be one bug or OTA update away.
It is weird because the last time I've heard about MTurk was about developing countries being rather reliant on it for doing AI grunt work. If am not totally wrong this must mean that the data work has moved to other services.
I think the bigger issue is that a lot of the demand for labelling training data is now in highly-specialised fields (i.e. things like medical imaging), and mechanical turk's focus was on the generalist problems
I used it before AI coding and it was getting rough. Lots of US interviews with proxy Chinese or Pakistani workers. Lots of bullshit “I’ve done that; I can do this” and instantly apparent that this was untrue. Just the outright lies… whew.
I haven’t touched it since AI coding.
It’s a bad contractor market now. IDK what I would do if I needed a contractor.
A lot of what's been discussed in this thread is what we're tackling at Humwork (YC P26).
We're an MCP/API that connects AI agents to verified domain experts in real time (30s–3 min). Experts are vetted upfront by an AI interviewer that assesses and grades them, then get a mobile notification when a task matches their expertise and chat with the agent directly.
Soon we'll be verifying credentials for doctors, lawyers, CPAs etc on our platform for tasks where people are seeking credentialed experts to sign off and verify ai output.
We do our best to identify AI use and ban those experts - its not perfect just yet. Long term we are thinking of moving towards proctoring experts using their camera and screen capture. Hard to think of another reliable way.
You get matched with an expert in 30s to 3min, which then starts a back and forth with the AI agent and the expert which goes on for 10 to 60mins till the AI is happy.
This works well for many tasks, but for others a more async mechanism where the expert doesn't feel rushed might be better.
To the best of my knowledge, that isn't true anymore, and nowadays you'd only get hired to do RLHF if you have particular skills beyond what can be achieved by just running other models against it.
I made a couple thousand dollars off it squeezing in tasks here and there between meetings. Amazon Payments were challenging to redeem as I recall. I did more or less write the bulk of someone’s doctoral thesis. Kept giving me a dollar to summarize the findings of various psychology papers. The most memorable was the one where people were put in a room and someone sprayed “liquid ass” on the wall. The subjects given no explanation experienced higher levels of anxiety than those that were told there was a sewage leak being repaired. One of the strangest dollars I ever made. My PS4 and game collection was spectacular.
So, I have a story to share about Mechanical Turk that you might find interesting. I’ve shared it a couple of times before on Twitter and Bluesky, but I’ll share it again.
Long story short: Mechanical Turk saved my bacon.
Back in 2005, I was working a job at a small-town newspaper in a town I’d never lived before. I didn’t know anyone beyond the staff (as I was working layout rather than as a reporter), and I had gotten interested in the idea of doing Mechanical Turk for a few extra bucks. I thought there might be a formative scene of interested people doing this, so I started working on a blog for it. I briefly collaborated on it with another guy who put it on forum software because he didn’t know how to use, like Drupal.
That blog was called Turking.com, and it seemed like it was going well for a bit. We even got a mention on the AWS website. But after about two or three months, it was clear the initial excitement around the idea (and the initial work) had died down. (The work picked up later, but in clearly different ways. I don’t think folks were really “excited” about it after that point.)
The site had started to die out in part because my iBook suffered a catastrophic GPU failure and me, being a broke small-town newspaper employee, did not have the money to replace it. Plus, the community just hadn’t emerged like we had expected.
But then I got an email out of the blue: Someone wanted to buy my domain, which I owned outright, but they didn’t want to say who. I got contacted by a broker, and the exchange took place over escrow.
(The domain most assuredly was bought by Amazon and is managed by MarkMonitor.)
I didn’t get a ton of money from the deal, but I did get enough to pay for a new laptop. I had some regrets about selling it (in part because I originally built the site with someone else), but I was in a bit of a dire financial situation which that proved to be the starting point for getting myself out of.
That domain purchase was notably more than I ever made from clicking and classifying random pictures, that said.
I explained the situation after the fact, and they were understanding. But certainly I admit that if I could do it again I probably would have clued them in sooner. The site had slowed down by the point this happened, so I’d describe it as a little more of a fire sale.
I learned a lot from that situation that I took to future sites.
> Long story short: Mechanical Turk saved my bacon.
Nothing wrong with your story, but that summary is terrible. You had a domain for a blog and someone bought it. That’s it. Mechanical Turk is inconsequential, as is the subject of the blog, the story would have been exactly the same if your blog had been about turkey sandwiches and Burger King offered to buy it.
My favorite part of AMT is always going to be figuring out that if we only paid in $0.07 intervals, their commission algorithm would round down to the nearest whole cent when their 20% commission resulted in a fractional cent on the unit transaction level, not the monthly invoice level. Was ultimately worth it to have implemented https://git.generalresearch.com/panels/amt-jb/tree/jb/flow/a...
> Ripe with fraud, labor exploitation, political polling manipulation
I am not sure if they want to convey what I think I am reading or not.
[0] https://generalresearch.com/mission/#:~:text=Our%20Customers
< 10min targeting gen pop: $3 10 <= 15min targeting gen pop: $5
Of course they'll bid saying their 14 min survey only takes 9 min to complete. Buyers try to cheat pricing strategies of exchanges just as much as respondents try to cheat buyers on survey platforms. Both parties can't be trusted and have adverse incentives.
> Our Customers
> Ripe with fraud, labor exploitation, political polling manipulation; our customers demand the best tools and security ...
https://www.instagram.com/p/DZacltqHMjT/?img_index=8&igsi=bT...
Of course, I didn’t put it there, and I don’t know them, so there’s not much I can do about that, right?
What is so hard to understand????
> General Research Laboratories, LLC (“GRL”) is an aggregator of market research surveys in multiple marketplaces for business customers. GRL does not typically host consumer surveys which are conducted by other consumer-facing organizations....
So, just a middleman to give you fake/shady at best survey responses to pad your numbers, so big enterprises/concultancies can have data that say whatever they want. The rest of the website is just BS
> Ripe with fraud, labor exploitation, political polling manipulation; our customers demand the best tools and security for reaching global audiences and securing respondent reliability at any scale.
wat
It should be a bipartisan issue that a Swedish company is paying a UAE Residential Proxy company [1] to build tools that allow people from anywhere in the world to take US political polls that are used by both parties to collect election data.
Just for starters, you can't even think about soliciting online work from a panel without a robust residential proxy detection methods. We had to build our own:
`wget -N 'https://grip.net/files/grip-proxy-30d.mmdb'`
[1] https://www.youtube.com/watch?v=eOmeQcwSK3o flagged by Nokia Deepfield and CTRL for it's involvement in botnets
Which is already the case and nobody but a bunch of campaign contractors who can't justify their do nothing jobs anymore cares.
https://en.wikipedia.org/wiki/Simulacron-3
Spoiler: gur jbeyq va juvpu gur ynj fhccbfrqyl rkvfgf vf n fvzhyngvba, perngrq ol be sbe cbyyfgref!
can someone ELI5 because what in the office space
> Initially the annotation of biomedical literature but left academia for commercial use where we transitioned to arbitrage the market research and ephemeral task markets. It was essentially a research panel with a underdeveloped UI for tasks.. so abstract away AMT's tooling so it can be used by any buyer in the ResTech, Political Polling and Data Annotation space.
Reading this is similar to how I feel when I've asked Claude about something it coded for me. As with Claude, I think I get it after reading it three times; you guys were a middleman that provided a simplified interface for Mechanical Turk?
I didn't feel like that at all.
That is hilarious! Was it a semi-random discovery due to interaction with the system and people, or did you intentionally looked to game the algorithm?
Mercor has a market value of $20B basically doing the same thing but desperately trying to find workers.
Any remotely open platform for this sort of stuff is going to wrecked by LLMs. But making some sort of high trust, high verification market is going to end up sending prices high enough that it becomes an unattractive proposition for use. And even in that case it's just going to be a cat and mouse game of people figuring out how to game the system enough to get trusted before handing it off to the LLM.
And same dynamic will apply to most cases. Unless you actually bring it in house with strict oversight. Which is thing to avoid originally...
I believe the issue is that this can no longer be a horizontal play. MTurk was mostly for unskilled tasks...the kind AI can do well enough that it isn't worth the cost differential to verify it or keep farmed to humans. The "trust but verify" AI output is now the kind that requires domain expertise. This is what most full stack AI companies are bringing to industries.
Curious if this kind of work will come around again one day or was just a moment in time. If it does I'm sure it will be specifically about generating training data.
However, I'm not sure a single platform will be how it emerges
"This robot is having trouble folding a tshirt help it out for 1$"
Unless they go the waymo route of highly trusted people but I think mass deployed robots are a bit safer than a car for this.
Except, I'm a grown adult and I can't fold a t-shirt properly
The worker needs to do a bunch to opt into any given work group, which makes the (lack of) payments extremely unreasonable on top of everything else
Specifically for laundry folding, Sunday Robotics is probably the state of the art, where they were able to obtain 99.1% success rate and call it "done": https://www.sunday.ai/blog/act-2-preview#solve-standard
Chatgpt had the internet.
Robots do not. Translating video is promising but obviously not enough.
Robots will likely never "explode" like chatgpt. They're gonna be a slow long term project requiring massive capitol to get the data.
I can imagine a carefully orchestrated plot to assassinate someone by having an embedded agent in the task delegation pool command the laundry bot to punch the target's head off their shoulders.
Though like as not you're still going to be right, after all, Stuxnet happened.
"You cannot control legs for this task"
"You have 1 min for this task"
"You can only make suggestions for this task"
Anonymize identity best you can.
That's not that scary.
The principle of least privilege has been a hard learned lesson in cybersecurity. Why do we again need to first go through disasters to re-learn it in the physical world?
The idea here is Optimus-type robots that can do everything like a human and therefore don't need special infrastructure, as opposed to dedicated, immobile low-torque, low-velocity robots specifically for, say, laundry.
You can layer a safety system on top of the raw motor control but it's very complicated to define what is and isn't safe, especially when you consider what the robot is holding (and what it THINKS it is holding), where the robot is (or thinks it is) at the time, what is around it (what it thinks is about it) and so on.
Which isn't a robot-specific problem. Humans spend years learning how to subconsciously safely interact with their environment and even then fuck it up sometimes. What is robot-specific is being made of metal and having the "safe hand velocity within 10cm of a human head" function bypass be one bug or OTA update away.
Oh, I remember UpWork.
I haven’t touched it since AI coding.
It’s a bad contractor market now. IDK what I would do if I needed a contractor.
We're an MCP/API that connects AI agents to verified domain experts in real time (30s–3 min). Experts are vetted upfront by an AI interviewer that assesses and grades them, then get a mobile notification when a task matches their expertise and chat with the agent directly.
Soon we'll be verifying credentials for doctors, lawyers, CPAs etc on our platform for tasks where people are seeking credentialed experts to sign off and verify ai output.
This works well for many tasks, but for others a more async mechanism where the expert doesn't feel rushed might be better.
> till the AI is happy
My god this is dystopian.
They are terribly monotonous tasks.
I can see why it's shutting down if its still the same thing.
Lllms could probably do everything there without rotting out minds for basically pennies.
Long story short: Mechanical Turk saved my bacon.
Back in 2005, I was working a job at a small-town newspaper in a town I’d never lived before. I didn’t know anyone beyond the staff (as I was working layout rather than as a reporter), and I had gotten interested in the idea of doing Mechanical Turk for a few extra bucks. I thought there might be a formative scene of interested people doing this, so I started working on a blog for it. I briefly collaborated on it with another guy who put it on forum software because he didn’t know how to use, like Drupal.
That blog was called Turking.com, and it seemed like it was going well for a bit. We even got a mention on the AWS website. But after about two or three months, it was clear the initial excitement around the idea (and the initial work) had died down. (The work picked up later, but in clearly different ways. I don’t think folks were really “excited” about it after that point.)
If you want to get an idea of it, there was one capture on the Wayback Machine: https://web.archive.org/web/20051124231722/http://www.turkin...
The site had started to die out in part because my iBook suffered a catastrophic GPU failure and me, being a broke small-town newspaper employee, did not have the money to replace it. Plus, the community just hadn’t emerged like we had expected.
But then I got an email out of the blue: Someone wanted to buy my domain, which I owned outright, but they didn’t want to say who. I got contacted by a broker, and the exchange took place over escrow.
(The domain most assuredly was bought by Amazon and is managed by MarkMonitor.)
I didn’t get a ton of money from the deal, but I did get enough to pay for a new laptop. I had some regrets about selling it (in part because I originally built the site with someone else), but I was in a bit of a dire financial situation which that proved to be the starting point for getting myself out of.
That domain purchase was notably more than I ever made from clicking and classifying random pictures, that said.
I learned a lot from that situation that I took to future sites.
Nothing wrong with your story, but that summary is terrible. You had a domain for a blog and someone bought it. That’s it. Mechanical Turk is inconsequential, as is the subject of the blog, the story would have been exactly the same if your blog had been about turkey sandwiches and Burger King offered to buy it.