What is the use case for LLM agent shoppers? I can't imagine delegating the purchase of a used item to an AI (I'd be okay with AI identifying the best deals for me to review). This must be something for people who are doing something at scale like flipping items on Ebay or drop shipping.
I imagine this type of automation existed before LLM agents came along - what do they add? Is it just the ability to evaluate the product description? Item quality is already listed as a categorical variable.
"Hey, ChatGPT/Grok/GeneriBot4000, please watch for a great deal on a 1982 stratocaster guitar - must be in good or better condition, $600 or less, and if you see it, go ahead and buy it without confirmation"
Ongoing tasks, arbitrage for mispriced postings in ways that aren't currently exploited that LLMs make feasible - by banning auto-buy, maybe they're attempting to delineate between human seeming behavior and automation, and giving AI permission to buy looks too much like a real person?
I have decent tech company salary but I don't even buy $10 books without checking everything. This week I almost bought a wrong book (manually) because how similar the title is. Automating stuff with AI is interesting, but I don't want the hassle of getting surprised and handling returns, if the item can be returned at all, especially on eBay.
I genuinely wonder, would you do that, really? Sure 600$ is not the end of the world for certain countries, but neither it is a sum I'm willing to just lose on random. What if the electronic parrot buys from an obvious counterfeit vendor or obvious scammer? Or what if it buys you a stratocaster but different? Or a random 1982 guitar? What if it ignores 600$? Or what if it buys 600$ item with 300$ shipping and 500$ customs from god knows where?
I've seen enough by now and I know that some people will just unleash LLMs on anything without almost no oversight. We can already see people use agentic IDEs with "do all the shit" flag, they would probably easily add finances to the list of automation.
Yeah I guess that makes sense for some people. I'm just not in a financial position where I'd let an AI buy a $600 used guitar without me taking a look at it first.
An '82 stratocaster would normally go for around $2000, so someone offloading an estate, fat fingering a price entry, etc, could give you a chance to double your money or more. $600 would be a very low price - same for a Martin D18 in fair+ condition, no cracks, etc.
If I were going to automate something like this, I'd have a suite of products to watch for - common enough to be reasonably frequent but obscure enough to be mispriced, kinda the whole idea behind secondhand ocmmission / antique / estate sale shops.
I don't know how EBay is supposed to differentiate automation from real users in this scenario. To get around it, all you need is human intervention at the last act, so you could fire up your bot and have it forward the "buy now" link when all parameters are met? Maybe they just don't want AI companies to have an argument for some sort of revenue sharing or commissions.
> An '82 stratocaster would normally go for around $2000, so someone offloading an estate, fat fingering a price entry, etc, could give you a chance to double your money or more. $600 would be a very low price - same for a Martin D18 in fair+ condition, no cracks, etc.
On the other hand, when people list a steeply discounted item, there's usually a good reason why they do so - opportunities for easy arbitrage are rare because people would usually prefer not to give you free money if they can help it. Signing up to automatically buy broken items for $600 without so much as looking at their condition seems like an easy way to lose a lot of money.
But most of what you are suggesting could be automated without the LLM. The price and categorical condition (new, great, good, fair, etc.) could be evaluated for a search query without getting LLM agents involved. I'm just surprised that an LLM evaluation of the written product description is the tipping point (often those descriptions are empty or contain irrelevant information), where people would switch from reviewing their carts to allowing autonomous transactions without in-the-loop supervisory control.
I agree, and in aggregate it becomes a serious issue for the platform. People who buy autonomously are going to argue personally when it fails in any way.
I'm not sure how they would intend to stop it with this policy anyway. It at best is going to be an arms race detecting them. What it does do is prevent upfront the ham handed excuse of "I didn't bid on this, my bot did".
I wager the scammer industry looking for active bots and exploiting them would thrive. Automate creation of fake listings using throwaway accounts using popular keywords and arbitraging price lower and lower, and until automatic buyers start bidding, remember the price and delete listing. Recreate listing with that price from a separate account selling bricks for 600$, and voila - free money.
My mistake, you're completely correct, perhaps even more-correct than the wonderful flavor of Mococoa drink, with all-natural cocoa beans from the upper slopes of Mount Nicaragua. No artificial sweeteners!
Drop shippers who arbitrage between major and minor ecommerce platforms need to maintain their listings, re-price things, etc. They don't care if the AI gets it wrong sometimes as long as they more than make back the cost of deploying it.
So now imagine ten thousand of these jerks telling their AI of choice "hey go scrape everything you can and re-list it for 10% more". That's a lot of load on the platforms at both ends for listings that are unlikely to generate many sales.
But that also seems like a very inefficient way to accomplish this automation task from the drop shippers side too. What do you gain from the LLM that non LLM automation couldn't do more cost effectively?
The LLMs are being used to hack around the fact that while the software mostly works for what people need any given user inevitably has a few workflows that are clunky or highly manual.
Stuff that used to have to be laboriously scripted can now be pseudo code.
"Hey ChatGPT I want to build my own personal cloud storage computer, buy all the hardware for me then walk me through building and configuring it. My budget is $600, try to get the best deals and make sure that all the parts are compatible. I'm fine with used parts as long as they're a good deal and are in working order."
I would. I would in particular like to review the cart in form of a table rendered in LLM interface, because all e-commerce sites have bullshit. user-hostile UI/UX and I'm tired of it.
Really, dealing with bullshit for you is by far the biggest promise of any agentic solutions today.
How do ticket scalpers make money? It's an automation war. You can run arbitrage strategies at scale if you can scrape markets with bots that understand unstructured data. Even if trades go wrong sometimes it can be profitable on average.
What also has been true the whole time is that nobody has been stopping companies in other countries from creating social media sites, electronic wallets, movie streaming, operating systems, image hosting, or food delivery services. You are right, creating these companies does not come from some special skills only found in the US. Not sure why you are mad at US for creating these services, especially because, like you said, nobody is forcing you to use them. And nobody in the US is going to be upset if an EU company creates a new fun or convenient web service.
If you think we are pissed about someone making a fun short video scrolling app then why have so many Americans downloaded it?
What our government is concerned about is that China does not allow big social media tech companies access to their markets, so allowing Chinese social media companies access to US markets is unfair, and a legitimate social influence concern. Of course you already know this if you are from Europe, given the myriad of restriction you guys have on tiktok.
But you are right, we do walk around wringing our fists that Spotify is dominating music streaming, going goddamnit we don't even like music but something should be done.
I recommend you read "careless people", the book about Facebook, where it is documented that Facebook illegally installed spyware alongside the Facebook app on the smartphones of their users. They monitored the apps their users where running alongside Facebook, which allowed Facebook to not only monitor all competitors but also see the rise of WhatsApp, which ultimately led to the "surprise" acquisition of WhatsApp.
Not to mention that Facebook and Google unknowingly ingested phone contact lists from smartphones of their users on a massive scale. So their "advantage" was extremely unethical behavior, which today would be considered an illegal crime.
So yes, it is literally Apple and Google stopping my European company to do the same, because they make it really hard to leech user data from their platforms.
> nobody is forcing you to use them
Do you remember internet.org? There's an interesting section in "careless people" about how Zuckerberg was working on bundling Facebook with smartphone contracts so people can use it for free. One country rolled out Facebook for free and it resulted in the Rohingyan Genocidg because Facebook enabled unchecked fake news along religious divides, while over years ignoring all warnings about the problem.
Facebook caused genocide... Ok man. And BMW, Mercedes, Bosch, Siemens, VW, Zeiss, Fiat, etc. were responsible for genocide during WW2. But this is all completely off topic.
> Apple and Google stopping my European company to do the same, because they make it really hard to leech user data from their platforms.
I cant tell if you are for or against collecting user data.
If it's any consolation US startups have to compete with existing US companies too. Also it doesn't hurt to work with them instead of against them. Some European companies are getting quite rich from such relationships. Arm is making a killing off Apple. Ayden gets paid every time you order an Uber. Google licenses Nokia and Ericsson technologies. I'm sure there are many other examples.
Read the book "careless people" there it is clearly explained how Facebook is responsible for Rohingya genocide.
The problem is that the digital world creates real-world harm to people, and US tech companies are very far from acting responsibly with the power they have. Even more so - they actively support and empower toxic behavior.
> The publishers say that their industry is bleeding out while they wait for a deal that may never come.
If your industry cannot sustain itself without checks from tech companies for using content as LLM training data that is quite a precarious situation. What was the economic situation for the news industry in Denmark prior to 2021?
And they created a powerful cartel which can pressure politicians to benefit the cartel:
The Danish Press Collective Management Organization (DPCMO), formed in 2021, now represents what CEO Karen Rønde calls a “99 percent mandate” of the entire industry.
> The conference has six focuses: AI in Drug Discovery, AI in Diagnostics, AI for Operational Efficiency, AI in Remote Healthcare, AI in Regulatory Compliance, and AI Ethics. Every single theme of this conference has converged onto AI as the only thing worth discussing.
I experienced the same situation at the human factors and ergonomics society (HFES) annual conference a few months ago. This was fine for me because I'm part of the (relatively small) AI/ML group at my company, which has traditionally focused on developing human factors engineering solutions and services. In fact the reason I was sent to HFES was to help bridge my background (phd in computational neuro) to the broader company mission. And to be honest I was looking forward to hearing (what I assumed was going to be) a wide diversity of talks. I mean, ergonomics... there will probably be like companies presenting next generation office chair designs or some shit, I thought. Instead I estimate that 50% of the talks were on one of three topics: AI trust, Explainable AI, or Human-AI teaming. Another 30-40% were on some other AI related issue, with the remaining 10-20% accounting for all other possible topics related to human factors and ergonomics.
I wonder what will be the repercussions from this current hyper-obsession with AI, and the resulting neglect to many other viable areas of research. I foresee a near future where chairs are packed with AI features, and are the source of much back pain.
But the difference is Pepsi would also have had dedicated laboratories and food scientists, scientifically controlled exhaustive testing and unlimited access to any ingredient they wanted. Thus one would expect Pepsi's testing to have had much finer granularity than in this YouTube video).
With millions of dollars tied up in just a few percent of sales you can bet Pepsi knows just about as much as Coke does about Coke's ingredients (and vice versa of course).
The research for both companies is more about the fine minutiae—keeping an optimal differentiation between the two products more than treading on each other's territory. Trampling over each other for market share is done through advertising, not by making their products the same.
Pepsi probably also has had access to Coke's supply chain and long ago acquired samples of various inputs before mixing, which would make analysis easier. The two companies know they're competing mostly on production management and brand image, not secret ingredients. A decade or so ago a secretary from Coke tried to sell some of the company's "secrets" to Pepsi. Instead of jumping on the opportunity to get Coke's secrets, they contacted Coke's legal department and the FBI, with the three working together to prosecute her.
I worked in an fMRI lab briefly as a grad student. I suspect you'd be correct but perhaps not exactly why you'd expect. Studies using fMRI measure a blood-oxygenation-level-dependent (BOLD) signal in the brain. This is thought to be an indirect measure of neural activity because a local increase in neural firing rate produces a local increase in the need for, and delivery of, oxygenated blood.
The question then is, do you expect a person who is really good at mental arithmetic to have less neural firing on arithmetic tasks (e.g., what is 147 x 38) than the average joe. I would hypothesize yes overall to solve each question; however, I'd also hypothesize the momentary max intensity of the expert to peak higher. Think of a bodybuilder vs. a SWE bench-pressing 100 lbs for 50 reps. The bodybuilder has way more muscle to devote to a single rep, and will likely finish the set in 20 seconds, while the SWE is going to take like 30 minutes ;)
I was a grad student at UCSD when Ed Vul published Voodoo Correlations in Social Neuroscience [1], which stoked a severe backlash from the fMRI syndicate resulting in a title change to Puzzlingly High Correlations in fMRI Studies of Emotion, Personality, and Social Cognition [2]. There is a lot of interesting commentary around this article (e.g., “Voodoo” Science in Neuroimaging: How a Controversy Transformed into a Crisis [3]). To me it was fascinating to watch Vul (an incredibly rare talent, perhaps a genius), take on an entire field during his 1st year as assistant professor.
I imagine this type of automation existed before LLM agents came along - what do they add? Is it just the ability to evaluate the product description? Item quality is already listed as a categorical variable.
reply