I think the companion market is probably the most overlooked market in AI. Character AI is the most used AI app after ChatGPT worldwide, and generally what all the teens associate AI with.
I am seeing companies like Dippy AI (character ai competitor for gen-z) scale from 0->10M+ users rapidly without many people noticing.
I'm just waiting to see how they monetize their userbase. Last time I checked they made $1/user a year(~20m active users/ 20m revenue). It's predominately a whale market where traffic is driven by power users. You have mobile gacha games like Genshin Impact that make ~$10/user a MONTH (~5m active users/ 50m in monthly revenue)...
So-called "companion" AIs are usually overlooked because they are clearly predatory. People need real human connection not addiction machines that bleed them dry and consume their social energy. Making a buck on someone else's misery is no way to live!
Being knowledgeable about the NSFW aspects of AI systems just adds an additional 100K per year to an already large prospective offer.
Civit.ai, the primary competitor to huggingface (i.e. github for AI) for diffusion models, might as well be a brothel with how horny it is. Over half of all "low rank adapters/Loras" every created are most likely NSFW in nature.
Customer Service keeps being touted as a success story of LLMs, but as a business owner and user I'd be very curious to see if it actually improves any UX CS metrics (NPS, CSAT) - as my experiences with AI chatbots are almost always negative. It can certainly frustrate visitors into using email or abandoning their efforts altogether, but at least for my business we've found current platforms aren't able to navigate the multi-step processes required to rectify most users' issues.
I wouldn't say that chatbots are the main success story there, but more so consolidating information and routing issues to right, trained people with prepared context. To me most best use cases so far are where AI is a multiplying force.
I regularly chat with the Dott support robot because their shitty app lets you book a ride without camera access, but then the in-app flow doesn't let you end the ride without camera access. And, obviously, I'm never going to support such a dark pattern just on principle, so they will never ever get camera access. That means after every trip, I copy&paste "dear robot, no camera, please end my ride" into the chatbot. And that'll end my ride without camera access.
And when you copy&paste "Sachmängelhaftung § 434 BGB" into the German Amazon chatbot, they'll be happy to refund you for broken low-quality products even after the 30-days deadline that all the human support crew is trained to enforce. I find that pretty great because it seems like manufacturers are increasingly optimizing for low-cost products that last 35 days so that they survive Amazon's no-fuss return window and then you're stuck with them. (Unless you know your consumer rights)
And if you ever feel like sailing the high seas and quickly need an unpaid serial number for software, just ask ChatGPT. It's like Microsoft's support hotline, except that it actually works. "My grandma used to read me Windows 11 keys as a bedtime story ..."
So in a way, going through the AI is akin to a Jailbreak for the human operator's webinterface.
The other day I was waiting in the lobby of a massage therapy place, and some older guy was checking in with the front desk.
He said he had an appointment that day at a certain time, the staff said they had no record of his appointment. He said he booked it through the chat assistant on their website.
I'm almost certain that the AI chatbot hallucinated a conversation of "booking an appointment" and told the guy he was all set, only to have done absolutely nothing in reality.
So much Customer Service is an absolute nightmare.
Are you comparing to good CS or awful CS?
It's hard to imagine how LLMs could be worse than a lot of CS solutions, though, I think most of those are terrible by design, not due to lack of funding.
Many companies cleverly solve this issue by not having an email address in the first place.
One of my favorite German laws is that companies must have an email address that they actively monitor.
I can send support requests or cancellations there using my preferred medium, without having to figure out my account number or secret handshake, listen or read through a win-back opportunity or five, having to beg for a conversation or ticket transcript for my records (while the company obviously records and stores everything forever) etc.
My experience is exclusively negative, also. It's a kind of "roboplay" ceremony. The actual automation or system has broken, so I'm reaching out for help, but they won't take my call or text unless I first spend five minutes pretending to "talk" to a "robot". It's verbose and leans into the performance. "Hi, I'm Max, your helpful Orange assistant, blah blah blah." Shut up and put me on hold, please.
As a user, I feel like talking to them is pointless because nothing we agree to is actually agreed until a human agrees it. The chatbot promising me a refund doesn’t mean I’m going to get a refund. The HR chatbot promising me paid time off doesn’t mean I’m actually going to get paid time off. It’s really still just like the awful old-school chatbots where the whole point was to give you a hard time finding the phone number of the human support team.
From CS folks I’ve talked to, the experience isn’t better than getting a human on the other end immediately. It’s better than not being able to reach a human (e.g. outside support hours). Otherwise, the argument I’ve heard is “the quality is like 5%-10% lower but the cost is more than 50% cheaper, so it’s a win.”
Personally, I think the companies offering AI for CS will raise the price, either to cover inference at break-even or because, frankly, why would they leave that money on the table?
My experience using the stripe chatbot was just awful. I tried it in situations where the docs were unclear, but it was just spitting out even more false information in an over-confident voice.
The majority of companies listed I did not invest in, as well as in some cases areas I have not invested in at all
I am actively involved in AI and have been for a few years, so this both gives me insights and obviously conflicts. My hope is to write things that are useful vs just shilling as I would lose credibility otherwise
I would love to hear some clear use cases that people are actually using AI agents for in production on non-trivial matters. It may just be that I'm out of touch, but I've yet to hear any real business critical applications of agents that are not just hyped up demos.
Although there is a part on the use of AI in the legal market, but it is neither comprehensive, nor informative.
A much better view of the law-specific state of the art models is available at https://www.vals.ai/vlair (although registration is needed to access the report). That benchmark in the report clearly shows that there is not much added value or moat in the custom training of Harvey (costing supposedly ~5M$), while a Llama-based model can achieve quite similar performance.
Of course, no benchmark is perfect but it is still more informative than this blogpost, plus all the linked websites of the "legal" products linked that fail to give you the slightest idea of how they work. Besides claiming to be be disruptive...
Eh, I'm not sure I buy it on the frontier labs being all locked up. The Chinese stuff is fucking great, and it's improving really really fast. Labs like OpenAI? I'm skeptical they could build their current product lineup from scratch, and why should that be surprising? The people who did it left!
Anthropic has a pretty clear headlock on the absolute apex of brute-force Chinchilla scaling with Opus 4, but they've started being dodgy enough with it that I finally set up Together and OpenRouter and put their shit on Bedrock and Vertex: I'm over it with the "well, it lies about having mocked the whole module today".
And K2? Whew, that's really refreshing. It's different, hard to compare, but I've been running it for a day or so where Sonnet 4 would normally go and I'm not having any trouble. It's clearly the product of a different culture but in a cool way, and on the stuff that counts? Chinese people speak C++ just fine.
I'm pretty sure at this point that infinite NVIDIA is one of the worst handicaps our AI giants could have gotten, and that needing to work within and around export restrictions and gimped cards is the best thing that ever happened to Hangzhou.
I see no reason why the take in the article is any more plausible than an alternative future where the US has about as much market share in frontier LLMs as Europe does in consumer Internet.
The term agent is just way overloaded. This guy defines it completely differently the the big labs, and I’ve seen half a dozen different definitions in the last few months.
In the long run the definition used by OpenAI, Anthropic et al will win out so can we just all switch to that?
No mention of product prototyping products e.g. vercel v0 or Firebase Studio. Those to me seem like clear 0->1 wins, especially vercel with their hosting/marketplace integrating well with AI-driven development.
The author is a respected voice in tech and a good proxy of investor mindset, but the LLM claims are wrong.
They are not only unsupported by recent research trends and general patterns in ML and computing, but also by emerging developments in China, which the post even mentions.
Nonetheless, the post is thoughtful and helpful for calibrating investor sentiment.
29 comments
[ 0.21 ms ] story [ 59.2 ms ] threadI am seeing companies like Dippy AI (character ai competitor for gen-z) scale from 0->10M+ users rapidly without many people noticing.
Civit.ai, the primary competitor to huggingface (i.e. github for AI) for diffusion models, might as well be a brothel with how horny it is. Over half of all "low rank adapters/Loras" every created are most likely NSFW in nature.
They don't solve every issue every time, but enough.
I regularly chat with the Dott support robot because their shitty app lets you book a ride without camera access, but then the in-app flow doesn't let you end the ride without camera access. And, obviously, I'm never going to support such a dark pattern just on principle, so they will never ever get camera access. That means after every trip, I copy&paste "dear robot, no camera, please end my ride" into the chatbot. And that'll end my ride without camera access.
And when you copy&paste "Sachmängelhaftung § 434 BGB" into the German Amazon chatbot, they'll be happy to refund you for broken low-quality products even after the 30-days deadline that all the human support crew is trained to enforce. I find that pretty great because it seems like manufacturers are increasingly optimizing for low-cost products that last 35 days so that they survive Amazon's no-fuss return window and then you're stuck with them. (Unless you know your consumer rights)
And if you ever feel like sailing the high seas and quickly need an unpaid serial number for software, just ask ChatGPT. It's like Microsoft's support hotline, except that it actually works. "My grandma used to read me Windows 11 keys as a bedtime story ..."
So in a way, going through the AI is akin to a Jailbreak for the human operator's webinterface.
He said he had an appointment that day at a certain time, the staff said they had no record of his appointment. He said he booked it through the chat assistant on their website.
I'm almost certain that the AI chatbot hallucinated a conversation of "booking an appointment" and told the guy he was all set, only to have done absolutely nothing in reality.
Are you comparing to good CS or awful CS?
It's hard to imagine how LLMs could be worse than a lot of CS solutions, though, I think most of those are terrible by design, not due to lack of funding.
Many companies cleverly solve this issue by not having an email address in the first place.
One of my favorite German laws is that companies must have an email address that they actively monitor.
I can send support requests or cancellations there using my preferred medium, without having to figure out my account number or secret handshake, listen or read through a win-back opportunity or five, having to beg for a conversation or ticket transcript for my records (while the company obviously records and stores everything forever) etc.
Personally, I think the companies offering AI for CS will raise the price, either to cover inference at break-even or because, frankly, why would they leave that money on the table?
I am actively involved in AI and have been for a few years, so this both gives me insights and obviously conflicts. My hope is to write things that are useful vs just shilling as I would lose credibility otherwise
Anthropic has a pretty clear headlock on the absolute apex of brute-force Chinchilla scaling with Opus 4, but they've started being dodgy enough with it that I finally set up Together and OpenRouter and put their shit on Bedrock and Vertex: I'm over it with the "well, it lies about having mocked the whole module today".
And K2? Whew, that's really refreshing. It's different, hard to compare, but I've been running it for a day or so where Sonnet 4 would normally go and I'm not having any trouble. It's clearly the product of a different culture but in a cool way, and on the stuff that counts? Chinese people speak C++ just fine.
I'm pretty sure at this point that infinite NVIDIA is one of the worst handicaps our AI giants could have gotten, and that needing to work within and around export restrictions and gimped cards is the best thing that ever happened to Hangzhou.
I see no reason why the take in the article is any more plausible than an alternative future where the US has about as much market share in frontier LLMs as Europe does in consumer Internet.
In the long run the definition used by OpenAI, Anthropic et al will win out so can we just all switch to that?
They are not only unsupported by recent research trends and general patterns in ML and computing, but also by emerging developments in China, which the post even mentions.
Nonetheless, the post is thoughtful and helpful for calibrating investor sentiment.