The core problem is that these systems are just so incorrect in fundamental ways that they're effectively useless.
Imagine a buddy of yours tells you about an event he's pretty sure you'll be interested in. Why does he tell you about this event? Well, he knows your interests, what kind of things you enjoy, when you're free, who you might want to go to the event with, how much money you're willing to spend, how far you're willing to travel, when you like to go out... So when you're on the receiving end of such a suggestion it often feels great! It's like you've struck gold.
Now imagine your average 'AI' powered recommendation engine reading you a list of events. It doesn't feel magical. It doesn't even feel like it knows what the hell you enjoy doing half the time. Forget about knowing about your free time, budgetary restrictions, family restrictions, who you'd be able to go with; None of that stuff is even sort of in the picture. And it's all delivered to you in a voice that sounds like it would be as happy to kill you as give you advice. There's no lively back and forth on the logistics of the event. No feeling of discovery as you two talk it out, honing the plan that brings it from an abstract concept to reality.
I agree with you, and I cringed a little when I read the following from the OP:
> There are some fundamental reasons why conversational 3rd party platforms are hard.
In my mind the big fundamental problem here is the "3rd party". I'd love to have an "AI assistant" or an "AI buddy" that could watch everything I do and say and write and really get to know me super well... as long as I can be confident that I own and control everything it observes and learns. I sure as hell don't want a 3rd party involved! But alas, I don't see a way we get there that doesn't involve Amazon or Meta or Google or OpenAI sitting between me and my "AI" tools, at least in the short run.
Fwiw this is what I assume apple's long-term ai strategy is.
Let the hype-funded unicorns fight to develop (& end up commodifying) the tech and then design/sell selling devices that can support it locally. In that world, the AI assistant that you buy is a discrete piece of hardware, rather than a software treadmill.
Of course, this could mean that you end up on a hardware treadmill, but I think that's probably less bad, granted we can do something about the e-waste.
You're comparing a person recommending a single event versus an AI providing a list. In other words, proving OP's point.
GUIs provide information in 2D, letting eyes skim and bypass information that's not useful.
VUIs provide information in 1D, forcing you to take information in a linear stream and providing weak at best controls for skipping around without losing context or place.
Not coincidentally, this is why I absolutely hate watching videos for programming and news. If it's an article or post or thread, I can quickly evaluate if it has what I want and bypass the fluff. Videos simply suck at being usable, unless it's for something physical like how to work on a particular motor or carve a particular pattern into wood.
The person you are replying to is arguing the opposite - a future VUI could know your interests and just read you out the one relevant event rather than reading a list.
Alternatively it could summarise into “there’s a standup comedy gig, a few bands and various classes - what sort of thing are you looking for?” and then discuss with you to find the 2-3 events that are most relevant rather than reel off one big list.
And if it were a visual display of those things, I would have honed in on what I wanted and gotten my answer in the time it took me to ask what I was looking for.
It may be fantastic as an aid for low and no sighted people, but so long as I can read, a VUI is strictly inferior.
I would still prefer it to send it to me on my mobile device display. Voice interaction is nice for accessibility, bu the first method of control (whatever it is) is faster.
The point I'm trying to make is that this thing we're calling a 'VUI' is shit. There's no reason speech has to be this boring one dimensional thing. It's like the people that designed these things have never had a real conversation in their lives. When you're speaking with another person, or multiple other people, you're constantly exchanging cues that allow the other person to understand and re-calibrate what they're saying. These are verbal sounds, non-verbal sounds and physical movements. A crinkle of the forehead, a shake of the head, an uttered 'aaaaah' or a quiet verbal affirmation in support of what's being stated. It's not a single uni-directional stream of information, it's a multi-directional stream coming from multiple multi-modal sources at the same time.
None of these basic realities are accounted for in current technology. Instead we have these dumb robot voices reading us results from a preprocessed script that it thinks answers our question. No wonder the monkey part of our brain immediately picks up on the fact that this whole facade isn't just a lie, but an excruciating lie. It's excruciating because it's immediately obvious that there's nothing else 'there' to interact with. Even when speaking to another person over the phone, there's a huge amount of nuance you can pick up on. Are they happy? Are they sad? Are they frazzled? Are they in a rush? Are they relaxed? And you automatically calibrate your responses and what you say in the conversation based on all of these perfectly obvious things. Normal humans automatically calibrate what they say, how they sound, what they suggest based on these cues. It works really well!
There's no reason voice stuff has to suck. It has worked pretty great for humans for thousands of years. We're evolutionarily tuned to it. It's just that all the technology we've created around it totally sucks and people are delusional if they think it's anywhere near prime time.
This is all technically possible, but then there's also the privacy/security aspect. Many who would actually be into a solution like this won't be too hot to share the necessary information in the first place. And with good reason: companies with the resources to provide a decent experience don't have the best track record of protecting user data, and ensuring only the user has control of it. The privacy conscious would rather self-host, and end up losing out on capabilities in the process. So it's a sort of catch-22.
Part of the big challenge here in my mind is that companies are reluctant to put data into the world for others to consume in a friendly way - if, say, event organisers, put data out as open-API's, there would be the opportunity for a self-hosted or "convenient" third party (ala Google, Amazon) to create conversational experiences on top of it - private user and privacy uninterested user is well served (as is "event seller" as it's exposed to more people). But, as long as we're stuck with systems having to pull data by web scraping, no one can build a good solution that could work for either scenario.
I'm the case of events, I'm not so sure. Ticketing platforms want to drive sales via easy discovery and wide distribution. Take Eventbrite as an example (disclaimer: I used to work there). They
> Now imagine your average 'AI' powered recommendation engine reading you a list of events.
I think the issue is that having AI just fetch information catered towards any human is not using AI at all. I'm sure the hackathon groups pitching it all started with an idea of building a highly trained AI system whereby the recommendations are meaningful reflections of whatever information it has from you. Unfortunately for them, the most lucrative part of their plan neglected problems with both how to create an AI pipeline that takes many piecemeal inputs, along with millions having missing values represented some way, and renders meaningful outputs and neglected that success would only reap a massive backlash from privacy advocates.
In the end, their plan for a super-intelligent life assistant turns into just fetching event lists from facebook (or elsewhere) without even using the demographic data it has.
I agree, the real deal would be an AI assistant that actively learns and remembers everything important about you, and can access and utilize all that information when having a conversation, while also having access to external services, your calendar and so on.
I wouldn't be comfortable running that as a cloud-service though. Should be open source and run at home on my own machine.
Right, because if someone else is running it, you can be sure you're getting served thinly veiled ads, something someone else wishes you'd want, rather than something you really like.
You need to remember that most people on this site subscribe to the ideology that growth is the only thing that matters. They're Michael Douglas 'greed is good' type of people wrapped up in a spiffy technological veneer.
Any decision that doesn't make the 'line go up' is considered a dumb decision. So to most people on this site, kicking Sam out of the company was a bad idea because it meant the company's future earning potential had cratered.
> You need to remember that most people on this site subscribe to the ideology that growth is the only thing that matters
I'm not sure that's actually true anymore. Look at any story about "growth", and you'll see plenty of skeptical comments. I'd say the audience has skewed pretty far from all the VC stuff.
That’s unfair. The issue is poor governance. Why would anybody outside OpenAI care how much money they make? The fact is a lot of people now rely in one way or another on OpenAI’s services. Arbitrary and capricious decisions affect them.
My best guess is they turn off the commercial operations that are costing them the most money (And that they didn't want Sam to push in the first place) and pump up the prices on the ones they can actually earn a profit from and then try to coast for awhile.
Or they'll do something hilarious like sell VCs on a world wide cryptocurrency that is uniquely joined to an individual by their biometrics and somehow involves AI. I'm sure they could wrangle a few hundred million out of the VC class with a braindead scheme like that.
I was able to visit this place last year, it was great! I spent all day wandering around and playing. I would have stayed even longer if they weren't closing, hah.
I'm afraid I don't have a lot to add to this conversation but I have to say I just love Tailscale. I don't often run across software that feels so right and when I do it's a great surprise. Every time I see a new feature they're releasing I'm always impressed at how adept they are at targeting modern pain points.
I grew up and got into software by messing around with self-hosting web servers and game communities as a kid. As time has gone on I felt like we had lost some of the magic of easily sharing your machines and your creations with other people. We have a ton of services where you can now deploy and share your creations, but we've moved further and further away from direct sharing. There were plenty of good reasons why this has happened, with security being the most obvious factor, but it still makes me a little sad. I want my things to be able to talk to each other no matter where I am. I want to be able to invite my friends in and have access to my stuff.
Tailscale makes all of that quick, easy and awesome. I think it's really neat, makes me feel like a little nerdy kid again.
> I'm afraid I don't have a lot to add to this conversation but I have to say I just love Tailscale.
Strongly seconded. In my last company we used TailScale in some medium-advanced configurations, and from the dead-simple basic stuff up though some of the trickier stuff it's just a joy to use.. It's making much better networking practices highly-accessible and I'd bet ends up making the Internet a more secure, better organized system as a whole.
They run an amazingly transparent engineering process, for example their issue page (https://github.com/tailscale/tailscale/issues) is a model of transparent, responsive, involved open development. They embrace cool, modern, quirky stuff like NixOS (https://tailscale.com/blog/nixos-minecraft/). It's just generally really high-quality software developed with a very cool "hacker" philosophy.
TailScale is IMHO the coolest place to work right now, and something that almost any software company should look at if they do any networking.
If there's anything not to love, I can't see it. :)
Tailscale is cool, but if we focus on the product that this post discusses, Funnel won't give you the ability to use your own domain name. Cloudflare Tunnels will do that though. I will continue to use Tunnels.
That’s great to hear! I’ve been turning TS off and on when accessing services to make it through the day, but as soon as the battery use goes down (to plain wireguard app levels) I’ll be using it for DNS as well. Then it will truly be TS all the things for me.
There's a number of things. (at least four kinda five things you might mean by "keepalives", and the answer is all of them, so yes, and then others.) The same code that runs on iOS and Android also runs on Linux servers (where it was originally developed) where battery or perfect network efficiency wasn't really a top concern. When that code was moved to mobile, a few efforts were made to improve behavior on phones but not enough.
We're also working on measurements to make sure we objectively fix things and don't regress later in the future once it's fixed.
> As time has gone on I felt like we had lost some of the magic of easily sharing your machines and your creations with other people.
> I want my things to be able to talk to each other no matter where I am.
What isn't easy about forwarding packets destined for port 80/443 of your public IP to the local service in question and being a part of the public Internet like things were from the start?
Using Tailscale is the opposite of self-hosting, you're bringing someone else's third party service in, and adding more complexity and another point of failure.
> What isn't easy about forwarding packets destined for port 80/443 of your public IP to the local service in question and being a part of the public Internet like things were from the start?
- Not every home internet service gets a publicly routable IPv4 address anymore (e.g. CGNAT)
- Not every home internet service gets a static IPv4 address so folks have to handle DynDNS
- Not everyone is comfortable exposing their home network IP address in DNS (Tailscale only shares the endpoint IP once the endpoint is auth'd onto the network)
- Not everyone is comfortable configuring heavy auth/fail2ban/app layer safeties (Tailscale makes the services uncontactable unless you are auth'd into the Tailscale network)
- Not everyone is comfortable/can be bothered configuring Wireguard in highly dynamic environments
> Using Tailscale is the opposite of self-hosting, you're bringing someone else's third party service in, and adding more complexity and another point of failure.
Self-hosting need not be a zealot position - rather one can pick and choose what makes sense for them. Tailscale allows you to build your own network where all the nodes are auth'd (and tailscale lock means you don't even need to trust their keys by default) and non-public internet routable but still globally reachable from known safe devices. This can actually make folks more comfortable with self-hosting their own stuff since it removes so many other considerations. There is also headscale if folks want to self-host the coordination server.
Some argue that a third party service adds complexity and a point of failure. I'll point out that configuring a self-hosted publicly exposed thing from scratch for the first time has a rabbit hole of unknown complexity to the uninitiated. A tool like Tailscale can remove some of those complexities allowing focus on others.
>- Not every home internet service gets a static IPv4 address so folks have to handle DynDNS
For anyone who has only this specific problem out of your list, one solution is to get an HE tunnel. It's what I do.
If my ISP ever gets off its ass and implements IPv6 like it promised three years ago, I'll consider using that directly, though its current indication is that the IPv6 addresses will be dynamic for non-business customers which defeats the purpose.
I have gigabit fiber and it's IPv4 only. My ISP blocks incoming ICMP messages so I can't set up a HE tunnel. I used to use Route48, but they shuttered due to abuse, so I don't know what to do anymore.
Wireguard config is few lines (interface addresses, keys, AllowedIPs, post up and down). Simpler than SSH. You can run it on a cloud instance close to users.
Tailscale is still simpler and provides additional features. A small team or startup will appreciate Tailscale’s access controls.
> What isn't easy about forwarding packets destined for port 80/443 of your public IP to the local service in question and being a part of the public Internet like things were from the start?
Most of the evil in the world currently can be traced back to NATs and dynamic IPs.
In a more general sense, I think these compromises were made available because of a consumerist attitude towards the internet. Yes, we had a real issue with ipv4 exhaustion, but it if it affected businesses who couldn’t even host a website anymore, would this really have been an issue still? More likely people said that these things were an ok workaround because consumers don’t need X anyway. Sometimes these smart hacks engineers are so good at coming up with invalidate crucial invariants about the systems we love.
> As time has gone on I felt like we had lost some of the magic of easily sharing your machines and your creations with other people. We have a ton of services where you can now deploy and share your creations, but we've moved further and further away from direct sharing.
This is interesting, as it hasn't been my experience on the hobbyist side (game servers, personal projects, etc). ngrok, localtunnel, tunnelmole, rathole, tunnelto, zrok, et al. If the use case is just sharing something you built thats behind NAT / on a private subnet, there is no shortage of solutions.
I constantly read good things about Tailscale, as well as to a lesser degree Cloudflare, that I think I'm missing out.
But I've experienced so many times that companies change things and this can mess up the workflow or infrastructure really bad, adding days of work to implement an alternative.
With your hype, how much do you trust that you can rely on Tailscale? Should I feel safe when giving them control?
Any company can take a turn for the worse, and any time you've got SaaS deep in your stack there's risk there.
I can only say that I worry about TailScale growing up to be evil the least of basically every SaaS company I've ever used. They seem extremely serious about making the interaction a "win/win" and keeping it that way as they grow.
The examples in the article seem to be making the point that even when the AI cites the correct context (ie: financial reports) it still produces completely hallucinated information.
So even if you were to white-list the context to train the engine against, it would still make up information because that's just what LLMs do. They make stuff up to fit certain patterns.
That’s not correct. You don’t need to take my word for it. Go grab some complete baseball box scores and you can see that ChatGPT will reliably translate them into an entertaining English paragraph -length outline of the game.
This ability to translate is experimentally shown to be bound to the size of the LLM but it can reliably not synthesize information for lower complexity analytic prompts.
They look different because they ARE different. One was shot on film and the other was shot digital and then they used Filmbox to apply the film effect. It's two different pieces of source material, which means they had to act/film the scene twice.
I did the same "wait these are different source materials" double-take.
I think it's deliberate that they are different enough to notice. It reinforces the point of the comparison, because it would be impossible to do the comparison correctly using the same source material. It makes you consider what you're looking at.
Great analogy, and it's what I think about every time I hop on Netflix.
The old Netflix was an absolute gold mine. There was sooooo much good content and almost none of it was their own. The place was overflowing with great movies from every genre. It had all these great movies because they were the only game in town and the studios hadn't yet identified them as a competitor. It was a novelty. "Hey sure, movies on the internet. Why not! Here's our entire back catalog. Go nuts."
Once everyone realized this wasn't a novelty the IP holders began to pull back. First it was increased rates, which meant Netflix could have fewer of the top items available for streaming at any given time. Fine, kind of annoying that I can't find everything I want anymore but there's still a lot of good stuff! Then the megacorps that own all of the studios began to develop their own platforms to cash in on the money train and stopped licensing their content to Netflix altogether. Instead they began to pile the IP up to be released exclusively on their own platforms. Netflix knew this was going to happen so they went absolutely nuts on the spending, trying to produce enough of their own content to fill the gap left by 60 years of the top IP that was (mostly) banished from their platform.
The spending didn't work because it turns out making high quality IP is really, really hard. Think of all the film and television IP created over the last 60 years. Not the hits, I'm talking about everything. Now think of the hits. Those are like, what, 1% of the total? Netflix can't fill their catalog to compete with who they used to be. It's not physically/creatively possible. That means they will never live up to the initial experience of using the site.
As a consumer, I'd prefer they shift to become something closer to HBO/Apple TV+. A company that focuses on high quality, carefully curated productions. At least then I would have a good reason to continue subscribing. As it is, I don't find much value in the endless stream of schlock they've been producing to try to fill in for the lost IP. Quantity doesn't have a quality all its own when it comes to entertainment.
Netflix is the one that got greedy and over estimated their position.
Serving video is an absolute commodity. They should have been getting pennies on the dollar if that. A simple distribution service for studios who should have been thought of as their customers. But no, making a steady, useful, profitable business is not enough.
Taking a big cut from an easily replaced service, and using the money to fund a direct competitor to their customers sure was something. Took a lot of chutzpah, is about the best I'll say for it.
I guess the question is: What's the alternative to defining models? Just passing around DB cursors directly or maybe pulling the results out of a cursor and storing them in a dynamic collection?
I don't disagree that models are pointless if they're not doing anything other than aping the DB. Why waste your time creating another layer of required translation if it doesn't buy you anything?
With that said, I do think there are some benefits to using statically typed models in languages line C# with larger codebases. It seems like it makes it easier to refactor the application code.
As somebody who loves databases, I think the real issue that folks are meaning to critique isn’t actually using a “model.”
It’s using an ActiveRecord-style ORM (or any ORM) without grokking what lies beneath.
A database layer IS a model. It’s just not a class or an object.
ActiveRecord is a really nice trick when it works … but it can create some really performance-killing side effects.
Ruby’s Datawrapper ORM and its siblings in other languages requires understanding both sides (the object system and RDBMS) but can let you get your class/object semantics to play nicely with your database.
And just passing around database connections and arrays of hashes can get you awfully far.
But, if you want to not think about the database layer, ActiveRecord-style ORMs are a real win for developer ergonomics.
And that’s part of the win of Rails/Django/etc. You can live in a single mental model (classes/objects with references to each other) and ignore the database layer.
Except when you can’t.
One reason (not a criticism) that NoSQL can be such a win is that the semantics are closer to class/object semantics. So you’re not trying to manipulate data with an abstraction that doesn’t quite fit.
But most of our projects aren’t Twitter or FaceBook or Google or anything else functioning at galactic scale.
The rules nurses have to deal with around things as asinine as taking PTO are AMAZING. They’re required to put in PTO requests months in advance and the hospital can and will say “Sorry, denied. We don’t have enough people…” As they are intentionally creating skeleton crews of nurses to wring every ounce of profit out of the business.
My mom was a nurse, my aunt was a nurse, my sister is a nurse and my best friend’s mom is a nurse. I really can’t believe anyone continues to be a nurse given the insane working conditions these folks have to put up with. Twelve hour shifts, overflowing with patients, watching newcomers earn more than seasoned veterans… When I compare it to my laid back software engineering job it’s like I’m living in an entirely different universe. The hospital industry is a hugely demoralizing place.
The hospital I work at requires physicians to file their schedules 8 months in advance. The only deviation from that is for emergencies. Unofficially there’s a lot of flex for them, but that’s the official administration line.
Yeah, my mom was a small town nurse it was the same even there. I gave them a bit more slack because it was a hospital serving like 2000 people (so not really a high profit place) but even there, there was a lot of last minute "Oh no! People didn't show up for their christmas shifts, could you come in please!"
The core problem is that these systems are just so incorrect in fundamental ways that they're effectively useless.
Imagine a buddy of yours tells you about an event he's pretty sure you'll be interested in. Why does he tell you about this event? Well, he knows your interests, what kind of things you enjoy, when you're free, who you might want to go to the event with, how much money you're willing to spend, how far you're willing to travel, when you like to go out... So when you're on the receiving end of such a suggestion it often feels great! It's like you've struck gold.
Now imagine your average 'AI' powered recommendation engine reading you a list of events. It doesn't feel magical. It doesn't even feel like it knows what the hell you enjoy doing half the time. Forget about knowing about your free time, budgetary restrictions, family restrictions, who you'd be able to go with; None of that stuff is even sort of in the picture. And it's all delivered to you in a voice that sounds like it would be as happy to kill you as give you advice. There's no lively back and forth on the logistics of the event. No feeling of discovery as you two talk it out, honing the plan that brings it from an abstract concept to reality.
It's just dead and lifeless and shitty.