How Siri, Alexa and Google Assistant Lost the A.I. Race | Hacker News Reader

How Siri, Alexa and Google Assistant Lost the A.I. Race

nytimes.com

20 pointsfortran7748 comments

Fundamental misunderstanding of what chatgpt does versus virtual assistants, coupled with a bizarre declaration that ChatGPT has “won” even though it has yet to establish itself in any actual daily-use scenarios.

ChatGPT demonstrates a huge potential. It will take time for companies to integrate it in their products and for it to start impacting everyday lives of big number of people.

Siri also demonstrated great potential, but that never materialized. Not only that, but the service as it is in practice has been improving incredible slowly over time.

Siri might be maxing out as just a UI for basic things playing music, getting weather, and sending text messages.

As you say, the original vision, where it would help you in all kinds of activities, we are still far away from.

But Siri hasn't even maxed out that UI. It consistently messes up in terms of basic requests, both in terms of not hearing me correctly.

Then there's also the answers that are not even wrong. It's especially bad at sports. You ask for things like when a game is playing and it will return the current (or final) score. That is literally the last thing people want when they ask for a game time. If they wanted a score they'd ask for a score.

And this is trivial to fix. Just stop returning scores when people don't ask for scores. There are countless examples like this in different domains. It just hasn't improved.

I agree. Siri is incomprehensibly bad at basic tasks, especially given the massive engineering staffing and resources at their disposal. I'm down to just using it to play songs on Apple Music, and it fails maybe half the time.

I agree. But it hasn’t yet “won” against Siri or other assistants. It might, but it hasn’t yet.

A lot of people (including me) use ChatGPT daily.

I ask it silly things like "is it defence or defense" all the time, multiple times a day! Love my `ai` cli script, haha. Only thing missing is speech recognition and tts output for the basic knowledge queries.

Also looking into writing a prompt that turns "timer one hour 30" into set_timer(5400) and other functions but not got the bandwidth just yet.

I imagine I'll ask GPT for an intent-in-json-format and go from there, I'll need to do the programming for things like timers and todos. I just need it to convert natural language into function names and parameters really.

I now only use Siri to set a timer when I put a load of washing in and it manages to get that wrong sometimes.. always appreciate my one hour thirty timer being interpreted as "set an alarm named Timer for 01:30am" and making me jump when I'm trying to sleep haha, daft robot. It wouldn't take much to do better.

... and in a way that is cost effective or cost attractive?

I use it daily. Now what?

I wouldn’t say the race is won or lost at this point, just that we are on the edge of a bubble that is revealing decent, more affordable technology like in the dot com era. But for me at least the GitHub copilot and cpt-3 have been entirely underwhelming except for things I use fiverr for. I don’t understand what is causing all the hype. Would someone please show me something technically impressive or explain the steps to make it useable like the input engineering? It is just too deep a rabbit hole to go in to casually get a good answer from it for me. It requires a setup that I must be overlooking. Also, I am unimpressed mostly by hugging face. Skimming the article and it didn’t stand out. It takes a lot of inputs to get one good output, which requires a lot of expensive equipment and electricity.

What a weird article. Given that there are yet to be devices like siri that use these new language models, it's fairly nonsensical to compare a "chatbot" like chatgpt to a device like Siri/Alexa/Google that actually takes commands verbally and performs real actions on them.

NYT quality is really awful these days.

I've used Siri around 10 times today, I haven't used ChatGPT 10 times to date.

It's about being the right tool for the job. I don't catch a flight to do my groceries, even though a car needs me to drive it, only fits 5 people, and doesn't go thousands of miles on a single tank. However an analogous headline would be HOW AEROPLANES BEAT CARS.

And I’ve used Siri 20 times since inception, but used Chat 10 times, at minimum, today.

For you, it might not be the right tool for the job, but it is for me.

While missing the point, you've proved the point.

It's that Siri/Cortana/OK Google/Alexa are not a replacement for ChatGPT, and ChatGPT is not a replacement for those.

They're different tools and claiming they're in a race with each other is disingenuous.. like your reply, so maybe you should write for NYT too?

I watched the “Introducing GPT-4” on the OpenAI YouTube channel (https://youtu.be/TxkJMX0KyS0) and you see top of the line MacBook Pro’s and Apple Watches all over the place in their offices. I bet most of them have iPhones too.

The moral of the story is, all Apple care about is selling you hardware. It has made no dent in their business or brand that Siri is awful.

It hasn't made a dent because people had very limited use of AI assistants so far. With the progress that ChatGPT is showing, that can rapidly change and Apple can become old news - same as Nokia and Motorola, that once dominated the mobile market.

The point is that Apple's line of business is not directly tied to its AI assistant (nor is Amazon's [ecommerce/cloud] or even Google's [ads/cloud]).

If it seems worthwhile, Any one of them can license the new generation of AI assistance from someone or spend a bunch of cash to build an adequate competitor on their own.

ChatGPT is an exciting example of this new technology, but has almost no "moat" so far. And for as much as Microsoft/OpenAI can establish a moat, the way they'll monetize it is through licensing anyway.

Apple is a hardware corporation with superior products than most of its competitors. How is one new piece of software a threat to its business?

Smartphone becomes mainly an interface to an AI in the cloud. Doesn't matter if you use an expensive Apple device or a cheap Chinese phone. People stop buying expensive Apple devices.

I am pretty sure Apple is seeing the limitation of just being a hardware company.

Why do you think they are pushing so hard on their services and advertising?

Their services and advertising is highly and completely integrated into their hardware. In other words, the software exists solely to serve the hardware side. They became the most valued company in the world by selling hardware, I don't think it is fair to say their strategy is "limited".

The parts that they are pushing the most are still hardware. Cars, ARM chips, phones, Macs. The OS and App store for example are extensions of the hardware and are not competitive with similar products from other companies. For example, Mac OS is competitive because they have hardware level optimization with the M series chips such as acceleration for x86 simulations. The software or the OS itself is nothing noteworthy if you compare it to Windows or Linux.

Apple, Google, Amazon can build a similar large language model easily enough within a year. ChatGPT is useful for a limited set of tasks, the real gains will be integrating this sort of technology into other software stacks to do more than just creative writing. This will take adding additional task specific machine learning as you're seeing Microsoft attempting to do. But Microsoft should also be a bit of a cautionary example, it shows that it's not actually that "simple" to get a language model to work reliably for new complex tasks.

Apple can’t. Not sure about Amazon but Google definitely can. Apple’s machine learning capabilities are extremely limited.

How do you plan to use chatgpt if you have no phone and no computer?!

Smart speaker, smart TV, smart watch, smart glasses, tablet, etc.

and guess who makes them? hint: it's not OpenAI

Well, I lost confidence in the NYT quite a while ago but I read this article anyway.

I disagree about Apple: Siri on my Apple Watch does everything I want it to: using the on-board Apple Silicon neural cores, it accurately processes my speech input to set times, open email or messages, and many other things that I want to do even with no real keyboard input.

I think that Google also effectively uses hardware tensor support for the same kind of thing, at least on their phones.

I also think that the AI assistant requirements are different between watch vs. phone vs. tablet vs. laptop vs. smart car.

I would hope that even though they use some common code and deep learning models, that each platform is treated individually.

Now, on my laptop, absolutely Bing/ChatGPT search, the ChatGPT Plus web app, and custom applications written with LLMs + LangChain + Llama-Index, etc. are relevant to developing AI assistants for laptops.

EDIT: and a few years ago Apple did a fantastic job integrating deep learning (and other ML) into CoreML making it easy for developers to integrate this tech into their apps.

Siri, Alexa, and Google Assistant "lost" because they aren't assistants. They are voice user interfaces. They have discreet modes, requiring discreet vocal commands for controlling timers, controlling media players, creating schedule entries, or ordering products.

An assistant would be able to take an instruction like "In twenty minutes, play Lullabye by Billie Joel, and turn the lights off when it's done." That combines functions in useful ways, that none of the previous generation of assistants can do.

ChatGPT can't do that either, as it is cannot control things. It can generate text, but there's very little likelihood we'd want to allow LLMs to manipulate systems until they're far more reliable.

Where it will be useful is as supercharged autocomplete. I fully believe LLMs will pop up all over the place where content generation is the task (like Copilot).

Theres a simple equation that I think is true:

Time to integrate LLM into existing home assistant >= Time to build home assistant from scratch using LLM.

They lost it in a very simple way:

Research Scientists and Managers refusing to let their importance be impacted. Existing NLP paradigms preserved at all costs, conservative technical decision making. Building something that can be delivered in 4 months, as opposed to long term technical vision. The people running the important NLP orgs in Alexa have been there since alexa started. There is no fresh air, just constant pushing for feature upon feature. No room for bottom up innovation, road maps are built and then delegated to worker bee coders.

I have first hand knowledge, that in Alexa right now, there are year long+ initiatives to productionalize ML models that are completely obsolete compared to GPT-4. Models that are still being figured out, are already borderline worthless.

The only moat is the number of devices already in peoples homes. A startup armed with a ChatGPT model could take down alexa very quickly if they could distribute speakers to homes.

Lets not forget the panic and the existential crises of careers built on old NLP knowledge, in both academia and industry. It turns out the solution to NLP is just "scale". All of the complicated heuristics for getting old school NLP solutions to work will be thrown away

How would GPT be useful for Alexa? I don't actually want to chat with the speaker I just want it to find my keys and set timers

Alexa is a complicated rube goldberg machine of ML models. For instance, when asking, "play songs by michael jackson", it runs an entity resolution query to figure out who michael jackson is, and then retrieve his music. There are a ton of systems along the way to solve this. Tons of data pipelines, formatting, autiditing, rules based matching etc. But all of those brittle expensive systems can be replaced by a single LLM.

>I don't actually want to chat with the speaker I just want it to find my keys and set timers

Not to be snarky, but do you expect it to read your mind?

You have to communicate with it somehow. Now that you're using spoken language to communicate you need to actually decode the sounds into text, and then you need to decode the meaning and intent from those spoken words.

I have never seen a computer system as good at figuring out intent from comment statements at the Chat/Bing GPTs. So yea, at least to me it seems very useful.

I have a seen proper prompt translate commands into API calls. I imagine it’s doable even just with self hosted solutions like Home Assistant or homebridge.

Editor: Here's a racy headline, do you think you can make an article out of it?

Lost? These companies may have lost the AI software race, but they have quite a moat. They created a new hardware category, a large 3rd party ecosystem (in the case of Alexa), sold a shitload of mostly useless "smart" devices and now they just need new software and suddenly they're... not mostly useless! Who else can sell hardware right now with "new" AI tech?

If GPT-4 was hardware exclusive, it would be a different story.

Amazon lost a ton of money with the Alexa effort. They hoped people will use Alexa to order stuff on Amazon, which did not realize. A few months ago the Alexa department was downsized. With the huge progress in chat bots that ChatGPT demonstrated Alexa's future in uncertain.

So Amazon writes a big check to Microsoft and has ChatGPT integration in all the stupid little pucks that Microsoft that doesn't have. Or writes a big check to Google. Or writes a big check for AI researcher payroll and reserves some AWS oomph to train their own LLM. They're not the ones earning buzz this week, but they're actually very well positioned.

Positioned for what? Alexa is not making money for Amazon, how is writing a big check going to help?

> Amazon lost a ton of money with the Alexa effort.

As did everyone else! I got both my Google Nest devices for basically free. My point was that they laid the groundwork, they have a customer base and while unprofitable this might be the "silver lining" if they can figure out how to charge for it.

I doubt the race is over, its only a matter of time before Apple, Amazon, and Google integrate large language models into their assistants.

Most people I know have some home assistant, they are genuinely useful for some actions around the home, such as playing music/radio, weather, timers, controlling smart home devices.

The race is over?

For Alexa it is. None of the others have shown any improvement or added value in years. Id say they have a long uphill battle, and these companies are already looking to trim as much fat as possible. This is a new epoch and I don't think the old gate keepers stand a chance.

I just assumed that nobody wanted "their AI voice" to be recorded saying any kind of wrongthink or anything else uncouth, until they sufficiently lock that stuff down (if such a thing is even 100% possible).

Kinda extreme to declare a winner when the race just started. But I really really really hope this forces Apple to make a Siri + LLM feature. Even a 7B model would work better than Siri today.

Which public and private companies are doing well with AI? I’m thinking more along the lines as a long term investor.

pfft sure.

google/apple/amazon etc can plug their assistants into chatgpt with minimal efforts.

They haven't yet because they need to tweak it for profit.

Reply on news.ycombinator.com