Although tools like ChatGPT, GitHub Copilot, and other stochastic parrots are useful in specific contexts, AI still has nothing of "I"; on the contrary, what they really do is generate text that makes sense (grammatical text, as opposed to ungrammatical texts), which was already possible with Prolog in 2011.
If you think LLMs are no more capable than Prolog, I'm not sure if you'll ever be impressed by anything.
There are so many things bad about this contentless article.
- It's not a scam, I paid $199 for the device and I have one
- The hardware is nicely designed and fun to use, the UI for interacting with the LLM is better than other options (my grandma could play with it easily)
The action model idea for training UI usage as a mechanism is an interesting idea.
The author is just so incredibly wrong about the shift in capability from prolog and GOFAI to now it's almost a caricature.
Just because the device exists and appears to work doesn't mean it isn't a scam.
For example, if someone sells you a machine that they claim is powered by magic but actually there is a hidden physical mechanism, you can claim the machine still performs the function, but they have misled you about the workings in a way that has huge implications on the capabilities of the product.
That is what is going on here. Look at MKBHD's review, the main potential he pointed out was the LAM functionality and the possible future uses of that. It turns out that entire thing was faked, and the "magic" that was the LAM is a "hidden physical mechanism" of duct tape'd together manual automations
> That is what is going on here. Look at MKBHD's review, the main potential he pointed out was the LAM functionality and the possible future uses of that. It turns out that entire thing was faked, and the "magic" that was the LAM is a "hidden physical mechanism" of duct tape'd together manual automations
I suspect the plan is/was to replace this with an actual LAM after release or possibly they thought they'd have it ready for release but ran into the 'last 5%' issue that seems to plague AI automation of anything. Definitely not an ethical move, but I can see how someone who bought into the AI hype might think it was a reasonable gamble to become a billionaire.
When I watched the keynote/preordered it seemed obvious the LAM UI training was work-in-progress. The core of the device was a new UI for interacting with an LLM via voice in a way that's better than the phone. It does this.
I think the people calling it a scam are the same group that just complains about anything. The author drawing an equivalence between prolog and LLMs is a lot of evidence in favor of this sort of bullshit, I just have a really low tolerance for this kind of argument and person.
I'm not delusional - I think the Humane device is extremely disappointing (and their marketing was way more misleading imo), but I'm glad they built and shipped something/tried something in this space. I don't believe the rabbit complaints are earnest it's just cool to hate new stuff for some people. It's the same batch that hated the iphone in 2007, just tedious and uninteresting.
> I think the people calling it a scam are the same group that just complains about anything. The author drawing an equivalence between prolog and LLMs is a lot of evidence in favor of this sort of bullshit, I just have a really low tolerance for this kind of argument and person.
Thanks! I`m not a huge fan of people like you too.
You're looking for the Juicero: a real tangible (over)engineered product that certainly did a thing, but nowhere near what the manufacturers said it aimed to do.
Juicero might be an example, though I'm not sure if they made any specific large claim that was totally false, unlike in this case. Perhaps if explicitly Juicero claimed that some special magic was going on during the squeezing process, but I don't think they did
They claimed that you needed the large amounts of force from their machine to squeeze the juice from their packets. People found that you can open them up and squeeze the juice out by hand because the machine isn't actually doing any sort of fresh juice squeezing.
I mean, it is a bit of a scam though right? I can pay for an NFT, 'own' the NFT, and still have been scammed.
In the case of this device, it told users it has functionality that, plainly, it does not have. It marketed itself based on these features. That is a scam.
They say it's going to be free forever with no subscription, but they have to pay for chatgpt API calls. Even if you forgive them for overhyping their chatgpt wrapper, they're still a ponzi scheme.
I still think of NFTs like the art market. If you bought the thing and you are happy with just owning what you got then there isn't a problem.
On the other hand if you bought with the expectation to sell in the future for a profit, you are trading on the perceived value. Scams rely on making you believe that the future accepted value will be much higher than what it will actually be.
That would be analogous to creating the expectation of absent functionality in the Rabbit.
In both cases the fraud is not in the item being sold but in the misrepresentations about it.
I think why this matters is because these days it feels like everything is misrepresented. I can buy a graphics card and be happy with its performance even though it is almost guaranteed to be well below the manufacturers promised level.
What is the returns policy for the Rabbit? Can people get a refund if they don't like it?
'in 2011' is weirdly recent though isn't it, not even to mention specific? Was there some particular advancement in Prolog based NLP around then that those of us with only a basic introduction to it (and not using it in an NLP course) wouldn't be aware of how capable it is in that regard?
Because I agree with you that sounds ridiculous to me, but also my level of Prolog is like
and dimly recalling what a 'cut' is. (I'm exaggerating a bit but I've barely used it since university, not even 100% sure about that syntax)
So I'm prepared to believe the state of the art for ChatGPT-like thing done in Prolog is a lot more impressive than I might have expected if someone asked me yesterday.
In fact, when I compared it to Prolog, it was exactly with the view that, back in 2011, it was already possible to define a sentence/text as grammatical or ungrammatical with just a few lines of code. What LLM models like ChatGPT do is generate text based on corpora distributed across the web, grouped, tokenized and trained with a general purpose, but, in the end, they still need the same rules as Prolog to determine whether or not they can regurgitate the text to the user.
The problem is that most people can't see that rabbit r1 is a deceptive product, at least. chatGPT (and Gemini, Claude and many others) doesn't do this (it doesn't trick its users into thinking that the product does one thing, but it actually does another).
I know how they work, and I think it's a really good job. But in the end, they still have to perform a gramatical check (like Prolog sentences) to be sure the text should be sent.
As I say, I know i pushed the limits, that's on me.
Ironically, the "stochastic parrot" could very likely put together a better argument than this ridiculous comparison with Prolog and 2011 tech.
Sure, markov chains are also picking the next token, but so am I when writing this comment. Am I just a "stochastic parrot"? Or is it the author that is parroting other people's opinions without giving them any thought?
Impressed? Here are few things that impressed me in the long decades since I am online.
In 1995 we discussed the HP48G calculator on Usenet and Dave Arnett who designed the calculator chimed in.
When my uncle left illegally Hungary in 1981 for the US, communication was sparse. We went to my grandmother for Sunday lunch and wrote a letter, together. Answers came in like two months. End of the 1980s phone calls began to happen but supremely expensive and short. By the time my grandfather passed in 2011 at the tender age of 98 he spent easily an hour every day video calling over Skype for free with his son.
I wintered out in Israel in 2007. It was almost impossible for me to get around on public transit as I do not read Hebrew. I also spent more than a week at the turn of 2015/2016 in Israel: Tel Hazor, Tel Megiddo, Avdat, Mitzpe Ramon, Eilat. I used public transit for all that, thanks to the smartphones with GPS and maps and real time transit instructions it was trivial.
I am easy to impress. You just need to knock down barriers of communication.
When these LLMs roared onto the scene I was neither impressed nor was I entertained. I was frightened. Nothing has happened since which would have proven us wrong, to the contrary. Australia leads the way by banning deepfake porn. More of that please, most especially ban the use of deepfakes of people who run in elections and the mass generation of texts about those.
As an aside, I enjoy the translate-in-camera functionality of smartphones very much so I am not against all AI -- it just needs to be used wisely.
I mean, LLMs are pretty different than deepfakes, but are "fear" and "being impressed" mutually exclusive?
I agree that deepfaking stuff with politicians is extremely concerning, but that doesn't really detract from the fact that Stable Diffusion is pretty cool. I was very impressed the first time I tried out Midjourney and Udio.
tbf, I did this comparison with Prolog thinking more about the generated text bt LLM models than the tech behind it. My fault. But, I still think that rabbit r1 it's a SCAM in the sense that is misleading the users.
I get where that comment is coming from. As far as I can tell when asked to generate text what all of the popular AI products are doing is the equivalent of smooshing together the first page of google results for the same prompt, xor'ing the shit out of it, then running the soup through a Word grammar checker, and then slowly typing out the results at like 300 baud and everyone is losing their shit over it.
Asking AI to do anything that requires "I" has it falling flat on its face in a cruel mockery of the word "I" and I just don't understand the hype. Well, that's a lie I do understand the hype and the mechanism by which piles of cash and gigawatthours of energy are being burned through and it makes me sad.
When all of the upcoming virtual brand ambassadors prove to be embarrassing failures maybe a splinter of reality will penetrate the hype-reinforced skulls of all of the "visionaries" funding this nonsense.
I don't know if I completely agree with that summary; ChatGPT (and Claude and Gemini) also has some degree of memory and context as well. It can at least to some degree remember what I and it typed and adjust things based on that. I would consider that some level of "intelligence", even if it's kind of baby intelligence.
ETA:
I should point out that it doesn't just "smoosh google results together", and it can actually do pretty interesting transformations beyond a simple "Grabbing the first result". If I ask ChatGPT to give me ten unique Java homework assignment problems, it will give me exactly that, and from my experience they actually are unique, at least they don't show up immediately when I search for similar things on Google. I can then ask it to give me those questions in LaTeX so that I can render it into a pretty thing. I guess you could argue that that is just an AST, so fair enough, but I think it's pretty easy to see why people are losing their shit over it.
It also has for more than a year hooked into Wolfram Alpha, so it addition to usually correctly parsing your problem, it can then send that parsed problem to a more objective source and get a correct answer.
ChatGPT has been an immense timesaver for me. It's been great to generate stuff like homework assignments, or to summarize long text into something more palatable. I don't automatically trust its output obviously, but it's considerably more useful than a Markov chain.
I'm always surprise by this kind of article or comments of people who don't know anything of how LLMs work or what they can do. The problem is that it requires, as it is the case for most tools, some learning curve. Prompting is not always straightforward, and after using these models for a while, you start discerning what should be prompted and what won't work.
The best example I have is a documentation that I wrote in Word that I wanted to translate in Mardown on a GitHub site (see https://github.com/naver/tamgu/tree/master/documentations). I split my document into 50 chapters of raw text (360 pages) and I asked chatGPT to add Mardown tags to each of the chapters. Not only did it work very well, but I also asked the same system to automatically translate in French, Spanish, Greek and Korean each of these chapters, keeping the Markdown intact. It took me a day to come up with 360 pages translated into these languages with GitHub ready documents. So the electric consumption was certainly high for this task, but you have to compare it to do the same task by hand over maybe a few weeks of continuous work.
Every token emitted is a full-pass through the network, with the prompts and previous tokens (sent by you and the AI) given as input.
And I agree that there is certainly a capacity for reasoning, no matter how flawed it is. There is plenty of evidence of AI solving novel problems 0-shot. Maybe not 100% of the time, but even if you have to run it 100 times and it gets it right 75% of the time in pure reasoning problems, it's doing better than randomness.
I completely agree. I'm not philosophically-brained enough to know how to define "intelligence", but I do think that ChatGPT qualifies as at least "intelligence-lite".
Yeah, as a distiller of collective knowledge as captured over the first 30 years of commercial internet, it does exactly what it should. That's not always "right" but it still provides huge value even if all it does is practically filter and distill, leaving you to nitpick or correct.
It doesn't have to be "magic" to displace a lot of things people spend time on every day.
I don't get this take... what is the scam part? It's a real device that people can buy, and it does what it was purported to do... maybe it's not as amazing as people hoped, but does that make it a scam?
It does not do what it was purported to do though.
It was supposed to have a built-in AI, including an AI to navigate websites that is just a bunch of manually written playwright scripts, and plenty of capabilities that are present in the marketing videos and promised by the CEO that simply do not exist.
By that logic, Theranos wasn't a scam either. The hardware existed and it delivered a result. It didn't do what was claimed however and the results were routinely faked.
It was claimed that the R1 would navigate an app like a person would. That it wouldn't matter if the UI changed because the AI would figure it out the same way a person would. It follows a script and breaks when the UI changes.
It was claimed that it would be faster than ChatGPT. The majority of it is a ChatGPT wrapper.
A product exists, sure, but I'd be surprised if anyone feels it met expectations.
Can't find the interview now, but I remember watching it and yes they specifically said that because it is an AI, rather than just an automation script, it is intelligent and will not be thrown off by site redesigns or CAPTCHAs (they have later said that they won't handle CAPTCHAs also).
Turns out that it is just an automation script and it cannot deal with site redesigns or CAPTCHAs.
Edit, just found they have made this claim also which simply doesn't exist at all:
> The R1 also has a dedicated training mode, which you can use to teach the device how to do something, and it will supposedly be able to repeat the action on its own going forward. Lyu gives an example: “You’ll be like, ‘Hey, first of all, go to a software called Photoshop. Open it. Grab your photos here. Make a lasso on the watermark and click click click click. This is how you remove watermark.’” It takes 30 seconds for Rabbit OS to process, Lyu says, and then it can automatically remove all your watermarks going forward.
Regarding its “learning” - it is still a model that needs data. The best you can expect is it will take actual UI sessions (as in users interacting with the website) for specific tasks to build its scripts, and as with any current “large” model it’s not going to update in realtime based on user input alone.
Sure but that’s all in the future. All of the selling points of this device are in future tense. The “model” does not seem to exist, but it’s being “worked on”. Their client app was taken apart and there is nothing interesting there. Their servers were hacked into, and made to run Doom which is funny, and there is no trace of any AI model there.
One of their former engineers gave a statement that LAM is just a marketing term and nothing like that exists.
If all the selling points are in future tense at what point can we call it a scam?
Edit: also the founder’s previous gig was a crypto scam that also promised AI on the blockchain
There is evidently a LAM of sorts given the nature of the queries it can answer. It is able to use agents - something like langchain or ChatGPT tools - in order to perform tasks that may be dependent on other tasks.
The problem is their LAM sucks, and is likely no more than just a task builder prompt on GPT (instead of a model specifically tuned for generating these tasks) using lang chain for resolution. They also have limited tooling, and some of it is already broken.
As for it being a scam. I definitely don’t see how you can offer lifetime ChatGPT with no subscription. So unless they are going to bring in additional revenue somehow it is effectively a ponzi scheme.
Scam generally implies some form of deception, and I think selling a product that claims to have proprietary AI and not just wrap ChatGPT when it does not have proprietary AI and in fact wraps ChatGPT certainly qualifies as deceptive.
Some of the advertised features like Uber seem to be super buggy or permanently "temporarily" unavailable. I wouldn't call it a scam per se but it's more of a beta or early access product.
It’s like people who say ‘Apple products are a scam’. If you don’t like them and think they’re overpriced, it’s simple: don’t buy them. Provided there’s no deceit, there’s no scam.
It's supposed to be a model that navigates apps and websites using ML, but instead it's just plain old, manually written scripts, that will break when a website or app changes.
I’m with you. People saying it uses scripts - so does every other ai - what exactly are they expecting? Why is this device being singled out over literally any assistant.
They announced their LAM AI as being able to navigate websites and apps by itself, and not needing any adjustments when the sites change or new sites are needed due to ML.
Instead of that, they shipped a bunch of manually written playwright scripts for a small number of websites and apps... With no AI involved.
I see from your other comments you are saying this was supposed to be some AI that can navigate websites and perform tasks with no user input.
Maybe you had some unrealistic expectations about what they were offering, as it has been quite clear they only offer limited integrations. It was even widely discussed that it was just some langchain (or similar) driven agents - apparent to anyone who has read even basic information about the product.
New generation of tech. Same old problems scraping webpages.
> In the linked video below, Coffeezilla shows how actions fail when applications change their graphical interface - like Doordash, the US version of iFood, which altered a small hamburger menu in its interface, breaking the Rabbit R1’s action.
I don't know about anybody else, but the words "doordash", "hamburger", and "menu" made that sentence difficult to grok. I thought "How can altering the contents of a hamburger (food) menu (list of foods you can order) in a food ordering app break the device the app is running on?"
Edit: yay for downvoting me for being unable to grok a sentence on the first try. That's truly the spirit of HN. /s
A hamburger menu is a specific UI element (the one in many apps that reveals an app's functionalities and is rendered as three stacked horizontal lines). Now that I've told you that, you should be able to grok it.
Calling everything a "scam" is quickly evolving into a scam of its own.
It seems everyone on YouTube is doing clickbait thumbnails and Yellow Journalism in an effort to build a get-rich-quick "creator" business ultimately selling ads for junk D2C products and VPNs.
Soon we'll have folks doing hour-long YouTube exposés about the people who create hour-long YouTube exposés in a recursive loop. The winner will be Manscaped.
Of course the HN gods have smote this post off the front page, as they did with the post criticising Microsoft for producing a crappy LLM hype product.
It seems recently that any post on HN criticising LLMs gets instantly wiped off the front page. This seems to coincide a little too neatly with YC's announcement that almost all of their funding choices this year will be "AI based" companies
Humane brooch(mentioned by author), Alexa, Siri, whatever, also fails on all these points, or doesn’t offer the respective functionality. Why is the author giving them a free pass?
What is wrong with the tooling using playwright scripts for services that don’t provide an api? What does the alternative offer? What is this magic ai automation the author thinks exist supposed to use?
This whole article is just singling out one interface from many, and doesn’t understand the very basic backend implementations these agents, including ChatGPT plugins, Siri, Alexa, etc, use.
The company promised there is a "LAM", or "Large Action Model" which uses AI to perform actions. The claim is that it can perform actions on-the-fly that the Rabbit devs never considered, rather than simply turning NLU into a fixed set of actions as with Siri, Alexa, etc.
Please reconsider; the article is rather poor, yes, but the reasons you cite for flagging are not any better. Just watch this single minute of the Coffeezilla video, doing a comparison of the marketed features and the actual results. https://youtu.be/zLvFc_24vSM?si=91sG63QKUhFpx44J&t=18
The lam clearly exists, as that is what maps from the LLM to actions, otherwise it wouldn’t be able to do anything.
The actions are definitely faulty, as per the door dash example. I can’t get Alexa to play my last audiobook consistently. They all suck. Amazon own audible and still can’t get it right.
It IS faster than chatGPT. Even if it’s using GPT for inference the transcription and TTS latency are less than is available from OpenAI - at least at present.
Rabbit is just a “better” understanding alexa with less actions IMHO. They all suck, so why pick on this one specifically? Amazon charges $50 for an echo dot which cant even answer a basic question.
:/ ok, fair I guess, difficult to argue on _expectations_, but you, me and the next (informed) guy know this is a scam by any stretch of the imagination. I really would like to save even a single person those $200.
If you're still sure that the LAM is present in their product, that's fine, but my view (and I understand how plugins, LLM, AI and other related things work) ist thre are no LAM in the rabbit r1.
Nowhere do they say the LAM will magically infer how to perform a task from the UI. ALL their documentation says it learns from actual UI sessions. That means someone clicking through the site/app to fulfill a task.
EDIT: I think a end-user doens't need to read the documentation to understand how the "magic" happens, this user just want the buy something that works as advertised.
I do not need to read a junk article like that because I already know how a LAM would work. It will turn an intent into an action. If you have been following anything about agents and LAM in general you would know this. Here is a good podcast giving a primer on what they do and how they work: https://changelog.com/practicalai/254 - granted you can’t expect the general public to know the “magic” but we are supposedly more technical here - and I’ll stress that my original point was against singling out rabbit when the alternatives do EXACTLY the same thing, and they promise no less
Or, you can nsay that an automation is a LAM (it's not).
With the definition of the Silvio Savarese’s article (from the podcast you indicate):
> To be clear, an LAMs job isn’t just turning a request into a series of steps, but understanding the logic that connects and surrounds them. That means understanding why one step must occur before or after another, and knowing when it’s time to change the plan to accommodate changes in circumstances. It’s a capability we demonstrate all the time in everyday life. For instance, when we don’t have enough eggs to make an omelet, we know the first step has nothing to do with cooking, but with heading to the nearest grocery store. It’s time we built technology that can do the same.
**
The definition provided by Silvio Savarese highlights the ability of a LAM to not only transform a request into a series of steps but also to understand the underlying logic that connects and surrounds these steps. This includes the ability to adjust the plan as circumstances change.
Based on this definition, claiming that rabbit r1 is a LAM-oriented assistant seems to be inaccurate. If it does not demonstrate the ability to understand and adapt to contextual changes in a logical and effective manner, it cannot be classified as a genuine LAM.
For a true LAM, it is crucial that the technology not only follows a predefined sequence of steps but also understands the logic and purpose behind each step, adjusting as necessary to achieve the desired goal. If rabbit r1 does not meet these criteria, its classification as a LAM indeed needs to be reviewed.
And, with that in mind I can assure you that rabbit r1 it's not a LAM oriented assistant as they claim.
I don’t know exactly how sophisticated the setup is - it’s likely some tooling around langchain or similar - but it evidently DOES do this given the nature of some of the queries that it resolves.
You are suggesting that a LAM must route itself around a critical failure in its tooling. Maybe you also expect a LAM to grow arms and water your plants for you? You are taking an experts definition and projecting some extra magical requirement onto it to dismiss r1 having a LAM.
The r1s LAM sucks, for sure, but it evidently exists in some form.
As for it being a scam overall - I don’t see how they can offer ChatGPT for life with no subscription, so unless they have some other revenue stream they won’t be around for long.
