Rivian Assistant Makes Siri Look Overqualified

[Noplacelikeloam]

I genuinely appreciate Rivian trying to bring AI into the vehicle experience. That’s ambitious. That’s forward-looking. That’s the kind of thing that makes you feel like you’re living in the future.

Unfortunately, after using Rivian Assistant for several days, I can confidently say the future apparently runs on a 2017 Bluetooth speaker and mild confusion.

I’m honestly baffled by how far behind this feels compared to literally every modern LLM on the market. Claude, ChatGPT, Gemini — all of them can hold context, remember what you just said, and generally behave like they’re awake.

Rivian Assistant, meanwhile, has the memory retention of a golden retriever in a laser tag arena.

Example:
Me: “Find that song I like.”
Rivian Assistant: finds it
Me: “Great, play it.”
Rivian Assistant: “Absolutely. Here’s a completely different song from three playlists ago that nobody asked for.”

It’s like talking to a very enthusiastic intern who got hit in the head with a Roomba.

And this isn’t some obscure edge case. This is basically the entire experience. Every interaction feels like the assistant is waking up from a nap, hearing one random noun, and making a panic decision.

At this point I’ve gone back to Alexa, which feels like admitting defeat in a futuristic society. That’s like buying a spaceship and then commuting via horse because the autopilot keeps driving into lakes.

What confuses me most is: why reinvent the wheel here?

Why not integrate Claude or another mature LLM provider? Because right now this feels less “AI-native EV experience” and more “two Raspberry Pis zip-tied together behind the dashboard running Ask Jeeves.”

Surely the token costs can’t be THAT bad. I refuse to believe we’re rationing GPU cycles like wartime butter.

And the frustrating part is that the vision is actually great. An intelligent in-car assistant should be a killer feature for Rivian. This SHOULD be the perfect environment for contextual AI.

Instead, my truck currently has the conversational abilities of a haunted GPS unit.

I want this to succeed. I really do. But today, Rivian Assistant feels less like “the future of driving” and more like your uncle discovering ChatGPT after two margaritas and saying “watch this.”

Sponsored

 

[electruck]

There are definitely some bugs and growing pains are to be expected with the first release of something as complex as what Rivian has undertaken here. Rivian is indeed using existing models under the hood. But rather than one single massive LLM, they also have an orchestration layer that, among other things, determines what can be executed locally vs what has to be farmed out to cloud servers (Google "Rivian Unified Intelligence").

The context issue is an interesting one. Sometimes the context is there but it, for whatever reason, the Assistant doesn't leverage it. Yesterday, I asked the Assistant to do 2 different things from one prompt... "Hey Rivian, do A and then do B". Action A was implemented but then action B was dropped. so I asked: "Hey RIvian, did you forget about action B?" The Assistant responded somewhat defensively with: "no, I didn't forget, I you need you to confirm that you want me to perform action B." Well great, it clearly knew there was something else that needed to be done and it knew that it needed a confirmation from me to do it. The problem is, nowhere along the way did it ever prompt me for that confirmation (something it has no trouble doing when the prompt only includes a single action to be executed).

 

[Noplacelikeloam]

There are definitely some bugs and growing pains are to be expected with the first release of something as complex as what Rivian has undertaken here. Rivian is indeed using existing models under the hood. But rather than one single massive LLM, they also have an orchestration layer that, among other things, determines what can be executed locally vs what has to be farmed out to cloud servers (Google "Rivian Unified Intelligence").

The context issue is an interesting one. Sometimes the context is there but it, for whatever reason, the Assistant doesn't leverage it. Yesterday, I asked the Assistant to do 2 different things from one prompt... "Hey Rivian, do A and then do B". Action A was implemented but then action B was dropped. so I asked: "Hey RIvian, did you forget about action B?" The Assistant responded somewhat defensively with: "no, I didn't forget, I you need you to confirm that you want me to perform action B." Well great, it clearly knew there was something else that needed to be done and it knew that it needed a confirmation from me to do it. The problem is, nowhere along the way did it ever prompt me for that confirmation (something it has no trouble doing when the prompt only includes a single action to be executed).

That’s actually the most fascinating part to me technically — the assistant clearly has latent awareness of conversational state, but the orchestration layer seems unable to reliably operationalize it into coherent behavior.

What you described is classic “context without agency.” The system knew Action B existed, knew it required confirmation, but failed at the most human part of the interaction: proactively managing the conversation flow.

That’s the gap between “LLM integrated into a product” and a true agentic system.

And honestly, this is why people underestimate how hard the last 20% of AI UX really is. The model intelligence is often already there. The orchestration, memory persistence, tool routing, interruption handling, and conversational recovery logic are what separate a demo from something people trust daily.

 

[forestwalker]

There's definitely some rough spots, but I have found it FAR more useful than Alexa, and just as capable as Siri so far.

Rivian is also far more pleasant to interact with. She even told me to enjoy the snow up at the lodge when she told me the weather at my destination.

 

[Biturbowned]

Directionally correct, but some friction that I hope continues to get sanded down. Work in progress and I’m not ready to shut the door on it. YMMV.

 

[impulsoren]

…determines what can be executed locally vs what has to be farmed out to cloud servers (Google "Rivian Unified Intelligence").

I have read about this “executed locally” thing before but have yet to discover anything RA can do locally. Is there something it can do without a cloud connection?

 

[electruck]

I have read about this “executed locally” thing before but have yet to discover anything RA can do locally. Is there something it can do without a cloud connection?

No idea, I don't have any inside information. Purely speculation here but more than likely it varies by hardware with R1G1 have little or no local capability and R2G1 having the most local capability.

 

[Mos Eisley]

No Rivian on the road currently has (nor ever will have) their upcoming RAP chip that comes along with the "205 GB/s of memory bandwidth that Senior VP of Electrical Hardware Vidya Rajagopalan described as key for AI applications." - Jose's article from this morning. It's maybe the one thing (more than LIDAR) that makes me wanna wait for the GEN release in R2...

 

[UnsungZero_OldTimeAdMan]

This is only the first iteration of a public release. Give it time. What versions are each of the AI platforms you named?? Not first, right? Why compare based on false equivalences? Why expect RA to be perfect or near-perfect from the get-go? Have we forgotten tech MO of "run fast, break things"? Such practice in the tech world isn't new. Look at MS Office suite of applications. Every single release is faulty and still shit today. /s

 

[lefkonj]

I was pleasantly surprised how it did what I needed it too. Texting both reading and responding as well as navigation inputs. I barely touch my Spotify as it just plays through playlists with ease. Siri does these things fine as well. It isn't perfect but it is enough for me.

 

[@Andoo117]

I switched back to Alexa after my kids got mad that they couldn't ask it to play their songs from amazon music anymore, RA lasted for about 1 hour in our vehicle. I didn't find the voice commands that RA can do better than just pushing the buttons on the screen. I loved electrucks post, couldn't have said it better, 100% agree with him and all the other posts. Too early to tell if RA will be great. A cool feature would have been celebrity voices like TomTom use to do in the past.

 

[ukyank]

The ‘thinking time’ is way longer than any other similiar type AI I’ve used

 

[tate16t]

I genuinely appreciate Rivian trying to bring AI into the vehicle experience. That’s ambitious. That’s forward-looking. That’s the kind of thing that makes you feel like you’re living in the future.

Unfortunately, after using Rivian Assistant for several days, I can confidently say the future apparently runs on a 2017 Bluetooth speaker and mild confusion.

I’m honestly baffled by how far behind this feels compared to literally every modern LLM on the market. Claude, ChatGPT, Gemini — all of them can hold context, remember what you just said, and generally behave like they’re awake.

Rivian Assistant, meanwhile, has the memory retention of a golden retriever in a laser tag arena.

Example:
Me: “Find that song I like.”
Rivian Assistant: finds it
Me: “Great, play it.”
Rivian Assistant: “Absolutely. Here’s a completely different song from three playlists ago that nobody asked for.”

It’s like talking to a very enthusiastic intern who got hit in the head with a Roomba.

And this isn’t some obscure edge case. This is basically the entire experience. Every interaction feels like the assistant is waking up from a nap, hearing one random noun, and making a panic decision.

At this point I’ve gone back to Alexa, which feels like admitting defeat in a futuristic society. That’s like buying a spaceship and then commuting via horse because the autopilot keeps driving into lakes.

What confuses me most is: why reinvent the wheel here?

Why not integrate Claude or another mature LLM provider? Because right now this feels less “AI-native EV experience” and more “two Raspberry Pis zip-tied together behind the dashboard running Ask Jeeves.”

Surely the token costs can’t be THAT bad. I refuse to believe we’re rationing GPU cycles like wartime butter.

And the frustrating part is that the vision is actually great. An intelligent in-car assistant should be a killer feature for Rivian. This SHOULD be the perfect environment for contextual AI.

Instead, my truck currently has the conversational abilities of a haunted GPS unit.

I want this to succeed. I really do. But today, Rivian Assistant feels less like “the future of driving” and more like your uncle discovering ChatGPT after two margaritas and saying “watch this.”

Within the mobile app, there’s a setting for Rivian assistant called Memory. This appeared disabled for me and I assume is the default. Perhaps you should enable this if not already.

Within the mobile app on the main page click on your initials in the bottom right corner. Then click on Rivian Intelligence where you will find the option to enable Memory.

 

[Noplacelikeloam]

Within the mobile app, there’s a setting for Rivian assistant called Memory. This appeared disabled for me and I assume is the default. Perhaps you should enable this if not already.

Within the mobile app on the main page click on your initials in the bottom right corner. Then click on Rivian Intelligence where you’re find the option to enable Memory.

Well… turns out you were right

Memory was disabled in the app the entire time. Enabled it and hopefully this will be better.

That said, I think this actually highlights the bigger UX issue: if memory is critical to making the assistant feel intelligent, it probably shouldn’t be off by default. Most users will assume the AI is bad, not misconfigured.

 

[tate16t]

Or, Rivian should have made this more visible, especially since its turned off.

Sponsored