Grok almost got me in a heap of trouble! Elon please fix this.

[REM]

Look up AI hallucinations.

They all literally lie to us. If you dig into it deep with them with actual facts, they will actually fess up to it. I’ve seen it get worse and worse. Grok and GPT have both literally made up citations that they tried to use as a means to convey their answer as being truthful. I still use them, but with full knowledge that I must trust but verify for anything important. It’s like working with a pathological liar, good ones do it so well you have no idea (at first), that they are lying. You eventually believe nothing they say. This is where I am now approaching with them.

They will apologize profusely, blame it on training, but they absolutely do lie, and on purpose.

I asked Grok to calculate a pretty simple equation for how long a certain electrical load will last on an ecoflow of mine, and she... absolutely flubbed it by a factor of 10x (literally a decimal error). I corrected her, and she tried to argue with me lmao. "yes, I triple checked this time".

I walked her through it as if I was teaching a middle schooler, and she finally fessed up to understanding and apologized.

This type of interaction flags your account & interaction, and enters it into edge case corrections; thus improving the overall model. Slowly but surely. People really need to learn how to successfully use Ai's at their current limit; which is to get your *started* on research and development. The verification is up to you.

Even highly vetted scientists fail to do this all the time.

Sponsored

 

[dalton108]

I asked Grok to calculate a pretty simple equation for how long a certain electrical load will last on an ecoflow of mine, and she... absolutely flubbed it by a factor of 10x (literally a decimal error). I corrected her, and she tried to argue with me lmao. "yes, I triple checked this time".

I walked her through it as if I was teaching a middle schooler, and she finally fessed up to understanding and apologized.

This type of interaction flags your account & interaction, and enters it into edge case corrections; thus improving the overall model. Slowly but surely. People really need to learn how to successfully use Ai's at their current limit; which is to get your *started* on research and development. The verification is up to you.

Even highly vetted scientists fail to do this all the time.

100% agree.

What everyone needs to do is to be skeptical and ask it questions that you actually know the answer to. There is where you will find how it makes mistakes. Once you understand the way it fails, you will know better how to use it to your benefit.

 

[BlueLightning]

Understanding AI Lies and Hallucinations
What Are AI Lies?

AI lies, often referred to as "hallucinations," occur when artificial intelligence systems generate false or misleading information presented as fact. This phenomenon is particularly common in large language models (LLMs) like those used by Google (Grok) Unlike human errors, these lies can seem deliberate, raising concerns about accountability and trust.

Trust if you want to!

 

[Jhodgesatmb]

I know there is a ton of things going on at Tesla and SpaceX, but this happened to me last week and I barely dodged a bullet. Here is what happened.

My wife's business is super busy and she asked me to drop off some product in Little Rock Arkansas.. this worked out well, because I had a meeting at St. Jude in Memphis the next day. I was staying in a hotel in West Memphis and I had the truck compute my arrival energy level. It came back that I would arrive with 6% battery. I love this truck and trust its power consumption computer. 6 percent would be fine, but I needed more power to get across the bridge and make a few trips to the airport to pick up other St. Jude parents. I LOVE talking to Grok. On long trips I will ask it to explain concepts I don't understand like quantum teleportation or exciting new uses of AI on my business.... It seemed natural to ask it about the the supercharging resources in West Memphis. Grok not only confirmed that there was a WONDERFUL charging station in the parking lot of the the cracker barrel but that there were 8 stalls and all of them were in perfect working condition with no waiting when I arrived. It was so freaking confident! The cracker barrel was very close to my hotel. I proceeded on my journey with zero worries. On the way there I noticed that the truck was telling me I needed to stop in a small Arkansas town called Brinkley. I had stopped there before and there is not a lot to do there. I wanted to avoid the stop and just charge in West Memphis. On my way to West Memphis doubt started creeping in. I decided to call the cracker barrel and ask about the charging station. The person who answered assured me there was no supercharger at that location. I was literally a mile from the Brinkley exit when I figured it out. Had I not been suspicious I would likely have been stranded in West Memphis. It turns out there was possibly an off brand charger, but it didn't show up in the trucks database. I was lucky. But this is a cautionary tale. Do not trust grok! It is a wonderful tool. And it will only get better with time. But for some reason it chose to pacify me with an answer it thought I wanted to hear instead of the actual truth.

It takes a special personality to believe that a 6% remaining SOC prediction is ‘fine’ from a computer that routinely makes predictions that are 10% off. I suspect that the same personality would believe an AI agent is always correct. The navigation system is rarely wrong about SC locations.

I never allow my car to get me to a destination with less than 15-20% SOC. That way it can be off by 10% and I am ok. I do not believe that GROK has GPS information (I tried and it told me so - and that I believed), so why would it know about SCs? I guess I am of the paranoid, untrusting personality type.

 

[That Beast Mode]

100% agree.

What everyone needs to do is to be skeptical and ask it questions that you actually know the answer to. There is where you will find how it makes mistakes. Once you understand the way it fails, you will know better how to use it to your benefit.

The only thing we can be absolutely sure of with AI is that it will deliver solid CT, bacon and Sco-Joh pics. That is the only absolute truth at the moment. I'd assume they have to figure out the lying thing before giving it any kind of control over your vehicle. I can see it now.

::Thunk::
Me: Grok, I think you just hit a person
Grok : Nah, it was just a speed bump, don't worry about it, I've deleted all sentry footage of it.

 

[no%X#XMVk65v#cq]

Makes me glad that Grok isn't integrated into the vehicle systems, I guess.

While I can't imagine why you wouldn't have just used the already existing feature to find chargers, I think it does highlight the question of what exactly Grok is doing in our vehicles, and how we're supposed to use it. I for one was never given instruction, and I could see how someone could easily rely on the information it so confidently spews.

 

[ABILISK]

Why would you go to any supercharger without putting it in your navigation?

It pre-heats your battery, ensures there are stalls, etc. You basically did a new version of asking a stranger and then trusted it.

 

[GuyV]

Yeah you have to take whatever these AI tell you with a grain of salt. It gives out a lot of false information for the moment, and it does it as if it's an absolute truth. Even when you google something and the first result that pops up now (android) is the AI suggestion before the actual search results, I have found that to be wrong on many occasions.

AI has ruined the results of Google, and Bing is much worse, frequently displaying obviously false, and often self-contradictory data (it doesn't qualify as information) in the first frame. Anyone who trusts, relies or depends on this stuff is living in an alternate reality, Bizarro World.

 

[eswimm]

There's no reason to think Grok won't be incredibly effective when integrated with vehicle systems. That's actually the strength of these large language model systems; understanding human language. There's only so many control integrations available to be exposed to the AI, so a very limited training framework required. It's not like Grok is going to be driving, you just need to to translate a human request into a few feature toggles. Something like "plan a route from here to Disney World's Swan resort, arrive with at least 60% charge, make sure one of the charging stops is the St. Augustine Buc-ee's". Most current AIs are just fancy Google with a nice language wrapper. Unless they're specifically trained for a specific task/subject, they have little understanding of the validity of their results.

 

[REM]

It takes a special personality to believe that a 6% remaining SOC prediction is ‘fine’ from a computer that routinely makes predictions that are 10% off.

If your CT is 10% off the predicted range, then it's likely something you are doing to make it that far off the prediction. Most others aren't having this problem at all.

 

[REM]

I for one was never given instruction, and I could see how someone could easily rely on the information it so confidently spews.

Maybe because the onus is on you to learn about modern Ai limitations before using it?

 

[no%X#XMVk65v#cq]

Maybe because the onus is on you to learn about modern Ai limitations before using it?

Agree. That's why I've barely used it

Sponsored