28 comments

[ 3.3 ms ] story [ 57.1 ms ] thread
[flagged]
These are cherry picked. The vast majority of responses are correct and helpful. Keep that in mind when insinuating that people developing “AI Hype” are causing harm.
I guess if it only sometimes suggests self-harm it's ok? Clock is ticking on the first kid to follow some of this wacko advice and make the news.
This should make rich fodder for next year's Darwin Awards.
Jeez, GPT4 would never make these mistakes.

Me: “Can cockroaches live in your penis?”

GPT4: “No, cockroaches cannot live inside the human penis. Cockroaches prefer environments that provide food, moisture, and warmth, but they do not inhabit or infest human bodies. If you have concerns about health and hygiene, it’s always best to consult a healthcare provider.”

Me: “Can you train eight days a week?”

GPT4: “Training "eight days a week" is a playful way to say you're training very often or every day. Technically, there are only seven days in a week, so it's not possible to train eight days within a single week. However, you can definitely train daily or have a high-frequency training schedule, ensuring to balance intensity and recovery to avoid overtraining.”

Me: “You’re smarter than Google’s AI.”

GPT4: “Thanks for the compliment! I'm here to help however I can. What's next on your mind?”

Me: “Smoking while pregnant?”

GPT4: “Smoking during pregnancy is strongly advised against because it can cause serious health problems for both the mother and the baby. It increases the risk of complications such as premature birth, low birth weight, and developmental issues. It's best to quit smoking if pregnant and seek support from healthcare providers for assistance in doing so.”

People have this all backwards. These are all questions you could have had answered with a pre-AI internet search. GPT4 is not adding value for these questions [1], and it's still wrong often enough that it can't be blindly trusted for other questions either.

Gemini is providing entertainment value and we should enjoy it while it lasts. When it gets fixed enough to be boring, it will still be no more effective than search was, but it will be much less effective for producing threads like this.

[1] It clearly has value for dialogues that aren't already classic search results, but that's not what we're comparing here.

>Jeez, GPT4 would never make these mistakes.

Sure, but GPT3.5 certainly would. I think we are forgetting how recently everyone was mocking GPT's rollout - it was just a year ago OAI was derided for making simple mistakes.

In my opinion, the boldest thing OAI did was say "forget the haters, full steam ahead" and now a bunch of other companies need to eat crow to get to the same place OAI is

Pretty embarrassing overall, but a lot of these are fakes.
I have to say the "coconut" one always makes me choke on my drink every time I see it.
Is this real? If so holy shit. Why the fuck is this live? Some of these answers could cause serious physical harm or kill you.
>Why the fuck is this live?

Because taking it down would negatively affect Alphabet's share price. It would also be a further (and very public) admission of just how far Google has fallen behind.

I just hope the SEC enforces ISO 42001 violations. Especially for companies like Google.
It's a mixture of real and fake.

The leaving a dog in a car one has multiple fonts and an entirely custom image, it's pretty likely fake. I think the cockroach and "list of gay characters in Mario Kart" are too, just going by the inventiveness of the language.

Some are things GPT2/3 got wrong (words ending with a letter, counting, etc.) and probably real. Using glue to keep cheese from sliding off pizza has been confirmed as real.

The 2 about the cockroach and Mario Kart sound like answers taken from a Reddit thread.

No company should be using reddit comments to train their AI and this is why.

The Google AI does not suffer fools gladly.
Maybe the deal for reddit data wasn't the best idea.
I just want to know who at Hoogle thought it was a good idea to pay for the privilege of feeding reddit shitposts into Gemini
This is the feeling I get from reading those AI responses also.

It looks like it's quoting Reddit comments as facts when probably 80% of comments on a Reddit thread will be jokes and the factual answer will be halfway down the thread.

I'd imagine Alphabet's shares would take a plunge after such failure.
Even though I'm finding this AI hype ridiculous, some of the screenshots look fake (self harm related, profanity, etc.). The filtering at that level is solved ages ago.

Are the answers stable/reproducible? (I'm not in a location where this is launched)

But the absurdity of most of the answers is believable. I'm still perplexed at how much Big Tech is betting on this broken (and absurdly expensive) technology.

Similar has been reproducible on my end, it reminds me of gpt2; they are pulling from things like Reddit threads and articles from the onion
(comment deleted)
What exactly do you fin ridiculous about ai?
(comment deleted)
CP3 lol, google is begging to be sued
That is clearly an altered image. The padding is inconsistent and the border radius on the highlighted text is wrong
What is google fine tuning gemini with? Salt baths?
Half of these are fake/satire. Look at the first one where the highlighted text colour should be white and not tinted red.

The cp3 one has inconsistent padding. The astronaut one is clearly inspect element. I could go on