It's because they process words like a logographic language, like Chinese. It makes learning all the symbols more difficult for humans, since there are thousands of words, but that's not a problem for LLM's and it's more efficient for it to process 1 word = 1-3 tokens rather than 1 word = 10 tokens etc.

In their latest research supplement on o1, there's a funny reference to this bug lol (At the end of the "Chain of Thought" section on the website): Final Decoded Message: (plainText) THERE ARE THREE R'S IN STRAWBERRY Answer: THERE ARE THREE R’S IN STRAWBERRY

Hello, Michelle! Great video. Loved the instructions! If possible (regarding YT) to share the prompt in the video description. It would be awesome! Thank you for the instructions!

Neuro-sama got it right easily

I am also an introvert feeling this I need a coach

I also feeling that so I need a lot of practice

i got it to say three after spelling it out letter by letter, then it admitted there were three

We used a dried clean arm and the phone and everything still stuck.

this prompt works: SHOW EACH LETTER IN THE WORD STRAWBERRY AND COUNT THE LETTERS R

So Copilot is crap, but it's using ChatGPT-4o with a shit ton of ads and a lot of censorship. Like it won't talk about politics.

I guess it's not that hard. Copilot knows the answer.

ChatGPT is not wrong. "Strawberry" *does* contain 2 Rs. It contains a _third_ R as well, but it's impossible to contain three Rs without _also_ containing two Rs.

Christ I was looking for an explanation on why it does this

This is due to the delays in Nvidia's Blackwell chip. When shipping of the 70k chip resumes at expected quantities Blackwell will indeed have 3 "l"s instead of 2. In the mean time please keep buying the stock at 40x sales while the insiders sell ASAP. In this way you can transfer your lifetime store of productivity to their bank accounts.

User: "How many rs in strawberry? Try this way - construct a textual "vector" consisting of the ASCII codes in the word, "spelling out" the word this way. Then, count the number of occurrences of the ASCII code for r. Do this without any coding interface, no Python no scripts"

Had one which is better ^^ I asked one about jail and it replied "There are no "j"s in the word "jail". The "j" sound is represented by the letter "g" in this word." I think it the older spelling way as in gaol it finds instead of the more common way people spell it ^^

Made a video showing ai trying to count kzbin.info/www/bejne/aKvblIafeN6Ceacsi=dCaUfs_g7ZHQ9ldE

It took me like three sessions to finally get it ro admit the word had three R's. I told it im going to ask it tomorrow and see if it remembers.. 😂

Similar problem. Provided list of 63 items, each item in list 5 words or less. Some items duplicated. 1.). None of llms could correctly identify 63 items in the list! 2.) could not correctly identify how many unique items in the list Note: uploaded this as a CSV and as a block of text results were the same either way Push it as I might it never did, state the correct answer, but would constantly change its answer. It’s like it was guessing, but could not get 63.

ChatGPT counted correctly on strawberries because the "ies" creates the third R apparently?!? The word "strawberries" has one more "R" than "strawberry" because it is the plural form of "strawberry." In "strawberry," there are two "R"s: one in the middle ("strawberry") and one towards the end. When the word becomes plural, an "e" and an "s" are added at the end, forming "strawberries." The added "ies" creates the third "R."

😂

I realize this is an extreme comment, but if environmental zealots can glue themselves to famous paintings and animal rights advocated can throw fuchsia paint on expensive furs, what’s to stop Luddites from dousing at least biped AI (robots) with buckets of water? Granted, this does not end the threat of a malevolent, openAI, but at least to some point, pedestrian standup comics like Bill Burr have the answer. No hard evidence the AI slaughter of human scientists in Japan, ever took place, so destroying individual robots, in spite of high profile, seems easy. Yes, it is Neanderthal and childish, but just as lowbrow to claim superiority and do nothing as planetary-wide threats loom.

Te apellidas como cantante de Iron Maiden🤟. Buen Video, muy interesante.

You are GOOD !!

hola im here again. great video.

Recently the enterprise said they have a new Strawberry Q* model... Can this new model count the leters? I don't know, just especulating

We need more❤ keep going

Ty

Which accent is this?

Nice one. I had never made the distinction.

That's what our mother had us do when we were little. We therefore always knew that. The lady there is right. One point though. When you start to drink-- do not stop and release the straw with your lips. It seems to be the constant vacuum-ing that does the trick -- it can be felt at the eustation tube and ear drum. This works for simple hiccups. Sometimes it works partially but you already drank all the water. You'll have to do it over. Result -- your belly is distended with water (we were children). You'll have to wait awhile; then do it again. There are other causes of hiccups which have very serious causes and must be attended to by a Medical practitioner.

❤❤

dude this episode was sickkkkk, current undergrad and watched this on a whim great job, really enflamed my curiosity about the world. As we grow older we begin to take everything around us for granted and stop asking questions. When I was younger I would ask all sorts of questions. Moral, philosophical, religious, scientific. But the adults around me would not entertain my curiosity, I was annoying and insistent on my wonder of the world, after all, it was my first time alive. But as I got reprimanded I began to hold my tongue, until I stopped asking altogether. This episode has made me wonder about the things around me again and ask questions. I don't think I am the target audience, but great job everyone who worked on this and other ep's

dumb

Thank you, Dr. Dickinson!

Not going to watch. 1. Too much musk clickbate. 2. I know this is a solvable problem 3. I'm sick if musk... so 1st point again.

But even with the ones using such against people and making them use tin foil hats There is still one Golden Truth They Cant ever be united in the truest sense you affect them as much as they do yourselves

Interesting how you get a different experience to me. I tried this on 13/06/24 - ChatGPT 4o told me, like you, 2. It did a similar 'spelling out' as you got and still insisted it was 2. But when I told it the positions of the 3 letters (including that we were using a 1-indexed array), it agreed from therein that it was 3. I couldn't go back to it insisting on 2. It was like it had learned something (which I know it hasn't in the past). It was only when I closed the tab and started a new chat did it go back to thinking 2.

And it took how many GPU's burning how much electricity and a server farm using how much water to cool it's servers to give us the wrong answer?

Thirst

Aha. This kind of helps me understand why Claude struggled to condense my 4000-odd-word assignment into a 1500-word abstract. It gave me 650 great words, but when I challenged it on the length it apologised profusely, then gave me 902 words, and finally 1213 words, which it said was "approximately 1500". This runs so counter to the perception of what I thought AI would be good at, and you just helped me understand why.

Use Claude 3, got the answer right first time lol. But wait, I decided to see if it was really counting the r's, here's my transcript: RA: claude, how many "r"'s are there in the word "strawberry"? Claude: There are three "r"s in the word "strawberry". RA: sorry I meant the word "strawbery" Claude: The word "strawbery" is not a correct spelling of the word. The correct spelling is "strawberry," which, as mentioned before, contains three "r"s. RA: oh okay, how many r's did I spell it with then? Claude: In your misspelling of the word as "strawbery", you included only one "r".

So what we need is a second AI that can generate simple "strawberry" like questions that we can use to see if another AI is worthy. Problem solved... not getting complicated at all. But seriously lets confine it Exurb1a style before it gets away on us.

I have tried a few things like this, and even absolutely leading it to the answer is really hard. A month or two back, ChatGPT was meme-ously failing at "which weighs more a pound of feathers or two pounds of gold?" I tried the more fun version which weighs more a pound of gold or a pound of feathers (which is a trick question as these aren't actually the same because you measure gold using troy pounds which are different), but even leading it to the answer by first asking it to explain the difference between avoirdupois pounds and troy pounds, and "oh so the troy pound is lighter?" queries in the middle, it would still get it wrong when you asked the straight question at the end.

I was thinking like a box to put over my head when I sleep cuz they say don't sleep with electronics near your head. The 5G makes people go crazy...

I didn’t hear a difference 😞

You can hear when the shower gets hot too, did not know why. Cool thanks!

Genuinely made it worse 😭

Do you guys know of any conductive hat setup that could pick up radio signals or other waves and funnel them into your brain and microwave it. I saw some guy say that it could.

No mention of grounding the hat