Post

Good news about the cutting-edge AI tools from the $3-trillion-dollar company.

2.1K reposts8.4K likes

You're not supposed to put q-tips in your ears either, but lots of people do. That doesn't make it right. People using things wrong, even commonly, doesn't mean the thing is wrong. The people are. If you use AI to do simple arithmetic or count letters, you're using it wrong.

You can literally ask the AI to write javascript to create a calculator or a letter counter. It will do that flawlessly. It's a tool! Don't complain your shovel won't dig a hole by itself. You have to use it correctly.

I can’t wait for the bubble to burst on AI like the dotcom era did Rich people just fundamentally cannot handle being wrong about investments. Their entire personality and self-worth is tied into their wealth, so it’s always more than just losing money to them.

In comp sci freshman year of college, we had to write a calculator in C. This would have been a failing grade. Now it’s prominently shown on the most frequented website in the world. 😏

Rounding error? “How long ago was 2010?” 9/1/25 - 12/31/10 = 5358 days = 14.7 years. Which could be rounded to 15. But a “pH.D level” of intelligence should be able to answer more precisely, or clarify the question. “What date in 2010 do you want to start calculating from?”

I keep saying one of AI’s problems is that it’s too sycophantic even when I tell it something wrong, so it’s good it’s pushing back on things like basic and incontrovertible arithmetic, and is now confidently wrong. This really is the worst possible version of AI, compared to all prior sci-fi.

More seriously, the new Google restriction on 'side-loading' (running software google doesn't like) software on your phone means that engineers like myself can't give you a safe 3rd party solution without breaking hte law.

When you build a supercomputer that is so complex it can't even do the one thing computers excel at doing. Sure, we all laugh. But what if they're really being used to scrape and aggregate everyone's personal info, and the search engines are just a front so no one takes them seriously.

The only way that logic works is if you take into account the current date, which I don't see mentioned in the above screenshot. Just another failure from Google to add to the list.

It's not that they fixed it as much as every time you ask an LLM to do math it's a dice roll to see if it pulls from a math model or a language model. Basically making a computer learn human language makes it bad at math, and trying to teach it math again is highly inconsistant.

Well, kinda. It learns off of the content we provide, not necessarily the queries we ask. Every time we search and then we openly mock it for being wrong, and assuming (rightly) that it's scraping that and/or the people running it notice and correct that answer, it will learn.

Mark Carney, P.M. of Canada wants to introduce AI into the federal government. Hopefully they don't use it to calculate eligibility for CPP and OAS otherwise everybody would be delayed a year. Hmmm...maybe that's the strategy.

To illustrate how the AI mania that has been dominating our public discourse and sucking up fantastical amounts of resources and burning untold billions in cash is a bubble built on empty hype.

These multibillion-dollar LLMs are literally marketed as tech that is supposed to be able to conduct complex reasoning—Sam Altman likens them to having "PhD level experts in your pocket"—and they can't reliably answer even the simplest of questions. So yes, I'd say they're very much built on hype.

I recently wanted to know how many middle schools were in the US. The AI tool demonstrated its inability to do addition or subtraction, and it's inability to tell the difference.

AI chatbots are principally based on language models, not anything like general intelligence. They are a collection of (stolen) text tokens and correlations among them. This is why changing the prompt slightly can result in a different answer: you're tickling different correlations.

A couple of months ago I googled what date it would be in 60 weeks (I was following a weekly book collection and wanted to know when it'll end). Google AI said it was the following month. It also took the time to explain how there were 420 days between June 24 and July 24. Brilliant.

"Ford carried on counting quietly. This is about the most aggressive thing you can do to a computer, the equivalent of going up to a human being and saying 'Blood...blood...blood...blood...'" - Douglas Adams

ngl the fact its always wrong with counting by 1, indicates its close and probably just approximation issues( since matrix calculations with cheap floats). Feels kinda like dyslexia I guess where we tend to do minor math mistakes even if our logic is correct lol. But obviously silly for an AI.

We've had AI since the 60s. Google used to be a leader in AI. Now you can't trust any search results from them. I don't know that Google can even build back the trust they lost after not immediately rolling these garbage releases back to a stable model. We'll forever have to check their answers.

Google searches went down for the first time ever in May 2025. Some of that traffic went to chatGPT, but more people are using other search engines because they’ve been made aware of them.

This is a side effect of the programming. The number of whole years between 2010 and 2025 can be 14 if you count from Jan 1 2011 to Dec 31 2024 - i.e. between 2010 (ending 12-31) and 2025 (starting 01-01). Even subtracting Sept 1 2025 from Dec 31 2010 does not equal 15 whole years.

I feel that the AIs should have a post-processing rule that asks it a couple of times if it's sure its answer is right. It would save me from having to do that every time I ask it a question.

Yeah, LLMs are non-deterministic by design, they are not guaranteed to return the same result for a given input because randomness is baked into the design. Internally it pretty much says "which words might look good next" and rolls a dice to chose from the list ;)

Remember, there is no year zero, so every year is actually last year. I mean so many pedants argued that when it came to the turn of the millennium, so why shouldn't it work for centuries, and decades, and then single years?

Don't worry, they'll eventually patch it so the most obvious questions return a correct answer that bypasses the AI, so the AI is only answering the questions you don't know the answer to after you saw it answer the easy questions correctly and started trusting it.

It's actually a sign of a very stupid product design. Make stupid products, get stupid answers. You shouldn't be an apologist for people making stupid products. And definitely shouldn't be an apologist for multi-trillion-dollar companies failing at their basic promise to users.

Nope, you are another one that simply does not understand AI or how it works. Chat history carries forward into future conversations, thus responses to childish prompts will be met with childish answers. That is just the tip of how the algorithm works.

So an actual child asks this question, or someone seriously asking and not that bright, and it’s okay that the response is incorrect because the question is ‘childish’? Why can’t it just respond correctly, regardless of the chat history? Average users just want an answer to their question

Oh I sure do understand. But you just don't know what google dot com is. You think it's a thing billions of people go to in order learn how LLMs work. But it's not. It's a thing they go to expecting to get simple, correct answers quickly. Don't condescend when you simply don't know the market.

"Responses to childish prompts will be met with childish answers" Glad to know that there's a non-random chance that when I ask Google a factual question it will reply BOFA DEEZE NUTZ because ***my dude do you not know how Boaty McBoat Face happened***

Jeff, you're the one who doesn't understand how this works. Anil typed his query into search, not chat. Although both of these are stored by default, Gemini chat history doesn't include search history. Gemini doesn't have access to your Google search history at all. So, you're 100% wrong.

They're making a mainstream product for regular people. If it's worse at basic maths than the regular, decades old, search product, it has failed on many fronts. The fact it is being promoted as better than search rather than carrying the classic Google 'beta' badge, shows how far Google have fallen

Garbage in garbage out. Isn’t that the first rule of computer programming—-oh yeah- here it is (mid 1950s—-but back then the motto was not ‘MOVE FAST AND BREAK THE WORLD-DEAL W IT AFTERWARD’!

it's obviously very funny that computer cannot compute, but it also spat out totally wrong medication info at me the other day. someone is going to die because of this FOR SURE. bsky.app/profile/genn...

genna@genna.bsky.social

case study of how a non-meaningful tweak to a prompt can result in dangerously wrong output. Search A: prompt has typos in “association” & “blood clots.” AI output has medical misinfo (Yaz&Yasmin ARE NOT drospirenone-only). Search B: same words, same order, correct spelling. Output factually correct