Hacker News Reader

Comment by smallstepforman | Hacker News Reader

Google search is giving us a taste of AI summarised results, and for simple things its passable, but ask a serious question and you get good looking garbage. Yes, I know its early days, but looking at the current output quality we have nothing to worry about. It will be used as calculators, offload some menial repetetive task which can be automated, but the next gen of developers will still be tasked to solve complex problems.

View on HN

Google AI the other day told me that tinnitus is listed as a potential adverse reaction of Saphnelo.

Only it damn well isn’t. Anywhere. Not even patient reports.

The problem with AI is if it’s right 90% of the time but I have to do all the work anyway to make sure it’s not one of the 10% of times it’s extremely confidently wrong, what use is it to me?

This problem is has already gotten so much better. In my experience it's no longer 10% of the time (I'd estimate more like 1%). In the end, you still need to use judgement; maybe it doesn't matter if it's wrong, and maybe it really does. It could be citing papers, and even then you don't know if the results are reproducible.

Has it actually become that much better or have you let your standards and judgment lapse because you want to trust it?

How would you even know to evaluate that?

Ya I've had basically this question for a while. My assumption is that most of the time people search the internet to answer questions they DONT know the answer to.

If an LLM gives you a response to that question, how do you know if its right or wrong without already knowing the answer or verifying it some other way? Is everyone just assuming the ai answers are right a majority of the time? Have there been large scale verification of a wide variety of questions that I'm not aware of?

Case in point.

I purchased a small electronic device from Japan recently. The language can be changed to English, but it’s a bit of a process.

Google’s AI just copied a Reddit comment that itself was probably AI generated. It made up instructions that are completely wrong.

I had to find an actual human written web page.

The problem is with more and more AI sloop, less humans will be motivated to write. AGI at least the first generation is going to be an extremely confident entity that refuses to be wrong.

Eventually someone is going to lose a billion dollars trusting it, and it’ll set back AI by 20 years. The biggest issue with AI is it must be right.

It’s impossible for anything to always be right since it’s impossible to know everything.

github copilot gave me wrong instructions on how to set up github copilot client :D

Google's AI overviews is the single worst AI-driven experience in widespread use today. It's a mistake to make conclusions about how good AI has got based on that.

Have you tried search in ChatGPT with o4-mini or o3?

I don't use it that much, but I have noticed these AI overviews still seem hallucinate a lot, compared to others. Meanwhile I hear that Gemini is catching up or surpassing other models, so I wonder if I'm just unlucky (or just haven't used it enough to see how much better it is)

One is their state-of-the-art model, the other one's the best model they can run at scale and speed people expect from a search engine.

That sounds reasonable, and I don't doubt that it's part of the reason. Still, iiuc the solution to hallucination is that they can essentially train the model to recognize when it "doesn't know", and to say so in that case rather than just puke back the highest probability BS. I.e. it's a training time factor, not inference time, so it's not a fundamental cost issue, but more about priorities.

I had a couple of great examples of getting the exact opposite answer depending on how I worded my question, now I can't trust any of the answers.

Maybe the answer to your question was subjective?

The first set of questions was about code review timing in a workflow (background: was in a discussion about different styles of workflows):

Q1.1: "do most developers do code reviews before testing"

A1.1: essentially "yes most before..."

Q1.2: "do most developers do code reviews after testing"

A1.2: answer was essentially "yes most after..."

The 2nd set of questions was related to building a retaining wall. I've never used the type of wall block with a center notch and groove, only the type with a lip on the back. The center type creates a wall straight up but lip in back leans back a little. I was curious if one was more stable than the other:

Q2.1: "is retaining wall block with center groove more stable"

A2.1: essentially "yes, block with center groove is more stable..."

Q2.2: "is retaining wall block with back lip more stable..."

A2.2: essentially "yes, block with lip on back is more stable..."

When I finally found an engineering website with real details, the answer was that the lip on back was a little more stable due to the resulting angle of the wall. It also emphasized that the lip and the center groove are not factored in to stability calcs at all, they are only for alignment, gravity and friction of the blocks is what holds it in place.

It makes me wonder, why don't we want AI to be deterministic and make it nondeterministic on purpose instead? I hope someone can explain it to me...

Google Search AI is the worst and considering that AI is not a good alternative to search (the models are compressed data), I am not sure why Google has decided to use LLMs to answer questions.

Try asking Perplexity for a real taste. It works far better than google's--good enough to make searching fun again.

Try coding with Claude instead of Gemini. Those that do tell me it is well beyond.

Look at the recent US jobs reports--the draw down was mostly in professional services. According to the chief economist of ADP "Though layoffs continue to be rare, a hesitancy to hire and a reluctance to replace departing workers led to job losses last month."

Of course, correlation is not causation, but everyone white collar person I talk with is saying AI is making them far more productive. It's not a leap to figure out that management sees that as well.

Reply on news.ycombinator.com