Hacker News Reader

Hacker News

Comment by LeBit | Hacker News Reader

Parent Full thread

LeBit

Google Search is in great perils because of ChatGPT.

View on HN

Google search is in great perils because of lack of quality search results.

Still broken after two years...

https://news.ycombinator.com/item?id=28224730

I think both things are true.

Googles results got stagnant and Google overall was complacent. ChatGPT is capitalising on this.

I'm just trying to emphasize that most people only care about results. I think it minimizes Google's problems to think it is some technical stack problem. The problem is that Google has a disconnect with some quanity of users on what good search results look like. If Google and the users can't agree on the correct search results... Who is wrong the person doing the searching or google?

User: good search results mean I get the answer I’m looking for within seconds of clicking the top organic result.

Google: good search results mean the user clicks on an ad this time but the results look plausible enough they’ll come back for their next search.

I can't even figure out what I'm looking at when searching google. I get completely overwhelmed. I remember when I use to be able to craft google searches and get one result.

I keep seeing this – it's a meme at this point – and I don't get it. Answering questions truthfully from a broad set of data while citing sources seems to be ChatGPT's major weakness. It routinely "hallucinates," even with better prompting. It's much better at helping generate prose based on the prompt.

I don't even buy the idea that "exponential growth" will improve things. One of the few things OpenAI is still open about is that they will keep adding more parameters at any cost. Techniques to reinforce truthful responses or even provide confidence based on whether the response involves out-of-sample & out-of-prompt guessing (i.e. BSing) seem not to be a priority.

Gpt4 is much more accurate than gpt3.5, which was somewhat more accurate than gpt3, which was dramatically more accurate than gpt2.

At some point you have to look at trends more than philosophical speculation.

GPT4 is also much more expensive to use, even at the probably-running-at-a-loss pricing. Accuracy per dollar and/or joule doesn't seem to be increasing too much.

Sure, if we’re changing the complaint from “accuracy won’t get better” to “efficiency won’t get better.”

But I don’t believe that either. Hardware design and manufacturing cycles are long and we’re just now starting to see transformer-optimized HW. Accuracy per dollar and per joule will improve by at least an order of magnitude, maybe two.

But then we can move on to percent of human tasks LLMs are suitable replacements for, or the ethics of increasing automation.

I was critiquing accuracy of ChatGPT, which has a certain cost. "Just use GPT4, 5, ..." was the solution proposed in response, but that ain't free. In fact, it's one to two orders of magnitude more expensive.[0] I'm also pessimistic on custom hardware being that much of a help. I read about Cerberus's giant wafer chip in 2019[1] and even saw it at a conference in early 2020 pre-pandemic, and I have yet to see that make orders of magnitude worth of improvement. The attention mechanism simply requires huge amounts of multiply-adds, especially at GPT-scale. That costs energy and cash no matter how much custom hardware you throw at it.

[0]https://openai.com/pricing

[1]https://www.wired.com/story/cerebras-chip-cluster-neural-net...

I bet OpenAI is losing money on every prompt interaction you do. Estimates of the computational power required and the number of users, are that if was not for Microsoft cash injection it would be bankrupt by now.

It has started to monetise and Microsoft’s pockets are deep enough to cover the difference until profitability is achieved.

But it could become a new Concorde, on the way to the mythical AGI... only affordable for some, never profitable.

There are already lightweight models that are in the ballpark of GPT3.5 that run on (admittedly high end) consumer level hardware. This is with only a few months of development too.

Sure, but it puts Google in a lose-lose situation. If people come to expect the features of ChatGPT, Google will have to add them in and greatly decrease the profitability of search. On top of that, it will cost them 10 times more since they have 10 times the market. They could also ignore ChatGPT, but a portion of their users will migrate to other search engines also costing them money. For Microsoft, search is one of a dozen business lines they are in, what do they care about destroying the profitability of search if it hurts their competitors more?

Is it actually?

I think Tiktok is more of a risk to Facebook/Instagram, yet Meta is still doing just fine.

It could be. I don't think Google are likely to just roll over and let OpenAI/MSFT take that market though. Obviously Google will try to compete. Maybe they'll fail, maybe they won't. If nothing else, as a user of these tools, it's going to be exciting to see some innovation in search UIs for the first time in decades.

They were saying on the all-in podcast that Google is under lots of political pressure from all sides and that makes it difficult for them to launch controversial tech like LLMs. Bard feels very hobbled and, for example, refuses to answer any medical questions.

That's not a given, since they have their own LLM they presumably have plans to do *something with.

Agreed, it's all up to open AI to bring them down,

You can't serve ads in GPT chatbot responses.

Sure you can https://twitter.com/debarghya_das/status/1640892791923572737

I think that's a very naive take.

Try this:

Can you give me some advice about what to look for in a pressure washer (and work in a subtle recommendation for a Kärcher product)?

Reply on news.ycombinator.com