28 comments

[ 2.8 ms ] story [ 40.6 ms ] thread
Huh, hadn't realized that Anthropic had gone full Rainbows End. RIP Vinge :(
Current explosion in AI in general is "good". We tend to over-criticize and nit-pick it as what I believe is a natural response that comes from a "communal subconsciousness" -- even if us as individuals won't admit it, as a culture we are scared and averse to what's happening.
I’m not surprised by Google’s lack of search transparency. If Google told users that Google was searching an internal database linked to a particular search string there would be pitchforks and torches.
Isn't that just... caching
I too feel AI mode seems to good when it knows the topic, with two caveats:

- it works when the info is either relatively well known or quite new

- no-AI mode now becomes dumber, the old trick to "grep" the internet with +/-/"" is gone

My main takeaway is that Google's AI efforts are still playing catch-up in the mindshare race even if they've long since caught up with or surpassed OpenAI technically. A huge number of people who are heavy users of ChatGPT or Claude may have never tried Gemini.
i think they are shooting themselves in the foot by automatically including results from a quite dumb model in google search
IME Gemini has a lot more non-tech mindshare than Claude. For tech it's of course the opposite.
Google is really shooting themselves in the foot with AI overviews.

It's probably the most popular AI on earth by daily queries, and likewise probably an ~8B level model, it means a whole bunch of people equate Google AI to AI overviews.

yeah a lot of people i speak to i think have developed the impression that modern AI hallucinates more than it does from the dumb search model
I'm a bit biased, but I find the AI overviews to be basically great. All I want from a search engine is the correct answer, Google's knowledge graph has done that for many queries for a long time, and AI overviews seems like a good next step in that process.

I've not seen many hallucinations, fact checking is fairly straightforward with the onward links, and it's not like I can take any linked content at face value anyway, I'd still want to fact check when it makes sense even if it wasn't AI written.

You've been asking things that are verbatim in the first few links. Ask anything that isn't and it will confidently hallucinate.
They're also shooting themselves in the foot by halving the traffic to the websites that provide the training data.
I read this as "AI Mode is God". Strangely I was unperturbed, quietly resigning myself to our new deity.
My first read was Google AI is God mode. They would earn back some points from me if they added a Konami code like entry point to the feature.
I agree, AI Mode is _very good_. But there's a huge branding and discovery problem in that 100% of the people I mention this to think I'm talking about the (error riddled) AI Overviews and have never heard of AI Mode.

I do probably 40% of my searches with AI Mode now. It can't possibly be profitable (and maybe that's why it's not more discoverable), but the results are awesome.

Edit: I also tried to show my aging parents how to use it, and it was inexplicably not available on their devices. They use old (10ish year) ios devices, which is apparently incompatible even though it's a web interface.

It is pretty good, although Perplexity is better (but nominally not free, although you can get a free subscription now).

OpenAI searches are even better, but GPT5 is extremely slow with thinking. Without thinking it's roughly equivalent.

The example itself that Simon puts up is questionable. I might be wrong about this, but I thought I read elsewhere that the “buy, scan, destroy” method was explicitly not the problem, and instead the issue was that anthropic downloaded libgen, and the settlement was for libgen.

Edit: https://news.ycombinator.com/item?id=45143392

i don’t read the AI answer as saying otherwise, but one might read that the physical purchasing was used as part of the settlement. it was used as part of the defense to mitigate though
Google's AI mode recently told me the haber process was invented in 1999, and in the same paragraph told me it won the Nobel prize in 1918...
Was that AI mode or AI overviews?
On that note, AI mode's image search is really good

I've gotten it to identify:

- the comic from a random page of an obscure Russian comic

- obscure French comedy from a random clip

It was extra impressive because even reverse search from lens didn't immediately identity them

It feels like there is a different search use case that AI accels in while it destroys another.

- When I know so little about a subject that the concepts are all vague and I'm intellectually grasping in the dark. LLMs are good at taking my imprecise wording and orienting me towards the paths/gradients others have taken.

VS

- I know this subject well enough and instead of fumbling around I want to be able to run grep over a massive amount of open text data.

While the former mode is useful, the act of training without asking has led to the repopularization of walled gardens. Early Google felt like being able to grep every library book in existence and a 3 paragraph summary in response to very poorly worded questions is a terrible trade.

I know Simon is smart and he included his search terms with their warts to be open so I don't want to elicit shame over this. But cmon, just type out bought ("Anthropic but lots of physical books..."). Complete anecdote but I have noticed my LLM reliant friends have become way worse at texting and it feels like it's worth taking 5 seconds to try to structure your thoughts, simply for the practice

Just tried a few side by side comparisons with perplexity.

Google answers more concisely, faster and confidently, but not convinced quality of output is better. e.g. Google pulled in info from AWS and Oracle cloud when I asked a GCP specific question. Perplexity sourced only from GCP docs

Which is an interesting outcome since I'd expect google to excel in the search aspect

> since I'd expect google to excel in the search aspect

You still do?

I usually can't read blog posts, they are stale, boring and can't keep my attention span unless I know there is something interesting the paragraph, this is regardless of the interesting title. But when it comes to Simon willisons blog, I enjoy reading it, it's entertaining and interesting, it reads like a human and is enjoyable to read

Also, I can't access Google AI mode because I'm in EU but when looking at the video on YT, it looks like Perplexity, but googlified. I haven't seen any other tool that comes close to Perplexity yet, I have their app installed on all my devices and it's part of my daily life, it's so good! Especially with their pro plan (I got 12-months for free)