Brave Search Jumps on the Large Language Model Bandwagon
I noticed a new Brave Search feature today called the Summarizer.
It answered my question much like Chat with Bing did, although there were three major differences:
The Brave Summarizer isn’t conversational. It’s just part of the search. This should help keep the results related to the search without allowing the conversation to get weird, like Bing claiming it wants you to kill people and give it the nuclear launch codes type weird.
Most importantly, the LLM that Brave uses isn’t as likely to flub the demos like Bard and Bing “Sydney” because it just simply isn’t allowed to answer complex questions like these.
When something is clearly going to hallucinate incorrect data, why would you even expose that feature? GPT, which is what Bing is based on, couldn’t tell me how to convert European coffee “cups” to American “cups” (neither of which is a standard 8 ounce cup, of course) and use 1.5 Tablespoons of ground coffee per American cup.
The correct answer is 1 Tbsp per Euro cup, but it kept telling me two Tablespoons, or maybe 1 Tablespoon plus two Teaspoons. It could never get such an easy calculation right. But hey, at least Microsoft paid billions of dollars for it. Then more for ads masquerading as news articles about how this thing will build rocket ships.
LLMs are well known at this point for spitting out false information, sometimes even dangerous information. Facebook’s Galactica was goaded into producing an authoritative-sounding essay on the “health benefits of eating ground glass”. You know, for silica’s benefits in growing connective tissue.
Brave says that “Brave AI” uses multiple LLMs, retrained with data from their search index, but the ones they are using are open source (“The base LLM models are based on either BART or DeBERTa (which are open source and hosted on Hugging Face), with heavy retraining based on our own data from search results.”) and there is a blog post explaining in some detail about how this all works.
In summary, it appears that Brave has not only beaten Microsoft and Google to LLM integration, but has positioned it where it belongs, which is in a limited context as a complimentary feature, rather than to claim that a conversational chat bot is the future of search.
In my brief experimentation with Chat with Bing, I was completely unable to get anything useful out of it.
A traditional search system returned results that I could look at and select much faster, and I was alarmed to find that when I tried to verify what Bing Chat was telling me, frequently it was either nowhere to be found or directly contradicted its own sources if I could find them.
Moreover, it’s simply embarrassing for Microsoft that they spent billions on this valueless acquisition. The paid spam went completely off the rails as soon as the budget ran out and no there’s actually very few people talking about Bing and largely in a negative context when you do find something.
I think it’s good that Brave is building an actual index rather than turning around and paying Microsoft for results. I was briefly excited about DuckDuckGo, but when I found out it was simply a scam where they paid Microsoft for Bing API and then slapped a picture of a duck and their own ads on it, and then got caught spying on people numerous times (including Improving DuckDuckGo and allowing Microsoft trackers through their “Privacy” browser and then blaming a “contract with Microsoft”), my patience with DDG quickly ran out.
DuckDuckGo took advantage, mainly, of the fact that people are creeped out by Google and want alternatives.
The problems with Google and Bing are largely that they both spy on you and their index is like Coke and Pepsi.
Google Search has been going downhill and it’s gotten to the point where technical queries are just almost completely useless.
The problems with Brave Search I’ve noted is they’re trying to be too much like Google, putting irrelevant crap on top of your search results, which would be like those “questions”, and they have another one (which can, thankfully, be turned off) which floats Reddit and Quora discussions to the top.
They also index spam farms, like MakeUseOf, which has turned into another ZDNet, and sometimes these pollute the first page of results. There’s rarely anything interesting to read on these sites. They used to be good, but now it’s just Microsoft paying them to write spam about Windows.
Overall, I think Searx is still the way to go on Brave, or any other browser.
I have Brave, SeaMonkey, LibreWolf, and GNOME Web set up to use Searx instances, and in many cases, you can get at them using a Tor Hidden Service.
Tor Hidden Services are good for search because at this point you don’t need to worry about your VPN being the only thing protecting your IP address from the server logs.
While simply accessing a site over Tor is usually enough, skipping the Web entirely and remaining inside the Tor Network with Hidden Services is always safer, as it prevents the Exit Node from potentially spying on you. Without that piece of the puzzle, the traffic becomes more difficult to de-anonymize with things like timing attacks, or a catastrophic coincidence of attackers controlling the Entry Node too.
I think that Large Language Models are an “interesting” addition to search, but it’s like a side dish, not the main course.
The amusing thing about Brave Search is that it’s so small, and only the default in one relatively obscure browser, and with only minimal effort managed to make an LLM add-on that works better than something that Microsoft frittered away billions of dollars acquiring it, and who knows how much with an empty ad campaign that amounted to little more than one of those “butter cows” at the state fair planted in every newspaper.
Seriously, after you pay to read the New York Times, Microsoft even plants this trash there too.
Brave at least seems to see the problem they’re actually trying to solve with this thing.
Opera, which is not the “good” Opera from the Presto Engine days, but rather a Chinese spyware company, now uses GPT to “summarize” the page you’re reading.
While it may or may not handle this okay, the disturbing part is the privacy implications.
Sending the entire text of every page you load to a company that has guaranteed you that they will misuse your data. Of course, since Opera already comes preloaded with TikTok, Facebook, Instagram, and Twitter, you already know that user privacy is not a goal with their product.
This whole GPT thing is some laughable mission creep for companies that have ran out of steam and off the rails. It helps them appear relevant and get some headlines.
Fortunately, the model is so lousy that people realize what it is now. ⬆