If you take from the web, you should give back. Search engines like Google, Bing and Mojeek do that with hyperlinks; sending traffic back to the web-pages which they have crawled and indexed. The voluntary agreement underpinning them, expressed through website robots.txt files, has been based on the legal concept of fair usage. Read more…
2 Likes
Just tried the search summary on a query which I mentioned before AI excels at: difficulty of grasping keywords on an unfamiliar topic.
In this query, what I’m searching for is an idiom I’ve heard before, but I can’t remember the correct phrase, so I typed a bunch of words that I feel are related, and used the summarizer to check which article/s contain my target search.
Here are the problems I saw:
- Failed to surface the target search, “bite off more than one can chew”, found in the 2nd search result.
- The last statement is a hallucination. Those 2 references are neither discussing what the AI said (consequences of eating too much), nor are they related to each other.
- No related questions like in RAG search, which could have helped identify the correct keywords and narrow down the results.
On the other hand, here are some parts that I like:
- Using words like “could” help reduce the impact of hallucinations. Invoking a bit of uncertainty reminds users that the summarizer is not an authoritative source, and verifying the source is necessary.
- Referencing the sources (ex. “as mentioned in”) instead of just listing them reminds users that the summarizer is merely taking ideas from these sources, putting these sources in the forefront instead of in the background.
In general, due to the abundance of hallucinations, I prefer the AI to speak like an academic: explicitly pointing to references, and avoiding overconfidence.
Btw bizarrely, RAG search has a distinct response, contrary to what @Colin said before:
The sources are also different, where some have their order changed, and one unique (#7). Is this a bug or does RAG search work differently?
This isn’t a bug. RAG search is part of labs.mojeek.com, so does not have a Preferences option, being hard coded with some parameters, relating to language and location setting. It’s also using a different snippet length and snippet algorithm.
We have added a parameter &kaltem, so that you can change the LLM (Large Language Model) temperature setting in the URL. The current default is 0.5, and can be changed to for example 1.0 like this.
Mojeek =Should+AI+be+open+source%3F&fmt=summary&kaltem=1.0
A lower value makes the model’s responses more deterministic, predictable, and consistent. A higher value makes the responses more diverse and random. The range yoi can use if 0.0 to 1.0.
The example in the image, on the left has &kaltem=0.0, and on the right &kaltem=0.5
