pkjqpg1h

MiniMax-M2.5 is a SOTA large language model designed for real-world productivity. Trained in a diverse range of complex real-world digital working environments, M2.5 builds upon the coding expertise of M2.1 to extend into general office work, reaching fluency in generating and operating Word, Excel, and Powerpoint files, context switching between diverse software environments, and working across different agent and human teams. Scoring 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp, M2.5 is also more token efficient than previous generations, having been trained to optimize its actions and output through planning.

Because they don't really search or index quality content (it's very expensive and hard to do) and their search implementation really sucks, they don't do any real improvement.

The process is like this:

1. Take the user query and create 1-3 queries. For this process they use very stupid but fast and cheap models; because of that, sometimes they create very stupid search queries and, unlike a pro, they don't really know how to use search engines, like filtering, ranking, focusing...
2. Combine these search results (it contains slop AI-generated summary pages, YouTube videos, maybe forums, maybe Wikipedia...).
3. Use RAG with an LLM to find answers. LLMs will always try to find answers quickly, and instead of making a thinking loop in a long article they will use that slop page with a direct answer.

As you can see, there are many, many problems in this implementation:

• The biggest problem is citation: they cite confidently but it's wrong.
• They use low-quality data, like auto YouTube subtitles, improperly extracted tables and elements, content-farm sites, copycat sites, corporate blogs...
• Their search results are low quality.
• For the most important part (breaking down the user request) they use cheap, stupid models.
• They handle all data in the same context instead of parallel requests (which is very expensive)

It's still strange to me: we always say "they have all the data, all the money, all the hardware..." but they still can't create a better AI search than random FOSS developers.

It says

A proprietary service that combines GPS, Wi-Fi, cellular networks etc. to determine your location more accurately. It may connect to Google servers.

But I don't have Google Play Services, I use microG and microG don't use Google's services for location they use beacondb.net. Can we change this?