Impressive skills for its size, but this doesn't belong on normal peoples' phones.
This model is surprisingly good at coding, math, story writing, and even poetry.
However, the general population is souring on AI, as MS learned with copilot, and the more normal people use this AI model on their phones, the more they're going to hate AI.
I don't want to be negative, but how is it possible you guys don't realize what the consequences will be to the public perception of AI, hence the future of AI, if normal people start using AI models like this on their phones?
This model doesn't just regularly hallucinate about very popular core human knowledge, it vomits hallucinations. When people see this happen about things they care the most about, such as their favorite movies, singers, video games... what do you expect them to think?
This model has about a couple orders of magnitude higher hallucination rate when it comes to humanity's core popular knowledge than is required to secure the public's trust in AI.
My recommendation is to reduce the dictionary size to no more than around 50k, make separate edge AI models for each supported language, and even with the large amount of freed up space with only 50k vs 250k worth of embedding, it needs to be about 3 times bigger while preserving the same number of active parameters. It also needs a grounding compact relational database of core knowledge from the supported language to keep it from falling of the rails. And lastly, tasks like coding need to be almost entirely removed. >95% of the general population does not code, and coders would never use an edge AI for any coding task, especially with far more powerful coding models out there.
I beg of you Google, don't make the same mistakes Microsoft made. If you want to create ubiquitous AI models that run on the phones of the general population you simply must reduce the rate at which they hallucinate about core popular knowledge by about a couple orders of magnitude.