Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Grok is apparently taking off in India and has gone viral for its uncensored takes on politics and... other things.
Musk is capitalizing by charging 1/4th the price for Grok in India compared to else where in the world, while the Indian government is scrambling for political censorship. Expecting a deal between Musk and Modi soon. As expected, Grok is hitting its niche of uncensored AI well, given the lack of competition.
ModiG45 LLM
ARC AGI 2 has been released. Based on the leader boards, no modern LLM even comes close to solving the challenges:
That said, the problems appear to involve visuo-spatial reasoning, which for LLMs trained on text data, raises an apples & oranges problem - human reasoning isn't just linguistic, it is just as much, and perhaps even more so, visuo-spatial reasoning, which humans would possess even without language.
But at the same time, claiming you have an AGI without being able to do visuo-spatial reasoning is rather ridiculous. The test for these models *is* that they are able to generalize text-based learning to visuo-spatial learning through mapping between symbolic and visuo-spatial problems.
If they can't, fundamentally, do that, then they're not really thinking in a general way and, as such, cannot be considered AGI.
And of course, the ARC AGI foundation never said what method you should use. If LLMs can't do the job alone, then companies are always free to integrate a visual model.
That said, the problems appear to involve visuo-spatial reasoning, which for LLMs trained on text data, raises an apples & oranges problem - human reasoning isn't just linguistic, it is just as much, and perhaps even more so, visuo-spatial reasoning, which humans would possess even without language.
But at the same time, claiming you have an AGI without being able to do visuo-spatial reasoning is rather ridiculous. The test for these models *is* that they are able to generalize text-based learning to visuo-spatial learning through mapping between symbolic and visuo-spatial problems.
If they can't, fundamentally, do that, then they're not really thinking in a general way and, as such, cannot be considered AGI.
And of course, the ARC AGI foundation never said what method you should use. If LLMs can't do the job alone, then companies are always free to integrate a visual model.
That some impressive performance improvement. That's termed as 'minor' update, LOL
Remember the Nvidia CEO is talking about "agents" with "superhuman" capabilities in 2-3 years. That is still nowhere near AGI.
And the Microsoft CEO is fed up with OpenAI hype on AGI.
