A few weeks ago, in a viral YouTube video, The video was enjoyable, where Kerry from GenAI Nerds acts as the host and interviews friends. Luckily, Dusty from NVIDIA was present to reply to all the demanding questions. Looky right here:
I’ll level out that I don’t usually let folks {photograph} my aura that you see within the thumbnail right here. It’s very particular and requires delicate photographic tools for the complete hyper-spectral impact. Dusty is in one other quick section:
The Fascinating Bits
These are the primary movies on the channel. As of this writing, the channel has 1.4 million views and 331 thousand subscribers in just two weeks! To put that into perspective, the JetsonHacks YouTube channel is reaching the 35K subscriber mark in its tenth year. Wow, and congrats to the GenAI Nerds staff!
A Longer Reply
Throughout the lengthy interview, one of many requested questions was, “Why are LLMs unhealthy at math?” That’s a very attention-grabbing question that folks ask. The actual question is entirely different, after all. The question is, “Why would LLMs be good at math?”
A straightforward view of an LLM (Large Language Model) is that it is vitally good at guessing the following phrase given a context. It does this utilizing computational statistics. You might have a machine that reads vast amounts of textual content from books, articles, websites, etc. Massive LLMs have learned the whole public Web, after which some have seen it.
The machine learns patterns in the text and becomes good at predicting what phrase comes next in a sentence based on previous phrases. It doesn’t “know” what the words imply but is very good at predicting how phrases match together.
This course is way more mechanical/computational than this description. However, after this coaching course, the result is the LLM fashions you now know and love or that scare you to death. Folks appear to be on one facet or the opposite.
However, Why No Math?
That is the place where everything will get somewhat philosophical. Some folks argue that LLMs exhibit intelligence or not; it’s figuring out issues. It distills information way more than any particular person is aware of. On the opposite facet of the argument, we say that the LLM doesn’t really “know” something. The LLM represents a much less correct information retrieval mechanism than an extensive database. In addition, they make the case that till a machine is embodied, it doesn’t have actual world information. By embodied, they usually imply a robotic interaction within the bodily “actual” world.
The problem with math within the LLM is that there’s no illustration of how math works—possibly easy arithmetic with small numbers, but nothing algorithmic. Without information on how arithmetic works, getting a fitting reply by guessing what comes subsequently doesn’t work. Each of my math academics instructed me that I couldn’t simply think of the solutions. Or they graded my papers as if that had been true.
As a side note, these are the same killjoys who might fuss at me when taking several alternative checks. They got their feathers ruffled once I tested all the reply picks to hedge my bets.
The opposite concern is that there are a lot of numbers. Some folks have instructed me that there’s an infinite number of numbers. I’m undecided. Do I imagine them? Have they each sat down and made a critical attempt to depend on all of them? I feel not.
That’s why LLMs are attention-grabbing. Jeff Bezos says that LLMs are ‘not innovations, they’re discoveries.’ He’s not improper. There’s a whole lot of exploring to do.