4 ms·

The printing press takes your ideas that you already wrote and increases distribution. Deep research is writing the text for you. That’s completely different.

seanmcdirmid 2y ago
I disagree, it’s just another step in the automation of knowledge distribution that’s been going on for a long long time. Every step someone has lamented that we lost something, maybe starting with speech itself, and the invention of writing taking away from oral tradition, scrolls, books, the printing press, the World Wide Web, etc… I was just wowing to myself the other day that a local 70b 4q 40GB model on my Mac basically replaces Wikipedia for me and can be used offline in an airplane. A whole library can be replaced with a cube of knowledge that is just a wrapper around an LLM. Of course there is going to be resistance to any move up, there has never not been resistance.
- beeflet 2y ago
  It's like wikipedia except it spits out incorrect information at an alarming rate and has no references. Still, impressive that it only weighs 10s of GB when wikipedia alone is on the order of ~100GB
  - seanmcdirmid 2y ago
    Wikipedia isn’t a completely reliable source either. The LLM also includes a search engine. It’s weird since before I started playing around with it I thought it was just a fancy search engine and was grabbing sources from somewhere to shove into its context … but no, it’s all built in (although as deep research shows, it can do a lot better when it stuffs its context via research).
- porridgeraisin 2y ago
  Just saying that for 24GB you can get all of wikipedia losslessly compressed on your laptop.
- Foobar8568 2y ago
  While the 400B "knowledge" is reliable and replaces most of my googling, the 70B are not, nor their "reasoning". So until we can have these models locally, running at 10-20token /sec minimum... I am fairly hype up over gen AI but we are only at the start of this revolution, and will require a few more years for niche domains, or company knowledge (RAG doesn't cut it, fine tuning too expensive) and we aren't tackling all the media related. And because of hallucinations,which they require to work, they can't be reliable at 100% or be truly use without supervision. I feel it looks like internet back in the 90s, we were saying it as a new way to connect people, share knowledge, only to have echo chambers and porn distribution as few take away. I kid I kid.

seanmcdirmid 2y ago

I disagree, it’s just another step in the automation of knowledge distribution that’s been going on for a long long time. Every step someone has lamented that we lost something, maybe starting with speech itself, and the invention of writing taking away from oral tradition, scrolls, books, the printing press, the World Wide Web, etc… I was just wowing to myself the other day that a local 70b 4q 40GB model on my Mac basically replaces Wikipedia for me and can be used offline in an airplane. A whole library can be replaced with a cube of knowledge that is just a wrapper around an LLM. Of course there is going to be resistance to any move up, there has never not been resistance.

beeflet 2y ago
It's like wikipedia except it spits out incorrect information at an alarming rate and has no references. Still, impressive that it only weighs 10s of GB when wikipedia alone is on the order of ~100GB
- seanmcdirmid 2y ago
  Wikipedia isn’t a completely reliable source either. The LLM also includes a search engine. It’s weird since before I started playing around with it I thought it was just a fancy search engine and was grabbing sources from somewhere to shove into its context … but no, it’s all built in (although as deep research shows, it can do a lot better when it stuffs its context via research).
porridgeraisin 2y ago
Just saying that for 24GB you can get all of wikipedia losslessly compressed on your laptop.
Foobar8568 2y ago
While the 400B "knowledge" is reliable and replaces most of my googling, the 70B are not, nor their "reasoning". So until we can have these models locally, running at 10-20token /sec minimum... I am fairly hype up over gen AI but we are only at the start of this revolution, and will require a few more years for niche domains, or company knowledge (RAG doesn't cut it, fine tuning too expensive) and we aren't tackling all the media related. And because of hallucinations,which they require to work, they can't be reliable at 100% or be truly use without supervision. I feel it looks like internet back in the 90s, we were saying it as a new way to connect people, share knowledge, only to have echo chambers and porn distribution as few take away. I kid I kid.