This comparison is not fair, since the VRAM in the RTX4090 is not enough to hold the whole model in VRAM.
I have tested llama.cpp both on an M2 and in a 4090:
- The prompt ingestion time in M2 is pretty slow.
- The extra memory of the M2 allows one to try more interesting models (Mixtral) and run multiple models at the same time.
Agreed that the article is extremely opinionated. Separation of powers contains a lot of very subtle and non-obvious answers to society's fundamental organizacional sociology.
Particularly noticeable with regards to this subject is https://en.m.wikipedia.org/wiki/Federalist_No._78 where Hamilton posits that the Judicial power is the most frail one: with the executive holding the military power and the congress holding the money.
In my opinion, if we consider the simplest definition of Effective Altruism (EA), which is 'think about the most effective way to give to charity,' it differs from utilitarianism in a couple of ways:
Firstly, it's more pragmatic because it acknowledges our personal biases. Second, it's an individual practice, unlike utilitarianism, which is typically associated with broader societal or governmental actions.
I really appreciate this approach to altruism. It doesn’t force a moral standard on others. I like practicing it on a small scale, recognizing my biases. Just the act of thinking about how to best help my community is better than giving randomly. This approach also helps to see the world as a place with more love than what we often see in the news.
However, I have concerns about the more widespread version of EA, which seems to be about gaining as much money and power as possible while appearing altruistic. This, in my view, is more akin to a personality disorder, like narcissism, than to utilitarianism.
I've been using the term EA to describe this, but it might not align with what others think EA is. For me, it's a personal habit I try to maintain, but I don't let it define who I am.
This is completely false, what's your source? Also, it has never been the case. I recall that at some point around 2015, about 4/9 persons with commit access to the github repository (please note that this is a far fetched attempt to find out a metric for "controlled") were co-founders or employees at Blockstream, but right now I am not sure there are more than four or five, out of ~30.
It's not just you. Probably they added some sort of classifier at the beginning to understand whether they should send it to 3.5 or 4. In my (very opinionated, undocumented, and mostly unscientific) opinion, more complex queries generally hit the old model, with the slow chugging of tokens. For example, I just asked it to refactor a very horrible POC in python that was creeping into the 200 LoC and it did the job wonderful. The prompt was:
`Can you refactor this function to make it:
* More readable
* Split on different parts
* Easy to test
Consider the use of generators and other strategies.
(code here)`