airocker

8 ms·

61.
by airocker 2y ago
Maybe containers are a better way of isolating processes as mentioned in other comments.
62.
by airocker 2y ago
Its just maybe easier way to not have to go to the root shell.
63.
by airocker 2y ago
Or Kubernetes. Thats where a standard way of authentication/authorization should be there.
64.
by airocker 2y ago
NixOS at it again :)
65.
by airocker 2y ago
I have seldom come across unix multiuser environments getting used anymore for servers. Its generally just one user on one physical machine now a days. I understand run0's promise is still useful but i would really like to see the whol
66.
by airocker 3y ago
But what if there is a bad model that fine tunes with only tools?
67.
by airocker 3y ago
Are tools fine tuned in both or just prompt based? For a fair comparison
68.
by airocker 3y ago
is this also the ticker name when they IPO?
69.
by airocker 3y ago
Thanks! Great to know. I did not know llama.cpp could do this. It should be pretty straight forward to support, not sure why they would not do it.
70.
by airocker 3y ago
Thanks! This is great to know.
71.
by airocker 3y ago
OS was possibly much more complicated to write at that time than CUDA is to write today. And competition is too strong. It might be more briefer than even Sun.
72.
by airocker 3y ago
yes, not sure you can do better than that. You cannot still have one instance of LLM in (GPU) memory answer two queries at one time.
73.
by airocker 3y ago
More than one is easy: put it behind a load balancer. Put one ollama in one container or one port.
74.
by airocker 3y ago
I heard "Nvidia for LLM today is similar to how Sun Microsystems was for the web"
75.
by airocker 3y ago
According to this book: https://valsec.barnesandnoble.com/w/slow-down-time-the-power... , it basically happens because the number of snapshots you encode per second goes down. Snapshots are visual cues that we encode.
76.
by airocker 3y ago
I think a more general way to think about it would be to add any data and reduce weight. For eg, if we want to create geography vectors, we would add all geography data to fine tune and then take a difference. Now add this to any other mode
77.
by airocker 3y ago
Even if they have the time, my rule is that nobody who does not understand the complexity of the solution (not the problem or the what) should be allowed anything more than watch rights in jira and standups. I would love to hear one non cod
78.
by airocker 3y ago
We are not talking about a good manager here. We are talking about the one who can program manage to the level of lowest jira issues.
79.
by airocker 3y ago
Why not also be the developer if you can do that much already?
80.
by airocker 3y ago
Two years may or may not be enough "long term" depending on the product and the business.
81.
by airocker 3y ago
minikube does exactly that! Runs new versions of kubernetes inside a single Docker container.
82.
by airocker 3y ago
Maybe GKE and EKS should make LTS versions.
83.
by airocker 3y ago
Our cluster has reasonable complexity and yamls work well so far. Maybe we will see what you are saying if we scale further. It allows us to maintain a much more complex cluster with fewer developers. And the software is well tested/re
84.
by airocker 3y ago
This is sorely needed. It causes service disruption for weeks because of weird upgrade schemes in GKE (or any other vendor) which works on top of Kubernetes. There are too many options with non-intuitive defaults related to how the control
85.
by airocker 3y ago
(Increased influx === Modi is bad) is just plain untrue because the state of Punjab's police was never under Modi. In fact, separatist supported government has been in power in the most recent elections. Maybe US and Canada should inve
86.
by airocker 3y ago
I can write some code and follow along with the authors. But is there a resource which would ask for questions and check/develop the learners intuition? Like given a network what do you think will happen if we change sigmoid to Relu? A
87.
Ask HN: Is there a way to learn Neural Networks/LLMs through problem solving?
1 points by airocker 3y ago|2 comments
88.
by airocker 3y ago
Inference will not work on cpu, I see 100 percent cpu frequently when doing cpu inference and we had to change to use GPUs for a large client.
89.
by airocker 3y ago
Sounds like the number of employees at OpenAI in 3 years.
90.
by airocker 3y ago
I am willing to bet the author can walk into any IIT as a professor.

More ›