boroboro4

17 ms·

241.
by boroboro4 2y ago
This neither makes modification easy, or that’s how owner of the code does modification themselves, and that’s where the difference is.
242.
by boroboro4 2y ago
The access to modify (uptrain/finetune) these weights is the same between Meta & others, unlike with Word (where Microsoft has an advantage because they have code and can recompile it). I think this is the only thing which matters
243.
by boroboro4 2y ago
There is no connection whatsoever between hardware used to train models and performance of the models if hardware permits truly large models. Google's hardware does permit it.
244.
by boroboro4 2y ago
It does, but not necessary with LLMs, see https://news.ycombinator.com/item?id=41069829 for example - this is mixing Lean with neural network to automatically proof theorems.
245.
by boroboro4 2y ago
It worth noticing that number you're quoting is for embeddings between layers. If you split your model between 5 nodes you will need to send this 32kb 5 times. Also it's per token. Meaning if you process 1K tokens it turns to be 3
246.
by boroboro4 2y ago
It supports all of current datacenter GPUs. If you want to write very efficient CUDA kernel for modern datacenter NVIDIA GPU (read H100), you need to write it with having hardware in mind (and preferably in hands, H100 and RTX 4090 behave
247.
by boroboro4 2y ago
Effectively HIV stops replication in a presence of the drug, similarly to HIV drug regimes used today. However HIV is a very nasty virus - it’s called retrovirus because it writes itself into cell’s DNA (nucleus/core). So while those c
248.
by boroboro4 2y ago
Where did you get that it’s going to perform about half? Judging by official numbers it should perform around the same.
249.
by boroboro4 2y ago
H200 1979 TFLOPS includes sparsity, real one is twice smaller. So basically it’s on par performance to H200.
250.
by boroboro4 2y ago
Do you use same dataset to train / eval the model? Was the model used for example trained on GSM8K dataset for example?
251.
by boroboro4 3y ago
No one runs Cuda, everyone runs PyTorch. Which you can run on it.
252.
by boroboro4 3y ago
Name one use case where there is a difference between latency of 200 t/s (fireworks.ai mixtral model) and 500 t/s (groq mixtral)? Not throughput and not time to first token, but latency. Groq model shines at latency, not at the ot
253.
by boroboro4 4y ago
Both of these examples show GPT not barely approximating outputs (which doesn’t exist in real worlds for these inputs) based on training set but understands algortihms and able to apply them. I don’t believe our brains are doing anything
254.
by boroboro4 4y ago
I’ve just asked gpt3 to sum two large random numbers and it gave me correct sum of them. Then I’ve defined fibanachi like sequence (f1=1, f2=1, fn=fn_1 + fn_2 + 7) and it correctly gave me the value of 10th element. It’s not just statistica
255.
by boroboro4 5y ago
I don't think it will affect opaque types implementation at all, since they exist only in compile time. As for value classes - Valhalla is a reason (Scala's) value classes weren't removed from Scala 3 (in favor of opaque type
256.
by boroboro4 5y ago
While what you're saying true I'm pretty sure it's going to create a lot of confusion - these things share a lot - all of them have all the fields final for example, and feel like "simple" "data" classes.
257.
by boroboro4 5y ago
It's sad they introduced two kind of classes: "value" and these "kinds" together with both classes and records, java becomes much more complicated IMO.