spott

22 ms·

121.
by spott 2y ago
There is also this one: https://github.com/bpg/terraform-provider-proxmox
122.
by spott 2y ago
To me, the Alzheimer’s drug story is an argument for right to try laws. If the drug is available to the desperate, then the push by the public to shortcut the process goes away and the process can be done correctly. Those who are despera
123.
by spott 2y ago
They address that in the PEP. It would basically mean a new major version every year as they remove things that have been depreciated. Which isn’t really that useful.
124.
by spott 2y ago
> DataChain has no assumptions about metadata format. Could your metadata come from something like a Postgres sql statement? Or an iceberg view?
125.
by spott 2y ago
Tap a blue dot.
126.
by spott 2y ago
BPE is complete. Every valid Unicode string can be encoded with any BPE tokenizer. BPE basically starts with a token for every valid value for a Unicode byte and then creates new tokens by looking at common pairs of bytes (‘t’ followed by
127.
by spott 2y ago
The flip side is that when you are small, you probably don’t need all the fancy managed services that AWS offers. Simpler solutions can save you money and time.
128.
by spott 2y ago
How is it annoying?
129.
by spott 2y ago
Here “fast internet” is 500Mbps, and the reason is that quic seems to be cpu bound above that. I didn’t look closely enough to see what their test system was to see if this is basic consumer systems or is still a problem for high performan
130.
by spott 2y ago
Keyword there is “certain”.
131.
by spott 2y ago
The recommendation for periodic password change requirements is changing from SHOULD NOT to SHALL NOT, which hopefully will have an effect on things moving forward.
132.
by spott 2y ago
They did. > Elecrow's CEO has responded directly to Geerling and offered compensation to Geerling for the issue
133.
by spott 2y ago
They aren’t using the prompt caching on the query side, only on the embedding side… so you cache the document in the context window when ingesting it, but not during retrieval.
134.
by spott 2y ago
Yea, that is moving goalposts. A causal study in this case is basically impossible. So at the moment, we have correlation studies compared to your anecdote….
135.
by spott 2y ago
Is it? The same inputs have the same outputs, the only thing different is the performance. It is basically the same as memorization.
136.
by spott 2y ago
This is less complicated than you think. Civil engineering rules, safety margins and procedures have been established through the years as people died from their absence. The practice of civil engineering is arguably millennia old. Softwar
137.
by spott 2y ago
I’m curious: what is so clunky about the built in screenshot app? I don’t doubt you, I’ve just never had that experience, and I’m curious what I’m missing.
138.
by spott 2y ago
Instructor - great Litellm - just a wrapper around llm libs (OpenAI/claude/etc) Lmql - might not be necessary anymore with all the json modes out there. Just off the top of my head.
139.
by spott 2y ago
In waterpolo, you probably want the reflections from the bottom when treading water because it makes it easier to stay buoyant.
140.
by spott 2y ago
It sounds like a lot of the value here was in a more appropriate level of abstraction, not in the speed of the compute itself. Ray allowed them to optimize elements that spark didn’t, and that was what improved performance, not that spark i
141.
by spott 2y ago
Where are you going to put this box?
142.
by spott 2y ago
Does XProtect use kernel extensions? It also isn’t a separate product, it is bundled with the OS.
143.
by spott 2y ago
I don’t understand this world well enough to describe it well, but my understanding is that the security features in the kernel are different in purpose from the security features of something like an antivirus.
144.
by spott 2y ago
Wasn’t the whole regulatory argument that Microsoft was using kernel mode in their security software, while trying to relegate third party security software to user land? In that case, regulators stepped in and made Microsoft open up kern
145.
by spott 2y ago
Depends if you can fit the whole model into vram or not. If you can’t then you need some sort of gpu parallelism, and you need some sort of communication between the different gpus. But maybe that messaging is small enough that it doesn’t
146.
by spott 2y ago
Simply scaling to larger models, but retaining naive next token prediction, won’t be enough to get to reliable reasoning on truly novel tasks. Any gains will be reasoning by analogy to a larger number of reasoning examples. But I don’t thi
147.
by spott 2y ago
I expect that the AMD also looses out when multigpu starts to be required for it (which is arguably going to be for much larger models than for the h100, but a 70B parameter model with bf16 training is going to hit multigpu in terms of memo
148.
by spott 2y ago
Good to know. I went to recheck, because I swore I didn't see anything about that when I looked earlier, but they now say 3.2Tbps infiniband... not sure if they changed it or I was just blind. Thanks!
149.
by spott 2y ago
https://gpulist.ai No idea about how accurate that is, but if you want cluster pricing...
150.
by spott 2y ago
Only 200Gbps network per node. The infiniband stuff is 400Gbps per GPU (3.2Tbps per node).

More ›