sarjann

5 ms·

31.
by sarjann 2mo ago
I do think some of features in their harness seem interesting (workers in separate worktrees at once), recovery from crashes seem interesting.
32.
by sarjann 2mo ago
I don't see how not? A theoretical physicist can do all the thinking they want but if they can't test an idea against nature it's not super useful.
33.
by sarjann 2mo ago
dgx spark, nvfp4 so I have spare room for KV cache (context)
34.
by sarjann 3mo ago
Sounds like this would be expensive on the input token side with some saving on the decode side. For workloads with large documents / context that could be an issue.
35.
by sarjann 3mo ago
If it only switches when cache is cold, it seems like it'd lock in one model for the whole session? What use case are you imagining where cache eviction would happen during it? It might also be the case that a task might seem easy but
36.
Ask HN: Is there legal risk in AI memory?
2 points by sarjann 3mo ago|1 comments
37.
by sarjann 3mo ago
I guess the argument might be that it'll be faster than finding a spot. Slightly minor benefit is you don't need to remember where it's parked.
38.
by sarjann 3mo ago
Unfortunately them giving up coding means they have less traces to train on.
39.
by sarjann 3mo ago
I should try to rob a bank and if I get caught just return the money. No, there needs to be a penalty above what you get, otherwise it encourages people to take the free option of bad behavior. If they get caught they go back as though noth
40.
by sarjann 3mo ago
It could also be the case that by the time business adoption picks up a lot they might not be as compute constrained. Depends on rate of growth.
41.
by sarjann 3mo ago
No, equities naturally trend up with economic growth. RAM is up because of a supply shock, as new capacity comes on prices will drop, it’s a commodity.
42.
by sarjann 3mo ago
Seems pretty useful, often hopping between different quants. I wonder if this would work for different "branches" of a model. E.g. qwen 3.6 35b regular vs abliterated.
43.
by sarjann 4mo ago
I think this is also a reason why X has gotten worse. They pay people for engagement.
44.
by sarjann 4mo ago
I believe Nvidia chips have a secure way to run your model on other infra. https://www.nvidia.com/en-us/data-center/solutions/confident...
45.
by sarjann 4mo ago
Google tends to be very good at vision and smaller/ edge
46.
by sarjann 4mo ago
Isn't that what insurance is for? So you don't need to stash away a bunch of idle capital incase something goes wrong.
47.
by sarjann 4mo ago
Text autocomplete can write code, carry out actions (tool calls) and launch cyber attacks. It very much is a matter of national security.
48.
by sarjann 4mo ago
The problem is recursive self improvement creating a very difficult gap and the fact that power, compute has a lag from when you invest and when data centers come up. You also can't just spin up a research team out of nowhere.
49.
by sarjann 4mo ago
I wonder with these stories. Why are there so many individual country efforts? We know the scale needed with scaling laws / capital / energy. Most of these countries alone can barely compete (even large groups of them would strugg
50.
by sarjann 4mo ago
Until they choose to lock their models down (e.g. Fable). Building policy off copying isn't reliable.
51.
by sarjann 4mo ago
Asking the important questions
52.
by sarjann 4mo ago
I don't think token speed matters as much when a lot of tokens are needed to achieve a task. E.g. artificial analysis benchmarks where deepseek v4 is one of the biggest token burners to go through the benchmark.
53.
by sarjann 4mo ago
My understanding is google owns a pretty decent stake in spacex (around 6% end of 2025). Booking revenue now will probably help Spacex's valuation.
54.
by sarjann 4mo ago
I mean there is a little something called compute. And other complexity that comes like writing code to efficiently distribute a model across machines.
55.
by sarjann 4mo ago
I think it could help to maybe allow 10 searches free (without logging in) just for someone to see what the UI is like or show an example of the results page. It adds a bit of friction for someone to have to make an account and pay just to
56.
by sarjann 5mo ago
There is also the ability to fallback is one of the clouds degrades in performance.
57.
by sarjann 5mo ago
I used something similar for searching over wikipedia pages https://github.com/sarjann/WikiReader-rs Was pretty speedy
58.
by sarjann 5mo ago
I think it should be tested on goals. E.g. Crack this puzzle, fix this code so these tests pass. (A human can verify it doesn't cheese things).
59.
by sarjann 6mo ago
Isn't the solution in game theory when there are multiple turns to just repeat the move the previous person made. So if the other side overpumped by x1 amount then you pump an extra x1 the next turn / year (maybe multiplied some r
60.
by sarjann 8mo ago
Is there a pure stripe fund though? What if I just want exposure to stripe and not the rest of their portfolio?

More ›