jorlow

9 ms·

1.
by jorlow 2mo ago
Honest question when was the radio tech lora on the hn frontpage? Im slightly shocked this still creates such confusion every time
2.
by jorlow 6mo ago
I see this comment on every single LoRA post despite the vast majority of posts being about LoRA not LoRa. Can we please stop beating this dead horse?
3.
by jorlow 1y ago
I had a horrible experience with my XC90 plug in hybrid. It took months of back and forth with the dealership and thousands of dollars to find out that it's dangerous behavior (refusing to stop without flooring the break with all my mi
4.
by jorlow 1y ago
It's basically a blob in space. When you have millions of them, you can use gradient descent to minimize loss between them and source images.
5.
by jorlow 2y ago
I think this is just what a lot of websites looked like back then :-)
6.
by jorlow 2y ago
Does this compete at all with openAI's triton (which is sort of a higher level cuda without the vendor lock in)?
7.
by jorlow 2y ago
I wonder why browsers don't show a bit of UI that it's in compatibility mode and give a way to disable (or enable for other domains for dev/testing).
8.
by jorlow 2y ago
If a browser has too many compatibility issues, users will switch away. Outreach to the sites in question takes time and is often unsuccessful. Quirks is the pragmatic answer.
9.
by jorlow 2y ago
It doesn't really add up. The schedule was always uncertain. Why not postpone? Why a big announcement only a year and a half ago? Seems pretty cruel to all the people selected to take the trip...
10.
by jorlow 2y ago
Sure, it's weaker in telephoto. (I rarely shoot beyond 85mm.) Fwiw I never use portrait mode -- too artificial. The main lens still gets pretty nice shallow DOF tho
11.
by jorlow 2y ago
I have a pixel 8 pro and a canon r5 and honestly, unless I'm shooting at f2.0 or less, I generally prefer the pixel images. Not to mention they're instantly sharable.
12.
by jorlow 2y ago
To keep it "tiny". (IIRC geohot started it because he thought pytorch and others were bloated and a simple ml framework would be inherently better)
13.
by jorlow 2y ago
Note llama's feed forward is a bit different too: self.w2(F.silu(self.w1(x)) * self.w3(x)) I.e. the nonlinearity is a gate. https://github.com/meta-llama/llama3/blob/14aab0428d3ec3a959...
14.
by jorlow 2y ago
On the other hand, you have to be pretty sick to qualify for such extremely experimental procedures. And transplants are pretty intensive operations even when someone is otherwise healthy (and getting human organs). It doesn't seem far
15.
by jorlow 2y ago
I too feel a bit worried about AI making mistakes, but the thing is that a doctor rushing or an assistant can (and do) make mistakes too. These are all just cost vs quality of service tradeoffs. From my experience (having been in and out of
16.
by jorlow 2y ago
Sure, but OpenAI was not founded (or initially funded) to be an AI product company -- which is OP's point.
17.
by jorlow 2y ago
For me it's this + going overboard on minimalism (speed only on center screen, turn signals now buttons, wiper controls on touch screen). It's a shame because there's a lot to like about the model 3.
18.
by jorlow 3y ago
The article says "patients must undergo a preparatory treatment with a chemotherapy drug to remove any native stem cells that might remain in their bone marrow." It doesn't make much difference to the patient if it's the
19.
by jorlow 3y ago
Would love a tl;dw. How does he predict the fail will materialize?
20.
by jorlow 3y ago
It's a cool resource but all the examples have bitrotted to the point that I gave up trying to get them to work.
21.
by jorlow 3y ago
I think you're describing traditional scheduling. The point is that sometimes overheads from cache thrashing and context switches are worse than just spinning.
22.
by jorlow 3y ago
They definitely pay a large number of people a large amount, but they're happy to pay less when they can.
23.
by jorlow 3y ago
I don't know either way whether 20mph makes sense in places like Wales, but at least in central London I think the shift to 20mph speed limits is great. Based on what I've seen, I'd be shocked if it isn't providing a mea
24.
by jorlow 3y ago
Significant clipping happens there yes but more clipping happens when the "look" is applied and contrast is added.
25.
by jorlow 3y ago
Log is lower contrast so it's less likely to clip (be a fully saturated color or pure white or black). And clipping inherently limits your max dynamic range. Log also means a "look" is not baked into the image so, since you&#
26.
by jorlow 3y ago
Meta has more or less had this for many years -- besides the public shaming aspect.
27.
by jorlow 3y ago
> Eterprises don't care that much about code. And if they do, business continuity clauses may alleviate most of their closed source concerns. After all, how many financial firms switched to OpenOffice Calc from Excel? This is true o
28.
by jorlow 3y ago
AI (or probably more accurately, ML) is an efficiency technology. Surveillance is one thing it can make more efficient, sure.
29.
by jorlow 3y ago
[1] links to https://github.com/RadeonOpenCompute/ROCm/issues/2198 which has all the context (driver bugs, vowing to stop using AMD, Lisa Su's response that they're committed to fixing this stuff, a
30.
by jorlow 3y ago
He had abandoned them for about a week but then talked to the CEO and that got things back on track IIRC

More ›