jorlow

Honest question when was the radio tech lora on the hn frontpage? Im slightly shocked this still creates such confusion every time

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?

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 might) was due to a wire harness a critter had chewed on. Very similar response from Volvo in my case. No accountability and very low competence. Avoid Volvo.

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.

I think this is just what a lot of websites looked like back then :-)

Does this compete at all with openAI's triton (which is sort of a higher level cuda without the vendor lock in)?

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).

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.

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...

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

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.

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)

Note llama's feed forward is a bit different too:

  self.w2(F.silu(self.w1(x)) * self.w3(x))

https://github.com/meta-llama/llama3/blob/14aab0428d3ec3a959...

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 farfetched for unrelated complications to indeed be the cause of death.

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 hospitals a lot in the last year) it wouldn't surprise me if AI, even with occasional mistakes, would still be a net improvement in overall quality of care and outcomes.

Sure, but OpenAI was not founded (or initially funded) to be an AI product company -- which is OP's point.

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.

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 Lyfgenia itself or the chemo drug, if the chemo drug is a requirement. Right?

Would love a tl;dw. How does he predict the fail will materialize?

It's a cool resource but all the examples have bitrotted to the point that I gave up trying to get them to work.

I think you're describing traditional scheduling. The point is that sometimes overheads from cache thrashing and context switches are worse than just spinning.

They definitely pay a large number of people a large amount, but they're happy to pay less when they can.

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 meaningful improvement in safety for cyclists and pedestrians and lowering the severity when there is a car crash. And it seems like it also provides an extra nudge towards people using public transit rather than driving/ubering.

Significant clipping happens there yes but more clipping happens when the "look" is applied and contrast is added.

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're starting from scratch, it's 1) easier to tweak the images so you can cut between two cameras from different manufacturers without distracting differences and 2) you can give the image more of your personality.

As a general note, I've found that in the world of "cinematography", tech terms aren't used very rigorously and there's a lot of cargo cult which comes from the benefit of one tech being conflated as a benefit of something else. It's often hard to sift through the noise when learning.

Meta has more or less had this for many years -- besides the public shaming aspect.

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 of tools like excel, but for things like server infrastructure (at least in big tech companies) they are very hesitant to put something into production without at least source code access (much preferring being able to compile it themselves).

AI (or probably more accurately, ML) is an efficiency technology. Surveillance is one thing it can make more efficient, sure.

[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 comment that it's fixed)

He had abandoned them for about a week but then talked to the CEO and that got things back on track IIRC