Y
HN Search
Hacker News Search
new
|
comments
|
top
|
jobs
Majromax
searching PlanetScale…
1.
▲
2.
▲
3.
▲
4.
▲
5.
▲
6.
▲
27 ms
·
331.
▲
by
Majromax
3y ago
> If the near-the-money options are correctly modeled, how will you make money on them? In a hand-wavy sense, they're the hedge. A mis-priced option isn't completely mis-priced: some of the modeling is correct enough. The t
332.
▲
by
Majromax
3y ago
The comments on the pull request seem to be missing some context, not yet captured by the other HN comments on this story. If PlatformIO is asking for payment, just what is that payment intended for? In particular, it appears that Platform
333.
▲
by
Majromax
3y ago
You're seeing demographics at work, rather than the upthread theory of "demand destruction." If you look specifically at the prime-age labor force participation rate (25-54 years, https://fred.stlouisfed.org/g
334.
▲
by
Majromax
3y ago
> However, heat pumps are more than 100% efficient on a watt-per-watt basis as they are not directly converting electricity into heat, but rather are moving heat from one area to another. To reactivate this dead thread: I thought of that
335.
▲
by
Majromax
3y ago
> If they ever manage to make electric water heating efficient enough to just heat on demand I'd probably use it. It's more a matter of efficiency than throughput. Mark-one googling suggests that a shower head might have a flow
336.
▲
by
Majromax
3y ago
In Canada, signing treaties is a matter for the executive. However, the terms of the treaty are not automatically incorporated into Canadian domestic law – the words of a statute win out over a treaty. Sometimes a treaty can be implemented
337.
▲
by
Majromax
3y ago
6.2V AC? That's 9V peak-to-peak, so after rectification it would be 9VDC less diode drops.
338.
▲
by
Majromax
4y ago
To join in the pedantry, I think that the better phrasing would be "to second guess as you write." The sentence uses "as" as a conjunction, and the former phrase is "it short circuits [...]". That phrase uses
339.
▲
by
Majromax
4y ago
Highest posssible in which combination, though? If you’re fine tuning a model with N layers, then you could apply LoRA to any or all of them. Maybe it’s better to concentrate effort unevenly, in which case a uniform increase of adaptation
340.
▲
by
Majromax
4y ago
Only if consumers understand the metric. Otherwise, one product's metric is another product's meaningless buzzword.
341.
▲
by
Majromax
4y ago
Ideally? Something like RMS and peak-to-peak amplitudes below 120Hz and below 1kHz, when fed with good AC power. The former would be potentially noticeable in peripheral vision or with eye movement, and the latter could affect filming, pa
342.
▲
by
Majromax
4y ago
> how did you get incandescents to last 10 years? As a general answer, dimming. Incandescent bulbs are fantastically sensitive to applied voltage; Wikipedia's article on the subject ( https://en.wikipedia.org/wiki&#
343.
▲
by
Majromax
4y ago
> This seems like an excellent business opportunity. Consumer education is rarely an excellent business opportunity. Consumers are very good at comparing prices, and "incandescent watt equivalent" labels provide an understandab
344.
▲
by
Majromax
4y ago
> I have heard that newer LED Bulbs now have some kind of built in obsolescence. This is from someone I know who is an engineer working for the US military. BigClive covers this issue a fair amount on his Youtube teardown videos. It
345.
▲
by
Majromax
4y ago
LLaMA is the large language model published by Facebook ( https://ai.facebook.com/blog/large-language-model-llama-meta... ). In theory the model is private, but the model weights were shared with researchers and quickly
346.
▲
by
Majromax
4y ago
> Premiums should be based on accident rate of the individual driver, Fortunately, accidents are rare enough that most drivers do not have a statistically-valid accident history. Unfortunately, that means that the accident rate, for the
347.
▲
by
Majromax
4y ago
> He was convinced that if we trained the AI on enough data, GPT-x would become sentient. Is there enough data? As I understand it, the latest large language models are trained on almost every piece of available text. GPT-4 is multimoda
348.
▲
by
Majromax
4y ago
> In all of the examples it seems that there is information in the original (dark spots) that are getting boosted. That was the point of the Gaussian blur. By blurring the source image of the moon before taking its picture, there was n
349.
▲
by
Majromax
4y ago
> These magnetic switches need to get down to a more reasonable pricepoint. I think that's close to an inherent problem with the bill of materials. Taking a look at DigiKey, the cheapest in-stock linear hall effect sensor that I se
350.
▲
by
Majromax
4y ago
> In what world is revenue taxed before expenses. All of them, for capital expenditures. If I buy a widget-maker for $100,000 and use it to make a profit of $50k/yr on widget sales (net of staff and materials), then I don't
351.
▲
by
Majromax
4y ago
GBTC is a form of "locked bitcoin," and it currently trades at a discount to the equivalent amount of ordinary bitcoin. If redemption were possible, the profitable arbitrage trade would be: 1. Use cash to buy Y GBTC at $X 2. Rede
352.
▲
by
Majromax
4y ago
> These language models aren't actually smart, they're just good at parroting smart sounding things they've been taught (although one could argue humans often do the same thing); the problem here is they need more sources
353.
▲
by
Majromax
4y ago
> I don't know all the photography lingo, It's called "shutter angle," see e.g. [ https://www.red.com/red-101/shutter-angle-tutorial ]. The angle, as a fraction of a full circle, is the fraction
354.
▲
by
Majromax
4y ago
> There were attempts to create tsunamis with nukes. But I’m not sure how effective this sort of thing would be against a carrier. Not very effective. Tsunami waves move huge volumes of water, but in the open ocean they are also very lo
355.
▲
by
Majromax
4y ago
Aren't NaN-based operations computationally risky? Per Anger Fog's instruction latency guides, several microarchitectures seem to process NaNs and denormals through a slow, microcoded path rather than the fast FPU/vector pat
356.
▲
by
Majromax
4y ago
> Well that installer cutoff was likely waaaaay too high. Your heat pump should be fine to 30F, maybe even 25F before needing gas One possibility might be that the unit lacks de-icing circuitry. If that were cut for cost optimization, t
357.
▲
by
Majromax
4y ago
> "On background, we packed and shipped items left on impacted employees' desks. These items were shipped to him on the assumption that they were his personal items," it read. This… isn't what "on background"
358.
▲
by
Majromax
4y ago
> Eludes me why anyone ever gives a shit about small businesses. Because today's small businesses can become tomorrow's large businesses. Each new rule is a barrier to entry. Here, for example, zero-contamination rules for se
359.
▲
by
Majromax
4y ago
> If everything simply defaults to "may contain traces" then the notice is essentially worthless right? Not exactly. A business that produces no products containing sesame will not have risk of even trace contamination, so
360.
▲
by
Majromax
4y ago
> The false positive rate on quality facial recognition is actually very low. It depends on the purpose. For identity verification purposes, when you already have independent reason to suspect that someone is specifically Person X, then
More ›