svantana

15 ms·

181.
by svantana 3y ago
Wait, are we reading the same paper? What I'm seeing is comparable accuracy to unquantized models for <4B params, and nothing reported for larger models except resource consumption.
182.
by svantana 3y ago
That's much better, but still not enough. We also need to add that the host will open a door regardless of which door I initially choose.
183.
by svantana 3y ago
According to their definition, answering "I don't know" is also a hallucination. Even worse, the truth function is deliberately designed to trip up the models, it has no connection to any real-world truth. So for example, if
184.
by svantana 3y ago
As another software dev, I would pay big money for the "worst possible computer" that exhibits all of the glitches and issues that end users see. It's so annoying to get bug reports that I can't reproduce.
185.
by svantana 3y ago
Haven't you heard of "turtles all the way down"?
186.
by svantana 3y ago
By my counting, that document is ~ 373 km^2, which is much smaller than germany. It turns out the ruler was needed after all
187.
by svantana 3y ago
Not OP but these evaluations should be taken with a huge grain of salt. It's almost impossible to rule out data leakage of open datasets such as HumanEval.
188.
by svantana 3y ago
Indeed, 44100 may seem like a random-ish number until you look at its prime factors, 44100 = (2 3 5*7)^2. In other words, it has a lot of divisors - 81 of them in total, whereas 48000 only has 64.
189.
by svantana 3y ago
A very large fraction of music on youtube is also monetized by ads (for free users).
190.
by svantana 3y ago
(2017) Unclear what has happened in the 7 years since.
191.
by svantana 3y ago
Sure, that was the point - min(max()) forces you to give explicit preference to lo or hi, whereas with clamp it's up to the std library. I trust my users to bend my software to their will, but I don't want different behavior on ma
192.
by svantana 3y ago
That must depend on the platform and the surrounding code, no?
193.
by svantana 3y ago
I'm a heavy std::clamp user, but I'm considering replacing it with min+max because of the uncertainty about what will happen when lo > hi. On windows it triggers an assertion, while other platforms just do a min+max in one or t
194.
by svantana 3y ago
Counterpoint: "that should do the trick" is good enough for most people, but not for lawyers.
195.
by svantana 3y ago
The trick is to hide the answers to the test data with an authority that only reports your score, like Kaggle does. And then only allow a single submission for each new model to avoid data leakage. I find it a bit sad that this practice has
196.
by svantana 3y ago
Right, but if the perpetrator has to be on board, they run a pretty big risk of getting caught for a serious crime. While I imagine a called-in threat can have a similar effect, with much less risk.
197.
by svantana 3y ago
There is also geothermal, where some very interesting innovations are being made by Quaise Energy among others. Also, energy storage is emerging as an electricity grid stabilizer. The cost trend of solar+wind+storage is looking very promisi
198.
by svantana 3y ago
> but since they can only be higher frequencies than the original That's incorrect, distortion products spread both upwards and downwards in frequency. The easiest example to analyze is x^2, which corresponds to the spectrum convol
199.
by svantana 3y ago
There's a recent "real coding" benchmark that all the top LLMs perform abysmally on: https://www.swebench.com/ However, it seems only a matter of time before even this challenge is overcome, and when that hap
200.
by svantana 3y ago
That's a good one, and in a similar vein there's https://paperswithcode.com/ from facebook/meta that tracks papers and their accompanying github repos, along with their benchmark results.
201.
by svantana 3y ago
The only logic I can see is pretending to cancel the release, but then suddenly having a change of heart in the next fiscal year, when they are projecting to make a loss anyway. Alternatively, the cynic in me also has a theory - that this i
202.
by svantana 3y ago
Have you tried your model on this new benchmark from Princeton NLP? https://www.swebench.com/ It's more "real-world" than the toy-like problems in HumanEval, it would be interesting to see if you can crack so
203.
by svantana 3y ago
Couldn't it be just a shell of carbon or tungsten, but the bulk of material inside be molten sand, or something else "dirt cheap"?
204.
by svantana 3y ago
This looks like it has about 100x higher power density than what solar PV gives at peak. Combine that with 100% uptime and deployment pretty much anywhere, that are some pretty compelling advantages.
205.
by svantana 3y ago
Could this be used to create a nuclear power generator without moving parts? Some radioactive material in the center, some coating to absorb the radiation, and a shell of these cells to generate electricity.
206.
by svantana 3y ago
It is hopelessly weaselly. Take for example this Winograd scheme example from wikipedia: Q: The city councilmen refused the demonstrators a permit because they advocated violence. Who advocated violence? A: Woah, hold the phone! This questi
207.
by svantana 3y ago
It seems like Apple is intentionally dragging its feet in implementing new web standards in Safari (webusb, webgpu, webmidi, bluetooth, filesystem, etc) in order to keep native apps relevant (which is a big cash cow for Apple).
208.
by svantana 3y ago
One could say that if you don't want it solved, then it's not an actual problem.
209.
by svantana 3y ago
(2017)
210.
by svantana 3y ago
I found that in my (high quality) studio monitors, the audio sounded fine and hard to distinguish from 24kHz wav. But in headphones, the artifacts were pretty obvious. So probably some reverberation will do a lot to cover up artifacts. In t

More ›