joefkelley

14 ms·

91.
by joefkelley 10y ago
Do you happen to have a source for those stats? Not that I'm doubting you - I often make similar claims - but it would be nice to be able to back it up.
92.
by joefkelley 10y ago
Yes, this is very frequently done. Training is far more computationally expensive than evaluation in most cases. I don't know of any specific cases of this being down with a raspberry pi, but many phone apps, for example, have this sor
93.
by joefkelley 10y ago
Here's Knuth's opinion on 0^0: http://arxiv.org/pdf/math/9205211v1.pdf (on page 6) He makes a few pretty convincing points about why it's useful to define it equal to 1, while admitting that in some
94.
by joefkelley 10y ago
The one thing I haven't seen discussed in the serverless world is state as optimization. For instance, if I have a machine learning application that has to access a very large trained model, I would like to load that model into memory
95.
by joefkelley 10y ago
Yes, it's still very much illegal federally. It's still a Schedule I controlled substance. The federal banking system / IRS refuses to interact with marijuana businesses, the postal service won't mail ads for it, the DEA
96.
by joefkelley 10y ago
Awesome, thanks for clarifying. So does the training optimize some property of the "semantic" layer immediately before the final emoji prediction layer? Or does it just optimize accuracy of emoji prediction directly? And then the
97.
by joefkelley 10y ago
Why not train the RNN to directly predict emojis, instead of projecting everything to semantic space and picking the closest emoji? Seems like that would help with the problem of emojis with multiple meanings in different contexts. With thi
98.
by joefkelley 10y ago
I remember in high school having no shortage of teachers who recommended not guessing if you didn't know, to avoid the wrong answer penalty. I tried to argue with them initially, then settled for simply trying to re-educate my friends.
99.
by joefkelley 10y ago
Both components have been mentioned repeatedly, but I haven't seen the combination commented yet: Hugo on S3 https://github.com/jfkelley/blog
100.
by joefkelley 10y ago
The Chinese regulatory system had been preventing any actual integration of the two companies for some time. They finally agreed to allow it to proceed in late 2015, but sales and marketing have to remain separate for another 2 years.
101.
by joefkelley 10y ago
I would love to see an analytics database system built on this idea. Let me define my "incoming" tables, and then define how other aggregated or transformed materialized views should be computed, and automatically update the downs
102.
by joefkelley 10y ago
Off the top of my head, it would at best severely weaken the strength of the hash function. Instead of having to brute force to find an output that matches exactly, you would only have to brute force for one that was sufficiently close, the
103.
by joefkelley 11y ago
Cool! Yeah I poked around the druid site and didn't find anything originally, but this looks pretty promising. It's hard to tell how full-featured it is without getting hands-on, but I see something like this making Druid much mor
104.
by joefkelley 11y ago
If one of these databases included a really good SparkSQL connector (including great predicate push-down), I could see that being hugely beneficial to growth. Part of why people are using these systems is because you can connect in many dif
105.
by joefkelley 11y ago
Interesting analysis. I did some similar research a while ago around sparse matrix multiplication and came to a similar conclusion. The memory access pattern offered by CSR (and CSC) is startlingly effective.
106.
by joefkelley 11y ago
This seems like more of an antipattern to me. If you have to be thinking about linearization order, your hierarchy is probably too complex.
107.
by joefkelley 11y ago
This is really cool, but I'm not clear on how to actually install it. Or is it only available to a limited set of users? I see there's a "keybase fuse" command that might be related, but there's no docs for how to u
108.
by joefkelley 11y ago
As someone who considers himself very recently an adult, the quick and dirty and only-half-joking rule of thumb among my peers is whether someone is paying for their own cell phone plan.
109.
by joefkelley 11y ago
I've taken "non-3d" panoramas and viewed them in Cardboard, and that was cool enough. It is certainly the best way that currently exists to share the experience of being somewhere with someone else. I can't wait to try t
110.
by joefkelley 11y ago
I kind of don't understand this post. Surely Twitch is an example of the 1%, not the 99%? Not that I disagree with the end sentiment - I'm currently at a startup that was acquired as part of the 99%, but I don't think this is
111.
by joefkelley 11y ago
I wonder if there is an API to train this on a custom dataset. They have a few built-in models, but those seem pretty limiting in terms of uses cases. I can think of a lot more where you would want to be able to train on your own specialize
112.
by joefkelley 11y ago
The last two points are extremely debatable. I have used Google, Spotify, and Pandora, and in my experience, Google's hands-off auto-generated playlists are extremely high quality. They also have great radio/curated playlists via
113.
by joefkelley 11y ago
Don't have much experience with Presto, but I have used Hive to query Elastic Search. It works very very well for full-table-scan analytics-type queries. I expect PrestoDB would be similar. But when it comes to queries about smaller an
114.
by joefkelley 11y ago
One CPU cycle per row saves 27 seconds of one CPU's time. That 10 seconds was saved on every CPU in the cluster. So if there were 50 CPUs, that's 500 seconds, or ~20 cycles per record, by the original calculation. In reality, the
115.
by joefkelley 11y ago
No. You can mirror a Chrome tab; that's the best you can do. And it doesn't work as smoothly as a purpose-built app.
116.
by joefkelley 11y ago
You're getting closer. Regarding your electron-emitting-photons example: I would describe the photons being emitted as themselves in a quantum superposition of having been emitted/not emitted/emitted at different times, etc.
117.
by joefkelley 11y ago
I think the trouble arises when you try to connect that to consciousness and human experience. From my perspective ("me" being the version in me in the "current" universe or whatever), I only experience one timeline, and
118.
by joefkelley 11y ago
My understanding is that the gravity detectors would not detect 10% or 90% of an electron. One would detect 0% and one would detect 100%. You cannot measure the wave function directly. Attempting to do so collapses it. When that waveform co
119.
by joefkelley 11y ago
I also noticed this, and it seems like a pretty big flaw. It analyzed one of my emails as very confident because it saw the words "clear" and "sure", but in context, they were used in the phrases "None of this is cl
120.
by joefkelley 11y ago
Also check out Lytro, the company created based on this technology. They've got some pretty impressive demos on their website: https://pictures.lytro.com/

More ›