ekelsen

6 ms·

91.
by ekelsen 8mo ago
Can you show me where in the ieee spec this is guaranteed? My understanding is the exact opposite - that it allows implementations to return any NaN value at all. It need not be any that were inputs. It may be that JavaScript relies on it a
92.
by ekelsen 8mo ago
I guess at the bit level, but not at the level of computation? Anything that relies on bit patterns of nans behaving in a certain way (like how they propagate) is in dangerous territory.
93.
by ekelsen 8mo ago
I think it commutes even when one or both inputs are NaN? The output is always NaN.
94.
by ekelsen 9mo ago
private jets already solve these problems.
95.
by ekelsen 9mo ago
But for the private jet market, the reduction would be huge. They're already paying a premium to save time. The top end will pay an even larger premium to save even more time. I agree with you that for commercial, anything other than s
96.
Databricks Introduces OfficeQA Benchmark for Agents(databricks.com)
3 points by ekelsen 10mo ago|0 comments
97.
by ekelsen 10mo ago
sometimes not nearly so pleasant for them.
98.
by ekelsen 11mo ago
Of course that's true, that doesn't mean you should.
99.
by ekelsen 11mo ago
During the French Revolution, they tried to make a right angle have 100 degrees and even recomputed all new trig tables for this new standard. It obviously did not catch on :) https://en.wikipedia.org/wiki/Gradian
100.
by ekelsen 11mo ago
Google has the revenue to cover their spending and then some.
101.
by ekelsen 11mo ago
Is V1 recalculated based on the runway length at each airport / loading of the plane / etc.? If you have a long runway and a long runout, you have a much higher V1 than a short runway with tall clearance right at the end, right?
102.
by ekelsen 1y ago
No, but being short is completely disqualifying, so being tall is certainly a component of the physical traits that make you good at basketball. If you're 5'2" , it doesn't matter what other gifts you have -- you will n
103.
by ekelsen 1y ago
and all sorts of situations where they don't work. When they do work it's great, but if they don't and you rely on them, you have nothing.
104.
by ekelsen 1y ago
1. There's still overall fewer infections from high risk HPV types in these women. 2. It needs to be confirmed in ~10 years, but it seems very likely that women given the shots that protect against all high risk HPV types will see almo
105.
by ekelsen 1y ago
leapfrog!
106.
by ekelsen 1y ago
hmmm, the performance suggests it's not a particularly good implementation of one. (performance at sizes that fit in the cache being lower than performance that exceed it...) Also the statement in the text is just false ("Radsort
107.
by ekelsen 1y ago
Wouldn't it make sense to test radix sort? You could do it in one pass with 2 bits and it would degrade gracefully as the number of bits increased. A MSB bucketing followed by LSB passes would take care of the 5% random data case with
108.
by ekelsen 1y ago
I like "between each power of 2, there are the same number of numbers." So between 1/2 and 1 there are the same number of numbers as between 1024 and 2048. If you have 1024 numbers between each power of 2, then each interval
109.
by ekelsen 1y ago
A perfect score, nice!
110.
by ekelsen 1y ago
10488! https://www.johnayliff.com/games/seedship/index.html?blmOrai...
111.
by ekelsen 1y ago
Then the right calculation is to use FLOPs not bandwidth like they did.
112.
by ekelsen 1y ago
The math on the input tokens is definitely wrong. It claims each instance (8 GPUs) can handle 1.44 million tokens/sec of input. Let's check that out. 1.44e6 tokens/sec * 37e9 bytes/token / 3.3e12 bytes/sec&#x2
113.
by ekelsen 1y ago
I think it's just units. From wikipedia: "and its value in radians per day follows by setting Earth's semi-major axis (the astronomical unit, au) to unity, k:(rad/d) = (GM)0.5·au−1.5." the value given in the paper
114.
by ekelsen 1y ago
They paid too much? The invisible hand my friend...
115.
by ekelsen 1y ago
The point of buying a primary house is to be able to do what you want with it. And to know you can live there as long as you want. Economics dictates how valuable those things are and what premium they have over renting.
116.
by ekelsen 1y ago
I looked at the softmax kernel and the cast that it does from a float* to a float4* is extremely brittle -- it's trivial to break by offsetting the input slightly. Very likely a kernel for a standard library could not employ such a tri
117.
by ekelsen 1y ago
"the reference code is in the default FP32, and given a tolerance threshold (1e-02)" that's a huge tolerance and allows them to use fp16 operations to replace the "fp32" kernel.
118.
by ekelsen 1y ago
"FP32 is less common in modern ML workloads and often less optimized on recent hardware compared to FP16 or BF16, which may partly explain why it’s easier to achieve performance gains over PyTorch with FP32 kernels." People haven&
119.
by ekelsen 1y ago
Perhaps you took my comment too literally. Try it on a real neutral network, it doesn't work.
120.
by ekelsen 1y ago
now try the same experiment in 1 billion dimensions.

More ›