jmalicki

14 ms·

331.
by jmalicki 6mo ago
I'm a little surprised it's a recall - is there some expectation that it should have pressure release? Can you not sell simpler products legally?
332.
by jmalicki 6mo ago
Matrix multiplication on GPUs is non-deterministic. As are things like cumsum() https://docs.pytorch.org/docs/2.11/generated/torch.use_deter... This comes down to map reduce and floating point's lack of
333.
by jmalicki 6mo ago
Because you can't order many of your own labs, and then insurance won't pay for them.
334.
by jmalicki 6mo ago
Can you educate the rest of us by explaining your reasoning?
335.
by jmalicki 6mo ago
I suggested that. You should be able to look at algorithmic code and prove it correct without worrying about running it.
336.
by jmalicki 6mo ago
> never did learn how they understand code otherwise. This particular statement interested me. In code review, I always am a stickler that "if I need to run the code in my head to prove something about it for me, I want to see that
337.
by jmalicki 6mo ago
Not at all. It's called learning computer science. Just like you can do calculus without simulation, you can understand the semantics of a computer program without running it. It might make it harder, but running it is only a didact
338.
by jmalicki 6mo ago
Not FAANG. They just need to know you can leetcode extremely fast in arbitrary languages.
339.
by jmalicki 6mo ago
Bullshit. You need to learn to leetcode in psuedocode first.
340.
by jmalicki 6mo ago
Can sketchpad do this? (relatively simple, but showing what an LLM can do with a sketch with very little prompting, full transcript of further typing included) https://jmalicki.github.io/sp500-chart/
341.
by jmalicki 6mo ago
Are there any good papers on this you would suggest/specific search terms? I am vaguely aware of some stuff, but would love to study more, I don't quite understand what this is all about (but I do see how LLMs can do attention to
342.
by jmalicki 6mo ago
That is not safe in the same way. That is way more of a worse target than the existing us-east-1.
343.
by jmalicki 6mo ago
Even that isn't a serious data center (on the level of national security concerns). Wall St is run from Mahwah, NJ.
344.
by jmalicki 6mo ago
This is literally the probability distribution ML models are trained on. https://docs.pytorch.org/docs/2.11/generated/torch.nn.CrossE... You have a relatively small dictionary of tokens, each prediction has a
345.
by jmalicki 6mo ago
TIL Desmos. Thanks for the interesting info, seems super cool!
346.
by jmalicki 6mo ago
The US government's security priority should be moving critical us-east-1 services into Cheyenne Mountain.
347.
by jmalicki 6mo ago
The giant kilometer by kilometer hyperscaler data centers aren't in cities.
348.
by jmalicki 6mo ago
The relative lack of trucks is what would identify the data center. The only other buildings like that are warehouses, which have a lot more trucks going in and out relatively speaking.
349.
by jmalicki 6mo ago
As long as they're on opposite sides so the same Shahed missile doesn't hit them both.
350.
by jmalicki 6mo ago
> except for the 0.000000001% that leaves the facility as photons down a fiber Realistically you're getting photons returned too.
351.
by jmalicki 6mo ago
Just look on satellites for giant buildings with no cars or semi trailers parked in the parking lots. I wonder if data centers will have to start doubling as automobile junkyards to conceal themselves.
352.
by jmalicki 6mo ago
> First: There's the obvious "If the company is letting me do it, I'll be wasteful." This includes not clearing/compacting the context often. Opus now has a 1M context window, and quality is good to at least 200K
353.
by jmalicki 6mo ago
Behold the softmax: https://docs.pytorch.org/docs/2.11/generated/torch.nn.CrossE...
354.
by jmalicki 6mo ago
The cross entropy loss function is softmax . They are one and the same.
355.
by jmalicki 6mo ago
The central limit theorem is an extremely powerful justification. That doesn't mean it's considered whenever it's used, but it absolutely can be strongly justified (to the degree that other error measurements are only needed
356.
by jmalicki 6mo ago
The article also does not discuss the awesomeness of LLMs at negative temperatures! https://cavendishlabs.org/blog/negative-temperature/
357.
by jmalicki 6mo ago
Which models are not trained with the log softmax as the loss function?
358.
by jmalicki 6mo ago
So human language will improve and become more precise? I'm all for it, especially if we get more emojis in speech! Why is that sadly? Humans will learn to imitate their more intelligent betters.
359.
by jmalicki 6mo ago
With LLMs you can write your code by hand drawing a diagram on a touch screen.
360.
by jmalicki 6mo ago
I would love to see more work on beam search/Viterbi decodes rather than just greedy next token output.

More ›