kevmo314

10 ms·

121.
by kevmo314 1y ago
> Companies have the choice to amortize salaries if they want. I am curious, is there ever a time you would want this? Maybe if you’re operating at a loss?
122.
by kevmo314 1y ago
You've missed the point of the article, which in fact agrees with your anecdote. > It's equally common for developers to work in environments where little value is placed on understanding systems, but a lot of value is placed o
123.
by kevmo314 1y ago
Anything that involves sharing data with other people will run into issues around updating. If your API surface is a shipped sqlite db instead of an API call it's liable to be abused in so many ways.
124.
by kevmo314 1y ago
What's the incentive to finish the marathon? 0.05% of a 1% chance of a billion dollars?
125.
by kevmo314 1y ago
Many would argue that, indeed, said public CEO is spread too thin.
126.
by kevmo314 1y ago
I've found great benefit in voluntary code reviews. Engineers are self-aware enough that if they're at all worried about a change working they will elect for a voluntary code review. As a reviewer I also feel like my opinion is mo
127.
by kevmo314 1y ago
Yeah, go for it.
128.
by kevmo314 1y ago
Yeah, at a very high level it's similar to an actor-critic reinforcement learning algorithm. The rule text is a value function and one could build a critic model that takes as input the rule text and the main model's (the actor&#x
129.
by kevmo314 1y ago
That would be reinforcement learning. The juice is quite hard to squeeze.
130.
by kevmo314 1y ago
The people who don't think about such things probably wouldn't develop experimentally sans pytorch either.
131.
by kevmo314 1y ago
Nice work! I tried something similar a while back ago: https://github.com/kevmo314/tokie The takeaway I also found was that the running cost was really dominated by pretokenization (the regex). It's cool to see th
132.
by kevmo314 1y ago
I’m impressed with how helpful HN commenters are being despite the unilateral opinion about the pitch and readme. For what it’s worth, the title of the post is a pretty good pitch. Leaving it at “FOSS Alternative to …” would be a step in th
133.
by kevmo314 1y ago
Darn, well that's awkward, I also can't delete or edit the url :( The amusing result is on step 6 of the codelab.
134.
The 10th most common commit message is "can't you see I'm updating the time?"(codelabs.developers.google.com)
3 points by kevmo314 1y ago|2 comments
135.
by kevmo314 1y ago
> marginal value of each additional one is probably low This is the point of the paper. Specifically, that single embedding vectors are sparse enough that you can compact more data from additional vectors together to improve retrieval pe
136.
by kevmo314 1y ago
I think this is a pretty big limitation of the architecture (STT->LLM->TTS) they've chosen. The intonation around struggling to speak or difficulty with certain phrases is totally lost when the text is transcribed.
137.
by kevmo314 1y ago
There is a big gap between the other commenter's desire to write code by hand and air pollution. Look, as a programmer I also love writing code by hand, however it's childish to suggest that if I can no longer make a living doing
138.
by kevmo314 1y ago
I don't think they will but I also don't think abject squalor is the way to describe artisanal work. On top of that, I agree that capitalists continue to eliminate jobs but I also think they create jobs in other sectors. The reaso
139.
by kevmo314 1y ago
This article taught be about the QOTD protocol: https://datatracker.ietf.org/doc/html/rfc865 Cool artifact of the internet!
140.
by kevmo314 1y ago
> let those of us who wish to try anything other than boiling alive go do that Ok? Who is stopping you? Nobody here is prohibiting you from continuing to write code by hand and doing whatever you wish. Certainly if you're going to a
141.
by kevmo314 1y ago
That's fair but not a convincing argument if you're seeking to show that the pot is boiling. I don't believe it is and writing off the rebuttal is akin to "well it's boiling, just trust me".
142.
by kevmo314 1y ago
Yeah that's totally possible but I suspect if AI really does take over software, the number of people who will riot over software quality is going to be dwarfed by the number of people who are happier that they can get their chats GPT&
143.
by kevmo314 1y ago
> are often erroneously cited as analogous to efficiency gains Sure, I can believe that, but... > doing anything of value for their own self-preservation Even your own comment relies on some metric of value. I agree that profit motive
144.
by kevmo314 1y ago
> Actually, what are we going to do when everyone that cares about the craft of software ages out, burns out, or escapes the industry because of the ownership class setting unrealistic expectations on people? Nothing, I guess? There&#x27
145.
by kevmo314 1y ago
The same could've been said about Slack. It's easy to find reasons things failed.
146.
by kevmo314 1y ago
It's the equation in the diagram: π = 3 + 2/3 (1/4 − 1/20 + 1/56 - ...)
147.
Solving Character Prefix Conditioning with Coverage Beam Search(ninetyfive.gg)
2 points by kevmo314 1y ago|0 comments
148.
by kevmo314 1y ago
> I think learning CUDA for me is an endeavor of pain and going through "compute-sanitizer" and Nsight because you will find that most of your time will go into debugging why things is running slower than you think. This is so
149.
by kevmo314 1y ago
Unfortunately the choice of tokenizer is baked into the model. If you want to innovate on the tokenizer, you have to train a whole base model yourself to prove it's better which makes the barrier for innovation pretty high.
150.
by kevmo314 1y ago
This is an underrated reason why startups can be successful. Many large and even mid-sized companies' employees are encumbered in complexity indirectly because they need to engineer something to justify their jobs. I think a lot of dev

More ›