alyxya

6 ms·

61.
by alyxya 11mo ago
This seems to be a software bug and not something about model behavior, though the model is in some sense doing the wrong thing by internally evaluating what the echo command should output rather than saying what the output actually is. Edi
62.
by alyxya 11mo ago
> There is no memecoin for a67z. If you buy any "a67z" shitcoin, you'll lose all your money. I'm fully expecting someone to make this memecoin if it hasn't been made already, and for it to ironically be worth som
63.
by alyxya 11mo ago
Something about the camera perspective creates a skew that makes things feel artificial to me. It's a minor thing that bothers me, but I'd like the geometry to feel more like what I normally see. Video generation models tend to fe
64.
by alyxya 11mo ago
This seems like a good thing for him to get to fully pursue his own ideas independent of Meta. Large incumbents aren’t usually the place for innovating anything far from mainstream considering the risk and cost of failure. The high level id
65.
by alyxya 11mo ago
There's a difference between what a model is trained on and the inductive biases a model uses to generalize. It isn't as simple as saying training natively on everything. All existing models have certain things they generalize bet
66.
by alyxya 11mo ago
Personally, I think the direction AI will go towards is having an AI brain with something like a LLM at its core augmented with various abilities like spatial intelligence, rather than models being designed with spatial reasoning at its cor
67.
by alyxya 11mo ago
Maybe the solution is to counteract LLM pull requests with LLM first pass reviews. Obviously not a perfect solution, but LLMs should be good enough at detecting spam. This only needs to apply to first time contributors who haven't yet
68.
by alyxya 11mo ago
If someone wants to use keywords as a screening mechanism for resumes, they don't need a human recruiter to do that, and some automated tool is perfectly able to parse keywords. You could over-index your resume on the prospect that the
69.
by alyxya 11mo ago
There's a skills section that lists keywords. Personally, keywords mean relatively little to me, because I don't think of people's skill sets as being static, and anyone can learn anything.
70.
by alyxya 11mo ago
I tend to think resume advice is overrated. There's so much variability in how companies screen them, who reads it, what they care about, and how they get read. People tend to give advice based on their idea of what a good resume shoul
71.
by alyxya 11mo ago
To the people at the top, the job market is a statistic. They can't feel empathy on an issue they're so disconnected from, so they just think it's not their problem, or there isn't much they can do about it. Technologica
72.
by alyxya 11mo ago
Quality of the model tends to be pretty subjective, and people also complain about gaming benchmarks. At least context window length and generation speed are concrete improvements. There's always a way you can downplay how valuable or
73.
by alyxya 11mo ago
> I worry about other students being lazy and just ripping my hard work and claiming it as their own. Is the main concern that you have a class project you don't want other people taking credit for as their own? I wouldn't both
74.
by alyxya 11mo ago
The human interaction part is one of the main limitations to speed, where the more autonomous a model can be, the faster it is for me.
75.
by alyxya 11mo ago
It would be nice if there was more information provided on that page. I assume this is just the output token generation speed. Is it using speculative decoding to get to 1000 tokens/sec? Is there lossy quantization being used to speed
76.
by alyxya 11mo ago
It's a bit sad seeing how much focus there is on using courses and books to learn about compilers. > I’m not involved in any open-source projects, but they seem like a fantastic way of learning more about this field and also meeting
77.
by alyxya 11mo ago
There are a couple problems that come to mind. One is that a lot of people aren’t interested or willing to take a one sided interview with a machine. This is more true for candidates who have more options, and they’ll feel disrespected by t
78.
by alyxya 11mo ago
More generally, it's often worth learning and understanding things one step deeper. Having a more fundamental understanding of things explains more of the "why" behind why some things are the way they are, or why we do some t
79.
by alyxya 1y ago
Interviews are fundamentally really difficult to get right. On one side, you could try to create the best fairest standardized interview process based on certain metrics, but people will eventually optimize on how well they can do on the st
80.
by alyxya 1y ago
> The training budget for this model was $4,000, trained 60 hours on a cluster of 32 H100s. (For comparison, training an LLM of this scale from scratch typically costs ~$200k.) What they did is closer to fine-tuning, so this comparison i
81.
by alyxya 1y ago
This looks like one of many ideas for more efficient compute chips for machine learning. I'm waiting for the day some chip gets mass produced and works at scale for some large model and with sufficient reliability, but until then, I do
82.
by alyxya 1y ago
I tried warp last year and I wanted to like it but it just felt slower and more bloated. I don’t know if it’s improved since then, but I have a hard time seeing how the terminal is worth using. I’m ignoring price here and focused on value a
83.
by alyxya 1y ago
It’s the first release, so I expect it to get better over time. I didn’t care for ChatGPT when it was first released and thought it wouldn’t be trustworthy, but it’s much better now.
84.
by alyxya 1y ago
At a glance, Grokipedia seems quite promising to me, considering how new it is. There are plenty of external citations, so rather than relying on a model to recall information internally, it’s likely effectively just summarizing external re
85.
by alyxya 1y ago
Scaling applies to multiple dimensions simultaneously over time. A frontier model today could be replicated a year later with a model half the size, with a quarter of the FLOPS, etc. I don’t know the real numbers for optimization scaling, b
86.
by alyxya 1y ago
They can’t claim current models aren’t able to handle these problems if they didn’t use a setup similar to coding agents like Claude Code and OpenAI Codex. Using a suboptimal setup is akin to verbally telling a person the whole reasoning pr
87.
by alyxya 1y ago
The key point the paper seems to make is that existing benchmarks have relatively low complexity on reasoning complexity, so they made a new dataset DeepRD with arbitrarily large reasoning complexity and demonstrated that existing models fa
88.
by alyxya 1y ago
It actually skips the 8 in its repeating decimal. It’s better to think of 1/9^2 as the infinite sum of k * 10^-k for all positive integers k. The 8 gets skipped because you have something like ...789(10)(11)... where the 1 from the “10
89.
by alyxya 1y ago
If you set x = 0.123456..., then multiplying it by (10 - 1) gives 9x = 1.111111..., and multiplying it by (10 - 1) again gives 81x = 10, or x = 10/81. I’m not writing things formally here but that’s the rough idea, and you can do the s
90.
by alyxya 1y ago
I like to think of 0.987654... and 0.123456... as infinite series which simplify to 80/81 and 10/81, hence the ~8 ratio.

More ›