psyklic

14 ms·

121.
by psyklic 3y ago
FWIW something similar happened to me as well, dropped out of school to go full-time as a co-founder. We put in writing several times that we held similar numbers of shares, yet nine months in discovered my business partner awarded himself
122.
by psyklic 3y ago
Indeed, the OpenAI paper is more well-founded since the "ground truth" was not generated by GPT-4. However, both papers rely on black boxes instead of well-understood procedures. This places the papers on weaker scientific footing
123.
by psyklic 3y ago
Even research from OpenAI has attempted to use GPT-4 as quasi-ground truth (as a replacement for human evaluators). For example, their method in the recent paper "Language models can explain neurons in language models" [1] is: 1.
124.
by psyklic 3y ago
In the early stages, the main skill is just figuring out how to get whatever is needed done. You'll probably be wearing many hats, since there are only two of you. If the expectation is you'll build the product, you should have co
125.
by psyklic 3y ago
#3 and #4 say there exists a K for which the interior statement is true. K=2 doesn't make the interior true for either, so choosing it doesn't tell us whether #3 and #4 are true or false. However, if we choose K=1 then we prove #3
126.
by psyklic 3y ago
The space must be non-linear, as a consequence of the non-linear activation functions. A neural net is just a big math equation, and deeply embedded throughout are non-linear transforms which necessarily make the entire transform non-linear
127.
by psyklic 3y ago
This happened to me too. After I left our company, my business partner told new investors there was never a CTO. However, for nearly a year we had jointly agreed I was CTO -- and he routinely put it in writing. When the new investors found
128.
by psyklic 3y ago
Interesting analysis! Minor typo throughout: It's R^2 instead of R_2. Also surprised the dataset doesn't include Awair's products!
129.
by psyklic 3y ago
I don't see how arbitrary questions like this substantially show AGI. If there is a common solution, it could simply look up the solution. Also, AGI could be present just not in this very niche problem (that 99.9% of humans can't
130.
by psyklic 3y ago
What is the answer supposed to be? Doesn't seem like a simple IQ question to me. print(f'{0x110c8531d0952d8:066b}') EDIT: A browser extension hid most of the number from my view, so this answer is incorrect.
131.
by psyklic 3y ago
You can interleave the orderings. If there are two episodes (1 and 2), we only have to watch three episodes in a row: 121. Then, we've watched all permutations -- 12 and 21. The "14!" solution (where we watch 14x14! episodes)
132.
by psyklic 3y ago
Typically, we think of overfitting and underfitting as exclusive properties. IMO a large problem here is that the author's definition of overfitting is not inconsistent with underfitting. (Underfit indicates a poor fit on both the trai
133.
by psyklic 3y ago
It sounds like you possibly aren't that interested in the field and have some doubts about the PI. What about suggesting a small research collaboration first, like a summer project or paid as a summer research assistant? It is very com
134.
by psyklic 3y ago
Connections Museum in Seattle, WA (telecom focus) https://www.telcomhistory.org/connections-museum-seattle/
135.
by psyklic 3y ago
In independent consulting, you ideally need potential clients reaching out to you to achieve high rates/flexibility/exciting work/etc. It is possible to find these cold calling, but it will be more competitive and the work yo
136.
by psyklic 3y ago
The neural net gives a very noisy approximation of some ideal structure. So you're looking for what that ideal structure is, which very possibly does not require a neural network to represent.
137.
by psyklic 3y ago
I disagree that enjoying C automatically means you will like Rust. If you are obsessed with safety, Rust will be a nice solution. But Rust is a much more complex language than C, which will turn many C programmers away.
138.
by psyklic 3y ago
Another related one: https://kipp.ly/blog/transformer-inference-arithmetic/
139.
by psyklic 3y ago
Even with "just coding", a skilled coder can translate thoughts directly to code more precisely and faster than translating them first to imperfect English.
140.
by psyklic 3y ago
Note in CA, there are additional reporting requirements. This is not tax advice, but I believe passthrough single-member LLCs (including no-income) must file Form 568 by March 15 and pay a mandatory $800 annual fee by April 15. (Other types
141.
by psyklic 3y ago
https://handmadecities.com/
142.
by psyklic 3y ago
Jonathan Blow just tweeted: "Something to understand, about all the excitement around AI and how fast things seem to be moving... This is what computers in general used to feel like, across the board. New cool stuff was happening all t
143.
by psyklic 3y ago
Karpathy discusses viewing GPT as a finite-state Markov chain in this Twitter thread (and especially in the intro to the linked Colab notebook) : https://twitter.com/karpathy/status/1645115622517542913
144.
by psyklic 4y ago
In Python, variable names do not directly contain their values. Instead, in your memory (on "the heap") there are "boxes". You can visualize them like cardboard boxes that contain values. In Python, these boxes are calle
145.
by psyklic 4y ago
In recommendation systems in general, it's well known that recommending things based on recent activity far outperforms other methods (e.g. product category, past liked items, subscribed items, searched items). For example, Hulu showed
146.
by psyklic 4y ago
GPT-4 doesn't seem anywhere close to writing or intelligently modifying medium/large programs. This is what most professional programmers do. LeCun at least believes this is a fundamental limitation of AR-LLMs that can't be o
147.
by psyklic 4y ago
> I wonder how we can accelerate contributions to OSS projects. LLMs are impressive, but hacking on pre-existing complex codebases seems to be a current weakness.
148.
by psyklic 4y ago
I built my own CPU cluster with mATX AM3s connected via a switch, powerful processor/RAM and everything else was as cheap as possible. At the time, it easily saved money compared to AWS. The biggest issue I had was overheating. The sma
149.
by psyklic 4y ago
My initial contributions sometimes stem from noticing a bug while trying a library's examples. I often just fix the bug and submit a PR (e.g. for major memory leaks, crashes, and Windows incompatibilities). This can be useful to the pr
150.
by psyklic 4y ago
This is one of my favorite distance metrics* to show people! For example, perhaps one person likes Reddit and HN, while someone else likes HN and SO. Then their Jaccard Index would be 1/3, since they have one thing in common out of thr

More ›