gradys

8 ms·

31.
by gradys 3y ago
We don’t know if any spies are going to get burned or if any methods will be discovered and shut down. There was intelligence on specific things individual foreign leaders said in private. Combined with other info, that could definitely lea
32.
by gradys 3y ago
The model is indeed so overpowered that it doesn’t matter in practice. See the Sentencepiece paper for some discussion of the design decisions on stuff like whitespace.
33.
by gradys 3y ago
This sounds like a long running conversation. Are there problems with extending past the context window?
34.
by gradys 3y ago
One could imagine an alternative where the API manifest was provided privately to OpenAI in the developer console or where the plugin developer had to implement an OpenAI-specific API structure. Doing it this way is more, dare I say, open
35.
by gradys 3y ago
Indeed, as I write, this is sorta what I have been working toward my whole career. And it's not like this was a giant leap from what ChatGPT already showed it was capable of. People have already been doing stuff like this with LangChai
36.
by gradys 3y ago
ChatGPT with plugins! https://twitter.com/gdb/status/1638971232443076609
37.
Post-GPT Computing(grady.io)
252 points by gradys 3y ago|481 comments
38.
by gradys 3y ago
I'm not expecting this comment to do numbers, so anyone who is reading this must be feeling as affected by this announcement as me. Is software essentially solved now? I haven't been able to do much work since the announcement cam
39.
by gradys 3y ago
https://arxiv.org/abs/1703.04730 > How can we explain the predictions of a black-box model? In this paper, we use influence functions -- a classic technique from robust statistics -- to trace a model's predicti
40.
by gradys 4y ago
It’s harder because gradient descent is usually done with near synchronous updates of the parameters across the whole model on all training hosts.
41.
by gradys 4y ago
Any recommendations for learning to draw at the level depicted in this article?
42.
by gradys 4y ago
People not wanting to work is not the main reason we’re seeing more people find housing unaffordable. It’s the extreme rise in price of housing relative to income. I make dramatically more than my parents did when they were buying their fir
43.
by gradys 4y ago
That’s the effect of the drug.
44.
by gradys 4y ago
Why would that be good?
45.
by gradys 4y ago
Meet does both of those things.
46.
by gradys 4y ago
Much of the market is up, presumably because of the good CPI report.
47.
by gradys 4y ago
More relevantly, it means that the model parameter count, dataset size, and training procedure are balanced such that they can achieve equal or better performance as GPT-3 while being smaller.
48.
by gradys 4y ago
What do you (or others) see as its advantages relative to other similar offerings like those of AWS, GCP, Deno, Netlify, etc.?
49.
by gradys 4y ago
What fraction of people working on AI in general do you think it would be appropriate to have work on AI risk?
50.
Optimality is the tiger, and agents are its teeth(lesswrong.com)
1 points by gradys 4y ago|0 comments
51.
by gradys 4y ago
You're missing that in most high priced housing markets, you can't just build an apartment building anywhere you can build a single family home.
52.
by gradys 4y ago
Roam user here. My experience is that it's not about the graph as a whole, it's about the neighborhoods. I never look at the graph view and I don't meaningfully use the global structure of the graph for anything, but I find t
53.
by gradys 5y ago
You might like the paper "Do NLP Models Know Numbers? Probing Numeracy in Embeddings": https://arxiv.org/abs/1909.07940 Neural network models seem to encode an approximate notion of quantity in their represen
54.
Graphs and Embedding Spaces(grady.io)
1 points by gradys 5y ago|0 comments
55.
by gradys 5y ago
Definitely. It's got great learning resources like Clojure for the Brave and True, and it's very practical. There is good support for using it in server-side work, web apps, both frontend and backend, and now even lightweight scri
56.
by gradys 5y ago
I hope compojis makes it into some kind of standard. Some of the combinations are just silly, but often they help express an emotion that is difficult to express concisely otherwise. Maybe ligatures could work as a technical avenue?
57.
by gradys 5y ago
I would love to hear about the customizations my fellow HN readers use.
58.
by gradys 5y ago
The most surprising thing to me was how deep mammals and even birds manage to get! The diving depths of emperor penguins, elephant seals, and narwhals are astonding.
59.
by gradys 5y ago
Great write up! Using learned representations for content in personal knowledge bases seems like a huge missing piece of tools like Roam. This appears to go all the way to the other side of the spectrum, not supporting any explicit graph li
60.
by gradys 5y ago
Both terms mean something . Both are amorphous, change over time, and cast pretty large nets, but they serve as loose coordination points for many more specific ideas. That is a useful function that isn't about hiding the underlying m

More ›