zone411

27 ms·

331.
by zone411 4y ago
"The collage output"? What?
332.
by zone411 4y ago
Doesn't follow. Why would a tool built in Japan limit itself to Japanese? ChatGPT doesn't limit itself to English.
333.
by zone411 4y ago
No, not yet.
334.
by zone411 4y ago
Very interesting! This would give Japan a real competitive productivity advantage if ML tools are banned or severely hampered elsewhere. It would also mean that other countries would need to ban not just scraping but also the resulting code
335.
by zone411 4y ago
They do good work https://ai.facebook.com/blog/ but Google/Deepmind are ahead.
336.
by zone411 4y ago
You missed the point. It's obvious what it means to anybody in ML. But that's not how the term is used.
337.
by zone411 4y ago
I agree. This paper also has issues in the areas I'm familiar with (e.g. Jukebox: being "limited to English songs" is the least of its issues). I don't know why it was randomly submitted to HN.
338.
by zone411 4y ago
Here is a chart of Mathlib's growth: https://leanprover-community.github.io/mathlib_stats.html , and here is information about what parts of undegrad math are covered so far: https://leanprover-community.gith
339.
by zone411 4y ago
The large model weights are 14B, so at 16 bits per weight, it won't quite fit on one 3090 or 4090.
340.
by zone411 4y ago
This ship has sailed. Its broad, popular use will prevail. You can fight the good fight as an expert to advocate for the precise definition, but it's pointless. Maybe the term AGI has a better chance of being used specifically.
341.
by zone411 4y ago
Wouldn't they be more cost effective simply because we already have "dumb" roads in place? You could also gradually introduce them.
342.
by zone411 4y ago
I don't like "Foundation Models" because it's a term invented by Stanford and they're pushing it hard while not really doing all that much in the field.
343.
by zone411 4y ago
If you are training a model that requires this much memory, it will also require a lot of compute, so it would be too slow and not cost-effective. It may be useful for inference.
344.
by zone411 4y ago
He had other colorful pronouncements: 1991 "Microsoft is abusing its monopoly" 1993 "Wireless computing will flop permanently" 1994 "What the Internet needs now is more competition among greedy entrepreneurs" 1
345.
by zone411 4y ago
The grandparent is probably talking about the InstructGPT paper? But I don't remember seeing a preference for longer responses in that paper.
346.
by zone411 4y ago
I hope Google/Deepmind don't become Bell Labs or Xerox PARC of the 2020s. I'm seeing very strong research papers but no commercial execution.
347.
by zone411 4y ago
That's right. I see so many people get stuck thinking that the default setting without a proper prompt or context is all these models can do.
348.
by zone411 4y ago
Here is a fun example of what it can do: https://twitter.com/jayelmnop/status/1612243602633068549 .
349.
by zone411 4y ago
Yes, I have been reading a lot of papers in the last couple of years and the corresponding code, if available, is using PyTorch much more often than TensorFlow.
350.
by zone411 4y ago
Another data point that PyTorch overtook TF: https://trends.google.com/trends/explore?date=today%205-y&q=...
351.
by zone411 4y ago
Training with half precision (or mixed precision) is common.
352.
by zone411 4y ago
There is some criticism of this paper: https://twitter.com/TomHoltzPaleo/status/1611853690654150657 https://twitter.com/antrodemon/status/1611540542269317122 https://twitter.c
353.
by zone411 4y ago
It's possible to enhance mathematical abilities of LLMs by enabling them to externally run symbolic mini programs e.g. https://arxiv.org/abs/2211.12588 . It's also possible to create fact-grounded retrieval-en
354.
4th Gen Intel Xeon Scalable Sapphire Rapids Leaps Forward(servethehome.com)
2 points by zone411 4y ago|1 comments
355.
by zone411 4y ago
This is covered in the article, and they apparently think that it could be legal in two out of 300 courtrooms. You've linked to just one specific court's webpage.
356.
by zone411 4y ago
I tried this question, but I added "MEE Question 6" at the top so it has a better context and the answer seems somewhat better? No idea if it would get any credit though, maybe the author of this blog post can chime in. > It is
357.
by zone411 4y ago
None of this will work if you run the generated text through a smaller model designed to rewrite it. I'm betting on schools using expanded remote proctoring software to monitor and record students' homework process and flag potent
358.
Fast and Furious Movie Titles by 'Claude' from Anthropic AI(twitter.com)
2 points by zone411 4y ago|0 comments
359.
by zone411 4y ago
I don't understand: many more people would use it for free in Bing/Word/Cortana than on a website that nobody used previously.
360.
by zone411 4y ago
It gets a confusing between DALL-E, DALL-E 2, Stable Diffusion 1.x and 2.x. But Imagen is using some ideas introduced by OpenAI and they also published the influential GLIDE paper. To greatly simplify, my impression is that OpenAI's bi

More ›