sanxiyn

9 ms·

151.
by sanxiyn 2y ago
Fast inference can substitute for larger models in some circumstances. As you said, you can run multiple times. DeepMind had a detailed look, see https://arxiv.org/abs/2408.03314 .
152.
by sanxiyn 2y ago
I am not sure why not. From the article: "In the weeks to come we will be adding larger models such as Llama3-405B".
153.
by sanxiyn 2y ago
Doubtful. SRAM is not cheap, and this is entirely about SRAM vs HBM.
154.
by sanxiyn 2y ago
I mean, leptons are definitely real, not just concepts. An electron is a lepton, and an electron surely exists? You can spray individual electrons. If an electron is just a concept, it is in a sense that a tiger is just a concept, and in a
155.
by sanxiyn 2y ago
I didn't mean to suggest it as a competition to the method presented (LSR: latent space readout). It is old after all. LSR's use in evaluator LLM and working with small samples (because it works with linear direction) does seem no
156.
by sanxiyn 2y ago
I believe the paper to be cited is "The Internal State of an LLM Knows When It's Lying", published last year: https://arxiv.org/abs/2304.13734
157.
by sanxiyn 2y ago
You can have a CI builder using direct-minimal-versions to check this.
158.
by sanxiyn 2y ago
You can use #[non_exhaustive] if you want to avoid this.
159.
by sanxiyn 2y ago
Try property testing. Property tests written by myself routinely discover bugs in my codes.
160.
by sanxiyn 2y ago
This seems to be like kakasi for Japanese in that Japanese writing system does not separate words by spaces and one kanji can be read in multiple ways. As I understand the same is also true for cuneiform and this is an attempt to solve it.
161.
Can we improve the energy efficiency of EUV lithography?(arxiv.org)
2 points by sanxiyn 2y ago|0 comments
162.
by sanxiyn 2y ago
Re: NVIDIA. I wholeheartedly agree. Google/TPU is an existence proof that it is entirely possible and rational to do so. My surprise was that everyone except Google missed.
163.
by sanxiyn 2y ago
I think because he wants to "return to hands-on technical work", but that was difficult at OpenAI. As I understand, he is now managing alignment research at OpenAI. He is clarifying he can continue to do that, but he would like to
164.
by sanxiyn 2y ago
Better source for John Schulman directly from his account: https://x.com/johnschulman2/status/1820610863499509855 My read is "return to hands-on technical work" is more important than "AI alignment&
165.
by sanxiyn 2y ago
Note that unsafe blocks don't have limited blast radius. Blast that can be caused by a single incorrect unsafe block is unlimited, at least in theory. (In practice there could be correlation of amount of incorrectness to effect, but sa
166.
by sanxiyn 2y ago
> How do we even calculate the subsidy paid to fossil fuel companies by letting them externalize the cost of their mess onto the planet? I know this was rhetorical, but the standard method in the literature is Optimal Taxes on Fossil Fue
167.
by sanxiyn 2y ago
> The associated charging, inverter, and physical structure components aren't going to follow the same downward curve. I agree not the same downward curve, but it also has been on the downward curve, although different. Learning rat
168.
by sanxiyn 2y ago
AI is safe if it does not cause extinction of humanity. Then it is self-evident why it is important. The article does link to "Statement on AI Risk", at https://www.safe.ai/work/statement-on-ai-risk It is ver
169.
Transcendence: Generative Models Can Outperform the Experts That Train Them(arxiv.org)
4 points by sanxiyn 2y ago|0 comments
170.
by sanxiyn 2y ago
It works quite well as long as you don't care about battery.
171.
Instagram's Made with AI Tag Is Inaccurate, Misleading, and Needs to Go(petapixel.com)
3 points by sanxiyn 2y ago|0 comments
172.
by sanxiyn 2y ago
People often correct bad generation and I think such feedback is valuable. eg "Training Language Models with Language Feedback at Scale" https://arxiv.org/abs/2303.16755
173.
by sanxiyn 2y ago
Pharmaceutical companies know this by heart. What you are doing is learning about your drug candidates. Learning is not normally distributed.
174.
by sanxiyn 2y ago
US fought a war with China in Korea and they were evenly matched. Korean War is the best historical reference we have.
175.
by sanxiyn 2y ago
People are so gullible. Sam Altman deserves zero benefit of doubt. His words should be ignored, his words do not prove anything whatsoever.
176.
by sanxiyn 2y ago
Someone should do this for Llama 3.
177.
by sanxiyn 2y ago
I think the board should fire Sam Altman again. If the board does not fire Sam Altman it is not doing its duty, so someone should sue the board, or convince attorney general to sue. Attorney general of Delaware is Kathy Jennings <attorne
178.
by sanxiyn 2y ago
And all safety people are leaving, due to Sam Altman? And Sam Altman forced them to sign non-disparagement? And Sam Altman is lying about not clawing back equity, even if we know Daniel Kokotajlo gave up equity and has not made whole yet? T
179.
by sanxiyn 2y ago
Sam Altman is a grifter. We all should boycott Sam Altman. Repeat after me: Sam Altman is powerless if we all ignore him. Sam Altman is nothing without his people. He can't do anything himself. https://x.com/OpenAI&#x2F
180.
by sanxiyn 2y ago
If our standard is cool math, crypto is also full of cool math. (Have you read Vitalik's explanation of Quadratic Arithmetic Programs?) Our standard can't be that low. https://medium.com/@VitalikButerin/quadra

More ›