Ahh I fell for it. I was fairly convinced by the 100+ 40GB files as weights to download, but when I saw the paper was written in a word doc I remembered to check the date.
One interesting thing to note is that I could tell they used ChatGPT rather than GPT-4 to do the fake writeup. ChatGPT likes to generate paragraphs shown in their paper, GPT-4 will do full latex commands, inline citations, and even (a made up) bibtex for the citations.
I ask myself that every time I've submitted a paper. It probably doesn't help that GPA is an easy surrogate for self-worth, which smoothly transformed into publishing for a self-worth measure. And the fact that I don't have particular research I want to do (just that I want to do research), is something I'll have to consider more-- probably a sign of bad foundational motives.
To rephrase the answer, I'll give some extra context. I was a SWE for 2 years before my PhD, and I was financially stable and content. The choice to give up that salary for science was not hard, but it was significant. Now that I've reached the end of my current science road, it would feel like giving up to go be a MLE/SWE. Obviously I could do a post-doc somewhere, but that kicks the can down the road again. I'm pretty sure I'm not a narcissist... I'll ask my friends hah.
I've never cold connected to anyone. I wouldn't know how to go about doing it. My first idea is to find members of teams that cited one of my papers, but that makes me sound neurotic about papers, doesn't it. I've had the application page for OpenAI's position open for a couple weeks now, I'll give it a serious try tomorrow, and see if there's any team member that at least 2 degrees of separation from me.
It's strange, choosing to stick out a Ph.D. (and not just get a masters and be happy) really cemented the competitiveness in my mind for years. I knew from the start I'd never make it to Deep Mind, but I had figured my experience as a SWE and other internships would make me a competitive Research Engineer. Now that I'm applying for jobs and not doing as well as I expected, it's a huge emotional negative for me. I tell myself "Journey before destination" and that I really can focus on my family and enjoy my life, but it doesn't stop the hyper-rational part of myself from saying "but you could have been a self-actualized top research engineer and a family man." To answer you question of "Why fight?", it's that I cannot stop myself, despite trying.
Thank you for your perspective! I could definitely talk about all my work in extreme detail, though I have quite a hard time reaching the interview stage.
I have not put much effort into my GitHub (I only have the code for 1 of my 5 projects uploaded). I'll spend some time this week updating it, maybe that will help my woes.
For some extra context, I have been applying every ML research engineer/scientist job that would take a new grad.
I don't have any sense for what a hiring manager wants, it's especially hard with all the layoffs, I'm sure plenty of people are applying for the same research engineer positions. I certainly consider myself to be strong at coding. I've done a couple coding challenges on hackerrank that I do well on. But I have not been lucky at getting a phone interview yet. (I do have a personal website and CV)
The example here is a bit worrying for the peer review process. I am not looking forward to my "peers" reviewing my paper by putting it through LLMs and blindly copy pasting the output. I can already imagine emailing the Area Chair and saying "While reviewer 2 is detailed, the questions show a severe lack of basic understanding. We believe the contents are AI generated."
Then again, perhaps LLMs could simply be incorporated into the peer-review process, where after submitting your paper, you'd have to answer the AI's basic questions. As a reviewer, I could imagine a structured AI report for a paper being helpful in guiding discussion: "The paper compares to recent approaches X, Y, and Z. And the work is indeed novel."
I am going to write a lament about being a machine learning PhD (near) graduate. Whenever I see these very successful new PhD's, I just have an overwhelming sense of unfairness and sadness. Many of my papers were borderline rejects, delaying me months every time until there's no way for me to catch up with the likes of the author. I am attending a top 25 US school for machine learning (but not better than 25), and the magnitude of difference for the quality of PhD students is staggering. Rarely does anyone at my school graduate with 3 first author papers, let alone 5 (with so many orals!!). It consistently blows my mind how much better these top 10 schools are then the rest at producing papers: the only metric that matters. I've thought about applying to OpenAI as an engineer, which to me would be an amazing opportunity and very prestigious, but I don't want to waste my time: I only have 2 papers done (and God willing a 3rd one for CVPR in 3 weeks). Naturally the skill ratio between me and the author is at least 5:2 (probably worse, since I don't have any oral papers), but this doesn't stop me from feeling that I could contribute to OpenAI as effectively as any new PhD. I know if I were a hiring manager, I'd always go with the more proven candidate, but as the applicant I wish I could prove that my lack of publications should not be held against me. Perhaps this is a common occurrence for any engineer, hitting a ceiling for jobs in this manner, but that doesn't stop me from wishing differently.
I am at loss towards the tech-landscape right now. A year ago, everyone was talking about hot the market was and how you could easily switch jobs for a 40% raise. Now all the megacorps are doing lay-offs and here comes Nintendo with the opposite approach. Why are the business decisions between Ninendo and the tech-giants so different? Would anyone have been surprised to see "Nintendo will lay of 10% of its workers" ?
Using pixel-space or latent-space distances to measure if a generative model has simply memorized the training data is a common evaluation metric in ML literature. If the website is using the full training set of LAION 5-B used to train Stable Diffusion, then I find it extremely convincing that SD has generalized to the data distribution it was trained on, and does not simply regurgitate other people's artwork. If there is a way to find more similar images, I have yet to see it.
I think this website is evidence SD is a novel image generator, and rarely creates infringing images (at least in the way we thought of infringing before these kinds of AI).
This is a great website, but not in the way the authors intended. Based on some of the examples they explicitly provided, it is clear to me Stable Diffusion creates novel art. Here's a random example https://www.stableattribution.com/?image=a2666aee-0a1a-411b-...
I will admit this is a nice tool for verifying the creations of SD aren't pure copies, so I think it will be useful for a time. But as AI-generated images start to taint future datasets, attribution is going to be significantly more complicated.
Wow, these songs were impressively similar to my queries. I would love for a interpolation playlist feature, where I put in 2 different songs (from say, classic rock and EDM), and I could get 10-20 songs that slowly change from the start song to the end song.
It seems to me that research scientist is a post-postdoc in industry, so all these textboxes for cover letter (or "introduce yourself") are the way for a researcher to say "Here's how my extremely specific research area actually applies to your business objectives"
I am currently looking for ML scientist jobs (just about to get my PhD) and this tool is a life saver for me. I have been procrastinating cover letters for weeks, and I just wrote 10 letter tonight with this. Obviously the output is a little verbose and needs some tuning by myself, but it's significantly easier to edit down a already written cover letter than to write one from scratch.
The shape analogy doesn't really apply with modern language models. Each word gets its own context dependent high dimensional point. With everything being context dependent, simple transformations like rotations are impossible. A more accurate perception is that any concept expressible in language now has its own high dimensional representation, which can then be decoded into any other language.
This work is quite interesting, and I'm always happy to see such large improvements for memory usage and computation over existing high performant models. I imagine transformer-based models will be with the ML community for a very long time (like ResNets).
I will say that it is somewhat comical that an already existing paper titled "Hierarchical Transformers for Long Document Classification" isn't mentioned in the related work. But to be fair, the two papers are only modestly similar.
You'd be interested in learning about the origin and application of Chi-Squared tests (or Fisher's exact test). Detecting changes in population means, and making sure those changes aren't just random noise, is quite straightforward and can be done by hand.
After styleGAN-2 came out, I couldn't image what improvements could be made over it. This work is truly impressive.
The comparisons are illuminative: StyleGAN2's mapping of texture to specific pixel location looks very similar to poorly implemented video-game textures. Perhaps future GAN improvements could come from tricks used in non-AI graphic development.
They will have no free will but to accept the recommendations given to them. But seriously, most churches belong to a denomination. And practically all denominations have a specific alignment on every theological issue you can imagine. In any case, you could simply go with reformed theology beliefs and that'll work for most users.
It doesn't have to be so black and white on youtube's response. They could, for example, but a disclaimer pre-roll ad saying "Don't take medical advice from youtube." There's a difference between Brett's interview, where they discussed the scientific merits of a drug, and someone saying "Here's how much drug you need to buy at your farm store, and how to dilute it." In any case, the analysis for banning topics doesn't consider the risk of how many lives could have been saved if Ivermectin actually is pandemic ending. Some professionals speculate hundreds of thousands of deaths could have been prevented if Ivermectin was given back in November. I don't know either way, but surely the discussion is worth not being banned.
While randomized trials are certainly the gold-standard for determining if you should give a new drug to treat an arbitrary disease, that process is far too conservative during a pandemic where a huge majority of the world cannot access proper medical care. Ivermectin is already massively produced, used throughout the world, and cheap to manufacture (compared to new anti-Covid drugs). Even months ago before the benefits were known, the risks of taking the drug were very small, especially when the treatment guidelines were to do practically nothing.
Still, the many pieces of non-randomized evidence can still clearly point towards this drugs efficacy. The sudden drop in Covid cases in India as the clearest indicator to me.
I heard about Ivermectin for the first time this week and was shocked how much it seemed that I had stepped into conspiracy theory land. There seemed to be an overwhelming amount of evidence over the last year that this repurposed drug is effective at treating Covid during a variety of stages (preventative, first positive test, ICU admittance, and even long haul). But since there hadn't been a randomized trial the drug had, not only been dismissed, but heavily censored over social media. I still cannot believe how much the discussion has been silenced in the same way people discuss "covid vaxx causes magnetism". The fact the interview between Bret Weinstein and Pierre Kory was banned from youtube for non-WHO/CDC/NIH approved covid treatments is a tragedy. I truly hope in the coming years the censorship of this topic will be the catalyst for allowing free speech to prevail on social media.
It seemed like Ritesh was being recorded without his consent, so I don't think it's fair to take this interview(?) as a confirmation of election interference. It seems like he's saying google's results intentionally favor one candidate over another, which if we assume that to be true doesn't speak of Google's interference, it simply points to Google's search optimizing clicks. It's not like it's terribly difficult to find content in Google that is for either candidate or against either candidate.
These results seems quite suspicious to me. In all instances the author only received 1 email in reply. If you run these results through an A/B testing calculator, they aren't statistically significant. That's even the case for the 1% conversion compared to the shortest email 8% conversion. I like the thesis statement that short emails convert users better, but the experimental results certainly aren't strong enough to support that claim. I'd love to see an updated version of the article now that 3 years have passed and see what the conversion rate is.
I've often wondered if professors of a field (or PhD holders) are uniquely the least qualified to teach their area of expertise. A person who both cares enough about a topic and has an intuition such that they can understand their field, cannot convey their understanding of the material to students without the same care/intuition.
I don't think a university's primary purpose is teaching; I believe it's for research.
For example, a PhD prepares you for a life of research, not one of teaching. At my institution, advisors strongly prefer their PhD students are RAs and not TAs (research assistants who do no teaching, vs teaching assistants who might be responsible for an entire class).
If you look at a professor's job description, it's often 16 hours teaching duties per week, with the rest for research. Most professors I know rarely work less than 60 hours a week, so the majority of that time is on research related work.