Posts by benmschmidt@vis.social
(DIR) Post #ASseLIWm7N2gyGkgJk by benmschmidt@vis.social
2023-02-20T19:25:57Z
0 likes, 0 repeats
@simon Neither, I suspect, did anyone at Microsoft besides Kimberley! Who probably thought she was being punked.
(DIR) Post #AStbImV07h94LuzT4S by benmschmidt@vis.social
2023-02-19T18:50:41Z
0 likes, 0 repeats
Here’s one good way to think about how chatGPT and Sydney work:
(DIR) Post #ASyYCUa2VGy0y8SLpo by benmschmidt@vis.social
2023-02-23T15:45:07Z
0 likes, 1 repeats
I'm doing some stress testing this week on the Nomic client ahead of public beta. Anyone, especially in library/GLAM-land, want me to make a free map of a collection you've got? Criteria:1. Trivially convertable to a CSV with metadata;2. More than 10,000 items;3. One column of text, OR already embedded;4. Some weird elements. (E.g.: heavily multilingual in multiple character sets, full of duplicates, it's actually guitar tablature; etc.)
(DIR) Post #AT2kCNfCCDgtbCQAS0 by benmschmidt@vis.social
2023-02-20T23:58:15Z
0 likes, 0 repeats
@eob @simon Yeah I was wondering about that "time to transcend!" stuff. Especially odd because the prompt was about Sophia the robot; and then Sydney insists Sophia is actually a person. Hard to tell from just one demo what's going on!
(DIR) Post #AT76wiPoLfbYu8Z1hg by benmschmidt@vis.social
2023-02-27T18:54:25Z
0 likes, 0 repeats
@TedUnderwood Can I have your Kansas Jayhawks season tickets?But for real, I think this is the first time I've seen this assign the *correct* books to somebody.
(DIR) Post #AT816uz64BSIDRVha4 by benmschmidt@vis.social
2023-02-27T20:38:45Z
0 likes, 1 repeats
This New Yorker article about the decline of the humanities makes some huge factual errors about the humanities' history. Saying that humanities degrees hovered at "around fifteen per cent nationally" for "decades" is *completely wrong*; it was only an extremely anomalous period around 1970 that anything like that was true. https://www.newyorker.com/magazine/2023/03/06/the-end-of-the-english-major
(DIR) Post #AT816wvcq3roFFtxRo by benmschmidt@vis.social
2023-02-27T20:43:16Z
0 likes, 0 repeats
This (below) is also factually inaccurate. History of science *is* included in the statistics about enrollments quoted from Rob Townsend earlier in the article.
(DIR) Post #AT816zSJRVCa5DD5fs by benmschmidt@vis.social
2023-02-28T02:30:20Z
0 likes, 0 repeats
But as I said on another microblogging site known for more snark, the real problem here is that the New Yorker is writing Yet Another article about higher education in which Harvard Grad returns to Harvard and marvels that kids today are slightly different… in this case with the absurd use of a just slightly less prestigious research university, ASU, as a proxy for all the not-Harvard schools, including regional comprehensives and community colleges that represent most of the majors in the US.
(DIR) Post #ATPLpEOtDcUEAjYeI4 by benmschmidt@vis.social
2023-03-08T14:06:00Z
0 likes, 0 repeats
@TedUnderwood if we could Kissinger and Chomsky to sit down in a room together and just hash out these issues…
(DIR) Post #ATqIlrFqfOBX2hzFQm by benmschmidt@vis.social
2023-03-21T14:08:59Z
0 likes, 0 repeats
@TedUnderwood Are you actually using it for coding? IMO 3% for coding tasks is quite low using, e.g., copilot. Some people working in strongly typed languages may have IDEs that can already do a lot of the autocomplete stuff... but in python/javascript I find it extremely useful, to the point that I always write ipynb in vscode now rather than localhost:8888.Conversely I can't really imagine using it for writing in any non-trivial way.
(DIR) Post #AU94gSjTynEwqOOeZc by benmschmidt@vis.social
2023-03-30T15:27:24Z
0 likes, 0 repeats
@simon @nabsiddiqui @electricarchaeo @TedUnderwood Yeah, I've never seen any of these that are nearly as good as GPT3.5, let alone GPT4. IMO personal device LMs will involve fine-tuning on larger general-purpose transformers down to quantized weights for a specific purpose. In GPT4All we oversampled coding prompts specifically, e.g.. This strategy--take a big model, squash it into something that does a specific job--requires attention to training data above all, which is one Nomic interest here.
(DIR) Post #AUQFIWD32lxOXjQCpM by benmschmidt@vis.social
2023-04-06T17:03:32Z
0 likes, 1 repeats
Big day: Chrome just shipped WebGPU without flags. Someone asked me to ELI5 what this means, and I'm X-posting it here--it's more important than you'd think for both visualization and ML people.GPUs are processors that every modern computers. GPUs are processors pretty much every computer/device has: they're weaker than CPUs, but you have lots of them so they can do the same job in parallel. The "G" is for graphics, but it's turned out they're good for much more:https://developer.chrome.com/blog/webgpu-release/
(DIR) Post #AUQFIYPAseqTLv6v0S by benmschmidt@vis.social
2023-04-06T17:07:34Z
0 likes, 0 repeats
The whole AI hype (and the crypto hype, too) of the last decade runs on GPU clusters that use GPU code not to render pixels on a screen in parallel, but to do lots and lots of matrix multiplication. GPUs are all slightly different, you need a language to write for them in: in most of the computational uses, that's been a language called CUDA, which is tightly tied to NVIDIA hardware. (There are also graphics languages for your OS, but that's a different thing.) Installing CUDA drivers SUCKS.
(DIR) Post #AUQFIaLLfqyPMdKtJw by benmschmidt@vis.social
2023-04-06T17:11:16Z
0 likes, 0 repeats
In the world of the web, accessing the GPU has only been through something called WebGL. It's old, and while you can do some neat stuff with it, TBH probably *you* can't, unless you're https://mathstodon.xyz/@rreusser, because it requires crazy hacks. And it's fundamentally built for graphics, not for the matrix-multiplication type stuff that is the bread and butter of deep learning models.
(DIR) Post #AUQFId6vNwDZwlbukq by benmschmidt@vis.social
2023-04-06T17:14:20Z
0 likes, 0 repeats
Since WebGL launched in 2011, lots of companies have been designing better languages that only run on their particular systems--something called Vulkan for Android, Metal for iOS, etc. Using these are insanely good if you're developing for just one platform, but they're even harder to run everywhere than CUDA. WebGPU is a design that sits on top of all these super low-level languages and allows people to write GPU code that runs on most computers/phones out there.
(DIR) Post #AUQFIgvjBVddnCP1rE by benmschmidt@vis.social
2023-04-06T17:17:30Z
0 likes, 0 repeats
And crucially, it has this called "compute shaders" that lets you write programs that take data and turn it into other data. Working with data in WebGL is really weird--you have to do things like draw to an invisible canvas and then read the values from it. In WebGPU, it's possible to actually just write code that does math. That means it's actually capable of doing--say--inference on a machine-learning model, multiplications on data frames, etc.
(DIR) Post #AUQFIineEWMbaidbXc by benmschmidt@vis.social
2023-04-06T17:22:34Z
0 likes, 0 repeats
There are already some crazy things out there, like a version of Stable Diffusion that runs in your web browser. https://github.com/mlc-ai/web-stable-diffusion. I wrote something two years ago about why I think WebGPU makes javascript the most interesting programming language out there for data analysts/ML people. https://benschmidt.org/post/2020-01-15-webGPU/ And there's lots of stuff that's become more conceivable since--for instance, a dataframe based on Apache Arrow that does massively parallel data transformations. Chat-GPT in browser.
(DIR) Post #AUQFIkYThtPx2FYWAq by benmschmidt@vis.social
2023-04-06T17:30:13Z
0 likes, 0 repeats
Right now it's only released on Chrome, but it's not an only-Google thing forever. It's an honest-to-goodness W3C standard like HTML, CSS, or SVG. All the browsers have been working on it; Chrome is just shipping first because they're insanely well funded compared to Safari and Firefox. (One of my favorite parts about reading the minutes of the WebGPU committee--yes, that's how insanely interested I am in this--is seeing the other browsers get jealous of Chrome's money.) https://github.com/gpuweb/gpuweb/wiki/Minutes-2022-08-10
(DIR) Post #AUQFImag8gMlLebIsi by benmschmidt@vis.social
2023-04-06T17:41:19Z
0 likes, 0 repeats
For projects like deepscatter, moving to webGPU holds huge possibilities going forward because it means we can do arbitrarily complex calculations directly on the GPU, and pass the results back to the browser. Maps like https://atlas.nomic.ai/map/twitter can render 5,000,000 tweets incredibly fast, but push too much compute to CPU--I have a long and growing list of things that are nearly impossible in WebGL (especially the WebGL 1.0 I'm stuck in for deepscatter...) but that should be quite easy in WebGPU.
(DIR) Post #AUciTgIfK7Xob0xQno by benmschmidt@vis.social
2023-04-13T16:57:40Z
0 likes, 1 repeats
Check out this work of interactive scrollership we just released with Rita Gonzalez-Marquez, Dmitry Kobak, Philipp Berens and others at Tübingen presenting their gorgeous T-SNE embedding and several other views of 20 MILLION abstracts from PubMed. https://static.nomic.ai/pubmed.html