williamstein

One Example: you can get a "Home Equity Line of Credit" (HELOC) against your house, then use that money to do things, e.g., remodeling (the interest from such loans is in some cases tax deductible). The house you would own with a 3% interest rate is likely worth significantly more than the house you would own with a 7% interest rate, so the potential impact of a HELOC is higher.

Yes see https://basecamp.com/handbook/managing-work-devices#mobile-d...

They also don’t allow Windows or iOS.

One of many possible values to society is that punishment reduces the motivation of vigilantes. For example if person A killed person B’s child, then person B (or a mob encourage by B) might be less motivated to seek vigilante punishment if there is some chance that a legal system could punish person A instead.

I tried Mixtral via ollama on my Apple M1 Max with 32GB of RAM, and was a total nonstarter. I ended up having to powercycle my machine. I then just used two L4 GPU's on Google Cloud (so 48GB of GPU RAM, see [1]) and it was very smooth and fast there.

[1] https://github.com/sagemathinc/cocalc-howto/blob/main/ollama...

I think it's real. When I sign into my account, kernels don't run. Noteable is one of a couple of realtime collaborative Jupyter notebook environments, including Deepnote and Cocalc (my site). I know from chatting with Kyle Kelley earlier this year that they were investing strongly in ChatGPT integration, but evidently this didn't fully ship, since their "Upgrade to Pro" dialog says "COMING SOON: NoteableAI!" in bold. Kyle regularly posted a lot of exciting work they were doing on AI integration on Twitter (e.g., [1]).

[1] https://twitter.com/KyleRayKelley/status/1727430546354233763

New self-hosted customers will no longer get unlimited free access to Zulip’s Mobile Push. Fortunately, this doesn't mean they are making anything closed source!

They certainly claim it is SOTA for multimodal tasks: “Gemini surpasses SOTA performance on all multimodal tasks.”

Phind.com’s custom model (which is I think publicly available) is pretty good for coding tasks. It helps enormously that it uses RAG with web search. I use it all the time and also GPT-4, and GPT-4 is better, but Phind’s model is close and also very fast. I have nothing to do with the company, and am just a random (free) user.

It seems that for free public images, DockerHub has a limit of "200 pulls per 6 hours" , according to https://www.docker.com/increase-rate-limits/

I can't find any stated rate limits for ghcr.io for public images. Presumably they secretly exist, but are meant to be as high as Github can provide for free...?

Thanks!

What are the limits that DockerHub has, which are much worse than GHCR's limits for public images? (I'm genuinely curious.)

Also, it doesn't seem to be quite as simple to switch to as just prefixing ghcr.io. The default when you do that switch is to create private images (which cost and are not public), so you have to figure out everything related to making the images public, which possibly involves changing Dockerfiles to link the image to a public project, or other options. [1]

[1] https://docs.github.com/en/packages/learn-github-packages/co...

One challenge is that the power usage of a server is order(s) of magnitude greater than that of a laptop. This means the cost to do what you describe is significant, hence that has to be taken into account when trying to build a cluster that is competitive...

In the recent Painkiller miniseries [1], which is supposed to be about Purdue Pharma, what you describe above is basically how things worked. There's a key scene in which Rudy Giuliani calls in some favor, which results in a prosecutor being significantly influenced by the White House. I don't know if this actually precisely happened, but this portrayal on Netflix speaks to "how people thinks this works" and why.

[1] https://en.wikipedia.org/wiki/Painkiller_(TV_series)

I like https://www.slatejs.org/examples/richtext and found the source code very readable and adaptable to my needs.

I was a vegetarian for over 20 years, then I got a highly intelligent dog, who spent years trying hard to convince me to eat (more) meat!

I was a tenured mathematics professor in academia (at UCSD and Univ of Washington) from 2005-2019, and a full professor during the last few years. At 45, I left academia and started a small tech company, giving up tenure and the academic life. It has worked out for me so far, though it often felt like my instincts from academia are almost exactly backwards from the correct instincts for industry. Anyway, leaving the "golden handcuffs" of tenure was probably a somewhat unusual part for me. I tried very hard to stay X% with the university, but the intellectual property and other rules were too rigid to make that work for more than 2-3 years, even though the mathematics department was extremely supportive. My motivation for leaving academia was my inability to get funding to support development of open source software (in particular, SageMath), and I still aspire to one day create more value and have a bigger impact than I could in academia.

What is their "content policy"? One of my standard test prompts for these image generators is: "Can you draw an illustration for a web based collaborative environment for using Jupyter notebooks, Latex, Linux, GPU's for doing data science?" and no matter what variant on this I try, I get back "I'm sorry for the inconvenience, but I was unable to generate images for your request due to our content policy."

EDIT: "Latex" (pronounced "lay-tech", the document editing software from Knuth) is what violated the policy.

Math is often much more fun and compelling for some people when you both theoretically prove something works and then also convince yourself of the same thing via a numerical experiment. I’m pretty good at math proofs (pure math PhD, wrote some books and papers), but I still love to do numerical experiments. It’s fun, and you also set yourself up to be able to easily ask different questions that may be very hard to answer theoretically.

Cocalc uses a websocket and xterm.js to implement terminals on a remote server. Each terminal session corresponds to a file (with extension .term), so multiple clients can open the same session by opening the same file. If you type in one session then all sessions will see the typing at the same time. (Disclaimer: I wrote this. It’s way too heavy for this use case, but might be an amusing demo or proof of concept for somebody to play with before writing something new.)

I wish it was all online. I have long hosted scans of Grothendieck’s mathematics at https://wstein.org/sga/

I published a graduate level textbook on Modular Forms, and I also sometimes think I just barely know what they are: https://www.wstein.org/books/modform/

Could you provide some sort of status report about what Gitness actually is? According to the Git repo it looks like it started Feb 2, 2014, and Brad Rydzewski did the vast majority of the work on the project (e.g., 100x more than anybody else), and is still involved now. I downloaded it and tried it out (which is VERY easy to do), and the actual software seems to be a Typescript frontend and Go backend, with Monaco for a code editor, and nice diff view. I tried importing a very large repository, and it took a while, but did fully work, and browsing the repo after import was fast. It's Apache-2.0 licensed (thanks!). Was it open source from the beginning?

Yes

It's also a short, clean and easy to read shell script.

Related -- last month I wrote an implementation of something very similar to sshfs, but in Typescript over a WebSocket: https://github.com/sagemathinc/websocketfs

This might not age well: "OpenAI likely to go bankrupt by the end of 2024: Report" [1]

[1] https://www.livemint.com/ai/artificial-intelligence/openai-l...

> Mojo is a proper superset of Python.

Isn't it more that the plan is someday Mojo may be a proper superset of Python, but right now it is far from it? I just tried opening up the Mojo playground, copy/pasted the very first little example function from the official Python tutorial (see https://docs.python.org/3/tutorial/controlflow.html#defining...) and Mojo outputs a bunch of errors.

With Cython, our goal was to make it a proper superset of Python, and it was really difficult, but we got close.

Answer: "So, for us YC was a no-brainer."

You may be wrong: https://developer.nvidia.com/rtx/ray-tracing/optix