mbanerjeepalmer

I know the author has a long history here and I'm sure the post is a genuine reflection. But how is posting "I used my own product and it felt really good" anything other than self promotion?

We can assume first that at OpenAI he's going to build the hosted safe version that, as he puts it, his mum can use. Inevitably at some point he and colleagues at OpenAI will discover something that makes the agent much more effective.

Does that insight make it into the open version? Or stay exclusive to OAI?

(I imagine there are precedents for either route.)

Hyper relevant business news.

They even mark their own Arts & Culture email as spam: https://x.com/MBanerjeePalmer/status/1962538753328664693

I don't think you can say that AI-written can be reliably detected. Turnitin is only ~90% effective: https://teaching.temple.edu/sites/teaching/files/media/docum...

Set the TASKDATA to `./.task/`. Then tell the agents to use the task CLI.

The benefit is most LLMs already understand Taskwarrior. They've never heard of Beads.

Added to my list of converters: https://www.reddit.com/r/rss/comments/1ntvtb2/rss_converter/

I started with this: https://news.ycombinator.com/item?id=42271078

But wouldn't recommend going via Pandoc, for example for this reason: https://github.com/jgm/pandoc/discussions/10438

Some things like #show rules can be a bit confusing. And obviously LLMs can't really help you.

Alternatives to Opik include Braintrust (closed), Promptfoo (open, https://github.com/promptfoo/promptfoo) and Laminar (open, https://github.com/lmnr-ai/lmnr).

Note first that the piece is not about zacusca.net itself. It's saying 'You should go out of your way to build on a protocol. Even if the protocol is 'dead'. Perhaps I can clarify that.

Regarding Zacusca, I appreciate you explaining why it doesn't seem useful. First, it's not for everyone. Some people like reverse chron as an algorithm. Second, let me clarify how the pipeline works. But I think I explain the import/export poorly: - you export the feed addresses - you set up the categories - Zacusca continously polls the imported feeds, sorts them into the categories and exposes the categories as new feeds

So it's fire and forget. And you can keep your existing feeds as duplicates if you want.

Regarding copyright I believe that this is fair use. It lets individual readers print their subscriptions. It's indistinguishable from them using a printer at home, except that it's delivered automatically.

But yes it's a grey zone. It's making a copy of someone's copyrighted work without their permission. (Remember it was technically illegal to load songs onto your iPod.) And so that's why the book product (which brings together all posts from one publication) is for authors, since I think that strays more towards publishing rather than mere printouts.

Roo is less solid but better-integrated.

Hopefully I'll switch back soon.

https://zacusca.net

The point, though, is that if I'm following an ML arXiv feed and I'm interested in recommender systems, I really don't need to trawl through everything on interpretability, image generation etc. to get there, no matter how artful those papers are.

(A) Studiously prune your feeds like a bonsai. As the author suggest, follow the chain of trust to a small number of voices (for me, something like Stratechery, Simon Willison, Inner Ring).

(B) Realise that RSS is another form of 'tyranny', this time at the hands of the publisher instead of the platform, where the composition of your feeds, and therefore what uses your attention, doesn't correlate highly with what matters to you.

I can feel the pickaxes being unsheathed as I type this but...I have reason to believe my (and others') LLM/embeddings-driven products are a good solution.

Position (A) isn't tenable if FOMO matters. Paraphrasing another comment here: 'Following arXiv is part of the job'.

So let's say you adopt position (B). You recognise that everything that matters to you is distributed across some set of feeds. But only a small proportion of the total material in those feeds matters to you. If you can articulate what matters then you can let an LLM or embeddings model use their attention, instead of yours, on the low-relevance items by filtering them out.

Some options:

- https;//scour.ing

- https://feeds.fun

- https://zacusca.net (disclosure: mine, and still pretty janky)

Ha, that's almost word-for-word on https://zacusca.net. [0]

And Zacusca represents a third way (the first two being reverse chron and newsletters, which both put control in the publisher's hands). If you can articulate what you enjoy then you can filter the RSS according to that.

[0] I'm working on Zacusca. Other, frankly more developed, alternatives include: - https://feeds.fun/ - https://scour.ing/about