dreadsword

Not familiar with fever, but there is something similar buried at the heart of mine - the LLM clusters stories, and they get promoted to public when they reach a threshold of unique sources.

That threshold is a function of day of the week - on weekends when the news cycle is quiet, it lowers the bar --- tuesday to thursday its at its most restrictive.

Yes - Techmeme is definitely the archetype and a great product, and I have spent lots of time there over the years as well!

Cheers and thanks for the kind words! And yes - LLMs (at least o3-mini) do a great job as my editorial team - the site is 100% automated.

Cool - particle looks great - I really like how visual it is.

Distinguishing characteristics - personally I get value from the unambiguous timeline (no editorializing in /recent), and (as nice as the visual is) the non-visual, super simplistic presentation & the curated sources (...which I value b/c I curated them myself haha).

So bottom line is that DS will appeal to a certain kind of obsessive compulsive news consumer and synthesizer that wants the right balance of signal to noise ands a streamlined presentation that doesn't slow them down. I count myself among that group!

How about now? Story titles are still clickable links, but are black. Made the story count a link as well, and kept it blue as a visual cue.

I hear you - there's something to be done there. My initial thought was to stay as close to convention as I could (links are blue!), but as the RECENT list gets long, its definitely gets less scannable.

Thank you for the feedback!

What dark magic is this! Very cool repurposing of technology. If you don't mind me asking - what's the origin of the name Tommy?

Cheers and thank you! I'll reshare an earlier comment that I think answers your question - let me know:

Thanks so much for the kind words - its 100% o3-mini for clustering. I have zero editorial input as to what constitutes a cluster, what's "top" news, etc.

The one subtlety is setting up the LLM to understand whether a new story belongs in an existing cluster, or with > 1 neighbors, constitutes a new cluster. The challenge there is scoping the clustering window (hours of stories for consideration) and topic breadth to avoid creating Katamari-super-clusters that just end up with every story associated to them.

At this point I seem to have found a sweet spot re: the hours window, the frequency of processing, and the design of the prompt such that its working consistently.

Very few false positives in terms of spurious clusters being created, or potential clusters being missed.

Awesome - glad you're enjoying it and thank you for the kind words!

My "Stack" ---- LAMP + o3-mini for editorial tasks + Bootstrap for responsive front end. That is to say: Its old school, and painfully functional. But, light & fast.

Noted, thank you! I haven't put a tonne into readability, other than some basics - I prefer a serif'd font, and I made sure the background was easier on the eyes than #FFFFFF haha

Oh man, don't ask - not a dumb question at all. I'll reshare what I put in another comment that answers it, but bottom line is they're a design gap in the context of /recent.

You're right --- incoming & outgoing end up being redundant on the "Recent" view. Where they're (more) relevant is in the "Top" view where the LLM editor has picked a subset of stories to be categorized as top and incoming/outgoing are the ones that didn't make the cut, organized by timeliness.

Definitely a gap in design!

Many thanks for the kind words!

Cheers and thank you!

Thanks so much for the kind words - its 100% o3-mini for clustering. I have zero editorial input as to what constitutes a cluster, what's "top" news, etc.

The one subtlety is setting up the LLM to understand whether a new story belongs in an existing cluster, or with > 1 neighbors, constitutes a new cluster. The challenge there is scoping the clustering window (hours of stories for consideration) and topic breadth to avoid creating Katamari-super-clusters that just end up with every story associated to them.

At this point I seem to have found a sweet spot re: the hours window, the frequency of processing, and the design of the prompt such that its working consistently.

Very few false positives in terms of spurious clusters being created, or potential clusters being missed.

Cheers and thank-you!

Cheers - thank-you so much!

Fair enough - its honestly not something I expected anyone to be interested in enough such that an about page would be required.

At a high level, it reads RSS feeds from a number of sources, and uses LLMs to identify clusters of stories about the same thing, group them, tag them, and designate them a "top" story or not. That's it.

The biggest thing I've learned in all of this is that o3-mini is far and away the best at following instructions (for this use case). Periodically I'll cycle through the models available on Groq, and always come back to o3-mini.

Oh I should add: incoming will show stories ~20 minutes before they get picked up for "Top" inclusion, if they're going to make the cut, based on how jobs are scheduled.

Reading RSS myself, OLD SCHOOL: Cron Jobs. PHP. Hahaha! List of sources: At present, no; but if its of interest, it would not be hard to add.

I should also add - please post any recommendations re: sources to cover.

In the roadmap! RSS by tag - i.e.: for https://deadstack.net/tag/quantum And an RSS feed for /recent are both in progress

You're right --- incoming & outgoing end up being redundant on the "Recent" view.

Where they're (more) relevant is in the "Top" view where the LLM editor has picked a subset of stories to be categorized as top and incoming/outgoing are the ones that didn't make the cut, organized by timeliness.

Definitely a gap in design!

Cheers - thank you for the kind words!

Between this and the inscrutable nature of dark matter / dark energy it feels like were missing some pretty fundamental things.

Hah - the hypocrisy is delicious!

? What ? I'm a big advocate of RSS. But it has dropped off in the last decade.

I think that horse has left the barn. The golden age of RSS / Atom / Content syndication was ~10 years ago. Two trends have moved everyone away from it: (1) Paywalls and content monetization, and (2) content distribution via socials not blogs/websites. A decade ago, web-based content producers were incentivized to have RSS feeds to grow their audience; now socials provide the vector, and restrict what can be done w/ content.

Hopefully not a dupe...

I observed this first-hand: Switching from o3-mini to o4-mini was a big step back in accuracy, and also anecdotally slower. I rolled back after an hour of testing different things, having not observed anything better in my use cases.

OpenAI has recently introduced its new o3 and o4-mini AI models that are integrated into ChatGPT. Reports indicate these models not only hallucinate more often than previous versions but also demonstrate an uncanny ability to determine photo locations. This duality of increased factual errors alongside unexpected geolocation proficiency has raised concerns among users and experts regarding model reliability and potential privacy implications. The coverage highlights both the technical achievements and challenges of incorporating these advanced AI capabilities.

Taking advantage of a unique public health policy in the UK, a new study has found that receiving the shingles vaccine reduces dementia risk by 20%. The findings bolster a growing body of research linking the vaccine to lower numbers of dementia cases.