Keeping up with AI & tech

Technology has always moved fast in comparison to other industries and AI has only increased the speed.
It is therefore more important than ever to stay up to date and keep an eye on what might come next. I have always done that and over the years gathered more than 30 news sources that I frequent.

Those include general tech news pages, YouTube channels, blogs and a select few newsletters. In theory, there are also LinkedIn and X, but more on that later.
Obviously I do not read every article and watch every video and so on, as not every single piece of content is interesting/relevant to me. But even scanning the sources takes time and sometimes I miss interesting bits.

A while ago, I decided to automate the curation and build my own daily digest with content worth looking at. I would expect a lot of people have built something similar by now. This is how I did it.

Meet Perry

I have a personal agent called Perry, running on Hermes on a Mac mini at home. Our family has another agent called Bartimäus. What he does is a topic for a future post.

Naturally, I added keeping me up to date with tech to Perry’s daily tasks. He checks my sources, summarises what happened and puts together a daily digest based on my interests. Reading it takes around five minutes, plus the time I spend on articles I want to explore further.

I do not like summaries of summaries of summaries.
That is why I use Perry’s digest to give me an overview of what is important and then I read articles in full that I find really interesting.

Sources, preferences and delivery

The setup is pretty straightforward, I would say.
Hermes has access to a browser and my Obsidian vault. The vault contains my curated list of sources and a skill describes how I want Perry to handle the news:

• which topics are important and need detailed coverage
• where a roundup is enough
• how to handle duplicates
• what to do with stories that carry over from previous days

At the moment, for example, AI gets more detail, while news around Linux kernel or desktop developments (I game on Linux) get a combined roundup.
Those preferences may change from time to time. Every now and then I add another source or adjust how much space a topic should get.

For web pages, Perry visits the site, reads the content and gives me a summary with the most important points. For YouTube, I use Peter Steinberger’s summarize CLI, with Codex as the backend.
Perry also checks what is trending on GitHub to catch interesting projects and emerging trends.

He then structures and ranks the news according to my preferences. The short version arrives via Telegram; a more verbose version goes into Obsidian.
I am not sure yet if I will keep the verbose version in Obsidian long term.
So far, the biggest value comes from the short version, as I decide where to dive deep from there.
One benefit of a history in Obsidian is of course that Perry can check what was already covered yesterday or the day before, in case another source picks up a story later.

Getting the brief right

It took two iterations to get the Telegram digest to a point that was useful to me:

On the first day the brief was too concise. When I described to Perry what I wanted, I gave him some leeway to decide the output format.
He opted for a “top 5 articles to read” kind of format, which omitted way too many articles.
So I said “give me everything that fits my interests”, which obviously was going to be too verbose, but that was intentional.

On the second day I then gave instructions for what to cover only in aggregate and what to remove entirely.
I also compared Perry’s coverage with what I would have read on that day and found a subtopic I had not added to my preferences initially.

On the third day Perry delivered a news digest that was just right. I compared it with manual scanning again and was happy to find that everything I found interesting was covered.
I will compare occasionally in the future for further calibration, but right now I am very happy with this.

The day-two version also made me consider how to handle duplicates (remove them, duh) and content that carried over from the day before (only cover if no duplicate).
The model currently in use is GPT6-Luna at medium effort. So far, it works well and consumes very little of my ChatGPT subscription.

If you are wondering why I do not use Claude: I like the flexibility of Hermes, but I would not be able to use a Claude sub for it, so I use OpenAI.
I also get good results from their models for coding, which uses the bulk of my sub.
(Spoiler for a future post: Bartimäus uses MiMo 2.6 Pro right now and we are very happy with it.)

Alternatives and possible improvements

This setup is honestly super simple. It took maybe 30 minutes of active work to get it to a usable point.
There are of course other ways to build this and potential improvements.

For example, the whole workflow is always the same. I could have built a pipeline around well-known tools/frameworks. RSS could identify new articles where available, with content fetching for feeds that only provide an excerpt or a link. Other integrations could fetch new YouTube videos and so on.
Building such a pipeline is not hard. Fetching content outside of RSS is often brittle and error-prone though.
Letting Perry use a browser is just more simple and pretty reliable.

I could also add some sort of internal tracking to see which content I actually do consume and build a profile based on that. Then I could use that data to train an XGBoost ranking model and continuously refine the digest. My own Google News basically.
A pretrained text reranker could be another option, scoring articles against a description of my interests without first training a personal model.

I have my doubts, though, whether the latter two would materially improve the digest and be worth investing time in.
Building a fully fledged Google News for myself sounds kind of cool though, so maybe I will do that at some point. Just to see if it works at least.

Another idea is to make better use of the history in Obsidian by tracking the development of topics/stories over days and weeks. Maybe in the form of a weekly review on Sundays, as that is the day with the least actual news.

Take the hype cycles of AI as an example. Jev and decision models were all the rage for a week and then I heard much less about them.
A weekly review could help with looking into why the hype died down or what people are now building with those models.

While writing this post, I also realised that I did not even bother looking for prior art and off-the-shelf tools. I simply knew that I would get this to work in a matter of minutes with Hermes and Obsidian, so why set up another service in my homelab?

As already mentioned: this is how I built it. Others have probably done this in comparable ways with OpenClaw, Dots and other agents.

Sources not covered today

Content-wise, there are a few more immediate gaps.
The newsletters I subscribe to are not included yet. I am not sure if I will add them, as they mostly arrive weekly and I might also just unsubscribe from quite a few of them.

Twitter is also missing. Think what you will about that platform, but for the tech community, especially around AI, I still find it useful for catching news and developments early.
LinkedIn might be another source. Some people whose judgement I trust post there and make the platform less of a slop-fest.

Adding these should improve coverage, while still keeping the digest readable in five minutes most of the time.

For now though, I am happy with what I have.
Thirty minutes of work have saved me a lot of time already and I am yet to discover that Perry missed a story I would have found myself.
If you have not built such a system yet, I can highly recommend it.
Feel free to reach out to me and share your own.