There are many cases of harm caused by ML false positives.

There are also some cases of law enforcement successes caused by ML true positives, e.g. a RAF (Red Army Faction - https://en.wikipedia.org/wiki/Red_Army_Faction) terrorist gone into hiding was identified by a social media photo: https://web.archive.org/web/20240305044603/https://www.nytim... (although the success in law enforcement was actually not carried out by the police, but by investigative journalists/podcasters.)

The question we need to answer as a society is if we are willing to tolerate any innocent people to go to jail as the "price" of catching a few more criminals.

> I do definitely believe that a large number of ‘innocent’ people are criminals (by the letter of the law) without their knowledge.

Because of that (or rather, to sort out the mess), I always felt that citizens should have a right to be informed of every law that they are expected to obey so that at least in principle, they'd be able to comply (to be effective, plain language explanations would need to be included).

Imagine an app that told you, whenever you cross state boundaries, what is different in the law now from your previous location.

The original article doesn't dwell too much on the RAM limitation, but I agee with you that 8 GB is too little for the near future or even today.

I agree with most of the post's arguments, and most of the specs and limitations of the Neo would be okay with me, except there should be 16 GB RAM in 2026.

Apple could perhaps mitigate this somewhat by releasing a "slim" MacOS Neo version that is less bloated by pruning some features. Currently, the OS uses much of the available RAM for caching (I've seen "40%" of total OS RAM usage) to make the system faster, whereas 8 GB RAM permits only essential caching.

(Surely, the tough 8 GB RAM decision was influenced by the three factors 1. current DRAM cost and 2. limited DRAM availability considerations as of 2026, and 3. the massive Neo market size resulting from its attractive price tag, and this may get reconsidered in future editions.)

For scientific search experiments, you may like to consider using PyTerrier (which facilitates comparing multiple search model types - (sparse) vector space model; Boolean model; Binary Probabilistic Model; Support Vector Learning-to-Rank model; Divergence from Randomness model; (dense)embedding ranked retrieval models etc.).

Thanks for opening this up.

As was already said in one of the reference videos, it's impressive what one person can do.

But the next step is to define an architecture where authors can defined/implement plug-ins with particular modular capabilities instead of one big monolith. For example, instead of front-end (GUI) and back-end (feeds), there ought to be a middle layer that models some of the domain logic (events: surces, filters, sinks; stories/time lines etc.).

I would like to see a plug-in for EMM (European Media Monitor) integrated, for instance ( https://emm.newsbrief.eu/NewsBrief/alertedition/en/ECnews.ht... ).