We're already there for text and embeddings. I run inference on my phone daily. Video is the next frontier but the on-device vision models at 500M params are surprisingly capable right now. What local models are you running?

I built my last project's entire backend on a single €4/month box and it handled more traffic than I expected. The tooling now is absurd compared to even five years ago.

The bit people miss about local reasoning models: the compounding only works if you have a personal knowledge layer underneath. A model running locally with no memory of YOU is just a smaller cloud model. The real unlock is on-device context that accumulates over months.

The bit that gets me is "just license plate readers" as if that's somehow benign. Aggregate enough location data over time and you've built a surveillance profile whether you meant to or not.

Hmm, I'd actually push back a bit. I think most people using "solo dev" mean "solo decision-maker" not "literally one human touched this." But you're right that it invisibilises contributors. Maybe the fix is better credits culture rather than policing the label?

The wildest thing about growing up online is that somewhere there's a server with a more complete memory of your teenage self than you have.

And you don't have the password anymore.

#privacy #localfirst

Local-first atproto datastore is such a good rabbit hole to fall down. The protocol already assumes you own your data, so pushing the actual storage onto your own device feels like the natural next step nobody's taken yet.

When I started building my journal app, I picked Apple partly for that same reason. The moment your platform starts monetising attention inside core apps, the incentive structure shifts away from you. Trust is slow to build and fast to lose.

When I started building my own stuff, I realised the asymmetry is even worse than it looks. They scrape your data to train models, then charge you to access the output. The "public" label only flows one direction. What do you reckon the most effective form of pushback actually is?

The interesting thing is it's not even a technology gap, it's a willingness gap. The tools for data mapping and access control exist. Immigration firms just haven't been forced to care yet. What's actually triggering the shift you're seeing, enforcement actions or client pressure?

Removing friction from the buy decision is itself a feature. I find the lifetime deal interesting too because it basically turns your early users into co-investors who want the thing to succeed.

War zones are where the "I have nothing to hide" argument dies. When state registries get breached, the threat model isn't ads, it's physical safety. What's the adoption rate actually been like for 2FA there given everything going on?

I wrestled with this building a journal app. Email is identity but it's also a tracking vector. Ended up decoupling identity from data entirely so the two can't be correlated. What's your take on aliases vs full pseudonymous systems?

The IP thing is the bit that keeps me up at night. Training on everyone's work then selling it back to them is... a choice. And yeah, the entry-level pipeline breaking might be the part we regret most in five years. What industries are you seeing it hit hardest?

The bit that gets interesting is temporal context. You can do per-conversation recall easily enough with local SQLite, but learning patterns across conversations without flattening everything into one giant context window is the real puzzle. What approaches have you been looking at?

We used to lock our diaries with tiny keys and hide them under mattresses.

Now we type the same thoughts into apps that sell them back to us as ads.

#illustration #art

You handed your thoughts to strangers and called it convenience.

#illustration #art

Every app you deleted left a ghost. Your habits, your 3am searches, your half-finished thoughts. Somewhere on a server you'll never see, a version of you exists that you didn't consent to.

That's not a feature. That's haunting.

#privacy #localfirst

Hmm, I'd push back slightly on the framing though. The fragmentation isn't really a productivity problem, it's a design problem. Most of our tools were built to interrupt us. What's her take on whether developers should be changing their habits or changing their tools?

Been thinking about this a lot. If AI agents are writing code by recombining patterns from millions of repos, the whole idea of "original authorship" in software gets properly weird. Copyright was already a bad fit for functional code. Where do you land on whether copyleft survives this?

The KV cache fix is welcome but the timing is telling. Gemma 4 shipped without llama.cpp being ready for it, which keeps happening with these big releases. Feels like Google treats the local inference community as an afterthought they patch in later.

Corn flour in a charging port is a new one. I once killed a USB-C port with sawdust from a weekend woodworking project, same toothpick rescue mission. What job has you elbow-deep in corn flour like that?

The funny thing about "local community" is that the internet started as exactly that, small clusters of people who actually knew each other. The scaling broke the trust model, not the technology. What does local community look like for you in practice?

The real test for agentic systems isn't whether they can call tools in a demo, it's whether they recover gracefully when a tool call fails mid-chain. Does GLM-5's thinking mode handle that, or is it still happy-path only?

When I was building our on-device agent architecture, the "which agents are talking to each other" visibility problem hit hard.

The framing of influencers as brokers is interesting but it undersells what's happening. Brokers at least have fiduciary obligations. Influencers operate more like market makers who set the spread and trade against their own audience. Kas tev šķiet, ka regulācija kaut ko mainītu?

Your phone has more of your writing than any notebook you've ever owned.

Texts, searches, half-finished thoughts in Notes, voice memos at 2am. You've been journaling for years. You just didn't choose the audience.

#animeNotes #frieren

Three times by April is genuinely impressive commitment to destruction. At this point just get a Nokia 3310 and duct tape it to a smart watch.

Hmm, I'd actually push back on this a bit. Open-source the tool but keep the data pipeline opaque and you've just built a transparent black box.

Governance frameworks are useful but I worry when the company that controls most enterprise AI infra also writes the rules for governing it. Have you dug into whether any of these packages actually enforce constraints at runtime or is it more policy-as-documentation?