Shift; A Bad Sign For Self-Hosting AI; ADS-B WX

The weekend Bonus Drop explores the interplay between technological innovation and practical accessibility. As we take a look at how advancements in various tools and models, we’re faced with the stark reality that while – technology offers transformative potential – its benefits often require specialized knowledge or financial investment, limiting its reach to a select few.

In other news, WordPress informed me, today, that the Drop has hit 50,000 total views since moving to their platform. I have no idea how good or bad that is in context of the entire “newsletter ecosystem”, but I am thankful y’all keep opening, reading and not-unsubscribing! 🙏🏽 (I wish I saved the Substack stats, now.)

---

TL;DR

(This is an LLM/GPT-generated summary of today’s Drop using MLX LLM in OpenAI API compatibility mode, SmollM3-3B-8bit, and a custom prompt.)

• Shift browser aims to reduce cognitive load by organizing workspaces with email integration and app connections, offering layout customizations similar to Vivaldi. The subscription model (free tier sufficient for many) may be justified by claimed time savings, though the browser still has rough edges. (https://shift.com/pricing/)
• Peter Steinberger critiques self-hosting AI models, highlighting high costs and limited practicality for most users. While Claude Code remains strong for coding, the “OpenAI tax” of $6,000/month for sustained use makes private AI deployment impractical. (https://steipete.me/posts/2025/self-hosting-ai-models/)
• ADS-B WX uses aircraft ADS-B signals to create high-resolution wind maps via vector math and particle simulations. The approach outperforms satellite data and is accessible with $100 RTL-SDR hardware and free software. (https://obrhubr.org/adsb-weather-model)

---

Shift

We live in a time of browser Hades. The major players are all terrible, and up-and-comers — like Arc — have crushed our spirits and left us in a wasteland of AI-ridden monstrosities. On top of that I recently learned (thanks to this section’s resource) that we “knowledge workers” end up switching between various apps upwards of 1,000 times a day (burning over 4 hours).

Arc solved some of that problem, but they betrayed us. While we wait for some disruptive new players to polish their engines, there’s a new browser in town — Shift — that aims to help bring order to chaos, but steps a bit short of reimagining the browser like Arc did. While it does offer some layout customizations (in a Vivaldi-esque manner), one of the primary goals of the browser is to help us manage our multiple personas/identities and reduce cognitive load.

Shift’s solution feels almost obvious once you see it in action. The left sidebar organizes workspaces for different contexts – work, personal, that hobby project that’s slowly becoming a business. Each workspace connects to specific email accounts and integrates with over 1,500 web apps (Arc did something similar with a much smaller numer of apps). Shift makes it easy to toggle between Gmail accounts without logging out. Jump from work Slack to personal Discord without losing your place. Keep your vacation planning tabs completely separate from quarterly budget spreadsheets.

It feels like the designers are hoping their creations fundamentally changes the relationship between us and our digital tools. Instead of adapting our workflows to browser limitations, the browser adapts to how we actually work. The result is something that feels less like using software and more like having a well-organized digital command center (albeit with some rough edges).

The subscription model (the free tier is likely enough for everyone, which doesn’t bode well for the longevity of this browser) might give some folks pause. But if you can quantify the claimed time savings, it might be worth ~$150 USD/year.

I’ve installed it and am going to give it a go as the primary driver starting Tuesday ($SPOUSE had more Boston-based healthcare needs tomorrow, and I’m not going to be staring at glowing rectangles too much). I’ll repot back next Bonus Drop.

---

A Bad Sign For Self-Hosting AI

Peter Steinberger’s deep-dive into self-hosting AI models, especially in the wake of Anthropic’s Claude subscription changes, is a candid ledger of the realities of running advanced language models privately in 2025.

His dissatisfaction with new subscription limits for Claude Max ignited a thorough hunt for alternatives, pushing him to scrutinize model performance, tool flexibility, and – most sharply – cost. Anthropic’s Claude Code emerges as “still king” for coding workflows, praised for pairing a powerful model with highly tailored tools. But growing costs(—$6,000 for a month of hard, sustained use) prompted Steinberger’s exploration.

I could blather more, but the post did a great job walking through the realities of what I’ve been personally discovering: local “AI” is just not going to be something the average computer wielder will be able to make work or benefit from. Sure, there are very focused use cases that will work — something Pete Sergeant covered super well in “Get the hell out of the LLM as soon as possible”. But, I’m afraid there will be no choice but to pay the “OpenAI tax” if this is technology you want (or need) to work with.

---

ADS-B WX

If you’ve ever wondered whether you could do citizen science with aviation data, you’ll appreciate how accessible and clever one approach to this is.

Modern aircraft constantly broadcast ADS-B messages on 1090 MHz. These digital packets are unencrypted collision avoidance signals that anyone can receive with just an antenna and RTL-SDR dongle costing under $100. Using software like readsb, you can decode live data at home and see every plane’s position, speed, heading, and more.

What’s less obvious is that ADS-B contains enough detail to work as a high-resolution wind sensor. Each message includes ground speed and track angle from GPS showing how the aircraft moves over Earth, plus true airspeed and heading from onboard sensors showing how it moves through the air.

The trick is simple vector math. Take the difference between the aircraft’s ground movement and air movement to get the wind velocity at that location. Aggregate this from thousands of aircraft across Europe or any active airspace and you get a crowd-sourced wind map.

The model takes this further by gathering hours of data from services like ADS-B Exchange, then running a particle simulation. For each aircraft, 300 particles are seeded with the calculated wind vector and allowed to spread through random walk modeling. The map gets divided into squares where particle vectors are averaged together. More overlapping flights mean more robust wind estimates.

When verified against sophisticated global models like GFS, the homebrew approach performs surprisingly well. Wind fields at 11km altitude aligned closely with GFS data at the matching atmospheric level, matching both flow patterns and speed estimates over areas like the Mediterranean.

Temperature and pressure can be inferred too using methods from research papers, potentially augmenting existing weather datasets.

It’s pretty amazing that cheap radio hardware and free software let anyone build real-time, high-resolution wind fields. Elementary vector math combined with millions of crowd-sourced data points creates surprisingly realistic and accurate results that can outperform satellite-only datasets.

This is smart data plumbing that bridges aviation enthusiasm with meteorology. You’re not just seeing planes in the sky but using them collectively to reveal invisible wind patterns and compare against state-of-the-art weather models.

Check out the full details at “Weather Model based on ADS-B – Niklas Oberhuber“.

---

FIN

Remember, you can follow and interact with the full text of The Daily Drop’s free posts on:

• 🐘 Mastodon via @dailydrop.hrbrmstr.dev@dailydrop.hrbrmstr.dev
• 🦋 Bluesky via https://bsky.app/profile/dailydrop.hrbrmstr.dev.web.brid.gy

☮️