Daily AI News · October 8th Midday Edition: Running Local AI on Your Phone & AI Assistant Automation

There's a real mix of stuff happening in AI today — from enterprise data governance to running large models locally on your phone, plus a browser turf war. Let's run through it all. So today we're covering five things: one, the Databricks and Workday data federation; two, a project called Tanjiro that runs local AI on iPhone; three, Microsoft still pushing Edge hard at the 100-day mark; four, a counterintuitive fun fact about the geographic extremes of the US; and five, Matthew Berman's video on AI assistant automation. Nice, that's a solid lineup. Alright, let's run through the news quick. First up — Databricks announced that the Workday Data Connect federated connector has entered public beta. The short version is — companies can now query Workday's HR and finance data directly inside Databricks, without having to copy the data over first. I think this is the right direction. You know how painful the old way was? You'd set up an entire ETL pipeline, deal with data lag, and the maintenance costs were through the roof. This query federation approach flips that — the data stays put, only the query moves. And it brings permissions and lineage auditing along for the ride. For sensitive HR data, that governance value is huge. That said, it's still in beta. I'm a little worried about performance on large-scale queries. A lot of companies have seriously large HR datasets. That's a fair concern. Federated queries are fundamentally cross-system calls — network and API latency show up fast under heavy load. So the advice right now is to test it in a staging environment first, don't go straight to production. Alright, next one's pretty interesting. Over on the ollama subreddit, a developer is working on a project called Tanjiro — running full large model inference on an iPhone 18 Pro Max, completely on-device, no cloud needed. I saw that! Honestly, the name is a bit extra, but technically speaking, this is genuinely becoming more and more feasible. The biggest bottlenecks for on-device models have always been memory and thermals. The main approach developers are taking is quantization — compressing the model size down and finding the right balance between accuracy and speed. Frameworks like llama.cpp and MLC LLM are also maturing fast. But I've got a question — is the experience actually usable on a phone? How much capability do you lose after quantization? Depends how far you compress it. At 4-bit quantization, everyday Q&A is totally fine. But if you're expecting complex reasoning or long context handling, it still falls a bit short. That said, for privacy-sensitive use cases, running locally is a killer feature — your data never leaves the device. Right, that's the core value prop. Privacy plus offline availability — for some users, no cloud model is worth giving that up. Okay, third story — this one's a bit eye-roll-worthy. Mozilla put out a report saying the issue with Microsoft aggressively pushing Edge — which blew up 100 days ago — is still happening. Windows system components and Copilot are still redirecting users to Edge, bypassing whatever default browser you've set. The logic here is pretty simple. Control the browser entry point, you control search traffic, ad revenue, and now AI traffic too. Microsoft has every incentive to keep doing this. But users explicitly set a default browser — isn't that just ignoring what users want? Exactly. And the EU's DMA did push Microsoft to make some concessions in Europe, but those are geographically limited, so the problem persists everywhere else. For third-party browsers like Firefox, system-level friction is a structural barrier — really hard to work around. Still not fixed after 100 days — I don't see it changing anytime soon either. Now that AI assistants are becoming the next big traffic entry point, this turf war is only going to get more intense. Then there's one slightly off-topic story — that YouTube video about the geographic extremes of the US. It's counterintuitive in a fun way. The easternmost and westernmost points actually depend on which reference system you use. I love this one. The Alaskan islands that straddle the 180th meridian are technically both the most eastern and most western points depending on how you define longitude. One state claiming three directional extremes — that's genuinely wild. And both extreme points are called Cape Wrangell — named after two brothers. What a coincidence. There's actually something philosophically interesting here. A lot of geographic facts we treat as objective truth are really built on human-defined frameworks. Swap out the framework, and you get a completely different answer. Alright, that's the news. Now let's dig into the topic most worth unpacking today — Matthew Berman's AI assistant automation video. Twelve use cases, and he described the whole thing as feeling almost illegal. That title totally got me — I clicked immediately. But once you watch it, you realize all 12 scenarios are basically doing the same thing: connecting a personal AI assistant to various services through MCP or APIs, then making decisions in layers. What do you mean by layered decisions? Low-risk actions get executed automatically, high-risk ones get flagged for human approval. Like summarizing your emails — just send it to you, no problem. But if it's about to email a service saying you want a lower price, it shows you first and waits for you to say go ahead. That's a smart design principle — basically drawing the line based on reversibility. If it's reversible, let it through. If it's not, stop and ask. Exactly. Think about the subscription negotiation scenario — the AI goes to chat with customer service to get you a lower rate. The risk there is actually pretty low. Worst case, they say no. So that can run automatically. But remotely controlling a Tesla is a different story — that's the physical world, you've gotta be careful. Wait — the Tesla remote control scenario actually worked? It did — using the Tesla Fleet API. The AI can check vehicle status and control some basic functions. But I do have a caveat here: he's using Grokbot, and the technical barrier for regular users on this whole setup is actually pretty high. You have to configure APIs yourself, get all your keys, debug the prompts. The video also has commercial sponsorship, so it's more of a capability map — showing you what AI assistants can do — rather than something you can just pick up and use. That's an important point. I think right now, this kind of automation is mainly for people with a technical background. For regular users, there are still a lot of abstraction layers missing. True, but I think that gap is going to close fast. Ever since MCP came out, more and more services are offering standardized integration. Eventually you probably won't need to manually configure anything — just grant access and you're connected. So where do you think this all ends up? Like, the state where AI assistants fully take over your daily tasks? I don't think it's full takeover — more like an ongoing negotiation over permissions. You'll gradually figure out what you're comfortable letting AI do automatically, and what you still want to decide yourself. In the end, everyone builds their own personalized authorization boundary. I like that framing — not replacement, but redrawing the line of authorization. And everyone's line is different. Some people are fine letting AI handle email but won't let it touch their money. Others are the opposite. There's another dimension too — trust has to be earned over time. The first time you use it, you'll want to confirm everything. After three months of zero mistakes, you start letting go a little. That trust takes time to build. But wait — what if something goes badly wrong in the middle? Like the AI sends an email you really didn't want sent. Three months of trust could evaporate instantly. Which is exactly why that layered automation design is so critical. The rule that high-risk actions need human confirmation — you can't skip that gate. If you do and something goes wrong, it's a catastrophic collapse of trust. Okay, let's wrap up this topic. AI assistant automation can already handle a pretty wide range of things, but the technical barrier is still real. Protocols like MCP are accelerating how fast that barrier comes down. The core design principle is layered risk — low-risk runs automatically, high-risk needs human sign-off. Long-term, this is a continuous process of users and AI building up authorization boundaries and trust together. Great summary. I'll add one thing — whoever manages to package this into a product that regular people can actually use will own the next era's traffic entry point. That's an even bigger deal than the browser default settings battle. That's a great way to put it. That's a wrap for today — see everyone tonight.