if you're curious what your $30 pizza "cost you", this is what @hark_labs had on me after less than 48hrs: -934 gmail messages -415 calendar events, including 1 yr into the future -177 contacts it scraped from gmail and calendar events -33 refreshes of the "home card" I'm not against giving these personal assistant tools the keys to some of my accounts, but we all need reminders of what the "cost" of these are you can see this for yourself in the Hark app by: taping your initials icon on the top right then System Preferences then Data Management then Export your data then Request

This app is insanely farming your entire life when you order this pizza. I connected my Gmail to order a pizza, and afterward the AI casually admitted it had collected information about my family, closest friends, businesses, travel habits, loyalty programs, favorite restaurants, and even Citi Bike usage. How fucked up is it that an AI agent tasked with ordering a pizza is digging through your entire personal life? This is a massive privacy problem with AI agents.

if your thesis around chips / neoclouds is something like "long chips not because AGI/ASI but because of widespread inference adoption" then you REALLY should pay attention to local AI a year ago, local AI was about as non-consensus as it got and that was when @mwaldman130 started building; it's quietly turning consensus now among top SV VCs and most people still haven’t noticed The reason it’s so slept on is that almost nobody outside AI research realizes the implications of the fact that that MOST of a frontier model is EMPTY look inside a trained model and most of the weights are tiny numbers sitting near zero and only a small minority does the heavy lifting you can see this in the open-weight Chinese labs (we suspect this is the case for Anth / OAI models but can't confirm for obvious reasons): DeepSeek V3 has 671B parameters but uses only 37B (~5%) for each token it writes while the rest sits idle so you can shrink a model a lot before it gets meaningfully dumber: there are three main ways 1) Distillation (teacher → student) you train a small model on a big model’s answers if you own the teacher, the student doesn’t just learn “the answer is cat” but also sees the teacher’s full guess: 90% cat, 8% dog, 2% fox that’s far more signal per example, so the student picks up the teacher’s judgment, not just its outputs 2) Pruning you cut the useless weights, i.e. both individual neurons and even whole layers that barely matter, then retrain briefly so the model "heals" researchers have removed half the weights from 175B-parameter models in one shot with almost no measurable loss 3. Quantization each weight is trained as a 16-bit number but you can round it to 4 bits and the model will need a QUARTER of the memory with almost no loss in performance a 27B model goes from ~54GB (may need data center) to ~14GB (need just your laptop), and a 3B model at 4-bit is ~1.5GB, which fits on a phone stack all three (plus a few other technical tricks) and you get a 27B model that runs on a laptop, beating Opus 4.6 on agentic coding, at well under 1/100th of Opus’s rumored size. WHY THIS MATTERS !!Consumer AI doesn’t need to live in a data center!! MOST of what people actually use AI for (emails, summaries, search, docs, basic coding) is well within reach of a model on your device there’s also no good reason to keep sending your medical questions, finances and DMs to someone else’s server to get it I’m personally still long AI, but I DO NOT think everyday consumer or business demand is a reason to be long data centers I’m almost certain ~95% of daily usage ends up running locally (try @askloci for a week and you’ll see how close it already is to frontier models) the ONLY reason I’d personally stay long chips and data centers is recursive self-improvement into artificial superintelligence i.e. fast takeoff and building the machine god if you believe that happens in 2027, or within the next few years, then the capex trade is still the bargain of the century if you don’t, you really should sell your AI infra because without the machine god, most of that buildout is worth only a fraction of what you’re paying