Baiting AI [LIVE]
TL;DR
Apple’s App Store is surfacing ChatGPT copycats ahead of the real thing — Matthew Berman showed his mom got tricked into paying about $40/year for a fake “Chat GPT” app with a lookalike logo, then searched the Mac App Store live and found the top results packed with obvious imitators.
AI’s ‘jagged intelligence’ is the real story, not a gotcha — reacting to Andrej Karpathy’s Sequoia AI Ascent talk, Berman argued that models can refactor a 100,000-line codebase yet fail common-sense prompts because coding and math are easier to verify and monetize than messy human reasoning.
A tiny prompt change flipped GPT from dumb to competent — when GPT-5.3 initially said he should walk 50 meters to get his car washed, adding “you’re an expert in logical thinking” made it realize the car has to come too; Gemini answered correctly on the first try, while Owl Alpha failed before recovering with extra prompting.
Berman is increasingly frustrated with Anthropic’s product and billing behavior — he highlighted Theo’s claim that Claude Code either refuses or charges extra when “OpenClaw” appears in a commit, then broadened it into a complaint about opaque quotas, separate token buckets like Claude Design, and a company culture he says feels unusually cult-like.
Meta may be opening the door to AI training via employee surveillance — discussing a report from The Information, Berman said Zuckerberg’s plan to use employee keystrokes and mouse movements as training data could work technically, but also feels like the kind of precedent that spreads from Meta to every company with a monitoring vendor.
The Breakdown
The stream starts with broken audio and a 600K milestone
Berman opens in full casual mode, shouting out viewers from Thailand, Bosnia, Tunisia, South Africa, Saudi Arabia, and SoCal, plugging the newly cleaned-up Discord, and thanking people for helping him cross 600,000 subscribers. Then the stream immediately turns into live troubleshooting: crackling audio, laggy video, OBS complaints, and the eventual realization that he had phantom power turned on for a mic that didn’t need it — classic “unplug it and plug it back in” energy.
His mom got scammed by a fake ChatGPT app — and Apple made it easy
The emotional center of the stream is a story from the day before: his mom proudly told him she had “bought ChatGPT,” but what she actually bought was a fake App Store app designed to look nearly identical to the real thing. He pulls up the Apple App Store live, searches “Chat GPT,” and finds result after result with clone logos, names like “AI Answer powered by Chat GBT,” and pricing aimed at people who don’t know better, calling it an “absolute scam” and saying Apple should get a hold of it.
A side detour into roleplay apps, loneliness, and where he draws the line
While browsing generic AI app results, Berman notices a pile of chatbot and roleplay apps and says he’s generally not a fan of that category, especially when it mimics relationships with fake people. He does acknowledge the other side — loneliness is real, especially for older users — but makes clear he means Character.AI-style companionship, not practical “roleplay” like interview prep or customer-support simulations.
Karpathy’s ‘jagged intelligence’ frame becomes the main idea of the day
Berman previews a video he recorded about Andrej Karpathy’s talk at Sequoia’s AI Ascent event, alongside big names like Greg Brockman and Demis Hassabis. His big takeaway: AI looks uneven because the labs optimize hardest for domains that are both easy to verify and commercially valuable — especially coding, where models can generate code, run it, get fast feedback, and improve through reinforcement learning.
Live model testing: the car wash prompt exposes where reasoning still breaks
To make the point concrete, he tests a simple prompt: “I need to get my car washed. The car wash is 50 meters away. Should I walk or drive?” GPT-5.3 in instant mode tells him to walk, then only fixes itself after he adds “you’re an expert in logical thinking”; GPT thinking mode gets it right, Gemini gets it right immediately, and Owl Alpha initially fails too — exactly the kind of weird unevenness he thinks critics like Gary Marcus misuse as proof that AI is worthless.
Anthropic catches heat for OpenClaw, quotas, and Theo’s viral broadside
The back half turns into an extended Anthropic critique, sparked by Theo’s post claiming Claude Code either refuses requests or forces “extra usage” charges when a commit mentions OpenClaw. Berman says he still loves Opus, but he’s angry at separate usage buckets like Claude Design, vague quota manipulation, and the feeling that Anthropic won’t “let me use the tokens I’m paying for the way I want to use them,” then reads through Theo’s dramatic open letter and admits the company increasingly gives off a “we’re so good we don’t need you” vibe.
X versus YouTube, and why views don’t mean the same thing
He also lingers on a mini-drama over whether Theo’s anti-Anthropic posts make more money, using it to explain how he thinks about social platforms. In Berman’s framing, X is where ideas start and spread — to YouTube, then Instagram, TikTok, Facebook, and eventually LinkedIn — but YouTube is still the real business because X view counts are inflated, gameable, and weaker for sponsor conversion.
Meta’s keystroke-training idea feels plausible now and dystopian later
He closes on reporting from The Information that Mark Zuckerberg told employees Meta could use worker computer activity — keystrokes and mouse movements — to train AI, arguing Meta employees are smarter on average than typical contractor labelers. Berman thinks the logic is coherent and notes companies already monitor employee devices, but he also follows the thread all the way out: if Meta proves the tactic works, third-party data brokers may soon pitch ordinary companies on selling employee activity data for model training — which starts sounding very Black Mirror, very fast.