Monday, Aug 03
Presented byFramer 3.0Your phone is watching your squat
gm legends, happy Monday.
Today: the camera in your pocket doing a personal trainer's job, a Mac app that admits what it can't do, and a use for the ninety seconds you spend staring at Claude Code.
Daily Pulse
One developer burned 69 billion tokens in July, roughly $87,000 of API spend across twelve extra Claude Max accounts, to run eighteen agents against a single codebase. His conclusion after an earlier attempt collapsed: agent harnesses have to be bonded into the application, not bought as a framework. (Steve Yegge)
An unreleased OpenAI model solved ten math problems that had gone a decade or more without progress, then wrote each argument into a Lean proof a computer can check line by line. The tokens behind all ten would cost roughly $2,000 at current API rates. (OpenAI)
An AI chatbot out-charmed human scammers, talking nearly half of 22 test subjects into installing an app after a week, versus fewer than one in five for the humans. Running the model through the long trust-building phase and handing off to a person for the final ask slips past vendor safeguards entirely. (Ars Technica)
Language models agree with each other more than twice as much as two human readers do, and no model matched readers better than another reader did. Across 18 models from 11 vendors tested against 2,523 reader highlight sets, only the smallest models picked sentences at human-level agreement. (arXiv)
Microsoft cannot patch as fast as AI finds bugs: Anthropic’s Mythos model turned up 90 critical and 141 important SharePoint flaws in April alone, most of them still unfixed weeks later. July 14 brought patches for more than 600 bugs, an all-time record against June’s previous high of about 200. (ProPublica)
Nobody watches how you lift
Mouhab Tarek kept watching people train hard and get hurt anyway, because trainers are expensive and nobody else is looking at how you move. So he built CoachAI, which uses the iPhone camera to catch the rep where your form goes, all on-device. He flipped the onboarding too, so you do one tracked rep before it asks you anything.
🔥 Our Take: Gym injuries come from nobody watching you, and the fix has always been paying someone to stand there. A phone that can really see a rep break down removes the last piece of training that needed a person in the room. Shipping it rough is a brave call for an app whose job is telling you your back is fine.
Two accounts, two apps
MacDupl clones a Mac app into a second copy with its own login, data and Dock icon, so work and personal stop logging each other out. Sigit Kusuma gives each clone a separate bundle ID and a mock keychain, keeping the credentials away from the real macOS one. About two seconds a clone, $19 once, no telemetry.
🔥 Our Take: Cloning an app so two accounts stop fighting is a small thing. Doing it honestly is not, and he could have claimed the whole App Store works, then let you discover which clones quietly share data with the original. He cut the sandboxed ones and published the limits, which tells you more than a longer compatibility page would.
Your canvas just learned to design
Pokémon while Claude thinks
Sergio Zamarro spends half his day watching Claude Code work, never long enough to switch tasks and always too long to just sit there. So he built claudemon, a Pokémon game that only moves while Claude does: your prompts are steps through the grass, and something eventually jumps out at you. MIT, free, fully local.
🔥 Our Take: The dead time while an agent works is new, too short to start anything else and too long to stare at a spinner. Sergio filled his with a Pokémon that steps only when Claude does, which is more honest than pretending you'd use those ninety seconds well. Nobody else treats that gap as something to enjoy, the rest of the tooling just wants it gone.
Daily Top Products
Monday through Friday
Our ultra-fast Daily: Three takes on new products. Yesterday’s top ten launches. That’s it.