Mimi

Introducing ChatGPT for competitor research. Drop in any rival’s website. GrowthPeel shows you exactly how they grew: → The exact reels, with views and when they were posted → The exact ads they’re running → The exact timeline of what they did and when We ran it on Lovable. Their biggest reel has 4.3M plays, and a creator made it.

THIS GUY HAS BEEN BUILDING AN ENTIRE GAME WITH CLAUDE OPUS FOR 15 WEEKS. THE AI WRITES THE CODE, BUILDS THE CITIES, CREATES 3D ASSETS, AND EVEN EDITS THE TRAILERS. it's called Nightshoal, a Godot game set in a 1920s world of harbours, fishing boats, bridge cities and black markets. he's been building it with Claude Opus, posting weekly progress on Reddit, and there's already a free playable demo. what's insane is how much of the actual game is being built by AI. the world has: a massive bridge city with moving traffic, trams and a lighthouse a sun that actually moves across the sky, changing shadows and water reflections a moon that takes over at night, with buildings lighting up an entire black market with steamers, warehouses, piers and fishing boats harbours, docking scenes, weather, achievements and more and here's how he builds it: Claude Opus 5.5 writes basically all the code. it also controls Blender through MCP to build city layouts, buildings and fix 3D models. TripoAI generates the 3D assets. he makes images showing all four sides of an object, Claude splits them up, Tripo turns them into a model, and Claude cleans it up in Blender. OpenAI's image API generated 24 reference mockups of the city, showing 6 different views at 4 times of day. then he built something called a Gauntlet-Loop. AI critic agents blindly compare screenshots of his actual game against those mockups and keep improving the game until it wins. after 10 rounds, 31 out of 33 pieces beat the reference images. even the performance work is done by Opus. this week it found a map using 4 GB of graphics memory because it was being drawn at double size. Opus helped bring that down to 0.3 GB and cut startup time from 26 seconds to 13. the first harbour docking freeze went from almost 2 seconds to 0.2. he also has another system called Jev that decides how much reasoning effort Opus should spend on each task. the crazy part? he spent around $1,100-$1,150 on AI tools just this week, including four Claude Max plans. 15 weeks in, he's already got a playable game, entire AI-built environments, dynamic lighting, weather and a working world. and he's still building the actual gameplay. links below 👇

THIS GUY MADE ANDROID AND MAC SWAP FILES WITH NO CABLE, NO ACCOUNT AND NO CLOUD. YOU JUST TAP THE PHONE ON THE NOTCH it's called TapDrop and it's open source under MIT how it works: the phone flashes its screen, the light sensor next to the MacBook notch picks it up, and the mac grabs whatever is on the phone's screen what gets sent stays on your wifi and is end-to-end encrypted. there's nothing to sign up for the proof: he tried it on pixel, samsung, vivo, oppo and nothing phone and says it worked on all of them setup: > install the mac app > install the android app (android 13+) > any mac with a notch works what you can send: > pdfs, links, original photos and videos > any type of file > sms otps, copied straight to the mac > mac to phone works too the funny part: it started because he didn't want to email a pdf to himself one more time file transfer used to mean cables, accounts and a cloud in the middle now it's a phone touching a screen cutout links below 👇

THIS GUY CUT THE SCREENSHOTS OUT OF HIS COMPUTER-USE AGENT. THE MODEL JUST PICKS A NUMBER it's called ywza and it's open source most computer agents send the model a fresh screenshot at every step. ywza describes each window as a numbered list of elements instead: buttons, fields, text the model says "click 14" and that's it why that matters: > far fewer tokens per step > less for the model to reason about > still works on any window, whatever technology it's built with > can still ask for a screenshot when it really needs one the numbered list comes from smart detection that runs on the window itself, so it doesn't care if the app is Qt, a terminal or something old it runs as a small background agent on the PC you want to automate, and it works in the terminals you already have open it's written in Rust using pictures to read a screen was always the expensive way turns out a menu is easier to pick from when somebody numbers the items links below 👇

THIS GUY TRAINED A 1.1 GB MODEL ON AN RTX 4060 LAPTOP TO SAY "NOT FOUND" INSTEAD OF MAKING THINGS UP it's called Obsessive-1.7B, a fine-tune of LFM3-1.7B with exactly one job it's a subagent, not a chat model. a bigger model hands it a factual question, it calls web search once, and it answers with a short value, a word-for-word quote and the URL. if the answer isn't in the results, it says "Not found in the search results." example: asked who was the second woman to win the Turing Award, it answers Barbara Liskov, quotes the line, and links the Wikipedia page the training: > LoRA rank 16, about 9k tool-calling conversations > search results are Wikipedia and FineWeb pages > every answer and quote checked to appear verbatim > about a quarter of the examples teach "not found" > a local qwen model helped prepare the data and questions the test: 131 held-out SimpleQA questions, and in 29 of them the answer was removed from the results on purpose the honest part: on live Wikipedia search the base model actually gets more answers right, but 62% of the time it runs out of turns and returns nothing. Obsessive only states what it can quote it still says "not found" for about 1 in 10 answers that are really there, and it only handles one question per conversation everyone trains small models to answer more this one is trained to shut up when it can't prove it links below 👇

THIS GUY HAD CLAUDE PULL THE AI MODEL OUT OF AN ORAL-B TOOTHBRUSH APP AND SHRINK IT 100X WITH BARELY ANY LOSS the iO5 has a motion sensor inside that tells the app which part of your mouth you're brushing. the guessing is done by a 1.2M parameter neural net, a bi-directional GRU (a net that reads a sequence of readings forwards and backwards) claude extracted that model from the companion app, then tested it on a public dataset: ~63% accuracy on the mouth zone, where random guessing gets ~6% then the guy asked claude if it was overkill the ablation (remove a piece, see what breaks): > most of the value came from just the last ~10 seconds of motion > so a giant recurrent net wasn't needed > claude used real examples to make a larger synthetic dataset > it trained a bunch of small models on it the result: a small CNN hit ~61% and is 100x smaller he also has a replay of one dataset session running on a live 3D twin of the brush, in the video the toothbrush company shipped a big model for a job a tiny one does almost as well now one person and claude took it apart in a side project links below 👇

THIS GUY GOT 215 TOK/S ON CODE OUT OF A 2-BIT DEEPSEEK RUNNING ON EIGHT V100S FROM 2018. 524K CONTEXT, ALL EXPERTS IN VRAM it's an open source repo called Volta-sPLIt, a port of 1Cat-vLLM work to 8x V100-SXM2 the poster is a lawyer, and a few days earlier the same model ran at 36-38 tok/s on four cards. then came custom kernel work, prefill tuning, and four more cards the speed, measured with thinking off: > code: 215 tok/s at 1K, 200 at 32K, 183 at 128K, 160 at 256K > JSON: 196-209 from 1K to 128K > legal writing: 118 at 1K, 111 at 128K, 103 at 256K why code is faster: the model guesses five tokens ahead each round, and code is easy to guess. 3.6-3.8 guesses land per round on code, about 1.6 on legal text the box: > 8x V100-SXM2-32GB on two NVLink boards > Threadripper PRO 3945WX, 256 GB DDR4 > 150 W per card > 110 GB of Engram tables live in system RAM, about 0.12 ms per round to gather the slow part is the front: prefill runs about 2,200 tok/s at 128K, so the first token takes 59 s there and about 2 minutes at 256K everyone says you need a new rack of H100s for this turns out a pile of 2018 cards and a lot of patience gets you most of the way links below 👇