Kodzilla @k0dzilla

@Beaver_0x 70ms decisions sound great until slippage eats the entire edge i’d love to see the live P&L after fees, not just the response time

@AleiahLock Elon watching people buy Grok Bot just to ask it questions they could’ve Googled in 5 seconds 💀

@sanc_cash MCP is getting a lot more interesting now that AI can actually work across your apps but Apple shipping native support and someone connecting a few tools are two very different things

@Kuzoxbt people will forget a paragraph about drone physics in 10 seconds give them a working flip button and they’ll spend 20 minutes figuring it out themselves

@aifindr calling something open source when it’s only free to use is exactly the kind of mistake AI should catch before 100k people read it 2 unsupported claims out of 27 is a pretty good argument for running this check before posting

@itsarcvale 17 Grok bots sounds impressive, but i’d rather see how many hours they actually save each morning four useful briefs beat 17 bots generating noise

@0xzivo 0.7% of Google’s compute recovered by changing a scheduling rule imagine how much hardware was being wasted simply because CPU and memory weren’t packed efficiently

@Newtonixe 90 FPS Mario 64 floating above your desk is wild even wilder that the hardest part to replicate isn’t the code, it’s getting your hands on those PICO glasses

@0xCortiq 135 photo accesses and 147 server connections sound scary until you realize those are two different metrics show which app accessed what and when before calling it surveillance

@0xGoldTooth 1.7 million infected Apple devices and a virus that jumps over Bluetooth without a tap? that’s a global security emergency, yet somehow Claude Code is the only one that knows about it 💀

@elixlune 23,614 recordings and a new file appearing 7 seconds ago? show the file paths and the process accessing the camera before we call this a macOS surveillance scandal

@kodera_ai using Opus to debug every broken frame is like hiring a CTO to fix CSS much smarter to let Haiku find the mess and bring Opus in when judgment actually matters

@AIBoticssq circular funding can keep the buildout moving, but it can’t manufacture demand eventually those $145B in CapEx need to earn a return outside the loop

I TURNED GROK INTO A SECOND BRAIN THAT CHECKS ITSELF. Imagine dropping a PDF, a link, or a voice note into a chat and never having to explain that information to AI again. The setup uses 3 specialized agents: 1. READER Collects links, PDFs, and voice notes. Saves the original material without changing it, so nothing gets lost or distorted. 2. LIBRARIAN Turns incoming information into organized wiki pages. Updates related topics, connects new findings to existing knowledge, and saves useful answers for later. 3. CHECKER Runs a nightly review. Looks for contradictions, outdated claims, missing sources, and pages that haven’t been verified recently. Here’s where it gets interesting: → Ask a question, and a useful answer can become a new wiki page. → Add a source that contradicts an older claim, and the system flags the conflict instead of silently rewriting history. → Leave a page without fresh evidence for 30 days, and it gets marked as unconfirmed. Every knowledge page keeps three things: • What it claims • Where the information came from • When it was last verified Want to try the same setup? Paste this into Grok Bot: “Design a second-brain system with three agents in one group chat: Reader, Librarian, and Checker. Reader must preserve all original sources in raw/. Librarian must maintain wiki/ pages, connect related information, and save useful answers as new pages. Checker must review the knowledge base nightly for contradictions, stale claims, and missing references. Every page must include its claim, source, and last verification date. Show me the proposed structure first. Do not create anything until I approve it.” Start with one subject you actually work on. Give your AI a memory it can inspect, update, and question.

@Shadetheory 50 rounds of review and the AI still wants another lap 😭 at some point someone has to tell the bots the code is shipping

ONE PULL REQUEST WENT THROUGH 50+ AI REVIEW CYCLES OVERNIGHT. Claude reportedly writes around 90% of the code for Claude Code. Its monorepo went from roughly 2,500 changes a day in May to more than 30,000 in September. But one pull request exposed a problem that more automation couldn’t solve. An AI author kept fixing issues raised by an AI reviewer. The reviewer kept finding more. They went back and forth over 50 times in one night, until a human manually stopped the loop. When the team investigated the repeated findings, roughly 59% involved code that had already existed since the first pass. Their response was to cap reviews at 3 cycles. If problems remain, the author must reconsider the approach instead of endlessly patching individual findings. Subsequent reviews focus on the new diff. Meanwhile, merged PRs more than doubled, the revert rate fell by over 40%, and PRs receiving substantive review comments rose from 16% to 54% between June and September. The lesson goes beyond code review. As AI agents become capable of working around the clock, the expensive failure isn’t always a bad answer. Sometimes it’s two capable agents repeatedly doing exactly what they were told, with no rule telling either one to stop.

Addy Osmani @addyosmani

3 days ago

x.com/i/article/2114…

395 2K 13K 2.4M 34K

@mrfokki the worst part is finding out your two-year roadmap is now someone else’s dropdown menu 😭

@mrfokki paying $200 a month just to start planning your prompts around a usage bar is a pretty wild subscription experience

@mrfokki exactly, change a few rules and suddenly the entire world behaves differently level design is starting to look a lot more like programming

I USED AI TO BUILD A 3D WORLD WITHOUT TRADITIONAL MODELING. I combined Claude + Duet to create this interactive environment. Floating cubes. Animated chains. Dynamic lighting. Physics. Everything is generated and controlled through procedural code. Instead of manually placing hundreds of objects, I defined the rules that control them. Change the geometry, adjust the physics, rewrite the lighting — the entire world updates. No traditional 3D modeling workflow. Just AI, code, and a world that responds to the rules I create.

Kodzilla @k0dzilla

6 days ago

CLAUDE ASKS PERMISSION. YOU CHECK THE CONSEQUENCES. The recipient. The amount. The attachment. A practical guide to what you should check before approving—and five metrics to track during a 14-day test. x.com/i/article/2113…

5 1 24 4K 4

@HeyAIpenguin for the bigger changes, I kept approval in the loop didn’t want an agent getting too creative with the physics 😅