『📅 ThursdAI - Apr 2 - Gemma 4 is the new LLama, Claude Code Leak, OpenAI raises $122B & more AI news』のカバーアート

📅 ThursdAI - Apr 2 - Gemma 4 is the new LLama, Claude Code Leak, OpenAI raises $122B & more AI news

📅 ThursdAI - Apr 2 - Gemma 4 is the new LLama, Claude Code Leak, OpenAI raises $122B & more AI news

無料で聴く

ポッドキャストの詳細を見る
Hey Ya’ll, Alex here, let me catch you up. What a week! Anthropic is in the spotlight again, first with #SessionGate, then with the whole Claude Code source code leak, and finally with an incredible research into LLM having feelings!? (more on this below). And while Anthropic continues to burn through developer good will faster than their sessions, OpenAI announced a MASSIVE $122B round of funding (largest in history), Google released Gemma 4 with Apache 2 license - we had Omar Sanseviero on the show to help us cover what’s new, Microsoft dropped 3 new AI models (not LLMs) and PrismML potentially revolutionized local LLM inference with lossless 1-bit quantization! P.S - Oh also, something on X algo changed, I get way more exposure now, 3 out of my best 5 posts ever have been from this week + I got the coveted Elon RT on my Claude Code leak coverage. I’ll try to stay humble 😂 Anyway, let’s dive in, don’t forget to hit like or share with friends, and TL;DR with links is as always, at the bottom: ThursdAI - Highest signal weekly AI news show is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.The Claude Code source Leak: Half a Million Lines of “Oops”So here’s what happened. On March 31st, Anthropic shipped Claude Code version 2.1.88 to npm. Inside that package was a 59.8 megabyte source map file — basically a debugging artifact that contained the entire compiled source code. 512,000 lines of TypeScript across 1,900 files. The entire playbook for how the Claude Code harness works, including a lot of stuff that wasn’t supposed to be public yet.A researcher named Chaofan Shou spotted it at 4 AM ET, posted the download link, Sigrid (who came to the show) posted it on Github and within six hours it had 3 million views and 41,000 GitHub forks (This repo is the highest starred repo in Github history btw, with well over 150K Github stars). Anthropic started filing takedowns, but the internet being the internet, it was already everywhere. The source code is still on tens of thousands of computers right now. (I won’t link directly but there’s a website called Gitlawb, look it up) The community went absolutely wild digging through the source code btw, and they found some interesting things!KAIROS: Claude Code is going to become a Proactive Agent!This is the biggest take-away from this leak IMO, that like OpenClaw/Hermes agentic harnesses, Claude Code is already a fully featured proactive agent, we just don’t have access to this yet. With KAIROS, Claude Code will have it’s own daemon (will run independently from the CLI), will have a background ping system (hello Heartbeat.md from OpenClaw) that will make it wakeup and do stuff, will do “autodream” memory consolidation reviewing your daily sessions and fix memories, subscribe to Github, and maintain daily appent-only logs to show you what it did while it and you were asleep. This is by far the hugest thing, I’m excited to see how / when they ship KAIROS, as I said, 2026 is the year of Proactive agents! My Wolfred OpenClaw agent summed it up very nicely: Undercover ModeFor Anthropic employees working on public repos, there’s an Undercover Mode that auto-activates and strips all AI attribution from commits. The system prompt? “Do not blow your cover.” They really said “this is fine” about shipping internal tools to production while hiding from the world that AI wrote the code. Which, honestly, is kind of incredible meta-humor from whoever wrote that.The Buddy SystemMy personal favorite discovery: there’s a hidden Tamagotchi-style terminal pet called the Buddy System with 18 obfuscated species, rarity tiers (including a 1% legendary), cosmetic hats, shiny variants, and stats like DEBUGGING, PATIENCE, and CHAOS. If you activate it now, you can do /buddy and you’ll have a little companion judging your coding decisions. Anthropic shipped a game inside their CLI tool. Mine is called Vexrind and he’s sarcastic as f**k, I’m not sure I like it. Anti-Distillation ProtectionsThe code also revealed that Claude Code injects fake tool calls into logs to poison training datasets. If you’ve been backing up your .claw folders to train on the data; Stop. Pass your data through something like Qwen or make sure you’re filtering out the noise. (a Nisten tip)The Models That Don’t Exist YetBuried in the code are references to Opus 4.7, Sonnet 4.8, and a model called capybara-v2-fast with a 1 million context window. These haven’t been released. This is yet another confirmation of the leaked “Mythos” model that’s coming soon from Anthropic. Which btw, with Anthropic very rocky uptime lately, the tons of SessionGate issues, the leaked blog announcing Mythos, the leaked Claude Code oopsie, they are not having the best Q1 in terms of proving to the world that they are the safest lab out there. I hope they protect their weights better than they protect ...
adbl_web_anon_alc_button_suppression_c
まだレビューはありません