Tokens & Signals for 6/5/2026. We scanned ~1,200 Twitter accounts (1117 tweets), 13 subreddits (61 posts), Hacker News (13 stories), 3 newsletter posts, 5 podcast episodes, 245 Discord messages, and leaderboard data for you. Estimated reading time saved: ~11 hours.
* Anthropic says Claude now writes 80% of their internal code — and engineers are shipping 8x more than their 2025 baseline because of it. reddit.com/r/ClaudeAI/comments/1txisil/claude_n...
* Google dropped Gemma 4 with quantization-aware training (QAT), which basically means high-end model performance is now runnable on your local machine. x.com/GergelyOrosz/status/2062861559009820976
* Nvidia released the 550B Nemotron 3 Ultra as open weights — a massive model built specifically for agentic reasoning and long-context tasks. x.com/TheAhmadOsman/status/2062689767180153022
* SpaceX and Google quietly signed a $920M/month deal for 110,000 GPUs, making SpaceX one of the biggest cloud compute providers on the planet. reddit.com/r/singularity/comments/1txve3j/googl...
* Cloudflare confirmed that automated AI bot traffic has officially passed human traffic on the internet — and it happened ahead of schedule. reddit.com/r/OpenAI/comments/1txh6yx/bots_have_...
* @GaryMarcus on Anthropic's RSI warning: "The industry is starting to sweat the risks of recursive self-improvement as models get better at coding their own upgrades." x.com/GaryMarcus/status/2062699408974856265
* @karpathy on the new Stanford study: "Two bad agents don't make a good team. The problem is they optimize for different local objectives." x.com/karpathy/status/2062761794028875894
* Sakana AI is opening a dedicated Recursive Self-Improvement (RSI) Lab in Tokyo to study what happens when AI starts evolving itself. x.com/SakanaAILabs/status/2062948403815030850
* Nvidia open-sourced the Rubin NVSwitch BoM — and buried in the details is a surprising reliance on AMD EPYC 3151 CPUs in every rack. x.com/SemiAnalysis_/status/2062720812042371558
* Kling AI hit 100 million users and 50,000 enterprise customers in just two years. x.com/Kling_ai/status/2062912327385575895
Best to Build With Today
* Coding — claude-opus-4-8-xhigh-effort (LiveBench leader).
* Reasoning — gemini-3.1-pro (Arena ELO leader).
* Chat — gemini-3.1-pro (Arena ELO leader).
* Open-source — gemma-4-31b-it-qat (Best for high-fidelity local inference).
* Value pick — gemini-2.5-flash-preview (Top-tier performance at a fraction of the cost).
Deeper Dives
💼 Industry & Business
Anthropic: Claude writes 80% of internal code
Anthropic engineers are now shipping 8x more code than their 2025 baseline, with 80% of it written by Claude. This isn't a demo or a case study — it's internal data showing that LLMs have become the primary engine for high-level architecture work and debugging. That's a pretty wild thing to hear from the lab that built the model. 📱 Twitter · 💬 Reddit
SpaceX & Google sign $920M/month compute deal
SpaceX is now, effectively, a cloud provider. They're renting 110,000 NVIDIA GPUs to Google for $920 million a month — $26 billion annualized. The fact that a rocket company is becoming a major player in AI infrastructure tells you everything about where the money is flowing right now. 📱 Twitter · 💬 Reddit
AI agent traffic overtakes humans
Cloudflare confirmed it: bot traffic has officially passed human internet usage. And this isn't just scrapers running wild — it's the broader shift toward an agent-first web. If you manage infrastructure, things are about to get a lot more interesting. 💬 Reddit · 🎙️ Podcast
🧠 Models & Research
Stanford: AI coding agents are bad at teamwork
A Stanford study (CooperBench) found that multi-agent coding teams perform roughly 50% worse than solo agents. The culprit? Coordination breakdowns and agents that can't resolve conflicts. Turns out "just throw more agents at it" is not, in fact, a strategy. 📱 Twitter
Nvidia releases 550B Nemotron 3 Ultra
Nvidia's new 550B sparse MoE model has a 1M token context window and a hybrid Mamba-Transformer architecture. It's built for complex multi-turn reasoning and agentic workflows, and it's available right now under an open-weight license. Hard to complain about that. 📱 Twitter · 💬 Reddit · 🎙️ Podcast
Google releases Gemma 4 with QAT
Google's new Gemma 4 checkpoints (31B/26B MoE) use Quantization-Aware Training to preserve near-bfloat16 quality at 4-bit precision. Translation: high-end reasoning on consumer hardware is now genuinely within reach. 📱 Twitter · 💬 Reddit
Sakana AI launches RSI Lab
Sakana AI is opening a Tokyo lab focused entirely on autonomous AI evolution. Using evolutionary algorithms, they're researching how AI systems can redesign their own architecture without a human in the loop. Whether that's exciting or alarming probably depends on your disposition. 📱 Twitter
Anthropic: RSI risk warnings
Anthropic is publicly flagging the dangers of Recursive Self-Improvement and calling for a global pause on certain research tracks. It's sparked real debate about whether frontier labs are being genuinely responsible — or just doing PR. Probably worth watching how this one plays out. 📱 Twitter · 💬 Reddit
Physics: Entanglement builds space-time
New research suggests quantum entanglement might literally be what constructs space-time geometry. Very theoretical, but researchers thinking about the long-term future of physical computation are paying close attention. 🔶 Hacker News
🚀 Products & Launches
* Kling AI — Celebrated 2 years, 100M users, and 26 model iterations.
* Rubin NVSwitch BoM — Nvidia released hardware diagrams revealing a surprise reliance on AMD EPYC CPUs.
Closing thought: The industry is splitting in real time. Labs like Anthropic are pumping the brakes on recursive self-improvement over existential concerns, while others like Sakana are leaning all the way in. Meanwhile, the internet has officially flipped — it's a bot-first environment now. We're not just building AI anymore. We're living inside it.