Tokens & Signals for 8/11/2026. We scanned ~1,200 Twitter accounts (1361 tweets), 13 subreddits (63 posts), Hacker News (14 stories), 5 newsletter posts, 7 podcast episodes, 126 Discord messages, and leaderboard data for you. Estimated reading time saved: ~13 hours.
* SpaceXAI dropped "Grok Bot," a platform where AI agents live on virtual machines and autonomously handle your emails, coding, and scheduling. It's live now for Grok Premium/Premium+ subscribers. x.com/MTSlive/status/2087227768119673022
* NVIDIA released Nemotron 3.5 Lightning (30B), an MoE model that's 35% faster than Qwen3.6 35B — and it's built specifically for running agents on your own hardware. x.com/JensenHuang/status/2087184542050496763
* OpenAI finally brought the ChatGPT desktop app to Linux, with support for Ubuntu, Debian, and Fedora and native chat syncing out of the box. x.com/thsottiaux/status/2087254026232775052
* Anthropic is watermarking all Claude output to comply with EU AI Act transparency rules. Power users are already pushing back hard on what that means for ownership. x.com/kimmonismus/status/2087102975106019579
* @scaling01 on researchers extracting internal CoT reasoning: "Security teams are going to have a nightmare with side-channel attacks on hidden model logic." x.com/scaling01/status/2087202712710513041
* Unsloth Desktop just made training models like FLUX or LTX something you can do on Mac, Windows, or Linux through a simple GUI — no command line required. reddit.com/r/LocalLLaMA/comments/1vlj87v/introd...
* @paulg on industry turnover: "The revolving door in safety and ethics roles at major labs is becoming the rule rather than the exception." x.com/paulg/status/2086924675163345285
* Anthropic locked in a massive 20-year, $9.1B data center lease with Riot Platforms in Texas to scale their compute. x.com/MTSlive/status/2086973121555566600
* LTX 2.5 is here, with a reworked architecture that lets you generate consistent, multi-shot scenes in a single pass. reddit.com/r/StableDiffusion/comments/1vlqy46/l...
* Gemini hit 1 billion monthly users, making it the fastest-growing product in Google history. x.com/GeminiApp/status/2087229532382347277
Best to Build With Today
* Coding — claude-opus-4-8-xhigh-effort (Currently holds the lead for complex agentic tasks).
* Reasoning — gemini-3.1-pro (The top choice for raw math and logic benchmarks).
* Chat — gemini-3.1-pro (The general favorite for daily use).
* Video generation — LTX 2.5 (The new gold standard for consistent, multi-shot scenes).
* Open-source — NVIDIA-Nemotron-3.5-Lightning-30B (Unbeatable speed for local agentic workflows).
Deeper Dives
🧠 Models & Research
Researchers Extract Hidden Reasoning Traces
Researchers pulled off a side-channel attack that reconstructs hidden "chain-of-thought" (CoT) reasoning from proprietary APIs — just by analyzing latency and token probabilities. The cost? About $0.0007 per task to uncover the internal logic that labs like OpenAI and Anthropic actively try to keep private.
Why it matters: It proves "black-box" models aren't really black boxes. Hidden reasoning can be stolen, and that blows a pretty big hole in how labs think about model security.
� Twitter� Reddit� Hacker News
NVIDIA Launches Nemotron 3.5 Lightning
This 30B Mixture-of-Experts model only activates 3B parameters per forward pass, which is how it manages to run 35% faster than Qwen3.6 35B. It's specifically tuned for tool-calling and long-running agent tasks on local RTX hardware.
Why it matters: It's a high-speed engine for developers who want to run agents locally without needing enterprise-grade compute.
� Twitter� Reddit� Hacker News
Meta Releases Muse Glimmer 30B
Meta's new open-weights model uses a novel "DFlash" drafter to push generation speeds 2-4x faster. It brings serious creative synthesis capabilities down to consumer hardware.
Why it matters: It keeps pushing the ceiling on what open-weights models can actually do at the edge.
� Twitter� Reddit
🚀 Products & Launches
SpaceXAI Launches 'Grok Bot'
Grok Bot lets Premium+ subscribers deploy autonomous agents that live on virtual computers — logging into apps, managing emails, and coordinating multi-step workflows around the clock.
Why it matters: We're shifting from AI chatbots to persistent digital coworkers that just... keep going.
� Twitter
OpenAI Releases ChatGPT Desktop for Linux
After years of requests, OpenAI finally shipped a native desktop app for Ubuntu, Debian, and Fedora. You get full access to GPT-4o and o1-preview with native hotkeys included.
Why it matters: Linux developers have been waiting on this for a long time. The workflow gap is finally closed.
� Twitter
💼 Industry & Business
Anthropic's Multi-Billion Data Center Deal
Anthropic signed a 20-year, $9.1 billion agreement with Riot Platforms for 191 megawatts of data center capacity in Texas. That's a serious long-term bet on owning the physical infrastructure needed for future training runs.
Why it matters: It's a clear signal that Anthropic is playing the long game on compute — right alongside Google and OpenAI.
� Twitter
Anthropic's Ethics Head Resigns
Following a strategic shift toward intensive biology and health research, Anthropic's head of ethics has stepped down.
Why it matters: The churn in safety roles keeps happening. It points to real, ongoing tension between moving fast on products and keeping governance intact.
� Hacker News
Nvidia Assembles $500B AI Infrastructure Fund
Nvidia is teaming up with financial heavyweights like BlackRock and Apollo to mobilize $500B for data centers and energy infrastructure.
Why it matters: The serious money is now flowing toward securing the physical backbone of the AI era.
� Reddit
Funding & Deals
* River AI raised $1.1B at a $5B valuation, led by General Catalyst and AMP, to build out inference infrastructure.
* Anthropic secured a 20-year, $9.1B lease to scale proprietary compute capacity.
Launches
* Grok Bot — Autonomous agent platform for persistent AI coworkers.
* LTX 2.5 — Video model capable of multi-shot consistent scenes in one pass.
* Nemotron 3.5 Lightning — High-speed 30B agentic model from NVIDIA.
Closing thought: Between the $500B infrastructure fund and Anthropic's multi-billion dollar power lease, we've stopped asking "what can AI do" and started asking "how do we build enough power plants to keep it running." And the "black box" is clearly becoming a glass box — that stolen reasoning research makes one thing obvious: if you want to keep your model's thinking private, an API key isn't going to cut it anymore.