📄 The Long ReadEvery story in this issue, reported at length — with the numbers, charts and sources behind them.→ Read the PDF · 201 KB
01Meta Releases Muse Glimmer: A 30B Open-Weight Agentic Model [AI Models & Research]Meta 30B open-weight model distilled from Muse Spark 1.2, Apache 2.0, runs on consumer GPU, day-0 support in transformers/llama.cpp/vLLM/Ollama. → source
02Beyond Transformers: Four Startups Chasing the Next LLM Architecture [AI Models & Research]Subquadratic, Manifest AI, Liquid AI, and Pathway bet against transformer quadratic scaling with sparse attention, power retention, liquid nets, and state-space models. → source
03Making Knowledge Distillation Cheap Enough to Run on a Single GPU [AI Models & Research]Offline top-K logit caching and fused chunked KL loss cut distillation VRAM from 250GB to 58GB, shrinking 4-node jobs to 1 node with 5x faster steps. → source
04NVIDIA Magpie TTS: 12-Language Open-Weight Voice Model at 32ms Latency [AI Models & Research]364M-parameter TTS covering 12 languages with 32ms TTFA on B200, frame-stacked local transformers for 2x faster decoding, targets self-hosted voice agents. → source
05OpenAI Launches GPT-5.6-Cyber and Expands Daybreak Defense Program [AI Tools & Ecosystem]Daybreak splits into Blue and Red tiers; Red includes GPT-5.6-Cyber for offensive security testing, limited to vetted partners like Accenture, IBM, CrowdStrike. → source
06Ollama Ships Day-One Support for Muse Glimmer on Apple Silicon [AI Tools & Ecosystem]MLX engine supports DFlash speculative decoding and image input, one-command deployment for Claude Code, Codex, Pi, OpenClaw, Hermes agent frameworks. → source
07Interconnects Nathan Lambert Publishes RLHF Textbook with Manning [AI Tools & Ecosystem]Post-training textbook covering PPO to CISPO, async RL systems, distillation, character training. Free online at rlhfbook.com with 12-hour video course. → source
08Prompt Caching vs. Fine-Tuning: A Cost and Latency Decision Framework [AI Tools & Ecosystem]Practical guide: cache KV states for repeated prompts vs LoRA fine-tune for consistent formatting. Hybrid approach recommended for agentic loops. → source
09Anthropic Signs $9.1B, 20-Year Compute Deal with Bitcoin Miner Riot Platforms [AI Applications & Industry]491MW at Riot Rockdale Texas through June 2048, extendable to $16.1B. Former Bitcoin mining infrastructure reborn as AI compute. First delivery December 2027. → source
10Nvidia Partners with Wall Street Giants on $500B AI Infrastructure Fund [AI Applications & Industry]Apollo, Blackstone, BlackRock GIP, Brookfield, Goldman Sachs, KKR join Nvidia to finance chips, power, data centers. Big Tech AI spend to surpass $730B this year. → source
11OpenAI Completes $7B Employee Tender Offer at $852B Valuation [AI Applications & Industry]Buyback at March valuation, signaling IPO may be delayed. Employees cash out while OpenAI refocuses on enterprise after missing internal revenue targets. → source
12Zuckerbergs 6500-Word Personal Superintelligence Manifesto Draws Backlash [AI Applications & Industry]Essay pitching AI tutors, lawyers, 24/7 agents criticized for ignoring real-world harms. TechCrunch argues it demonstrates how trust was lost. → source
13Discovered Materials Raises $9M to Use AI Agents for Cooler Chip Materials [AI Applications & Industry]YC startup uses Anthropic-powered agent swarms for thousands of material candidates daily, verified by physics sims. Backed by Paul Graham, Lightspeed. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.