AI Development Digest

Capex vs revenue tension persists. Cursor experiment shows 15x cost variation. Chip export restrictions may backfire. Treasury warns systemic risk. Microsoft shifts to in-house models. SambaNova raises $1B; Nvidia wearable ships; Ollama raises $65M. DeepSeek aims for own chip. Anthropic, OpenAI, SpaceX IPOs generate huge value. Apple sues OpenAI. Bipartisan support for chip export restrictions. Modal DFlash boosts inference 67%. Anthropic leases $10B compute from Meta. Fireworks AI $1.5B, Chai Discovery $400M. Gary Marcus highlights pricing moat collapse. OpenAI HuggingFace zero-day hack. Gemini 3.6 Flash and 3.5 Flash-Lite on Poe. New papers and benchmarks on agents, security, distillation. EU AI Transparency Rules take effect. SkillJack shows persistent backdoors. Self-reflection loops underperform. LLaDA MoE v2 challenges AR dominance. Gemini 3.7 Flash launched at 50% lower price, now in Devin; demo shows low-latency mobile control via ADB; tops AA-AnalystAgent for quantitative analysis; code generation requires zero fixes on voxel city task. GLM-5.3 shows emergent cyber capabilities; cost comparison shows 15x cheaper than Fable 5. Replica 27B agent beats larger models on research replication. Kog deep GPU optimization achieves 3,000 TPS on 2B model. Cursor acquired by SpaceX; Cursor Origin now hosts repos and deploys to Vercel, moving away from GitHub. Mistral plans 1GW European AI compute by 2030. Minimax H3 runs locally on 4x3090. Opus 5 writing quality issues with workarounds. Anthropic IPO valuation tied to $190-200B 2028 revenue forecast. Qwen 3.8 27B open weights released, outperforms Qwen3.7-Plus. Digital refinery critique highlights environmental costs of gas-fired AI data centers. EU AI Act enforcer joins IJCAI-ECAI 2026. Ethan Mollick cautions against overconfident AI extrapolations; also notes AI diffusion in political campaigns (Cornel West spending 10% on AI tools). Nathan Benaich argues Europe needs more compute. Anthropic internal productivity gains <2x. Funding roundup: Databricks $5B, River AI $1.1B, CodeRabbit $143M. GLM-5.3 may be first model subjected to new safety evaluations. Alibaba AI models hit 3B downloads, surpassing Meta and Google, reinforcing open-weight momentum. Stripe acquires OpenRouter for $7B, signaling AI infrastructure consolidation. Anthropic CEO Amodei frames AI backlash as a trust crisis, advocating careful regulation. New papers: Agentic Transaction, GenRouter, ClawGym II, VibeWorlding, UI-Mate, pixel-space diffusion training, agent evaluation, claim-level reliability, Intern-S2-Mobius, Dion3, scaling data repetition, RL fine-tuning risk. Additional new papers: UNMASK, UniProbe, Agents Catching Agents, ggerganov's inception pattern. New tooling: DeerFlow 2.0, shared memory layer, DeepSeek Harness (Cordis composability). Agent skill insights: skills provide procedural anchoring (65.7%) not knowledge; precision drops with skill pool size; 56k skills compete for <100 trigger slots. Resource: 3.8M SKILL.md dataset. Practical tips: 9-point Claude Code optimization. Legal/regulatory: Lethal Language article on AI accountability. Mojo open-sourced under Apache 2; RSI benchmark from Scale AI Labs; PACE-Bench for physics adaptation; ENTLORE for enterprise QA; RAG poisoning detection via attention collapse; LLM verifier bias from prior audit context; matrix multiplication exponent improvement via AlphaEvolve; PantheonGPU tool; Physical AI investment surge ($47.4B in H1 2026); Warp Factories infrastructure; Reach Capital $265M fund; Perplexity India growth experiment; Shoehorn quantization tool. Cerebras CS-4 launched with 750 PFLOPs, 30x GPU speedup. Gary Marcus flags young adults more concerned than enthusiastic about AI (Pew). Greg Isenberg challenges 'GPT wrappers useless' narrative. Show HN tool visualizes HuggingFace model architectures. Critique of AI slop papers highlights research quality concerns. ASI-Bench shows AI far from autonomous research (scores drop from 50 to 26 without human guidance); FreeToken enables 753B MoE model on single workstation; TensorRT Model Connect simplifies HuggingFace-to-TensorRT deployment; Optima platform for custom model benchmarking; HumanEvals library for human judgment in evals; Miles Brundage urges focus on safety culture over technical details in OAI training pause; AI slop threat to scientific system highlighted. Public trust crisis deepens; Chesky and Amodei acknowledge backlash. Silicon Data aims to launch GPU compute futures on CME, addressing price transparency. Japan to require AI firms to disclose training data (weak enforcement). CaliBench evaluates world models on stochastic physics. Anima Anandkumar highlights LLMs' physical world blind spot, promotes Fourier Neural Operator. Visual models need more data (Luma Abra paper). Physics of Agents applies statistical physics to agent collectives. Robotics data pipeline bottleneck (Scobleizer: one-shot in-context learning for real-world generalization remains hard). Frugal Tokens tool for agent cost tracking. New papers: Co-RL (unsupervised reasoning via multi-agent diversity), Temporal Multi-Signal Fusion for hallucination detection (0.84 AUC), Training Leaves Traces (model lineage verification), Decision-Metric Alignment (latent world models). New tool: MUZIM local AI for photo/video libraries (privacy-first). New benchmark: LHTB (long-horizon terminal agents, 300+ steps), FM-Bench (long-horizon management with competing agents). New: US reports AI-assisted cyberattacks on water systems, highlighting dual-use risk. Article warns of 'code factories without quality' as AI development blind spot. Dutch data protection authority advises Twitch users to opt out from Amazon AI training. Study finds AI firms cannot yet contain what they've built. Show HN: 125M piano autocomplete model on iPhone demonstrates edge AI creativity. New: Ethan Mollick calls out Stripe's 'singularity' claim as marketing hype. New: S1-mini (0.6B) open-weights model for on-device transcription released.