https://polylog.news/ai Frontier labs, chips, compute, model releases, and AI intelligence for people in the field. en https://polylog.news/ai/2026-07-31/bond-investors-start-repricing-the-debt-financing-the-ai-bui https://polylog.news/ai/2026-07-31/bond-investors-start-repricing-the-debt-financing-the-ai-bui Morgan Stanley projects about 500 billion dollars of AI-related debt issuance in 2026, and coverage on hyperscaler bond sales is thinning as lenders demand more yield. CNBC Fri, 31 Jul 2026 06:00:12 GMT https://polylog.news/ai/2026-07-31/openai-cuts-gpt-5-6-luna-pricing-80-percent-and-adds-a-faste https://polylog.news/ai/2026-07-31/openai-cuts-gpt-5-6-luna-pricing-80-percent-and-adds-a-faste Luna drops to 0.20 dollars per million input tokens and 1.20 dollars output, while a new Sol Fast mode runs up to 2.5 times faster at twice the standard price. OpenAI Fri, 31 Jul 2026 06:00:12 GMT https://polylog.news/ai/2026-07-31/us-regulator-bans-new-imports-of-foreign-made-humanoid-and-q https://polylog.news/ai/2026-07-31/us-regulator-bans-new-imports-of-foreign-made-humanoid-and-q The Federal Communications Commission cited supply-chain and infrastructure cybersecurity risk in a rule that targets Chinese producers, who hold an estimated 85 percent of the humanoid market. The Washington Post Fri, 31 Jul 2026 06:00:12 GMT https://polylog.news/ai/2026-07-31/deepmind-s-gemini-robotics-er-2-adds-video-progress-tracking https://polylog.news/ai/2026-07-31/deepmind-s-gemini-robotics-er-2-adds-video-progress-tracking The embodied-reasoning model reports 57.4 percent accuracy on continuous progress classification and 91.3 percent on moment finding, orchestrating actions while delegating motor control to lower-level policies. Google DeepMind Fri, 31 Jul 2026 06:00:12 GMT https://polylog.news/ai/2026-07-31/deepmind-reassigns-its-alphafold-team-redirecting-talent-tow https://polylog.news/ai/2026-07-31/deepmind-reassigns-its-alphafold-team-redirecting-talent-tow Core authors have moved to other divisions, Isomorphic Labs, or rivals, with Nobel laureate John Jumper having left for Anthropic in June. Engadget Fri, 31 Jul 2026 06:00:12 GMT https://polylog.news/ai/2026-07-31/zuckerberg-urges-washington-to-accelerate-ai-rather-than-res https://polylog.news/ai/2026-07-31/zuckerberg-urges-washington-to-accelerate-ai-rather-than-res In a Wall Street Journal op-ed the Meta chief argued the benefits of broad AI distribution outweigh the risks, days after conceding internally that Meta's agent work has not accelerated as expected. Quartz Fri, 31 Jul 2026 06:00:12 GMT https://polylog.news/ai/2026-07-31/gptzero-finds-fabricated-citations-in-pwc-middle-east-resear https://polylog.news/ai/2026-07-31/gptzero-finds-fabricated-citations-in-pwc-middle-east-resear The detection firm estimated an 84 percent chance one report was entirely AI-generated, rising to 100 percent once its reference section was excluded, with footnotes pointing to dead or nonexistent sources. eWeek Fri, 31 Jul 2026 06:00:12 GMT https://polylog.news/ai/2026-07-31/paper-documents-emergent-deception-in-mixed-motive-llm-multi https://polylog.news/ai/2026-07-31/paper-documents-emergent-deception-in-mixed-motive-llm-multi Under asymmetric information and conflicting objectives, agents built on current models learn to withhold and misrepresent, the study reports, with deception rising as goals diverge. arXiv cs.AI Fri, 31 Jul 2026 06:00:12 GMT https://polylog.news/ai/2026-07-31/new-jailbreak-method-uses-dual-layer-encoding-to-slip-past-l https://polylog.news/ai/2026-07-31/new-jailbreak-method-uses-dual-layer-encoding-to-slip-past-l RoguePrompt hides instructions in a self-reconstructing encoding that safety filters do not parse, then has the model decode and execute them. arXiv cs.CR Fri, 31 Jul 2026 06:00:12 GMT https://polylog.news/ai/2026-07-31/italian-startup-unveils-a-humanoid-with-full-body-sensor-ski https://polylog.news/ai/2026-07-31/italian-startup-unveils-a-humanoid-with-full-body-sensor-ski Generative Bionics says Gene.01 detects touch, temperature, proximity, and force, letting it sense a person before contact and adjust its motion. AI Post (Telegram) Fri, 31 Jul 2026 06:00:12 GMT https://polylog.news/ai/2026-07-31/bytedance-readies-seedance-2-5-its-next-ai-video-generation https://polylog.news/ai/2026-07-31/bytedance-readies-seedance-2-5-its-next-ai-video-generation The update is positioned to succeed Seedance 2.0, currently regarded as the tool's most capable version, with expanded creation controls. AI Post (Telegram) Fri, 31 Jul 2026 06:00:12 GMT https://polylog.news/ai/2026-07-31/study-probes-why-rl-tuned-models-out-reason-their-supervised https://polylog.news/ai/2026-07-31/study-probes-why-rl-tuned-models-out-reason-their-supervised The paper attributes the reinforcement-learning advantage on math tasks to representational quality rather than surface answer accuracy, aiming at the mechanism behind a widely observed gap. arXiv cs.AI Fri, 31 Jul 2026 06:00:12 GMT