Perplexity AI Study Shows 87% Time Reduction, NVIDIA Expands AI Factory Partnerships, Open-Source Momentum Continues

Key Takeaways

1. Perplexity AI research demonstrates autonomous agents reduce task completion time by 87% and cost by 94% compared to pure search. Perplexity also integrates Similarweb data for market intelligence.
2. NVIDIA announces multiple partnerships with SK hynix, LG, Doosan, NAVER, and Hyundai to build AI factories and next-generation memory, signaling massive infrastructure investment.
3. AI film production costs drop dramatically: "Hell Grind" achieved 50x faster time, 660x less human effort, and 36x lower cost ($500K vs $18M), though quality remains mediocre.
4. Meta AI's user base grew 2.5x in two months but 30-day retention is only 4.5%, suggesting non-organic growth.
5. Open-source ecosystem advances: Hugging Face announces OpenEnv as unified protocol for training-testing coupling, Gemma 4 MTP merges into llama.cpp, and SUPER GEMMA 4 26B uncensored model released.
6. vLLM-Omni v0.22.0 upgrade supports NVIDIA Cosmos 3 world model, robot real-time API, production TTS, and expanded hardware coverage.

1. AI Application Efficiency and Cost Breakthroughs

1. A Harvard study in collaboration with Perplexity found that using Perplexity Computer autonomous agents reduces task completion time by 87% and cost by 94% compared to pure search, with higher user satisfaction and cross-disciplinary search capability. — via 1 2
2. Perplexity AI now integrates Similarweb data natively, enabling users to access market, audience, and competitive intelligence through native integration or official connectors. — via 1
3. RunwayML released Aleph 2.0, a model that automatically adapts uploaded videos to different aspect ratios by filling in remaining scenes. — via 1
4. Deedy evaluated the AI-generated film "Hell Grind" as a great technical demo but mediocre film quality: 50x faster, 660x less human effort, 36x lower cost ($500K vs $18M). Technical issues like character consistency, camera angles, and realism are solved, but editing, acting, and dubbing remain artificial. Deedy argues AI films are inevitable and currently at their worst/slowest/ most expensive stage. — via 1
5. vLLM-Omni v0.22.0 major update includes support for NVIDIA Cosmos 3 world model, robot real-time API, production TTS, faster image/video/diffusion models, and broader quantization and hardware coverage, with 339 commits and 124 contributors. — via 1

2. AI Infrastructure and Industry Investment

1. NVIDIA and SK hynix announced multi-year technology collaboration to develop next-generation memory for global AI factories. SK hynix will develop memory for NVIDIA Vera Rubin to Jetson Thor platforms and leverage NVIDIA Omniverse and CUDA-X to accelerate semiconductor design and manufacturing. — via 1
2. NVIDIA expands partnership with Doosan Group on physical AI, robotics, and AI factory infrastructure. — via 1
3. NAVER is building the GAK Sejong Data Center based on NVIDIA DSX platform, initially 55 MW, planned to reach gigawatt scale, serving sovereign and physical AI models for Korea and enterprise workloads. — via 1 2
4. LG Group and NVIDIA announce AI factory construction based on DSX platform for LG's robotics, autonomous driving, data center technology, and GPU cloud services. — via 1 2
5. Jensen met with Hyundai Motor Group leadership to discuss NVIDIA-HMG collaboration on mobility and physical AI. — via 1
6. Hugging Face announced OpenEnv, a unified protocol layer connecting training environments, models, and testing, jointly maintained by Hugging Face, Meta-PyTorch, Nvidia, etc. — via 1 2
7. Gemma 4 MTP officially merged into llama.cpp, supporting Gemma 4 QAT+MTP for lightweight fast configuration. — via 1 2
8. SUPER GEMMA 4 26B UNCENSORED GGUF v2 model released with 0/100 rejection rate, fixed tool calling, 90% inference speed improvement, runs on 16-22GB VRAM. — via 1
9. Nvidia models account for 30% of Hugging Face homepage models, indicating open-source return in the US. — via 1

3. Key Insights and Independent Signals

1. Meta AI's user base grew 2.5x in the last two months, but 30-day retention is only 4.5%, very low, suggesting growth is likely not organic. — via 1
2. Ethan Mollick advises that as AI dramatically lowers the cost of executing good ideas, good ideas themselves are not easier to find. He suggests storing the most difficult, valuable, and unusual ideas now as a huge opportunity. — via 1
3. Sam Altman finds a daily Codex selection plan (one person per day with 10x usage limit) interesting and potentially recursively looping. — via 1