New Waterfront Restaurant Planned for Mission Bay

- **Constitutional AI: This technique, used by Anthropic for its Claude models, reduces the need for constant human supervision by providing the AI with a set of principles to self-critique its outputs. [This automates the feedback](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEn-XhCtM8zNz27ptKbN6ZQjdqNNI4FCfeQwz3f7V6Bc6KLlyXB2BNrZWhxaOi6mf24zTDSqk_Rq7OUDM34LS4_VeZhomYvQ4jOszlkef92olHBMulBP0JDdxoyAFBUzCX_uWk=) process, making it more scalable and consistent than traditional Reinforcement Learning from Human Feedback (RLHF), which relies on sometimes biased or inconsistent human ratings for every output. - Synthetic vs. Human Data:** While synthetic data can be generated much faster and avoids privacy issues, it often lacks the nuance and contextual understanding that human-labeled data provides. Hybrid models, which use large amounts of synthetic data for initial training and smaller sets of human-labeled data for fine-tuning, have shown the best overall performance, combining scalability with real-world robustness. - **Agentic AI Evaluation: Evaluating autonomous AI agents requires moving beyond traditional LLM metrics to focus on task completion success, tool-use accuracy, and cost per task. [Key benchmarks for this](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHrOY2RHmp2A2KKxYRA4mAuokgf4TxD7evtER_bl3vqWwi3OeraTqCM5AG7xp6-_PloBcTGv32lf91g9bgp7OKTJZEgIQ-oPxdDrQoRSQaLU2CnEbQWdAWYLNpa5xdATPEKgRwJnvuYi-0NYnphNjp9QUrbebnzQUJpZ1If0RomGy4BoZCEHs8b8PTcS2XUS-dUUWoWi0gp7syY5eeO) include GAIA for general intelligence, WebArena for web navigation tasks, and AgentBench for multi-domain reasoning. [- Fundraising Climate](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEOoisTItdWgkeKwFBR8oTN-TXH-BX24fOHhRW4ctD8oR60RZqpZjArqmgam5IQMckdstWTWX3hzgRm71EICPmWXNBCCV6Ne8RUeqbhWUMkTtmYoAV6hkYyZleevE6U5e1WlpMmfhMcTpbam5VnQkEGY9NeyP8gYswoQmlqm8Fsu9Vk2BvVvgT-MnI3D_2VuS1OuXroh8D571NFLe4NVIhopp-zvlrTE5e4YvpKaDQjwffLJqsi-mxzvWPo3ngPTxtMAFKifeXWBjIL6uw=): In the first half of 2025, AI startups attracted 53% of all global venture capital dollars, with U.S.-based AI companies receiving 64% of domestic VC funding. [Investors are reallocating](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEIwxtxPWrHPIkGTT_K4Q6QsHGYJivV4MZ5NCGia7xjewCJq2XbfzsF6xWe48Vcwif4VWc5OZxHc1JtnPk7-0hEqtLn3CYqRVVk4H3hHqIr4dyMnzBLr7npxMM-Lh2v2ndm9y04n8DuSeSpRxBFcAG-yFo=) capital away from traditional SaaS, with AI infrastructure and cybersecurity companies commanding premium valuations of 25-30x EV/Revenue. [- GTM Strategy for](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFdCp1PP_e9CsPxpkmlTIB97-XIXGk3nNODyWkOhqSPNmZYz64GCdifiADeRpicnvOXurz4GxQv2HvFO31ZV7eWI0LGlnWLO1BTjxxTeCn4CDE5BrtNOwpEQvFiAurAHIMKwuT3s1bE3rOD6kYo-wSbOoPzPXr8oMtyuncWMHnYteJ4kACPBjfEn18aEeUSQ7Y9FkgXqw==) Technical Buyers: Selling to AI labs requires an education-focused, non-linear approach, as buyers often conduct up to 80% of their research independently before engaging with a vendor. [Effective strategies](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFdCp1PP_e9CsPxpkmlTIB97-XIXGk3nNODyWkOhqSPNmZYz64GCdifiADeRpicnvOXurz4GxQv2HvFO31ZV7eWI0LGlnWLO1BTjxxTeCn4CDE5BrtNOwpEQvFiAurAHIMKwuT3s1bE3rOD6kYo-wSbOoPzPXr8oMtyuncWMHnYteJ4kACPBjfEn18aEeUSQ7Y9FkgXqw==) involve providing sandbox environments, engaging in technical communities like Discord and Hugging Face, and mapping out internal champions within target organizations. [- The Future of the](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHKMHgl7XrQZrgn5BAYe3mV97ou8P2pOJFpV_nUfNxA-UHeLlj5noDRPCFmFxZNkWpDNgXWf8hGbenV59jkUHpyAN3CPY4lvyjF4bRbg73C81X_scGAuIwvHcZA161DBXYIX0EvU_xarfwREvbIdNkV5vhWYcF3PQo6pDEObAAf5ccaQWeUvA==) Data Labeling Workforce: As AI models become more complex, the demand is shifting from large-scale, simple labeling to more nuanced, expert-level annotation, often called "AI Tutors". [This creates career pathways](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFkSfk9XVlJ5VEvVKT0QDx7kW0Sb0HI4vm7Lv_0d-xmfieImAhCrKYQ5QsCuj7BVFlVEOZmnu_pTVVmbfzR6EORkoTOdVU2S1iyLVUuQrnim2Ji-3rdXBPeDQwzaSwzlxfWRxx6CEaeapfV5pQnHSxWu1A2oFF9tIaTbzHf5T32R3lPXbjq_KHOtDY=) for data labelers to advance into roles like quality control analyst and AI trainer, which involve fine-tuning specialized models. [- Reinforcement Learning](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE3qpG_0y_xf4gKNMChIESvZvEB7KrfDlxOS4iO0CAB2n3btxMEiTUKmp1e2Ww3MoplJBsuzlF-I9LyjXKpRrSSJz-4c07o01XpaMj6i4ZWzjVaXeDrZ25UKHZRXdGqq_nxdaA3i4kGdrtKnaU6O8oxLAQNQ4lp) from Human Feedback (RLHF):** RLHF is a multi-stage process used to align models with human values. It involves first training a base model, then collecting human preference data by having reviewers rank different model outputs, using this data to train a "reward model," and finally fine-tuning the original model based on the reward model's predictions.