San Francisco Seeks Applicants for City Government Roles

- Reinforcement Learning from Human Feedback (RLHF) is a critical process for aligning large language models, involving supervised fine-tuning, training a reward model based on human-ranked responses, and then further fine-tuning the language model using reinforcement learning to maximize the rewards predicted by the reward model. This process is moving from using large-scale crowd-sourced data to smaller, higher-quality datasets curated by domain experts to handle nuanced and specialized topics like coding or legal analysis. - Constitutional AI presents an alternative to RLHF by using a predefined set of principles, or a "constitution," to guide the model's responses. In this two-phase process, the AI first critiques and revises its own responses based on the constitution, creating a self-generated dataset for fine-tuning, and then uses a reinforcement learning phase based on AI-generated feedback to further align its behavior. This method aims to provide a more scalable and transparent way to instill desired values compared to the extensive human labor required for RLHF. - Evaluating agentic AI systems, which take actions and use tools, requires specialized benchmarks beyond traditional language model metrics. Key benchmarks include AgentBench for multi-domain reasoning, WebArena for web navigation tasks, and ToolBench for assessing the accuracy of tool usage across various APIs. Enterprise-focused evaluations also consider cost-efficiency and operational stability, as some agents can have a 50x cost variation for similar accuracy levels. - A significant debate in AI training revolves around using synthetic versus human-labeled data; while synthetic data offers scalability and cost-effectiveness, it often lacks the nuance and accuracy for context-sensitive tasks where human annotation excels. A hybrid approach is often most effective, using synthetic data for broad coverage and human-labeled data for fine-tuning and handling complex edge cases. Models trained primarily on synthetic data can see significant accuracy improvements with the addition of even a small amount of human-labeled data. - The go-to-market strategy for B2B AI startups is shifting from a focus on technology to demonstrating clear business value and outcomes. Successful strategies involve identifying a narrow Ideal Customer Profile (ICP), validating the GTM approach early, and creating tight feedback loops between product development and early customer interactions. Instead of broad marketing funnels, the focus is on mapping key roles within a target account, such as AI teams and internal champions, and using AI-powered tools to deliver personalized messaging. - The fundraising landscape for AI startups has seen a surge in investment, with AI companies capturing nearly 50% of all global venture funding in 2025, a significant increase from 34% in 2024. Foundation model developers like OpenAI and Anthropic have raised a substantial portion of this capital. Investors are increasingly cautious, prioritizing startups with clear product-market fit and a strong go-to-market strategy. - AI is projected to be a major force in the future of work, with the World Economic Forum estimating the creation of 19 million jobs and the displacement of 9 million over the next five years due to advancements in AI and information-processing technologies. This shift will place a high demand on skills related to AI-driven data analysis, cybersecurity, and general technological literacy. In response, 77% of employers are expected to prioritize reskilling and upskilling their workforce to effectively collaborate with new AI systems.