Custom Reward Functions for Multi-Turn Reinforcement Learning with Amazon Nova Forge
In brief
- Amazon's Nova Forge introduces a new approach to multi-turn reinforcement learning (RL), focusing on custom reward functions that guide AI models to learn correctly over extended interactions.
- Traditional RL can fail silently even when training looks good because subtle errors in reward design lead the model astray.
- Nova Forge addresses this by allowing users to define precise reward criteria, ensuring models behave as intended across multiple steps-like calling tools or recovering from mistakes.
- The key innovation is Reinforcement Fine-Tuning (RFT), which differs from standard methods by learning from evaluation signals on the model's outputs instead of relying on curated examples.
- This makes it easier for developers to train models iteratively through feedback, rather than needing detailed reasoning paths.
- Multi-turn RFT optimizes cumulative rewards across sequences of actions, leading to better generalization and performance in diverse tasks.
- For teams without the infrastructure to manage environments, Nova Forge offers a serverless multi-turn RL option.
- This feature is now generally available, providing flexibility for different user needs.
- Developers should focus on defining clear reward functions while Nova Forge handles coordination during training.
- The approach emphasizes safety by executing model-generated code securely within reward calculations and monitoring each component's contribution to ensure reliable learning outcomes.
- Looking ahead, this development could significantly improve the reliability of AI systems in complex, multi-step tasks-where subtle design choices can make or break performance.
Terms in this brief
- Reinforcement Fine-Tuning
- A method where AI models learn by adjusting their behavior based on feedback from evaluations of their outputs. Unlike traditional methods that rely on pre-curated examples, RFT allows for iterative training through user feedback, making it easier to refine model performance without needing extensive predefined reasoning paths.
Read full story at AWS ML Blog →
More briefs
Comparing AI Tools: A Developer's Perspective on Codex and Claude
A developer shared their experiences using Codex and Claude for coding tasks this week. They noticed that Codex generated simpler, cleaner Ruby code with fewer comments compared to Claude. Codex also felt more technical, like a precise tool, while Claude seemed more collaborative, acting almost like a colleague. The developer found Codex faster but noted it didn't save time overall due to the need for thorough testing and reviews. Codex excelled in creating efficient solutions but sometimes made mistakes, such as complex branching issues. In contrast, Claude was better at understanding context and handling tasks the way the developer preferred, especially when working with Jira. Codex's CLI approach for MCPs was praised for its straightforward authentication process. The developer plans to test both tools further, focusing on specific, smaller projects with Codex to leverage its speed. They'll also analyze if Codex's simpler output truly streamlines their workflow or if it sacrifices necessary complexity.
Waymo Reveals Custom Compute System for Autonomous Driving
Waymo has shared details about its custom computing system, designed to handle the extreme demands of autonomous driving. Unlike traditional systems that rely on human backup, Waymo's system operates entirely on its own, requiring ultra-low latency and high reliability. The system processes data in milliseconds, ensuring quick responses to road situations. It also features redundant components, so if one fails, the other takes over immediately. Over eight years, Waymo has boosted its compute power by 20x, allowing for faster decision-making. This innovation ensures safety and efficiency, even in tough weather conditions. As autonomous driving evolves, Waymo's approach sets a new standard for real-time, reliable computing on the road.
Uber's Pricing Puzzle: How Algorithms Influence Your Fare
Uber has dramatically changed its pricing strategy since it first promised affordable rides. Instead of fixed rates based on time and distance, the company now relies on algorithms that factor in real-time data like demand, location, and even rider behavior to set fares. This shift has led to significant price hikes-average Uber fares in the U.S. rose 83% from 2018 to 2022, nearly four times the inflation rate. Internal tests by Business Insider showed that identical rides at the same time could vary by up to 21%, while Consumer Reports found even wider gaps on certain routes. Critics argue that Uber uses AI to maximize profits, but the company denies using personalized data and points to higher costs and post-pandemic driver shortages as reasons for the surge. The question remains: How exactly do Uber's algorithms work? And what truly drives the high cost of rides? This issue highlights the growing role of technology in pricing decisions, raising concerns about transparency and fairness in the gig economy.
AI-Powered Dashcams Set to Revolutionize Storm Damage Response in Sarasota County
Researchers at the University of South Florida are developing an AI-powered dashcam system to help Sarasota County quickly identify and respond to storm damage. The system uses cameras in vehicles to spot issues like damaged pavement or blocked roads, alerting crews faster than traditional methods. It also includes route-planning tools to efficiently deploy resources where needed most. Additionally, an AI system will monitor social media for reports of hazards like flooding. Testing is set to conclude this fall, with the aim of making storm response more efficient and effective. If successful, this technology could be adopted by other communities nationwide.
AI Enhances Security Operations Centers with Wazuh Innovations
Wazuh has introduced AI-powered tools to boost SOC efficiency. The Wazuh AI Analyst, accessible on the Wazuh Cloud and through third-party integrations, automatically processes security data using Amazon Bedrock and Anthropic’s Claude. It provides actionable insights like endpoint protection status, alert volume, and posture summaries via email reports. This tool reduces analyst fatigue by streamlining repetitive tasks and accelerating decision-making. With cyber threats growing more complex, Wazuh's AI solutions help SOCs maintain situational awareness across diverse environments. As AI adoption continues to evolve, expect further advancements in threat detection and response capabilities.