Weekly AI Agents report

January 23, 2025

Articles

• Production LLM Security: Real-world Strategies from Industry Leaders Securing large language models in production environments demands a comprehensive multi-layered defense strategy. This involves addressing security challenges such as prompt injection attacks, data privacy concerns, and indirect prompt threats by implementing techniques like robust fine-tuning, adversarial training, and secure data management practices. Industry leaders emphasize the importance of input sanitization, output filtering, anomaly detection, and effective monitoring to mitigate potential risks. Additionally, secure deployment choices, access controls, privacy-preserving methods such as differential privacy, and human oversight are crucial for enhancing both security and functionality in LLM applications. These strategies collectively ensure robust protection throughout the LLMOps lifecycle while maintaining user trust through transparency and tailored organizational measures.
• Thoughts On A Month With Devin during testing, the AI tool Devin showed mixed results in its ability to independently perform complex programming tasks. While initially successful at automating simple tasks like data integration between Notion and Google Sheets, it struggled with more complex challenges such as project creation from scratch, often producing impractical or overly complex code. These issues led to inefficiencies and frequent failures across various tasks, including debugging, deploying apps, and enhancing existing projects, highlighting the tool's limitations in reliability and problem-solving compared to other developer-driven workflows like Cursor.
• Microsoft and OpenAI evolve partnership to drive the next phase of AI Microsoft has announced an evolution in its strategic partnership with OpenAI, continuing their collaboration from 2019 into a new phase centered on AI development through 2030. The agreement includes access to OpenAI's IP and revenue sharing arrangements, with Azure maintaining exclusivity for hosting the OpenAI API, which enhances Microsoft products like Copilot. This expansion also involves Microsoft as a major investor in OpenAI and features a recent commitment by OpenAI to build more capacity on Azure, supported by Microsoft’s right of first refusal on new infrastructure for AI research and model training.
• Gemini 2.0 Flash Thinking model scores 73.3% on AIME (math) & 74.2% on GPQA Diamond (science) benchmarks

• Master citations to build trustworthy AI Cohere introduces North, a secure AI workspace aimed at enhancing productivity and trust in AI through features like master citations that reduce hallucination risks. Their platform offers various models such as Aya Expanse for multilingual tasks and advanced retrieval tools like Embed and Rerank to improve search quality. Designed to empower workers across industries including financial services and healthcare, North integrates best-in-class security measures and is supported by Cohere's research lab focused on solving complex machine learning challenges.
• Run & Finetune DeepSeek-R1 introduces DeepSeek-R1, a powerful open-source AI model that sets new benchmarks in reasoning performance and is available for fine-tuning via the open-source package llama.cpp. The R1 models are large (up to 712GB) and require specific hardware setups; users can run them on CPUs with sufficient RAM or leverage GPUs like an RTX 4090 for faster processing. To facilitate use, Distilled versions of these models are provided in smaller sizes compatible with the GGUF format, and users are guided through installation and execution steps, including handling prompts correctly and managing hardware requirements.
• AI Agents in Cybersecurity: Cutting SOC Workloads by 30% in 30 Days Doppel has successfully reduced its SOC workloads by 30% in just 30 days using an advanced AI agent built on OpenAI's o1 model. This AI agent was trained to emulate the decision-making expertise of human analysts, incorporating thousands of historical decisions to enhance nuanced judgment and improve threat detection accuracy. As a result, it surpassed human-level benchmarks with lower false positives and faster response times, enabling security teams to focus on more complex tasks while maintaining efficient and proactive cybersecurity defenses.
• DeepSeek R1's recipe to replicate o1 and the future of reasoning LMs DeepSeek R1 represents a significant advancement in reasoning-based language models by employing reinforcement learning techniques to enhance performance and affordability, reflecting a shift towards more open and competitive pricing in the AI model landscape. The training of DeepSeek R1 involves a meticulous four-stage regimen focusing on foundational research, while also addressing challenges like unlearning preset response lengths during RL finetuning, which is crucial for refining model capabilities. Additionally, techniques such as few-shot prompting and advanced reinforcement learning strategies are utilized to improve emergent behaviors in AI models, with human preferences playing a critical role through methods like Group Relative Policy Optimization (GRPO) to ensure transparency and progress in community research on reasoning language models.

Tutorials

• Build a Multi-Agent System with LangGraph: A Complete Tutorial on Agent Orchestration guides users through building a multi-agent system from scratch using LangGraph, where a supervisor agent manages conversation flow and routes tasks to specialized agents. The system consists of various AI agents, such as web researchers and NL2SQL tools, that work together to process requests and return control to the supervisor until the task is complete. The tutorial also provides learning resources on building RAG applications and multi-agent systems using LangChain and other AI technologies, with a specific example of finding information about the founder of FutureSmart AI, Pradip Nichite.
• Building Chatbots with Memory Capabilities: A Comprehensive Tutorial with LangChain, LangGraph, Gemini AI, and MongoDB provides a comprehensive overview of developing a context-aware chatbot using several advanced technologies: LangChain for language model integration, LangGraph for state management, Gemini AI for natural language processing, and MongoDB for persistent storage. The goal is to create a system that delivers personalized and memory-retentive interactions.
• Scheduled Tasks in LangGraph Lance Langshan introduced a feature in ChatGPT called scheduled tasks, allowing users to receive daily reminders like inspirational quotes at specific times. This concept extends to his LineGraph projects, where he uses the Langra platform to schedule cron jobs for different agents and workflows, such as generating daily reflections with AI models and sending them to Slack channels. The integration of crons into these deployments facilitates ambient agents that run in the background, producing outputs like social media content or trend tracking updates at regular intervals.
• Foundations of Large Language Models discusses advancements in natural language processing (NLP) through the development of large language models like BERT and GPT. It emphasizes how these models, enhanced by self-supervised learning techniques such as masked token prediction and permuted language modeling, allow for more effective contextual understanding and task adaptability without extensive labeled datasets. Additionally, it explores various strategies to optimize model performance, including activation functions, efficient memory management, and reinforcement learning methods like Proximal Policy Optimization (PPO) for aligning AI training with human preferences. Overall, the focus is on improving efficiency, scalability, and alignment in NLP tasks through innovative pre-training techniques and optimization strategies.
• Implementing A Byte Pair Encoding (BPE) Tokenizer From Scratch focuses on enhancing Byte Pair Encoding (BPE) for token reduction through improved techniques such as weighted pair merging and dynamic vocabulary updates. It discusses challenges like calculating character pair frequencies and optimizing performance while maintaining thread safety in parallel processing. Additionally, it explores extending BPE to support multilingual capabilities with language-specific rules, caching mechanisms, and context-awareness to manage complex vocabularies effectively.
• Learn How to Build AI Agents & Chatbots with LangGraph! provides a step-by-step guide on using LangGraph to create a basic chatbot through defining nodes, states, and edges, with instructions on setting up the environment and utilizing APIs for Large Language Models. By demonstrating how to construct a simple StateGraph-based chatbot that responds to user inputs, it highlights LangGraph's capability in building both simple and complex AI applications effectively.
• 10 AI agent courses available now on Coursera
• Simon: Six short video demos of LLM and Datasette projects On January 22, 2025, Simon Willison reported on a new Datasette Public Office Hours session where Alex Garcia and other community members demonstrated various projects involving LLMs and Datasette. Six distinct demonstrations were featured, including Matthias Lübken's llm-logs-feedback plugin for storing feedback on prompt responses, Thomas Hughes’ llm-model-gateway and llm-consortium plugins enabling model collaboration via a local server or consortium approach, and Derek Willis’s Congressional Travel Explorer project analyzing travel data using Datasette. Nat Knight introduced llm-questioncache, which caches similar question answers using embeddings, while Simon Willison highlighted improvements to Datasette's Enrichments with features like progress bars for operations on tables. The session concluded with Alex Garcia showcasing collaborative plugins for Datasette Cloud: datasette-write-ui for editing data, datasette-pins for customizing user home pages, and datasette-comments for commenting functionality.
• Understanding and Effectively Using AI Reasoning Models explores how language models have evolved from basic next-word prediction to handling complex reasoning tasks through techniques like "Chain of Thought" prompting and reinforcement learning. These advancements allow models to process more structured reasoning and receive feedback for improved accuracy, leading to rapid benchmark achievements in AI development. While these models excel at deep reasoning and complex problem-solving, their applications extend across various domains such as medical diagnostics, educational reporting, and trend analysis, with distinct advantages over traditional chat models due to enhanced scaling methods and integration capabilities.
• Deeplearning.ai: Building Towards Computer Use with Anthropic is a beginner-level course offered by Colt Steele, Head of Curriculum at Anthropic, available for free during the DeepLearning.AI platform beta. The course focuses on learning how to utilize Anthropic's AI models for computer interaction, including image reasoning and tool use, enabling an LLM-based agent to perform tasks like web searches through interface navigation. Participants will explore effective prompting techniques, prompt caching, multimodal capabilities, and API interactions with Claude, aiming to build AI applications that can execute tasks on a computer by combining these elements into a functional AI assistant. Suitable for those with basic Python knowledge interested in Generative AI and AI research principles, the course offers comprehensive insights into Anthropic's approach to AI safety, alignment, and interpretability.
• Building a fully local "deep researcher" with DeepSeek-R1 DeepSeek's R1 introduces a novel open-source reasoning model that combines fine-tuning with reinforcement learning to enhance AI scalability in reasoning tasks. It employs a two-stage training strategy involving generating multiple problem-solving attempts using GRPO, followed by fine-tuning on non-reasoning examples to balance strong reasoning skills and general language capabilities. The content also highlights advancements in reinforcement learning for efficient model deployment on personal devices, explores the management of "think tokens" during model interaction, and describes iterative evaluation processes in large-scale models.

Papers

• Titans: Learning to Memorize at Test Time advancements in neural network architectures, particularly focusing on the Titans model, which incorporates a novel neural long-term memory module. This module enhances attention models' ability to utilize historical context, outperforming traditional Transformers and recurrent networks in tasks such as language modeling and time series analysis. Key features of the Titans architecture include adaptive forgetting gates, optimized write operations, and dynamic surprise metrics, all contributing to better memory management and improved task performance by efficiently handling long sequences and balancing memorization with generalization.
• Towards System 2 Reasoning in LLMs: Learning How to Think With Meta Chain-of-Thought Advancements in large language models focus on enhancing reasoning capabilities through Meta Chain-of-Thought (Meta-CoT), which involves integrating complex problem selection and external computational tools. Research from 2020 to 2024 highlights improvements using reinforcement learning, meta-learning, and human feedback, with efforts directed towards self-correction, efficient exploration techniques, and effective instruction following. Studies address challenges in model training datasets and propose solutions like asynchronous model training for better throughput, despite memory trade-offs. The "Big MATH" project exemplifies these advances by creating large-scale, verifiable math problem datasets to improve reasoning accuracy and reliability in language models.
• Predicting the Performance of Black-box LLMs through Self-Queries delves into predicting large language models' performance using black-box methods and follow-up prompts to enhance reliability across various configurations. It introduces QueRE as an innovative technique that uses tailored questions and confidence scores for accurate predictions, detecting adversarial influences, and assessing model robustness in diverse settings. Additionally, the content discusses recent AI research spanning 2022-2024 on topics like interpretability, safety, adversarial learning, and performance optimization across multiple domains, emphasizing mathematical techniques to analyze data perturbations and prompting strategies' impacts.
• Evolving Deeper LLM Thinking "Mind Evolution" is an evolutionary search strategy using language-based genetic algorithms and feedback mechanisms to enhance large language models' problem-solving capabilities, particularly in natural language planning tasks like meeting and travel planning. It combines broad exploration with deep refinement, achieving superior success rates without formal task specifications or extensive fine-tuning, as demonstrated on benchmarks such as TravelPlanner. The approach involves iterative refinement through AI-driven feedback, fitness evaluations, and periodic resets, outperforming baseline strategies by effectively meeting complex requirements and scaling well with complexity.
• UI-TARS: Pioneering Automated GUI Interaction with Native Agents UI-TARS represents an advanced native GUI agent model enhancing human-like screen interactions through superior perception, action modeling, and system-2 reasoning. It leverages large language models like GPT-4 for adaptive, scalable task handling, integrating perception, reasoning, memory, and action into adaptable systems. These developments focus on improving decision-making in complex environments by addressing data scarcity challenges and optimizing task execution with structured reasoning processes, ultimately aiming for autonomous learning with reduced human oversight.

Tools

• Holistic Agent Leaderboard is a standardized, cost-aware platform for evaluating AI agents, developed by the SAgE team at Princeton University. It addresses current evaluation challenges such as inconsistent harnesses, lack of reproducibility, and inadequate cost tracking by offering leaderboards across various benchmarks like SWE-bench, USACO, and Cybench. HAL comprises two main components: a central leaderboard platform for assessing agent capabilities with cost-controlled evaluations and detailed failure analysis, and a standalone evaluation harness that supports easy integration of custom agents and tracks costs efficiently. Designed to serve diverse user groups including developers, researchers, and procurers, HAL aims to enhance visibility, reproducibility, and understanding of agent performance across benchmarks, fostering innovation and collaboration within the AI ecosystem. -- LLMOps Database is an open-source framework for managing machine learning (ML) pipelines and deployments.
• The King RAGent RAGent is an AI-powered research assistant application built with Streamlit that utilizes vector databases and external APIs to provide detailed answers to user queries. It features PDF upload for document processing, AI-driven query handling, web search integration for accuracy enhancement, and a dry run mode for testing without real API calls or database operations. The project encourages contributions via pull requests on GitHub, is licensed under the MIT License, and supports setup and usage through straightforward commands.
• Build a clone of Perplexity with LangGraph, CopilotKit, Tavily & Next.js provides a comprehensive tutorial on building a Perplexity clone using LangGraph, CopilotKit, Tavily, and Next.js. It details creating AI agents with the LangGraph framework and integrating them into applications using FastAPI for backend services and Next.js for user interfaces. Key steps include setting up API keys, installing dependencies with Poetry, defining workflow graphs, managing state transitions, and configuring server settings.
• coding-agent is an open-source web application designed to assist developers in coding tasks by utilizing Next.js and CopilotKit, powered by a LangGraph agent. It simplifies development processes with features like generating responses and offers a setup guide involving API keys, cloning repositories, and managing dependencies using Poetry and npm. Additionally, GitHub's suite of tools like Actions, Security, and Codespaces are highlighted for automating workflows, enhancing security, and creating instant dev environments, catering to various company sizes and use cases across industries.
• deepval DeepEval is an open-source framework by Confident AI for evaluating large language models, offering features like G-Eval and hallucination detection to enhance code quality and security on GitHub. It supports integration with CI/CD environments and provides tools for continuous evaluation, management, and testing of LLMs in applications such as chatbots. The framework includes CLI-based testing with parallel execution options and integrates with platforms like LlamaIndex and LangChain for logging, debugging, and optimizing evaluations, all under an Apache-2.0 license.

Models

• Sky-T1-32B-Preview is a 32B reasoning model trained from Qwen2.5-32B-Instruct with 17K data, performing on par with o1-preview model on both math and coding tasks. Developed by the NovaSky Team at UC Berkeley, this model uses supervised fine-tuning on the data with a batch size of 96 and achieves competitive results in various evaluations such as Math500 and LiveCodeBench. The model has been used in multiple spaces and has over 7,500 downloads in the last month, demonstrating its popularity and usefulness in text generation and conversational tasks.
• Ray2 Luma Ray2 is an innovative video generative model that transforms text instructions into ultra-realistic, production-ready videos using advanced multi-modal architecture. The content showcases a variety of dynamic scenes ranging from nature's power to surreal artistic expressions involving humans and everyday objects, crafted by diverse artists through cinematic techniques like lighting effects and camera angles. Additionally, the content describes imaginative scenarios featuring futuristic elements and creative macro photography, highlighting Luma Ray2’s role in enhancing visual storytelling for photographers available in Dream Machine and soon in the Luma API.
• HF: DeepSeek R1 the most powerful reasoning open-source model - available in GGUF, original & 4-bit formats. Includes Llama & Qwen distilled models.
• Ollama: DeepSeek R1
• UI-TARS-72B-DPO UI-TARS is a next-generation native GUI agent model designed to interact seamlessly with graphical user interfaces (GUIs) using human-like perception, reasoning, and action capabilities. Unlike traditional modular frameworks, UI-TARS integrates all key components—perception, reasoning, grounding, and memory—within a single vision-language model (VLM), enabling end-to-end task automation without predefined workflows or manual rules.