Ideas Worth Exploring: 2025-05-06
- Charles Ray
- May 6
- 5 min read
Ideas: Max Woolf - As an Experienced LLM User, I Actually Don't Use Generative LLMs Often
Max Woolf, a Senior Data Scientist at BuzzFeed with extensive experience in text generation models, reflects on their use of Large Language Models (LLMs) both personally and professionally. Max Woolf notes that despite being critical of modern LLMs, they have utilized them extensively over the past decade. Here's a summary:
Interaction with LLMs: Max Woolf predominantly uses LLM APIs directly rather than user-friendly frontends like ChatGPT.com, as it provides more control. Techniques like prompt engineering to enhance output quality and system prompts to set generation constraints are always integrated.
Professional Use: At BuzzFeed, Max Woolf has employed LLMs to quickly solve various problems:
Automatically categorizing thousands of articles into a new taxonomy.
Generating unique labels for distinct semantic clusters of articles.
Creating a grammar-checking tool using the BuzzFeed style guide.
Personal Use and Ethics: Max Woolf doesn't use LLMs for writing their blog or personal projects due to ethical concerns about authorship misrepresentation and potential hallucination issues, especially when discussing recent tech events.
Coding Assistance: Max Woolf finds LLMs helpful for coding tasks like regular expression generation, complex queries with specific libraries, and logging metadata during model training. However, they prefer asking LLMs ad hoc questions over using in-line code suggestion tools due to the latter's distraction factor.
Future of LLMs: Despite the author's nuanced views on LLMs, they believe that OpenAI's collapse wouldn't mark the end of LLMs since open-source models and dedicated hosting providers can fill the gap. They conclude that using LLMs is like forcing a square peg into a round hole at times, but it's sometimes worth doing for quick iteration.
GitHub Repos: Ladybird - a truly independent web browser
Ladybird is a truly independent web browser, using a novel engine based on web standards. It is a brand-new browser & web engine. Driven by a web standards first approach, Ladybird aims to render the modern web with good performance, stability and security.
Ladybird is in a pre-alpha state, and only suitable for use by developers
Ladybird uses a multi-process architecture with a main UI process, several WebContent renderer processes, an ImageDecoder process, and a RequestServer process.
Image decoding and network connections are done out of process to be more robust against malicious content. Each tab has its own renderer process, which is sandboxed from the rest of the system.
GitHub Repos: Union - hyper-efficient zero-knowledge infrastructure layer for general message passing
Union is the hyper-efficient zero-knowledge infrastructure layer for general message passing, asset transfers, NFTs, and DeFi. It’s based on Consensus Verification and has no dependencies on trusted third parties, oracles, multi-signatures, or MPC.
It implements IBC for compatibility with Cosmos chains and connects to EVM chains like Ethereum, Berachain (beacon-kit), Arbitrum, and more.
The upgradability of contracts on other chains, connections, token configurations, and evolution of the protocol will all be controlled by decentralized governance, aligning the priorities of Union with its users, validators, and operators.
GitHub Repos: The Hugging Face Agents Course
Introduction to Agents: Definition of agents, LLMs, model family tree, and special tokens.
Fine-tuning an LLM for Function-calling: Learn how to fine-tune an LLM for Function-Calling
Frameworks for AI Agents: Overview of smolagents, LangGraph and LlamaIndex.
The Smolagents Framework: Learn how to build effective agents using the smolagents library, a lightweight framework for creating capable AI agents.
The LlamaIndex Framework: Learn how to build LLM-powered agents over your data using indexes and workflows using the LlamaIndex toolkit.
The LangGraph Framework: Learn how to build production-ready applications using the LangGraph framework giving you control tools over the flow of your agent.
Observability and Evaluation: Learn how to trace and evaluate your agents.
Use Case for Agentic RAG: Learn how to use Agentic RAG to help agents respond to different use cases using various frameworks.
Final Project - Create, Test and Certify Your Agent: Automated evaluation of agents and leaderboard with student results.
Ideas: Lorenzo Franceschi-Bicchierai - Apple notifies new victims of spyware attacks across the world
Lorenzo Franceschi-Bicchierai indicates Apple recently sent notifications to several individuals informing them that they were targeted by government spyware. Two recipients have come forward: Ciro Pellegrino, an Italian journalist from Fanpage, and Eva Vlaardingerbroek, a Dutch right-wing activist. Both received messages indicating a high-confidence warning of targeted attacks using mercenary spyware.
Pellegrino disclosed that the notification stated attacks were happening in 100 countries. Vlaardingerbroek shared Apple's alert on her social media, calling it an attempt to intimidate and silence her. Neither replied to TechCrunch's request for comment. The specific spyware campaign linked to these notifications remains unknown.
This isn't the first time Apple has sent such warnings; they did so twice last year for unspecified spyware attacks. Pellegrino is the second Italian journalist this year to receive a notification, following his colleague Francesco Cancellato, who was targeted by Israeli company Paragon Solutions in February 2023. Subsequently, two Italians from Mediterranea Saving Humans also revealed being targets of Paragon.
Apple and WhatsApp have previously notified users about similar spyware attacks, indicating an increasing trend of tech companies alerting individuals to potential security threats. Despite Apple's notifications this week, neither the company nor any other sources have provided further details on the current targeted campaigns.
Ideas: SWE Quiz - 5 Core Distributed Concepts Every Developer Should Know
The article underscores the importance of five fundamental ideas for developers working with distributed systems: Consistency, Availability, Partition Tolerance (CAP Theorem), Scalability, Fault Tolerance, Latency and Network Communication, and Distributed Consensus.
CAP Theorem: This principle states that a distributed system can only guarantee two out of three properties: Consistency (all nodes have the same data), Availability (every request receives a response), and Partition Tolerance (the system continues functioning despite network partitions). Systems must choose between being CP (Consistent, Partition Tolerant), AP (Available, Partition Tolerant), or CA (Consistent, Available) based on application requirements. MongoDB, for instance, allows configuring read and write concerns to lean towards either consistency or availability.
Scalability: This refers to a system's ability to handle increased load. There are two main types: Vertical Scaling (adding resources to an existing server), with limits due to hardware constraints, and Horizontal Scaling (adding more machines/nodes), which is crucial for distributed systems. Netflix uses Amazon EC2 Auto Scaling Groups to automatically adjust service capacity based on traffic patterns.
Fault Tolerance: In distributed systems, failures are inevitable. Strategies to ensure continued operation include Replication (making copies of critical data), Failover Mechanisms (switching to a backup system/server upon failure), and Graceful Degradation (continuing operation at reduced capacity).
Latency and Network Communication: Latency, the time taken for data to travel between components, can be minimized through strategies such as Data Locality, Caching, Asynchronous Communication, and efficient serialization. Google Cloud CDN reduces latency by serving content from locations closest to users.
Distributed Consensus: This involves ensuring all nodes in a distributed system agree on a single value or decision, even with failures or network partitions. Common algorithms include Paxos, Raft (used by etcd for Kubernetes), and Zab (ZooKeeper Atomic Broadcast). The FLP impossibility result proves that no consensus algorithm can guarantee both safety and liveness if one node might fail.
Understanding these concepts is vital for developers to effectively design, develop, and troubleshoot distributed systems.
Comments