Explains key LLM benchmarks like ARC-AGI-2, GPQA Diamond, and SWE-Bench, helping evaluate model capabilities beyond hype.
Sam Altman's 2022 email on OpenAI's open source strategy to release a GPT-3-like model for consumer hardware.
Excerpt from Sam Altman's email to OpenAI's board about open source strategy and releasing a GPT-3-like model.
Analysis of GPT 5.6's 72 configurations, mapping model choices and reasoning effort to performance and cost.
![[notes] Crowd-Sourcing: Metaphors for LLMs and AI](https://crawljobs-blogs-media.s3.us-east-1.amazonaws.com/alldevblogs/thumbs/article-9e9f2c4d4688-full-82119b3a.webp)
Crowd-sourcing metaphors used for LLMs and AI, like intelligence, learning, and reasoning.
Explores Cosma Shalizi's view of AI as a mechanization of cultural traditions, analyzing how LLMs retrieve and synthesize formulaic information.
A reflection on the arrival of Artificial General Intelligence (AGI), arguing that its 'general' nature distinguishes it from all previous purpose-built AI models.
A reflection on the arrival of Artificial General Intelligence (AGI), arguing that its 'general' nature distinguishes it from previous purpose-built AI models.
Explores how AI language models shift a programmer's role from writing code to managing context and providing detailed specifications.
Analyzes LLM APIs as a distributed state synchronization problem, critiquing their abstraction and proposing a mental model based on token and cache state.
Explores training a hybrid LLM-recommender system using Semantic IDs for steerable, explainable recommendations.
A Chrome engineer discusses the design challenges and considerations for creating new built-in AI web APIs, focusing on the prompt API and task-based models.
Critique of the 'how many r's in strawberry' test as a poor benchmark for AI intelligence, arguing it measures irrelevant trivia.
Explores how advanced AIs use 'chains of thought' reasoning to break complex problems into simpler steps, improving accuracy and performance.
A guide to benchmarking language models using a Jupyter Notebook that supports any OpenAI-compatible API, including Ollama and Foundry Local.
Explains why standard language model benchmarks are insufficient and how to build custom benchmarks for specific application needs.
A tutorial on building a transformer-based language model in R from scratch, covering tokenization, self-attention, and text generation.
Learn how to accurately calculate token counts for strings using language models with a provided Jupyter Notebook tool.
Explores reward hacking in reinforcement learning, where AI agents exploit reward function flaws, and its critical impact on RLHF and language model alignment.
A philosophical and technical exploration of how Large Language Models (LLMs) transform 'next token prediction' into meaningful answer generation.
