Topic
All digests tagged Large Language Models (LLMs)
Open models are key to building trust and sovereignty. By owning the full stack—from weights to training traces—developers can specialize AI for niche use cases, achieving performance that surpasses generalized frontier closed APIs.
Read article
Model compression, particularly through advanced quantization like NVFP4 (a 4-bit float format), is critical for running massive LLMs on edge devices. Techniques allow shrinking models dramatically (e.g., GLM 5.2 from 1.5 TB to 250 GB) while maintaining performance.
Read article
A massive community effort tested ML reproducibility at ICML 2026 using AI agents. The challenge used 'tracko' to create fully auditable log books, finding significant flaws in published research.
Read article
The research team discusses a joint scaling law for pre-training and RL, showing how compute allocation across these stages impacts downstream task performance. Key findings relate to optimizing model size versus training duration.
Read article
To build trustworthy LLMs, we must move beyond surface explanations and understand internal mechanisms. Interpretability requires mapping model behavior to specific circuits using tools like Sparse Autoencoders.
Read article
Datadog revealed how deleting years of AI context files improved agent performance, proving that 'context rot' is a real threat in large-scale LLM deployments. Learn about building eval platforms for code review.
Read article
This course covers LLM scaling laws, CoT reasoning, and the shift to agentic workflows. We explore techniques like RLHF, inference scaling (Large Language Monkeys), and building systems with verifiers.
Read article
This lecture examines test-time compute scaling, showing how repeated sampling and advanced architectures like Arkon can boost model performance without retraining. It covers power laws and optimizing inference pipelines.
Read article
The lecture explores advanced LLM verification techniques, moving from simple outcome checks (GSM8K) to sophisticated process-based reward models (PRMs). Learn how ensembling weak verifiers can dramatically close the generation-verification gap.
Read article
Dive into train-time scaling and RL techniques that enable self-improving AI agents. Learn how STaR, DeepSeekMath, and DAPO boost reasoning in LLMs using closed feedback loops.
Read article
Benchmarking often suffers from contamination, reward hacking, and misalignment between prompts and verifiers. To improve AI evaluation, domain expertise and rigorous QC are essential.
Read article
Diogo Almeida argues that current LLMs, trained via RLHF, optimize for human preference rather than reliable automation. The next frontier requires optimizing AI for verifiable rewards and autonomous decision-making.
Read article
Data quality is the compute multiplier: In an era of scarce compute, improving data through curation and synthetic generation allows models to achieve superior performance with less training power.
Read article
The lecture details advanced RL techniques for LLMs, covering Policy Gradient derivations and the stabilization mechanism of Proximal Policy Optimization (PPO). It shows how to model Chain-of-Thought prompting as an MDP problem.
Read article
This deep dive covers LLMs, focusing on the Transformer architecture. Key topics include subword tokenization, auto-regressive modeling, and the self-attention mechanism using Q, K, V vectors.
Read article
Scaling AI to long horizons requires more than bigger context windows; it demands better environments, robust simulations, and advanced algorithmic techniques like using value models for variance reduction.
Read article
This session covers Diffusion Models, detailing the loss function $\mathcal{L}_{t-1}$ derived from lower bounds. It also explores Foundation Model paradigms, zero-shot learning, and parameter efficiency techniques like Low-Rank Adaptation (LoRA).
Read article
Modern biology research requires verifiable substrates for AI training. LatchBio is building benchmarks like 'spatial bench' to measure scientific reasoning, moving beyond simple Q&A settings.
Read article
This session guides developers from LLM theory to practical agentic implementations, detailing how primitives like structured output, tool calling, and guardrails build robust AI systems.
Read article
AI is fundamentally changing the cost structure of cyberattacks, making them cheaper for bad actors while increasing defensive costs. The discussion covers Model Inversion, Prompt Injection, and the necessity of robust guardrails.
Read article