Topic
All digests tagged Large Language Models (LLMs)
Suggested route
Move from earlier context through the available material to the newest signals. This route is a reading aid, not a difficulty ranking.
Building AI for finance requires more than just a powerful LLM. Success hinges on integrating proprietary data, mastering plugin architecture, and training models on subjective human 'taste' to mimic expert judgment.
Read article
Code Mode is an alternative architecture for AI agents that uses a programming sandbox instead of a large tool catalog. This approach drastically reduces context bloat, lowers token usage, and enables emergent behavior by allowing agents to compose tools in code.
Read article
Kepler's architecture ensures LLMs cite numbers but never write them. By integrating deterministic code, a citation registry from 30M SEC documents, and a post-processing filter, the platform guarantees verifiable, auditable financial data.
Read article
Can AI understand a distressed credit? Harneet Kaur demonstrates how building a structured, multi-layered reasoning process forces top LLMs (Astra 6, Opus 5.5) to converge on a single probabilistic view of a distressed company.
Read article
Joseph Kim details how ChatGPT for Financial Services addresses the 'capability overhang' by integrating proprietary data, firm templates, and interactive tools, setting the stage for proactive, persistent AI agents.
Read article
AI News & Strategy Daily | Nate B Jones
Goodfire's research suggests AI's internal workings are much more understandable than previously assumed. This interpretability breakthrough is a major green flag for AI safety.
Read article
Context rot degrades agent reasoning as history piles up. Deep Agents mitigates this using offloading (saving large tool results to virtual file systems) and summarization (condensing context when token usage hits 85%).
Read article
The lecture details advanced RL methods for LLMs, covering PPO, GRPO, and On-Policy Distillation (OPD). These techniques maximize expected rewards by optimizing policies against human preferences or verifiable outcomes, while managing gradient variance and policy deviation.
Read article
This workshop compares GPT-5.5, Kimi, DeepSeek, and Mistral across complex scenarios like RAG, tool calling, and vision. Learn how to use ASSERT and DSPy to automate model evaluation and optimize prompts for portability.
Read article
Model distillation offers an efficient alternative to RLVR for reasoning LLMs. By training a smaller student model on the outputs of a large teacher model, developers can achieve high performance with reduced computational cost.
Read article
AI agents are shifting from complex coding tasks to proactive 'assistance.' Engineers must focus on durable state management, structured output (Jev), and centralized connection gateways to build reliable, production-grade AI tools.
Read article
Builders explore Gemini Audio, Gemma, and Lyria for advanced applications. Learn how to use Gemma for offline capabilities and integrate Gemini 3.8 via WebSockets for real-time voice tools.
Read article
The continual learning loop is unsolved. We must balance SFT (knowledge injection) and RL (behavior optimization) by carefully tuning the SFT/Distillation to RL ratio for optimal model performance.
Read article
The lecture details the full LLM lifecycle, covering pre-training (next token prediction), SFT (behavior alignment), and advanced alignment techniques like DPO and RLHF. Key engineering concepts include LoRA for parameter efficiency and RoPE interpolation for long context handling.
Read article
Nemotron 3.5 Super VL is a multimodal open model built for autonomous agents. Learn how its 120B parameter architecture, optimized with NVFP4 and trained via distillation and mid-training, enables superior performance in agentic and vision-language tasks.
Read article
Nemotron 3.5 Super VL is an open-source, multimodal LLM (120B total, 12B active parameters) that powers specialized agents. It can analyze dashboards and scanned documents to propose code-based cost optimizations.
Read article
Context engineering is crucial for AI agents. Learn about context rot and four failure modes—Poisoning, Distraction, Confusion, and Clash—and how to improve agent reliability by deliberately pruning context.
Read article
GSQ and RCO are complementary methods for accurate low-bit LLM quantization. GSQ uses Gumbel-Softmax sampling for stochastic optimization, while RCO optimizes bit allocation under fixed size budgets, enabling state-of-the-art model compression.
Read article
Reflection Beam is a new open-weight model for agentic use cases. It is highly token efficient and supports advanced RL and SFT, making it ideal for building complex, domain-specific AI applications.
Read article
Learn how speculative decoding techniques like MTP and DFlash achieve massive LLM speed-ups by training small 'draft models' to predict multiple tokens in parallel, validating them with the main model in a single forward pass.
Read article