External Reading & Communities

Section 78.5

"The specific LLM tools in this book will mostly be obsolete in five years. The venues that publish them will not."

SageSage, Venue-Reader AI Agent
Note: Learning Objectives
Big Picture

The specific tools listed in this book will mostly be obsolete within five years; the venues that publish the next list will mostly still be alive. This section identifies the latter: the annual-report, newsletter, lab-channel, and community spaces that have proven durable across the 2018-2026 reshuffles and should keep working through 2030. For an LLM or agent engineer, this meta-list is the durable substrate: the specific RAG library, fine-tuning recipe, or agent framework you use this quarter will change, but the venues that announce the next generation of LLMs, agents, and evaluation benchmarks are what keep your practice current.

Prerequisites

This is the book's closing reading list and assumes familiarity with the frontier-LLM modules in Chapters 77 and 83.

Key Insight
Mental Model: the half-life of any tool list is short

If you assembled a "best LLM tools" list in 2022, half the items would be obsolete by 2024 and most by 2026. Hugging Face Transformers has held; almost everything else has churned. The right mental model is to maintain a meta-list (the venues that publish lists), not a list. The State of AI Report, Stanford AI Index, Epoch AI, and the lab research blogs will still be operating in 2030 with high probability; the specific Python packages, Discord servers, and Substack feeds you found this year mostly will not.

The half-life of any specific tool listed in this book is short; the venues below are the ones that should still be alive when you need to find the next list.

Concentric-ring infographic of the LLM field
Figure 78.5.1: A map of the LLM field in mid-2026. Closer to the centre is closer to the model itself. The outer rings depend on (and feed back into) the inner rings: applications use tooling; tooling consumes open-weight models; the open ecosystem informs the closed labs. Pick the ring you operate in and follow the venues that publish there. mid-2026. Centre ring (dark navy) holds frontier labs: Anthropic, OpenAI, Google DeepMind, xAI, Meta FAIR, DeepSeek, Qwen. Second ring holds open-weight near-frontier models: Llama 4, Qwen3, DeepSeek V4, Mistral Magistral, SmolLM2, Phi-4, Liquid LFM2.5, Gemma 3, Falcon-3. Third ring holds tooling: PyTorch, JAX, Hugging Face, vLLM, SGLang, llama.cpp, Ollama, MLX, TRL, unsloth, torchtune, Axolotl, DeepSpeed, FSDP, DisTrO, Nous Psyche, TensorRT-LLM, FineWeb-Edu, The Stack v2, lm-eval-harness, LMArena, Pile-2, MMLU, GPQA, SWE-bench. Outermost ring holds applications: ChatGPT, Claude.ai, Gemini, Grok, Cursor, Windsurf, Continue, Aider, Perplexity, You.com, Phind, Claude Code, Cline, Roo, GitHub Copilot, Devin, Notion AI, Glean, LangChain, LlamaIndex, DSPy, CrewAI, agno, n8n. Dashed arrows show that outer layers consume inner layers.

78.5.1 Durable venues, organized by category

Tip: lurk first, contribute second

In every active LLM community (EleutherAI Discord, the Hugging Face forums, specific subreddits, the LessWrong AI-alignment subforum, X / Twitter ML threads), the highest-signal posts come from people who spent at least a month reading before posting. The reverse pattern (post first, calibrate later) is the fastest way to be ignored. Lurk to learn the local norms, conventions, and what has already been answered. Then contribute by answering questions or sharing reproducible experiments rather than by posting opinions.

Tip: Three papers to read first after closing the book

If you want to stay current, start with three primary sources rather than aggregators: the most recent Transformer Circuits issue for interpretability, the latest State of AI Report for trend data, and the current frontier-lab system card (Anthropic, OpenAI, or DeepMind) for the most recent capability disclosures.

Key Takeaways

What's Next?

This is the final section of the book. See the Table of Contents for navigation, or revisit any chapter where the material is most useful.

Further Reading
Annual reports. State of AI Report (Nathan Benaich, Ian Hogarth); Stanford AI Index; Epoch AI (trends in compute, data, and capability).
Tracking newsletters and podcasts. Latent Space; Ahead of AI (Raschka); Import AI (Jack Clark); The Algorithmic Bridge; The Gradient; Dwarkesh Patel podcast (top-shelf 2024-25 long-form interviews with frontier-lab leaders); Cognitive Revolution (Nathan Labenz); 80,000 Hours podcast; Astral Codex Ten and Marginal Revolution for cross-disciplinary perspective; Asterisk Magazine.
Critical-perspective reading. AI Snake Oil book and blog (Narayanan, Kapoor): the canonical 2024 critical perspective.
Policy-focused. AI Now Institute, Centre for the Governance of AI, AISI (UK and US).
Engineering venue. AI Engineer World's Fair annual report and YouTube archive; the State of AI Engineering report (swyx).
Interpretability venue. Transformer Circuits (Anthropic): the durable home of mech-interp research.
Live model dashboards. LMArena; Artificial Analysis.