Office Hours

Live sessions, documented for the cohort

Summaries and expanded write-ups from First Break AI office hours. Each session covers topics from the roadmap with live Q&A, debugging, and discussion.

Office Hours are live cohort sessions where we go deeper on roadmap topics, debug together, and discuss what learners are working on. Each session is documented here as an expanded write-up — not just meeting notes, but context and explanations you can reference later.

To join office hours: join the Fetchlens.ai Discord server — that’s where we share the meeting link, time, and reminders.

ImportantParticipation guidelines

Age: Office hours are recommended for learners aged 16+. Under 18: join only with parent or guardian permission. Please do not share personal information (phone, address, IDs, exact location, etc.) during sessions. See our Privacy Policy and Terms for the full details.

Live sessions

Cohort 01 runs 1 May 2026 — 31 July 2026. These are the live sessions for enrolled learners — the most recent session is at the top.

Cohort 01 · Session 1

8 May 2026

Cohort 01 kickoff. Three intuitions about LLMs (architecture, autoregressive, vocab-wide probabilities), why Qwen3 0.6B is small but structurally identical to large models, HuggingFace as the “GitHub for models” (Git LFS → Xet, model cards, scale), config.json and the Transformers library, markdown everywhere and Karpathy’s AI-managed knowledge base, the math-or-no-math question, and a live pure-C / GGUF demo with system prompt, --thinking, and --multi-turn flags.

Read session notes

Earlier sessions

Sessions held before Cohort 01 started. Kept here as reference — the same roadmap topics apply when you follow along later.

Pre-cohort · Office Hours

24 April 2026

NanoGPT speedrun infrastructure: modded-nanogpt repo walkthrough, 8x H100 cost and cloud GPU options, Modal for containerized training, Accelerate and DDP vs FSDP, sharding explained (parameters, gradients, optimizer states), training script demo, HF and W&B secrets setup, and troubleshooting pip/PATH issues on Windows.

Read session notes

Pre-cohort · Office Hours

10 April 2026

Learner LAN stock game (scope and shipping), Claude Code harness leak and safe learning, KV cache caveats, benchmark literacy, Gemma / Matryoshka, nanoGPT speedrun and Tyler Romero’s worklog, DDP and all-reduce, NVLink and DeepSeek-style optimization, Muon optimizer, RoPE/RoFormer in code, and walking through a distributed training script.

Read session notes

Pre-cohort · Office Hours

27 March 2026

Transformer architecture deep dive — why attention replaced LSTMs, self-attention and multi-head attention, decoder-only LLMs, dense vs MoE models, benchmarking, the three pillars of model development, data parallelism (DDP), and Project Watch: Speedrun and Auto Research GPT.

Read session notes

Pre-cohort · Office Hours

13 March 2026

GitHub collaboration (PRs, conflicts, rebasing), Qwen3 inference concepts (temperature, chat templates, speculative decoding, GGUF vs SafeTensors, quantization/precision), cohort-based learning, and Unsloth efficiency.

Read session notes