Deep Dives
One topic a day, explained properly: models and agents, research, system design, interviews, open-source projects, weekend builds and an LLM architecture lab, following a 64-week curriculum. RSS.
Models & Agentic
Mistral Large 4's open weights (1T total, about 52B active, 1M context) arrive with mostly self-reported benchmarks; the week's building block is the eval set, calibrated judge and trace that turn those claims into a go/no-go.
LabLab · Step 1, part 4
Part 4 of Step 1: ablate a tiny GPT one component at a time, measure the damage against a seed-noise baseline, and learn which pieces are load-bearing and which are decoration at this size.
Builds
Two three-hour builds from this week's retrieval theme: a harness that scores chunking, context lines and fusion on a labelled set (runnable, numpy only), or a clause-aware chunker for contracts.
GitHub
thedotmack/claude-mem stores what a coding agent did, then serves it back through search, timeline and get_observations; the interesting part is the retrieval contract and its size caps, which you can copy for any RAG system.
Interview
A practice project deep-dive in the style of a Meta Principal/Staff Architect loop: a RAG system you owned has plateaued, so how do you separate retrieval misses from insufficient context, and why can better recall make answers worse?
System
An end-to-end design for a policy and contract assistant at a professional-services firm: metadata-first retrieval, hybrid search with a reranker, per-document access control, and the numbers and threats that decide whether it ships.
Research
A 21 October paper stacks RAG techniques that each help alone: the top three retrieval gains predicted +46.6% NDCG@10 and delivered +17.2%, and the best pipeline changed as the corpus grew from 5K to 200K documents.
Models & Agentic
Anthropic's 2026-10-21 API version rejects citations on user-message text blocks, which hits RAG code that replays retrieved chunks; and the retrieval stack worth defending is hybrid search, contextual chunks and a reranker, measured on your own labelled set.
LabLab · Step 1, part 3
Part 3 of Step 1: swap characters for a BPE tokeniser, grow the model in steps, and read loss against compute using the Chinchilla and Kaplan scaling papers as the yardstick.
Builds
Two three-hour builds from this week's ideas: a small MCP-style gateway with per-identity tool listing, deny-by-default calls and an audit log, or a skill-preselection harness that measures how routing recall falls as the catalogue grows.
GitHub
mksglu/context-mode sits between an agent and its noisy tools: hooks reroute big outputs into a sandbox and a SQLite FTS5 index, so only an answer reaches the context window; the design is worth studying even if you never install it.
Interview
A practice question in the style of an Anthropic Senior AI/ML Engineer loop: a customer's agent has 140 tools across five MCP servers and is getting worse as it grows. Do you add autonomy, or take some away?
System
How Eon replaced ten local MCP servers per developer with one authenticated gateway, and the three design choices that decide whether that pattern holds up: identity, policy and tool listing.
Research
A 13 October paper shows that with 1,000 irrelevant skills installed, agents' recall of the skills they should load falls from 73.6% to 27.4%, and explicit preselection fixes much of it.
Models & Agentic
Managed Agents can now let the model write a program that fans out up to 1,000 agents, which blurs the workflow-vs-agent line and moves the interesting design question to who approves the plan.
LabLab · Step 1, part 2
Part 2 of Step 1: read nanoGPT's train.py line by line, then rebuild its warmup, cosine decay, weight-decay groups, gradient clipping and eval loop around your own tiny GPT, and break each one on purpose.
Builds
Two three-hour builds from this week's ideas: a paired-counterfactual harness that sorts every recovery step into rescue, harm or unchanged, or an offset-addressed session log with idempotent producers and epoch fencing.
GitHub
durable-streams/durable-streams is a small HTTP protocol for append-only, offset-addressed streams; the interesting design choice is treating the agent's token and tool-event stream as a resumable log instead of a socket.
Interview
A practice question in the style of an OpenAI Forward Deployed Engineer loop: design the inference tiers, cache layout and eval gate for a customer with live, near-live and bulk workloads.
System
Why long-running agents fit workflow engines like Temporal, where the model call belongs, and the event-history ceiling that shapes how you design the loop.
Research
A 30 September paper treats agent error recovery as a per-run causal decision: fork from the same state with and without a refresh, count rescues and harms, and route recovery only where it pays.
Models & Agentic
Opus 5.5 and Sonnet 5.5 return a 400 on tool_choice any/tool; the replacement is auto plus strict tool use, which changes how you design the agent loop's guarantees.
LabLab · Step 1, part 1
Step 1 of a public research ladder: grow a character-level GPT from a bigram table to a small transformer, watching the loss at each rung, with nanoGPT as the reference.
Builds
Two three-hour builds that turn this week's ideas (cache-aware routing, query-time context selection, suspendable agent tasks) into small Python systems you can run without a GPU.
GitHub
google/ax declares agent tasks, workspaces and models as manifests and runs them in sandboxes with suspend/resume; the interesting part is the decision to keep short-lived tasks out of etcd.
Interview
A system design round for an AI/Solution Architect role: an internal knowledge assistant on Google Cloud, with a framework, a model answer, and the trade-offs interviewers actually probe.
System
Round-robin across vLLM replicas turns one warm prefix cache into eight cold ones; llm-d's precise prefix-cache routing is a good case study in fixing that, and in what its benchmark does and doesn't prove.
Research
A 30 September paper keeps the attention an agent computes anyway as a tiny graph, then picks context for the next request with no model call, cutting compaction latency by about 95% against Codex-style summarization.