The library

Stage 5 of the path, and the other omission from the source roundup. One derivation carries the course: during decode your arithmetic intensity is numerically your batch size (parameters cancel), so against an H100's ridge of 295 FLOPs/byte a single user runs at ~0.3% of the chip's arithmetic capability. Then the constraint nobody budgets for — the KV cache, not the weights, sets concurrency: four H100s serving Llama 3 70B hold ~501 sequences at 1K context and three at the advertised 128K, and at high batch the cache reads more bytes per step than the weights do. Includes PagedAttention with the paper's own profiling (only 20.4–38.2% of KV memory in prior systems held real token state), the prompt-ordering rule that silently destroys prefix caching, and two widely repeated claims corrected: quantization "doubling throughput" (only if you quantize the KV cache too), and speculative decoding as a general speedup — it spends idle FLOPs, so it fades exactly where you need it. Ends in the Serving Simulator: predict the winning lever, then watch speculative decoding fall from 2.31× at batch 2 to 1.05× at full batch.

Stage 4 of the path, and one of the two the source roundup omitted entirely. Error-analysis-first practice: read 100 traces and open-code the failures before choosing a metric, because a rubric written first measures the failure modes of a generic system and yours is not generic. Then the technical core — your LLM judge is an unvalidated classifier, and the 1978 Rogan-Gladen estimator turns its pass rate into a number you can defend: a judge passing 80% of traffic with a 95% TPR and 55% TNR means a true quality rate of 70%, ten points of flattery. Includes the half most treatments skip — a mediocre judge also divides your standard error by J = Se+Sp−1, so it costs statistical power, and in a same-judge A/B the bias cancels while the noise amplification does not. Plus documented benchmark failures with numbers (57% of MMLU's analysed Virology questions had errors; 68.3% of SWE-bench samples were filtered out to build Verified) and, deliberately left in, one claim the course could not verify and therefore does not repeat. Ends in the Judge Calibrator: correct a number, price judge quality against sample size, and watch a plausible 4-point improvement fail to clear its own noise floor.

Stage 3 of the path. SFT, RLHF, DPO and successors — with DPO's closed form derived in four steps rather than asserted, so you can point at the exact line where the intractable normalising constant cancels (it depends only on the prompt, and Bradley-Terry needs only a reward difference). Then the part most treatments skip: what that closed form costs. It is off-policy, it assumes Bradley-Terry, and it can suffer likelihood displacement — the probability of the response you preferred falling while the margin improves, with mass moving to responses of opposite meaning (Razin et al.: preferring "No" over "Never" can raise "Yes"). Plus why fine-tuning cannot add knowledge, argued from the objective rather than from anecdote; why verified reward removed the ceiling that learned reward models impose; and the part of the DeepSeek-R1 story that gets dropped — R1-Zero's pure RL also produced language mixing and poor readability, which is why the shipped model has a cold-start SFT stage. Two interactive tools: a recipe diagnosis that will tell you to stop, and a displacement simulator where you can find the similarity threshold at which preference training starts working against you.

Stage 2 of the path. Scaling laws taught as unit economics rather than as facts about quality, and every number derived in front of you: C ≈ 6ND checked against Llama 3's published compute to within 0.3%, N = √(C/120) as mental arithmetic, and 16 bytes per parameter of Adam state before you store an activation. Includes the parts usually left out — that Chinchilla's headline analysis failed replication in 2024 (the fit was poor, the confidence intervals implausibly tight, and it contradicted the paper's own other two approaches, though the 20:1 rule survived), that Meta's own scaling fit gave ~41 tokens/parameter rather than 20, and that Llama 3 405B was trained essentially at its predicted optimum — a counterexample to "everyone over-trains now." Plus the H100 datasheet trap (1,979 TFLOPS is the sparsity figure; dense is ~989), the arithmetic-intensity ridge point that explains FlashAttention, and PaLM's finding that loss spikes came from a batch-times-parameter-state interaction, not bad data. Ends in the Training Run Planner: guess the compute-optimal model size before the tool computes one, then trade tokens-per-parameter against a priced serving bill and a data ceiling.

Stage 1 of the five-stage path. The forward pass with nothing hidden: attention derived from a Python dictionary in four steps, the residual stream as a sequence of additive edits rather than a pipeline of transformations, and where the parameters actually live. Includes the arithmetic most explanations skip — why the familiar "one-third attention, two-thirds MLP" rule is exactly right for GPT-2 and wrong for every current model (grouped-query attention and SwiGLU move Llama 3 70B to roughly 18/82), why removing the causal mask makes training loss fall, and what the 1/√d scale factor is actually protecting. Each interpretability claim is stated with its limit: layer deletion degrades gracefully for deeper layers only, and "each head has a job" is not supported for frontier models. Ends in the Forward Pass Ledger — commit to a guess about the parameter split, then get every tensor shape, an exact parameter ledger reproducing four published model sizes, and five ablation switches that show which sabotage changes the numbers and which changes only the behaviour.

A widely shared "best AI courses for 2026" roundup offers to build a learning path and never does. This builds it. Opens with a currency audit of all 13 recommendations, checked link by link: Papers with Code was sunset in July 2025 and its domain now redirects; arXiv-sanity is dead and its domain is now a job board; Full Stack Deep Learning froze at Spring 2023; fast.ai is still the 2022 recording; CS231n's free YouTube lectures are the 2017 offering and predate ViTs entirely — while its Spring 2026 syllabus is genuinely modern. Then the path itself: five stages (mechanism → training stack → post-training → evals → inference economics), each teaching the concept in plain language before pointing at the best current resource, and each ending in a clearing test phrased as "you can do X unaided." Plus the three areas the list omits — evals, inference optimisation and post-training — where most 2026 practice lives, a ladder of 8 papers with a falsifiable definition of "reimplemented it," and a module reading the whole thing as a curriculum design problem. 11 modules, plus a Path Builder that makes you guess your entry stage before eleven capability questions compute one, then produces a dated schedule from your real weekly hours.

Anthropic deleted 80%+ of Claude Code's system prompt for Opus 5 and Fable 5 with no measurable eval loss. Built on Thariq's July 2026 article: the six then→now shifts (rules→judgement, examples→interfaces, upfront→progressive disclosure, repetition→tool descriptions, CLAUDE.md→auto-memory, specs→rich references), plus applied modules on CLAUDE.md, skills, tool design, and a staged deletion playbook. 15 modules with quizzes & flashcards.

Inside Demis Hassabis's July 2026 manifesto calling for a "FINRA for AI" — the Anthropic Mythos & Fable export-freeze that triggered it, the three competing blueprints (Hassabis's SRO, Amodei's FAA, Altman's IAEA), the regulatory-capture and open-weight fights, a full governance history, key-figure profiles, and a timeline. 14 modules with quizzes & flashcards.

A curated course built around 15 must-follow X accounts across Anthropic, OpenAI, Google AI, Cursor, and xAI — who they are, what they teach, and what they're saying now. Auto-updated weekly.

A technical deep dive into the current landscape of generative AI: frontier models, benchmarks, architectures, and emerging trends shaping the field.

Comprehensive deep dive into the July 2026 model landscape: Thinking Machines' Inkling (975B MoE), the MoE architecture revolution, every major open-weight and closed model family, reasoning models, multimodal, small/efficient models, pricing, and what's next.

Explore how reinforcement learning techniques are applied to large language models — RLHF, reward modeling, and alignment strategies for production systems.

Run state-of-the-art AI locally: quantization, fine-tuning, deployment, and performance optimization with Google's Gemma 4 model family.

Building the next interface for learning — speech recognition, synthesis, real-time processing, and designing voice-first educational experiences.

Building trustworthy AI products through rigorous measurement — eval design, datasets, LLM-as-judge, offline vs. online testing, reading results honestly, and a production measurement culture.

Advanced course on building AI-native products: LLM integration patterns, RAG deep dive, AI agent architecture with MCP, fine-tuning, AI product design patterns, evaluation, MLOps, AI at Coursera, and building an AI-first engineering org.

Built on Anthropic's copilot-kit-ag-ui quickstart, read from source rather than from the README. What a standard protocol between an agent and a UI actually buys you — 33 event types, RFC 6902 state deltas, interrupt outcomes — and the four places it costs you: lock-in, five debugging surfaces, a silent state-divergence failure mode, and capabilities that are advertised but unimplemented. Then contextualised: a file-by-file verdict on migrating the Learn Imagine whiteboard tutor (short version: don't), and the build-vs-adopt case at platform scale. 10 modules, plus The State Desk: a simulator that locks the results until you commit to a state-ownership call and write down why, then runs it through refresh, interruption and a second tab — and a deterministic instrument for the reconciliation tokens nobody budgets for.

Hands-on guide to building autonomous agents with the Claude Agent SDK: tool use, multi-turn conversations, and agent orchestration patterns.

Design and deploy multi-agent systems — delegation, coordination, shared context, and scaling agent networks for complex workflows.

Comprehensive course on monitoring, debugging, and optimizing AI agent systems — traces, spans, metrics, and production observability patterns.

Build real-time AI applications with streaming, WebSockets, function calling, and low-latency voice and text interactions.

Deep dive into the Model Context Protocol and the full stack around it — primitives, transports, tool design, building servers, the agent harness, multi-agent orchestration, and security.

The definitive course on the hot new discipline (June 2026): designing automated, self-correcting loops for coding agents. Based on Addy Osmani, Boris Cherny, Geoffrey Huntley, and Peter Steinberger. 10 modules, 30+ code examples, 15K+ words, plus an animated explainer and in-module concept clips.

Boris Cherny's ladder for how a team grows into working with AI: Gated → Assisted → Parallel → Supervised Autonomy → AI-Native, and how leverage climbs from 0 to 1,000+ agents per engineer. Includes a 42-second animated explainer.

The practical guide to the 2026 coding agent landscape: Claude Code, Cursor, Copilot, Kiro, Devin, and more — effective prompting, real-world workflows, and building an agent-first engineering culture.

Agent architectures (ReAct, plan-and-execute, multi-agent), tool use, MCP, memory systems, guardrails, evaluation frameworks, cost optimization, observability, and real-world case studies.

Comprehensive prep for Anthropic's CCA-F certification exam: all 5 domains (agentic architecture, MCP, Claude Code, prompt engineering, context management), 120 practice questions, full mock exam with timer, and quick-reference cheat sheet.

Companion study guide covering all 17+ free Anthropic Academy courses: API fundamentals, Claude Code mastery, Agent SDK & subagents, MCP, prompt engineering, production patterns, and a CCA-F certification readiness checklist with study plans.

Inside the August 4, 2026 announcement: Account vs Virtual Wallets and the delegation model, the x402 protocol on the wire, cloudflare.pay handles over Web Bot Auth, the Monetization Gateway sell side, spending guardrails against prompt injection, and a full analysis of how a credential platform should play it — plus x402 vs AP2, Stripe Link Agents, and metered billing. 8 modules with quizzes & flashcards.

Cloudflare open-sourced the AI workspace it runs internally: agents, user-owned apps (Gadgets), and Gatekeepers — a capability-based security layer that makes human-in-the-loop asynchronous. Includes a full adopt/don't-adopt analysis for Coursera.

The complete Cloudflare developer platform: Workers, KV, R2, D1, Pages, Workers AI, Durable Objects, security, and why coding agents make it the new default for internal tools.

Built on Sean Goedecke's August 2025 essay, then stress-tested against the people who disagree with him — Fielding, Google AIP, Zalando, Stripe, GitHub and the RFCs. Resource modelling, compatibility, versioning, errors (RFC 9457), pagination, idempotency, auth blast radius, deprecation (RFC 9745/8594) and the GraphQL question. 10 modules, plus The Change Desk: a decision simulator that locks the results until you commit to a call and write down why, then bills you for the deprecation debt you deferred.

Distributed systems, microservices vs monolith, database sharding, caching strategies, load balancing, message queues, CAP theorem, and 5 interview-ready system design problems with solutions.

Internal developer platforms, CI/CD pipelines, Terraform & Kubernetes in production, observability with OpenTelemetry, SRE practices, feature flags, and database migrations at scale.

Production operations for AI systems — GPU clusters, model serving, inference optimization, cost management, and reliability engineering.

CTO-level deep dive into claims processing systems — event-driven architecture, HIPAA compliance, real-time adjudication, and system modernization.

Modern data stack, Snowflake/BigQuery/Redshift, dbt, Airflow, real-time streaming with Kafka & Flink, data mesh vs lakehouse, A/B testing infrastructure, and ML feature stores.

The 2026 CR26 overhaul, Certification Classes A–D, and the FedRAMP 20x fast track. Should Coursera pursue federal certification? Market sizing, competitor status, engineering implications, an interactive cost estimator, a readiness checklist, CMMC, and a 90-day decision plan.

Zero trust architecture, SOC 2 Type II, OWASP Top 10, supply chain security, secrets management, IAM, vulnerability management, and the Instructure/Canvas breach case study.

Anthropic’s stage-by-stage playbook for running plan, design, build, test, deploy and maintain with agents — taken seriously, then audited against the evidence it doesn’t cite. The position: it is a control-system document wearing a productivity document’s clothes. Its premise (“code is no longer the bottleneck”) is a defensible queueing claim and an indefensible productivity one — METR’s RCT measured experienced developers 19% slower while they believed they were 20% faster, DORA 2025 finds throughput up and stability down in the same population, and GitClear’s 623M-change corpus shows refactoring down 70% and duplication up 81%. Which makes the playbook more useful, not less. Covers the artifact chain as traceability with the cost removed (and the ordering property that is the only bit worth enforcing), the sentence most readers skim — “a skill is a control, though an advisory one” — and what follows from it, correlated blind spots when one model family authors and reviews, why the managed-settings sandbox block is load-bearing and the permissions block is ergonomics, and why most of the playbook’s own indicators measure the step that just got cheap. 10 modules, plus The Control Board: five real policy requirements where you must pick the enforcement mechanism and commit your reasoning before anything is revealed — it keeps a paper-control ledger, and in scenario five an auditor arrives.

Seven organisations stood up forward-deployed engineering practices between March and July 2026, four with nine-figure money — Microsoft's $2.5B Frontier Company, AWS's $1B unit, OpenAI's Deployment Company, Anthropic-backed Ode, plus Accenture and EY. Built from a practitioner's LinkedIn argument that they will all fail the same way, then fact-checked against primary sources: every claim graded Verified, Reported, Assembled, Opinion or Unsupported. Includes the $9B arithmetic taken apart, the rotation premise that no vendor has actually published, an attribution note on the piece's strongest line, and the contract terms a buyer should insist on. 8 modules, plus a Build / Buy / Embed decision tool that makes you commit to an answer before it computes one, then tells you which single assumption is carrying your decision.

Managing managers, org design (pods, chapters, guilds), hiring at scale, performance management, engineering culture, roadmap planning, reorgs, remote teams, and developer experience.

BATNA/ZOPA frameworks, VP/CTO compensation negotiation with real numbers, board presentations, executive presence, crisis communication, and a CPTO role negotiation playbook.

Strategic frameworks for integrating AI into education products — market positioning, user research, roadmapping, and competitive moats.

OpenAI and Anthropic have both moved toward an IPO, and neither has publicly filed an S-1 — verified against EDGAR full-text, company search and the CIK master file. What a confidential draft submission does and doesn't disclose, Rule 135, provenance tiers for a sector awash in unverifiable numbers, the reported economics of both labs, PBC and Long-Term Benefit Trust governance under public shareholders, and the commitment asymmetry (Coursera's filed $27.5M against OpenAI's reported $665B). 8 modules, plus The Inference Budget: a simulator that makes you commit to a three-year cost forecast before it will show you what your own assumptions imply.

A CTO-level briefing on Fanatics: the direct-to-fan empire, the trading-card economy, Collect vs Collectibles vs Live, the competitive landscape, marketplace/vault/auction architecture, and a first-90-days playbook.

A CTO onboarding briefing on Kiddom: the "human-first" thesis, the K-12 / HQIM market, the curriculum + LMS + assessment + AI platform, the partner-content moat and its fault line, competitors, a reference architecture, and a first-90-days playbook.

How an Italian startup built an $18B acquisition machine: the F-1 filing, the buy-cut-optimize playbook, 50+ acquisitions (Evernote, Vimeo, AOL), AI-first engineering with ~800 people running 500M+ users, and lessons for tech leaders.

A company briefing on Instructure (maker of Canvas LMS): the BYU open-source origin, the LMS market, the Canvas + Mastery + Parchment ecosystem, the IPO → Thoma Bravo → KKR private-equity story, the post-Blackboard-collapse competitive map, a reference architecture, the IgniteAI strategy, and a risk-and-playbook module shaped by the 2026 breach.

LBO mechanics, the operating playbooks of Thoma Bravo, Vista, KKR, Silver Lake, Francisco Partners & H&F, what CTOs experience under PE, the Instructure/KKR case study, financial engineering, PE success stories and failures, and a practical framework for evaluating PE-backed roles.

The same GPT-4 raised practice scores 48% and left students 17% worse on the exam they took without it — while a hints-only variant raised practice 127% and erased the harm. Built on the pre-registered PNAS RCT and a 2026 evidence review: the replicated pedagogy core and where it backfires, honest effect sizes against the 2-sigma myth, and a six-level architecture whose credential is an evidence chain that decays. 15 modules with quizzes & flashcards.

The certificate is a PDF, the verification is a support ticket, and the learner owns nothing. The full standards stack that fixes it — W3C VC 2.0, Open Badges 3.0, CLR 2.0, DIDs, OpenID4VCI/VP, status lists — plus wallet and issuer architecture, the assessment-provenance schema that is the real moat, x402-metered verification on Workers, GDPR/FERPA analysis, and a five-phase roadmap with cost estimates and explicit kill criteria. 12 modules with quizzes & flashcards.