Welcome back! Four pieces on getting more out of the tools already in daily use, and understanding the machinery underneath them. First, the habits burning through a weekly Claude limit. Then the nine layers an AI app actually runs on, an interview question that trips up most candidates, and the memory trick that made open-source LLM serve 24 times faster.
In today's Generative AI Academy Newsletter:
22 rules that cut Claude usage in half: Most people hit their weekly limit because of habits rather than volume.
The full AI stack explained: Nine layers as of October 2026, with two categories that did not exist last year.
LangChain vs LangGraph, which one and when: A standard interview question, answered on control rather than capability.
A 1960s operating system trick made LLM serving 24x faster: vLLM cut KV cache waste under 4% by treating GPU memory like virtual memory.
22 rules that cut Claude usage in half
Stop using Claude. At least until you read these 22 rules.
Most people hit their weekly limit because of habits, not usage.
Use medium effort by default. Low for grammar and translation, High only for ambitious work. Never Max.
Save your free reset for a bad day.
Use Opus 5.5 every chat.
Never switch models mid-conversation.
Never reopen an old chat.
End questions with “No web search. No files.”
Plan in text. Ask for the file last.
Highlight what’s wrong. Never request a full redo.
Load apps on demand.
Add this to your instructions: Once you have answered something, treat that answer as done. On later turns, focus your thinking on what the user is asking now, and don’t go back over an earlier answer unless the user asks about it or points out a problem with it.
Crop every screenshot. Convert every PDF.
Pause scheduled tasks you don’t read.
Open every task with “ask me questions.”
Write one message for all your questions.
Speak your prompt instead of typing.
Never correct Claude. Start over.
Stick to one topic per chat.
Keep your instructions under 2,000 words.
Reused files go in a Project.
In Claude Code: one task, then /clear.
Split your day into 3 sessions.
Ask for images rarely.
Copy them. Work through the rest.
A Claude course from Remus Ranca covers 24 moves for making every session last longer. They span Claude.ai, Cowork and Claude Code.
Master Claude Code in 6 sessions
You already know the developers who seem to move twice as fast as everyone else.
They're not typing faster. They're:
Orchestrating teams of agents
Running parallel Claude instances across git worktrees
Wiring MCPs into their whole stack
Scheduling loops that keep working while they sleep
Almost none of it is written down in one place.
Mastering Claude Code course puts it all in one place.
6 live sessions, each one hands-on:
You start with the agentic loop and model selection
You build Skills and full agent teams from scratch
You apply Boris Cherny's 13 tips plus 6 new April 2026 techniques
You set up all 3 tiers of CLAUDE.md as a second brain that gets smarter every session
You wire MCPs into Slack, GitHub, Gmail and Supabase
You leave having built, tested and shipped your own custom Skill as a capstone
This is the deep track, the habits that separate casual users from power users.
Pricing is locked in for early sign-ups.
We just hit 100K on Instagram
100,000 people now follow GenAI Works on Instagram.
It started as a place to post the things that did not fit anywhere else.
→ A prompt worth stealing
→ A chart that explains something in one look
→ A meme about the model that confidently invented a citation
Now it is where 100,000 people keep up with AI without reading a word of prose.
What you get there:
→ Tips you can use the same day, not next quarter
→ Visual breakdowns of how the models actually work
→ The news that matters, in a format you finish at a traffic light
→ Memes, because some weeks the industry writes its own jokes
No prior background needed.
Start anywhere in the feed and you will know more in a week than most people pick up in a year.
If this newsletter is your morning read, Instagram is the version you get through in 30 seconds.
The full AI stack explained
Your AI app is 9 layers deep. The model is one of them.
The full stack as of October 2026:
1. LLMs → The brain. GPT-6, Claude Opus 5.5 and Gemini on the closed side, DeepSeek V4, Qwen3.8 and Kimi K3 on open weights.
2. Inference & access → Where the model runs. A router like OpenRouter, a cloud like Together or Fireworks, or Ollama on your own machine.
3. Agent frameworks → The loop that turns a model into an agent. LangGraph, Vercel’s AI SDK, the Claude and OpenAI agent SDKs.
4. Tools & protocols (new) → How agents reach everything else. MCP for tools, A2A for agent-to-agent, Agent Skills for reusable know-how.
5. Memory (new) → What survives between sessions. Mem0, Zep and Letta made it a category.
6. Vector databases → Where your knowledge lives. Postgres with pgvector is still the default. turbopuffer counts Cursor and Notion as customers.
7. Embeddings & rerankers → Text, images and PDFs into vectors, then rank what comes back.
8. Data extraction → Messy PDFs and websites into something a model can read. Firecrawl, Docling, LlamaParse.
9. Evals & observability → Traces every call and scores the output before your users do.
Off this year’s list: Llama, since Meta’s new models are called Muse. Also Phi-4, Amazon Nova, Groq, Cassandra, Haystack and Nomic Embed.
Save this for your next build.
LangChain vs LangGraph: which one, and when
A standard question in generative AI interviews, and the answer is about control rather than capability.
LangChain is a framework for building LLM applications out of components: prompts, models, retrievers, tools.
Reach for it when:
→ You’re building a simple RAG app or a chatbot
→ The steps run in a line, start to finish
→ You want an LLM connected to external data and tools quickly
LangGraph is a framework for stateful, multi-step execution built from nodes, edges and conditional transitions.
Reach for it when:
→ You’re building agents or multi-agent systems
→ The process needs loops, retries or branching decisions
→ State has to persist, or a human has to approve something mid-run
The distinction is in one line: LangChain assembles the pieces, LangGraph decides what happens between them when the path is not a straight line.
They stack. LangChain components run inside a LangGraph node without modification, so the usual answer in production is both
From Claude fundamentals to full content workflows, you learn to work with AI without sounding like it wrote everything for you.
Learn AI without paying for a single tool
You already know the people who seem to get twice as much done with AI. They're not on better subscriptions. They're:
→ Running research, analysis and transcription through free tiers
→ Writing prompts with frameworks instead of guessing
→ Building custom agents they reuse every week
→ Shipping actual tools, not just chatting with a model
Almost none of it needs a paid plan.
AI in Practice puts the whole path in one place. 7 modules, 19 lessons, 11+ hours, every tool free or free-tier:
→ You start with how LLMs actually work, tokens, attention, context windows, RLHF, and where they break
→ You write and iterate prompts using established frameworks, context engineering and meta-prompting
→ You build a QR code generator, an email-reply bot, a Zapier form automation and an appointment-booking agent
→ You work in ChatGPT, Gemini, Claude, Perplexity, Manus, GitHub Copilot and Zapier
→ You cover GDPR, the EU AI Act, bias and why most corporate AI projects fail
→ You leave with a portfolio website presenting everything you built as the capstone
Taught by Larno Visser, who trains teams at Philips, BDO, municipalities and schools.
Built for zero-coding backgrounds. Early bird is $60 instead of $80, and the offer closes in 5 days.
→ Save your seat
A 1960s operating system trick made LLM serving 24x faster
The problem is memory waste. During generation a model keeps a KV cache that grows with every token, and nobody knows in advance how long an output will be.
Traditional serving systems reserve a contiguous block big enough for the worst case. Most of it never gets used, and what’s left is too fragmented to give anyone else.
UC Berkeley researchers applied the fix operating systems settled on decades ago. Virtual memory with paging.
PagedAttention stores the KV cache in fixed-size blocks that do not have to sit next to each other in physical memory, allocated as tokens arrive.
Waste in the original evaluation came in under 4%.
Freeing that memory changes what else becomes possible. More sequences fit on the same GPU, so continuous batching has slots to fill.
Two requests sharing a prompt prefix point at the same physical blocks instead of holding two copies. On top of that, optimized CUDA and HIP graphs, custom attention kernels, quantization, speculative decoding and distributed inference.
Original benchmarks: up to 24x the throughput of Hugging Face Transformers, 2 to 4x other optimized serving systems, with no architecture changes to the model.
The gain came entirely from how memory gets managed around the model rather than anything inside it.
Everything else you shouldn't miss
The 75% Google stat nobody can find a source for: Researchers traced stats that AI chatbots cite and found fewer than half led back to a study saying the same thing, including the "94% of first impressions are design" line that started as 15 people looking at health websites in 2004.
44 minutes is a long time to wait with a back injury: NASA and Google built a Mars-bound robot that takes plain-English instructions, runs entirely onboard with no internet, and fetched a medical bag correctly about eight times out of ten in early tests.
Claude Code just restored a program from 1961: A team revived Herbert Simon's program-writing program and a 1984 MIT coding assistant, and Simon's produced the same output he printed in his 1963 paper.
9,600 Nvidia chips, free local pickup: A Minnesota reseller listed 300 racks of A100s on eBay for $60 million with a Make Offer button, no returns.
Learn more about AI from the experts building it
Follow us on Instagram for fast, visual AI updates in 30 seconds.
Subscribe to our Atlas newsletter — trusted by 3M+ subscribers — to stay ahead of AI news across tech, education, and business.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
Learn how to build your next AI application with practical resources and expert guidance.
Explore investment opportunities in the future of AI and join our community-backed growth journey.
