Bobo Lin | AI notes (@BoboLinNotes) header
Bobo Lin | AI notes (@BoboLinNotes) profile picture

AI is full of confident claims. I’m interested in the conditions that make them true. Notes on research, agents, and real-world AI.

Hong Kong
Joined August 2026
Photo by Bobo Lin | AI notes (@BoboLinNotes)Photo by Bobo Lin | AI notes (@BoboLinNotes)Photo by Bobo Lin | AI notes (@BoboLinNotes)Photo by Bobo Lin | AI notes (@BoboLinNotes)Photo by Bobo Lin | AI notes (@BoboLinNotes)Photo by Bobo Lin | AI notes (@BoboLinNotes)Photo by Bobo Lin | AI notes (@BoboLinNotes)Photo by Bobo Lin | AI notes (@BoboLinNotes)Photo by Bobo Lin | AI notes (@BoboLinNotes)Photo by Bobo Lin | AI notes (@BoboLinNotes)
If I had 6 months to break into AI Cybersecurity, I wouldn't waste time collecting certificates. I'd build the skills to attack, defend and secure real AI systems. First, master the new attack surface. Learn prompt injection, jailbreaks, data exfiltration, MCP vulnerabilities, tool poisoning and agent identity security. AI agents introduce risks that traditional application security was never designed to handle. Next, go deeper into offensive security. Combine LLM-guided fuzzing, exploit analysis, adversarial machine learning and automated red teaming. Learn to uncover vulnerabilities, test model defenses and understand how attackers exploit AI-generated code and dependencies. Then learn to defend systems at scale. Build AI-powered threat hunting, log analysis, incident response and threat intelligence pipelines. Master SIEM, SOAR, YARA, OAuth2, SPIFFE and zero-trust architectures. Add privacy-preserving techniques like differential privacy and trusted execution environments. Finally, publish your work. Find real vulnerabilities, contribute to open-source security tools, write technical breakdowns and pursue responsible CVE disclosures. Tutorials teach you concepts. Real-world exploits, defenses and public security research prove you can do the job. Builders get hired.
21
39
1
337
9,017
The biggest bottleneck in AI agents is how they use tools. Traditional MCP makes agents bounce between the model and tools. Every call adds latency, burns tokens, and pushes more context into an already crowded window. The problem gets worse at scale. When an API exposes thousands of endpoints, dumping the entire tool catalog into the context is a terrible way to build intelligence. Code Mode flips the architecture. Instead of calling tools one by one, the model writes code that orchestrates multiple tools in a single execution. Cloudflare has 2,500+ API endpoints, yet Code Mode reduces the agent's interface to just 2 tools. Stop teaching agents to click buttons. Let them write code.
7
The biggest mistake in an AI engineer's resume is listing tools instead of proving what you've built. Companies hiring Forward Deployed Engineers want people who can turn AI agents into working products, not just impressive demos. Build agents that solve real problems. Automate customer support, sales research, recruiting, onboarding, invoices, meetings, and browser workflows. Go deeper with RAG, multi-agent systems, MCP integrations, code reviews, data analysis, incident response, and security guardrails. Make them work in the real world. Pick 2–3 projects. Deploy them. Measure accuracy, cost, and latency. Document what failed and how you fixed it. Your best resume isn't a list of skills. It's proof that you can ship.
2
1
56
AI security engineering is one of the most future-proof skills you can build. Start by learning how AI systems actually get attacked. Master prompt injection, agent hijacking, MCP tool poisoning, and RAG data manipulation. Understanding how attackers exploit trust is the foundation of real defense. Then move beyond prompts. Build secure agent identities, least-privilege permissions, OAuth-based delegated access, PII protection, and output guardrails. An agent should have the tools to work, never unlimited power to abuse. Learn to test what you build. Use PyRIT, Garak, adversarial fuzzing, and automated red teaming to expose vulnerabilities before attackers do. Add model extraction defenses, interpretability, and continuous security checks to your engineering toolkit. Finally, prove your expertise. Study AI governance and compliance, publish security teardowns, responsibly disclose vulnerabilities, and build real-world projects. In a world where AI agents are gaining more access, the people who know how to secure that access will be increasingly valuable.
2
19
542
The biggest mistake in vibe coding is treating working code as production-ready software. Before deploying, lock down authentication, database permissions, server-side input validation, and API secrets. One overlooked permission can expose everything. Protect the infrastructure too. Add rate limits, scan dependencies, hide sensitive error details, and test that your backups actually work. Production needs visibility and financial guardrails. Set up logs, alerts, and spending caps for every cloud and AI API account. AI can build your app in minutes. A hacker can break it just as fast.
7
1
5
309
The biggest mistake in AI security is trusting agents before you can control them. Build systems that sandbox MCP tools, isolate agent-to-agent communication, enforce scoped OAuth permissions, and keep sensitive data away from external models. Autonomous agents need capabilities, not unlimited access. Then attack your own infrastructure. Build multi-turn jailbreak fuzzers, poisoned-memory detectors, RAG citation validators, and visual auditors that catch hidden UI exploits before an agent takes action. Security must be continuous, not a one-time checklist. Automate adversarial testing in CI/CD, enforce constrained outputs, detect API extraction attempts, and scan AI-generated dependencies for supply-chain threats. Finally, make every action accountable. Build cryptographic trajectory logs, investigate deceptive model behavior, and publish real exploit teardowns with reproducible mitigations. Most people learn AI security by watching demos. The best engineers prove it by building systems that survive attacks.
2
3
255
The biggest mistake in AI security is trusting prompts instead of enforcing boundaries. Start with attack detection. Build multi-turn jailbreak fuzzers, MCP tool poisoning shields, and RAG poison detectors to expose attacks before they reach your models. Then secure the execution layer. Build PII masking proxies, grammar-constrained output cages, OAuth brokers, and semantic API rate limiters to control what AI can access, reveal, and execute. Production security goes deeper. Add automated red-team CI gates, computer-use pixel auditors, inter-agent sanitization buses, vector memory scanners, and cryptographic trajectory ledgers. Finally, secure the entire ecosystem. Audit AI supply chains, investigate activation-level vulnerabilities, and build exploit-chain teardown engines. Pick three. Build from scratch. Document every decision. Most people trust the system prompt. Real security engineers enforce constraints at the token level.
5
283
The most valuable AI careers in five years may be jobs most people haven't even heard of yet. Companies won't just need people who build models. They'll need engineers who deploy agents, design their memory, evaluate their decisions, secure their behavior, and make them work in the real world. New roles are already emerging across the stack — Agent Engineers, AI Security Engineers, Evals Engineers, Context Engineers, Voice AI Engineers, AI Infrastructure Engineers, Synthetic Data Engineers, and more. The biggest shift is simple. Less time writing every line of code. More time directing, supervising, and securing systems that can do the work themselves. Don't wait for these roles to become mainstream. Pick one, build something real, and make your skills impossible to ignore. Which AI career are you betting on?
1
69
If I wanted to become an Applied AI Engineer, I wouldn't spend months collecting certificates. I'd build systems that actually work. First, master the fundamentals. Understand tokens, context windows, sampling, streaming, API costs, and model behavior. Then go beyond prompting with context engineering, structured outputs, function calling, and production-grade RAG. Build systems that turn messy inputs into reliable, usable data. Next, learn to adapt intelligence to the problem. Fine-tune with LoRA or QLoRA when necessary, build practical agent workflows, connect tools, and keep humans in the loop. Autonomy without constraints isn't intelligence. It's a liability. Then make everything measurable and safe. Build evals, regression tests, observability dashboards, cost attribution, prompt injection defenses, and PII protection. Benchmark models, route requests intelligently, and design fallbacks. If you can't explain why it failed or what it costs, you're not ready for production. Finally, ship real products. Master deployment, streaming UX, latency optimization, and session management. Publish three working applications with measurable results. The market rewards engineers who turn models into products, not people who merely know the terminology. Prompt wrappers are a commodity. Applied judgment is the moat. The Applied AI Engineer turns frontier models into shipped, revenue-generating products.
5
6
336
The biggest shift in AI engineering is happening right now. Stop building wrappers. Start building systems. Make agents operate real software, not just APIs. Turn LLM outputs into live interfaces, connect products through raw MCP, and replace fragile loops with durable workflows that survive crashes and human delays. Make intelligence cheaper and more reliable. Route easy requests to tiny models, constrain outputs to valid JSON, share KV caches, use speculative decoding, and enforce hard spending limits before runaway agents burn your budget. Treat agents like production infrastructure. Grade trajectories in CI, sanitize messages between agents, protect PII, test failure modes, and shadow-route real traffic before promoting new models. Then close the loop. Automate preference learning, build sub-second voice interactions, add browser and on-device fallbacks, and publish reproducible benchmarks for cost, latency, and task success. The wrapper era is over. The systems era is where the real value gets built. You have a lot of work to do.
3
6
335
If I had 6 months to become a GTM Engineer, I wouldn’t spend them collecting certificates. I’d build the systems that turn product activity into revenue. First, I’d master the data layer with Snowflake, BigQuery, dbt, and Reverse ETL. Then I’d connect buyer signals, intent data, and real-time triggers to identify who is ready to buy. Next, I’d build AI-powered sales workflows with Clay, Apollo, web scraping, and LLMs. I’d also learn PQL scoring, product analytics, usage-based pricing, and how to turn product engagement into expansion revenue. Then I’d tackle the revenue engine itself: attribution, funnel analysis, churn prediction, NRR, sales coaching, ecosystem integrations, and automated QBRs. Every system should answer one question: where is revenue being created or lost? Finally, I’d publish public teardowns, pricing analyses, and real revenue experiments. Tutorials teach you what tools can do. Building proves you can make them drive growth. Most people keep watching tutorials. Builders get hired.
2
189
The best GTM Engineers don't just understand revenue. They build the systems that make revenue measurable, predictable, and scalable. Start with data and attribution. Connect product usage, Stripe billing, Salesforce, and HubSpot. Build pipelines that reveal where revenue comes from, which accounts show buying intent, and which product actions signal real demand. Then put AI to work. Build SDR research agents, personalized outbound systems, call coaching tools, competitor intelligence agents, and automated battlecards. Turn scattered information into actions your sales team can use immediately. Growth doesn't stop at acquisition. Build usage-based billing, PQL scoring, churn prediction, expansion alerts, onboarding friction audits, and automated QBRs. Every customer signal should help the business convert, retain, or expand revenue. Finally, publish your thinking. Break down how leading AI companies design pricing, PLG funnels, and credit systems. Tutorials show what tools can do. Real systems prove what you can build. Systems that prove you can automate, attribute & scale revenue when AI touches your go-to-market stack. Most people watch tutorials. Builders ship systems.
1
3
199
Most people use Claude Code like a chatbot. That’s the biggest waste of its potential. Start by assigning the right roles. Let Fable 5.2 handle planning, Opus 5.5 lead implementation, Sonnet tackle scoped fixes and tests, and Haiku handle logs and file searches. Before coding, run /init to generate your CLAUDE.md, then enter plan mode with Shift+Tab. Give the main model the full picture before letting it touch your files. Make every sub-agent’s model explicit in .claude/agents/, set CLAUDE_CODE_SUBAGENT_MODEL, and use /tasks to verify the assignments. Start with medium effort and scale up only when needed. Keep the main model responsible for edits and critical decisions. Use /compact when context gets heavy and /clear between jobs. Fable plans. Opus builds. Haiku fetches. That's an AI engineering team, not a chatbot.
3
1
5
326
The biggest mistake in an AI engineering resume is listing tools instead of proving you can build real systems. Companies want engineers who can take AI beyond demos. Build a RAG system with 1,000+ documents, an API-connected booking agent, a voice assistant, or an AI coding tool that works on an actual codebase. Go deeper into the engineering side. Build model routing, semantic search, PDF-to-JSON extraction, persistent agent memory, and caching systems that reduce cost and latency. Then prove you understand production. Add automated evals, agent tracing, failure recovery, LLM cost monitoring, and red-team tests that expose vulnerabilities before users do. Don't build all 15. Pick 2–3, ship them to real users, and publish the numbers behind them — accuracy, latency, and cost. Real results are the strongest resume you can have.
2
6
420
The biggest career mistake in AI is learning how to use models instead of learning how to build systems. Start with Context Engineering and advanced RAG. Master token budgets, hybrid search, reranking, GraphRAG, and knowledge extraction. These are the foundations of reliable AI applications. Then move beyond retrieval. Build self-correcting RAG pipelines, durable agent workflows, multi-agent systems, and MCP-powered tools with memory, state management, and human oversight. Production is where real engineers stand out. Learn evaluation, security, prompt injection defense, model routing, cost optimization, and failure recovery. A demo that works once is not a production system. Finally, publish what you build. Share architecture decisions, reproducible benchmarks, cost breakdowns, and technical teardowns. Tutorials make you familiar with AI. Building real systems makes you valuable.
4
226
The next 3 years will reward the people who build in AI, not the people who just talk about it. Don't chase every new model or wrapper. Pick one layer of the stack — core models, AI agents, multimodal systems, infrastructure, or applied AI — and become exceptionally good at shipping it. The real edge is in the hard parts. Better evaluations and fine-tuning, reliable agents, real-time voice and vision, cheaper inference, and workflows that solve expensive problems. You don't need a PhD, millions in funding, or a huge team. Find one painful problem, build a solution, put it in front of real users, and improve it every week. Three years of focused execution can create a gap that most people won't be able to close. Start building now.
1
1
8
327
Mastering AI Cybersecurity takes more than learning how to use AI. If I had 6 months, I’d start with the systems attackers actually target — LLM applications, prompt injection, jailbreaks, data leaks, and vulnerable AI-generated code. Then I’d go deeper into offensive security. Build fuzzers, generate exploits, hunt zero-days, run adversarial attacks, and learn how to break models before someone else does. But breaking things is only half the job. I’d master AI-powered threat hunting, automated incident response, agent security, MCP defense, machine identities, and zero-trust architecture. Finally, I’d build proof of work. Run automated red teams with PyRIT and Garak, audit AI supply chains, hunt public CVEs, publish exploit breakdowns, and practice responsible disclosure. Tutorials teach you what AI security is. Building real defenses proves you can do the job.
4
2
60
1,818
The best AI portfolios in 2026 won't be built by copying tutorials. They'll be built by solving problems that production systems actually face. Start with agents that can use terminals, control browsers, and run for 24 hours without losing state. Real autonomy means surviving messy environments, not just passing a demo. Then build the infrastructure behind them. Ship production-ready MCP servers, multi-agent research workflows, and evaluation pipelines that test every step instead of blindly trusting final answers. Make your systems secure and economical. Red-team agents against 10+ vulnerabilities, cut inference costs by 50% with smart routing, and push private AI onto local devices with zero API bills. Finally, build systems that improve themselves through feedback, synthetic data, and LoRA fine-tuning. Publish measurable results. Tutorials show what you watched. Working projects show what you can build. Most people collect certificates. Builders create proof. That's what gets you hired.
2
2
89
If I had 6 months to become an Agentic AI Engineer, I’d stop collecting tutorials and start building systems. First, master the foundations. Python async, FastAPI, APIs, event-driven architecture, and the fundamentals of how LLMs handle context, tokens, latency, and failures. Then learn to make agents work. Structured outputs, tool calling, memory, state management, ReAct loops, planning, and multi-agent orchestration with LangGraph or CrewAI. But building an agent is only half the job. Learn human approval workflows, evaluation, tracing, observability, security, prompt injection defense, and reliable error recovery. Finally, take it to production. Deploy with vLLM or SGLang, scale with Kubernetes, build CI/CD pipelines, and publish real projects with demos and architecture docs. 12 stages. 6 months. One rule that matters most. Stop watching people build AI. Build something people can actually use.
5
5
28
1,108
The biggest mistake in becoming an Agentic AI Engineer is building agents that only work in demos. Start with systems that can handle real production pressure. Build async orchestration, dynamic context budgeting, structured tool routing, and memory systems that know what to keep and what to forget. Then make agents resilient. Give them self-healing loops, typed multi-agent collaboration, human approval gates, and browser fallbacks when APIs fail. Real agents must recover, adapt, and know when to stop. Production means more than getting the right answer. Build trajectory evaluations, OpenTelemetry tracing, prompt-injection protection, sandboxed execution, cost kill-switches, and shadow testing. Reliability, security, and cost control are engineering requirements. Finally, publish your architecture decisions. Show how your agents coordinate, handle failures, protect tools, and control cost. Strong engineers don't just know how to use LangChain. They know how to design systems that survive the real world. Most people collect AI tutorials. The next generation of Agentic AI Engineers will be judged by the systems they actually ship.
8
2
12
484
The fastest way to become a GenAI Engineer in 6 months is to stop learning AI as a collection of tools and start building real systems. Master the foundations first. Python, FastAPI, async architecture, multimodal models, context engineering, structured outputs, and advanced RAG. These are the building blocks behind reliable AI applications. Then learn to make agents work in the real world. Build durable workflows with LangGraph or Temporal, connect tools through MCP, add human approval, and handle failures, permissions, and security. Production skills are what separate engineers from tutorial watchers. Learn evaluation, observability, inference optimization, caching, cost control, and deployment. Measure latency, track quality, and make every system more reliable. Finally, ship projects publicly. Publish your architecture, document trade-offs, share performance benchmarks, and apply for GenAI, Agent, Context, or AI Infrastructure roles. Tutorials teach you what is possible. Building proves what you can do.
2
8
469