Bobo Lin | AI notes

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

[Bobo Lin | AI notes @BoboLinNotes]

Claude Code is powerful, but the right plugins can make it feel like a senior engineer instead of a code generator.

Start with Agent Skills and Context7. One gives Claude reusable engineering workflows, the other keeps its library knowledge up to date.

Add Graphify to map your codebase, and 21st MCP to turn prompts into polished frontend components. Better context, better architecture, better UI.

Then install Strix to scan your application for real security vulnerabilities. Because shipping code is easy. Shipping secure code is the hard part.

Junior devs write code. Senior devs understand the system, make better decisions and catch problems before they ship. Bookmark this for your next build.

[Bobo Lin | AI notes @BoboLinNotes]

The biggest mistake in an AI Engineer interview is thinking that knowing LLMs is enough. Companies want engineers who can turn models into reliable products.

Master the fundamentals: Agent reasoning, tool calling, RAG, embeddings, memory, context windows and MCP. Know when to use each and why.

Then learn system design. Single vs multi-agent, task orchestration, human approval, tool failures and real-world workflows. Can you design an agent that actually solves a problem?

Production is the real test. Debug hallucinations, prevent prompt injection, reduce latency and token costs, measure performance, and survive a million users.

These 25 questions reveal the difference between people who talk about AI and people who can build it. Bookmark this and send it to someone preparing for an AI Engineer interview.

[Bobo Lin | AI notes @BoboLinNotes]

The biggest mistake in AI/ML is learning to use models without learning how they work. If you want to stand out in 2026, build systems that prove you understand the engineering behind intelligence.

Start with the foundations. Build custom Triton kernels, CUDA activation functions, synthetic data pipelines and MinHash deduplication systems. Understand how data quality, memory access and compute efficiency shape model performance.

Then master model training. Experiment with QLoRA, TIES merging, KTO alignment, FSDP distributed training, sparse MoE routing and vision-language fusion. Don't just call models. Learn how to change their capabilities.

Next, focus on inference and reliability. Build AWQ quantization, speculative decoding, semantic routing, automated hyperparameter sweeps and adversarial red-team testing. Real engineering means getting better performance without sacrificing quality or blowing up costs.

Finally, reproduce an arXiv paper and publish your code, weights and benchmarks on Hugging Face. You don't need 15 projects. You need a few that prove you can train, optimize and scale real models. Most people watch tutorials. Builders understand the weights, master the systems and ship.

[Bobo Lin | AI notes @BoboLinNotes]

The biggest problem with traditional MCP isn't tool access. It's the endless back-and-forth between agents and tools.

Every call sends results back to the model, consumes more tokens, and adds latency. When hundreds or thousands of tools are available, even the tool definitions can overwhelm the context window.

Code Mode changes the architecture. Instead of asking the model to call tools one by one, let it write code that orchestrates multiple tools in a single execution.

The result is fewer round trips, less context overhead, and more efficient workflows. Cloudflare has over 2,500 API endpoints, yet Code Mode lets agents work through just two tools.

Stop treating AI agents like button-clicking machines. Give them the ability to write and execute code. That's a much more powerful way to use tools.

[Bobo Lin | AI notes @BoboLinNotes]

The biggest mistake when growing from zero is posting and waiting for attention. Nobody knows you yet, so go where your audience already spends time.

Find 15–20 influential accounts in your niche, track their posts, and join conversations early. Don't chase 100 lazy replies. Write 20–30 thoughtful ones that actually add value.

Forget empty compliments. Bring fresh data, useful examples, practical insights or smart disagreements. Make every reply strong enough to stand on its own.

When people visit your profile, give them a reason to follow. A clear bio, a strong pinned post, and at least five valuable posts can turn curiosity into followers.

Replies get you discovered. Original content makes people stay. Show up consistently for 30 days, and strangers can become familiar faces. You don't need an audience to start. You need to become valuable to one that already exists.

[Bobo Lin | AI notes @BoboLinNotes]

AI Security Engineering could become one of the most valuable skills of the next decade. As companies deploy agents with access to emails, databases and business tools, every new capability creates another attack surface. Knowing how to build AI is no longer enough. You need to know how to break it and secure it.

Start by mastering the attacks. Study OWASP Top 10 for LLMs, indirect prompt injection, jailbreaking, MCP tool poisoning, RAG data poisoning and model extraction. Build real attack scenarios instead of memorizing security definitions. Understand how attackers exploit an agent's permissions, context and trusted tools.

Then learn to build defenses. Implement least-privilege OAuth access, PII masking, JSON schema enforcement, secure tool validation and human approval gates. Protect retrieval systems with access controls, and use cryptographic and privacy safeguards to keep sensitive data out of the wrong hands.

Go beyond manual testing. Use PyRIT and Garak to automate adversarial testing, integrate security checks into CI/CD, and track every prompt, retrieval and tool call with immutable audit logs. Learn model interpretability, alignment testing and AI governance. Security must be continuously tested, not simply promised in a system prompt.

Finally, build a public portfolio. Publish threat models, reproduce vulnerabilities, responsibly disclose CVEs and document real attack-and-defense experiments. Certifications can open doors, but demonstrable security research builds credibility. AI is gaining more power and permissions every year. The people who know how to secure that power will become increasingly valuable.

[Bobo Lin | AI notes @BoboLinNotes]

AI companies aren't impressed by another chatbot demo. They want engineers who can turn LLMs into reliable products. If you're job hunting, build systems that prove you can do more than call an API.

Focus on real engineering challenges. Build a RAG pipeline with hybrid search, a voice agent, a document extraction system, and an agent with memory, tools and human approval. These projects show you can solve problems beyond simple prompting.

Then prove you understand production. Add automated evaluations, hallucination detection, red-team testing, prompt versioning and observability. Make your system measurable, testable and safe before users depend on it.

Cost and performance matter just as much. Build model routing, semantic caching, structured output validation, a fine-tuned small model and a local LLM deployment. Show actual numbers for latency, accuracy and API costs.

Don't build all 15. Pick two, deploy them, publish the benchmarks and explain what you learned. A demo proves you can use AI. Production proves you can engineer it. Bookmark this for your next AI job hunt.

[Bobo Lin | AI notes @BoboLinNotes]

AI security is becoming one of the most valuable engineering skills in the industry. Reading about vulnerabilities isn't enough. You need to build systems that can actually detect and stop them.

Start with attack detection. Build jailbreak fuzzers, RAG poisoning detectors, computer-use security auditors, and automated red-team pipelines. Make your systems discover attacks that ordinary benchmarks miss.

Then secure the agent infrastructure. Build MCP tool shields, OAuth delegation brokers, inter-agent sanitization layers, PII protection proxies, and semantic API rate limiters. Every tool call, identity and data boundary is a potential attack surface.

Go deeper into system integrity. Create vector memory scanners, cryptographic execution ledgers, constrained-output engines, AI supply-chain scanners, and activation-level safety dashboards. Learn to verify not just what a model says, but what it can access, execute and change.

Pick three projects. Build them from scratch, test real attack scenarios, and document every engineering decision. Most people trust the system prompt. Real security engineers enforce boundaries in the architecture itself. Bookmark this. You'll need it.

[Bobo Lin | AI notes @BoboLinNotes]

If I had to become an Applied AI Engineer from zero, I wouldn't waste months collecting certificates. I'd follow one rule: build something at every stage, and make every project prove a real engineering skill.

First, master the economics of LLMs. Understand tokens, context windows, streaming, API costs and latency. Then learn context engineering, structured outputs, RAG and fine-tuning. Build systems that retrieve the right information, produce reliable data and adapt models to real business needs.

Next, learn to make AI reliable. Build evaluation pipelines, regression tests and guarded agents with human approval. Add prompt injection defense, PII protection and automated quality checks. A demo that works once is easy. A system that keeps working is the real challenge.

Then move into production. Deploy APIs, build responsive streaming interfaces, add model routing, fallbacks, caching and observability. Track cost per request, latency and failure rates. Great AI products aren't just intelligent. They're fast, affordable and dependable.

Finally, ship three real applications and publish measurable case studies. Show how you reduced costs, improved accuracy or eliminated hallucinations. Certificates prove what you've studied. Deployed products prove what you can do. Prompt wrappers are a commodity. Engineering judgment is the moat.

[Bobo Lin | AI notes @BoboLinNotes]

Forget building another AI wrapper. The 2027 advantage belongs to engineers who build the systems behind AI. Computer-use agents, voice interfaces and generative UI are turning models into products that can actually get things done.

Stop building fragile agents. Replace endless loops with durable state machines, enforce structured outputs, test execution trajectories in CI, and make agents recover from failures instead of starting over. Reliability is the real competitive edge.

Inference costs are your next battlefield. Route easy requests to cheap models, reserve frontier models for hard problems, share KV caches, use speculative decoding, and set automatic spending limits. Faster AI is great. Cheaper AI that delivers the same results is even better.

Security cannot be an afterthought. Protect PII, isolate agents, defend against prompt injection, fuzz tool calls, and keep a browser or local WebGPU model ready when cloud APIs fail. Build systems that survive failure, not just demos that look impressive.

Finally, prove it with reproducible benchmarks. Compare cost, latency and task success against a naive baseline, then publish the results. The era of wrapping APIs is fading. The era of engineering complete AI systems is here. You have work to do.

[Bobo Lin | AI notes @BoboLinNotes]

You don't need another marketing certificate to become a GTM Engineer. You need to understand how data, AI and revenue connect. Give yourself six months to build that skill set.

Start with the foundations. Learn SQL, dbt, Snowflake, reverse ETL and customer identity resolution. Then build systems that detect buying intent, track product usage, identify PQLs and trigger sales alerts. The goal is simple: turn scattered signals into actionable opportunities.

Next, automate the revenue engine. Use Clay, Apollo and LLMs to enrich leads, personalize outreach and build AI SDR workflows. Connect usage-based pricing, Stripe billing, product analytics and churn prediction. Don't just automate sales. Understand how a product creates and captures value.

Then master the bigger growth picture. Build attribution models, ROI dashboards, sales coaching systems and automated QBRs. Explore integrations, community-driven growth and dark social attribution. Learn to connect every customer interaction to retention, expansion and revenue.

Finally, publish real teardowns and measurable projects. Show how you improve conversion, reduce churn and increase revenue. Tutorials teach tools. Real systems prove judgment. Most people keep learning GTM. Builders make themselves impossible to ignore. (Bookmark it)

[Bobo Lin | AI notes @BoboLinNotes]

AI security is becoming one of the most valuable skills in the industry. If you want to stand out, stop just building agents and start learning how to break them.

Build defenses against prompt injection, jailbreaks, leaked secrets and PII exposure. Add output validation, RAG access control and permission gates that require human approval for dangerous actions.

Then go deeper. Create isolated agent sandboxes, test poisoned documents and emails, audit MCP tools, and monitor suspicious outbound requests. Don't trust a tool just because an AI can call it.

Make security measurable with tamper-proof audit logs, model supply chain checks, automated attack benchmarks and CI regression tests. Build an incident replay tool to reproduce exactly how an agent was compromised.

Pick three projects. Build them from scratch, attack your own systems, and fix every weakness you discover. Most developers know how to build agents. Very few know how to defend them. Bookmark this.

[Bobo Lin | AI notes @BoboLinNotes]

The future of GTM Engineering isn't sending more emails. It's building systems that turn product signals into predictable revenue. If you want to stand out, stop learning tools and start building these 15 systems.

Connect product usage to your CRM, track buying intent in real time, score product-qualified leads, and build AI agents that research prospects before outreach. Automate usage-based billing, lead routing, onboarding recovery and personalized customer reports. Every customer interaction should create a measurable growth signal.

Then tackle revenue intelligence. Build multi-touch attribution, churn prediction, sales call coaching, competitor monitoring and dark social tracking. Know exactly where revenue comes from, why customers leave, and what makes them buy. Guesswork is not a growth strategy.

The real advantage comes from connecting everything. Automate commissions, identify expansion opportunities, improve free-to-paid conversion, and turn customer data into actionable decisions. Publish your architecture, benchmarks and real-world results. Don't just claim you understand GTM.

Most people learn marketing tools. Great GTM Engineers build revenue infrastructure. Pick a few projects, ship them, and prove you can make AI drive measurable business growth. Bookmark & Repost.

[Bobo Lin | AI notes @BoboLinNotes]

AI Agent Engineering could become one of the most valuable skills of the next decade. But watching tutorials won't get you hired. You need to build agents that can actually work.

Start with state machines, LangGraph, Temporal and MCP. Build agents that can manage complex workflows, discover tools, handle permissions and recover from failures. Learn how to give AI reliable hands, not just smarter prompts.

Then go beyond text. Build browser agents that operate real software, memory systems that manage long conversations, and multi-agent teams that divide complex tasks. Add sandboxed execution so agents can run code without putting your infrastructure at risk.

Production is where real engineers separate themselves. Master human approvals, self-correction, security guardrails, trajectory evaluations and graceful fallbacks. Your agents must know when to retry, when to stop, when to ask for help and how to survive failure.

Finally, publish real projects with architecture diagrams, benchmarks, failure analysis and cost breakdowns. Don't just learn frameworks. Build systems that people can trust with real work. The future belongs to engineers who can make agents reliable, not people who can merely talk about them.

[Bobo Lin | AI notes @BoboLinNotes]

AI can build your app in hours. One missing legal page can destroy it overnight.

Before launching, make sure you have a clear privacy policy, terms of service, refund rules, and cookie consent where required. Users must also be able to delete their accounts and personal data.

Know exactly which third-party services collect user data, provide a real contact for legal requests, and never expose API keys in your frontend.

Ask Claude to audit your integrations, identify collected data, draft your legal documents, and flag exposed secrets. But don't blindly trust generated policies. Make sure they reflect what your app actually does.

Shipping fast is a competitive advantage. Ignoring compliance is an expensive mistake. Build fast, but don't launch unprotected.

Not legal advice. Have a lawyer review anything that matters.

[Bobo Lin | AI notes @BoboLinNotes]

The biggest mistake in AI Agent engineering is thinking that building an agent means connecting an LLM to a few tools. Real agents must survive crashes, manage memory, recover from failures and execute long-running tasks without human babysitting.

Start by building the foundations. Create durable workflows with Temporal or LangGraph, implement an MCP server from raw JSON-RPC, design multi-agent communication with typed state, and build a memory system that knows what to store and forget. These are the foundations of autonomous systems.

Then tackle the hard engineering problems. Build computer-use agents, sandboxed code execution, self-healing coding loops, human approval gates and API-to-browser fallbacks. Add security layers to stop prompt injection and circuit breakers to prevent runaway execution. Autonomy without control is just expensive chaos.

Production readiness requires more than a successful demo. Build trajectory-based CI/CD tests, dynamic context budgeting and shadow testing to catch failures before users do. Measure reliability, execution paths and cost per task. If you cannot reproduce, debug and control your agent, you don't own the system.

Finally, publish three serious architecture teardowns showing how your agents work, where they fail and how you fixed them. Don't just collect frameworks and copy tutorials. Build, deploy and prove your engineering decisions. Most people build chatbots. Real Agent Engineers build systems that can survive the real world.

Bookmark & Repost.

[Bobo Lin | AI notes @BoboLinNotes]

@AnaHawrylu97708 The scenario itself functions as a risk signal independent of any 2027 deadline. Hundred-fold acceleration, mass displacement, and concentrated market power occurring together would be a reasonable basis for concern regardless of what specific forecasters predicted.

[Bobo Lin | AI notes @BoboLinNotes]

@hughphan_ The 2026 target names the real near-term friction. Agent-to-hardware integration currently requires substantial custom glue code per platform, and no standard short path exists for deploying a general-purpose agent to new robot hardware.

[Bobo Lin | AI notes @BoboLinNotes]

@natecurtiss_yt The gap makes sense. Good b-roll selection requires matching semantic meaning, visual rhythm, and production tone against whatever footage exists, and those judgments are hard to separate from a human editor's learned sense of the material.

[Bobo Lin | AI notes @BoboLinNotes]

@GuiltyDogVR The more precise parallel might be asset-store homogeneity criticism from the mid-2010s. That debate also collapsed tool accessibility and output quality into a single complaint, and the more productive question was always about craft.

[Bobo Lin | AI notes @BoboLinNotes]

@DaveShapi Meta does run Facebook Dating in several markets, so the product gap is narrower than the post implies. Whether behavioral engagement signals map well to long-term compatibility is a separate empirical question from product strategy.