đ AI Agentic Workflow Andrew Ng
Youtube Video, via JS Party podcast episode Building LLM agents in JS
Notes
Non Agentic workflow: Do it start to finish. Mostly zero shot prompts.
Agentic workflow: Revise, iterative, reflect, use tools if you need to….
Four design patterns.
1. Reflect: Produce one thing and ask another chat thread with different system prompt to evaluate it. E.g. Create a code, than ask a rubberduck debugger to read it line by line, or run the test suite and provide the result to LLM generated the code to evaluate it. Or write a post, and ask an editor LLM to reflect it, and ask writer LLM to update the post on feedback.
2. Tool Use: Use tools and function calls, there are lot of that, even we have developed one for SourceSailor.
3. Planning: Like give a task to LLM and then ask it to plan the solution step by step, and then ask LLM to execute the plan. Take the task description, create a plan, break a plan to subtasks, and then use aider to execute the plan.
4. Multiagent Collaboration: Create multiple LLMs for one single task, and create an orchestrator LLM to collaborate between them. Like one LLM (Powred by sonnet) for figuring out style, another LLM (Opus) for generating the writing, another LLM (Opus or Gemini or GPT) to do the reflection and a haiku or 4o powered LLM which orchestrates between them.
Updates in blog
đ How does AI impact my jobs
Itâs not their fault. They enrolled in a masterâs program to get a job in tech. Why? Letâs be candid: tech has promised job security and agreeable (sometimes borderline perverse) financial returns for a couple of decades. Many tech employers also spin a yarn about âsaving the worldâ and coast on the reputational allure of âif you work here, youâre a genius.â Sounds great, doesnât it? Except, now that students have invested five figures of money in their tech education1, had their skulls crammed full of âinvisible hand of the marketâ propaganda, and counted on having secured their ticket to the party, theyâre seeing layoffs. Theyâre seeing exposĂ©s. Theyâre seeing compensation adjustments at the most lucrative tech companies. And theyâre seeing a news cycle that oscillates wildly between blaming AI for these changes and extolling it as the solution.
you learn to write Dijkstraâs from a blank editor so you can get the job at Twitter, but once you accept the offer you never actually do that. What you need to do is understand, update, and (optimistically) un***k existing systems written by other people.
Are large language models gonna cause programmers to lose their jobs? Not anymore than StackOverflow did, in my view. However, it’s going to change themâŠsomewhat.