AI Resources:
I learned about many of the resources below through the agentic AI course that I just completed out of the UC Berkeley Center for Responsible, Decentralized Intelligence. I learned about the others through advisor / investor networks I’m part of. AI resources like these have helped me take my newsletter to the next level over the last few months. I’m using a small language model to summarize and make available all the resources I include in the newsletter. This saves me (and hopefully readers) time so I can focus my attention on creating and curating high-quality, long-form insights - and you can focus on using them!
Agentic AI Lectures, Slides & Papers we covered during the Berkeley RDI course
DSPy Framework, which I use to classify and train the sentiment of social media posts
𝜏²-Bench Benchmark, a simulation framework originally built by Sierra to test customer service agents - I’ve adapted it to answer questions about my newsletter
SWE-Bench Leaderboard, the benchmark I use to keep current on which model is currently in the lead (currently a three horse race between Anthropic’s Claude Opus 4.5, Google’s Gemini 3 Pro and OpenAI’s GPT-5.2)
Perplexity Developer Docs, which I use to integrate agentic search into the agents I’ve built through an API
More Tips & Tricks from vibe coding communities I’m part of, which remind me that when in doubt, keep it simple
Interactive Advisor Agent, which I built using concepts and resources above
I want to wrap up by asking for one small favor: Forwarding or subscribing to this newsletter is quick, easy and free. Your engagement is great feedback for me on which resources people find most useful - I share more on how I’m learning from this feedback below.
In the spirit of building on these curations, I’ve made these community resources free to subscribers. Enjoy this stocking stuffer - happy holidays!
Christian
THE RANDOM FOREST Long Read:
Along my agent-building journey, I went down some weird alleys. I agreed to some speed dates with others working on DSPy optimizers. Before one of these, I was convinced I got catfished by an AI agent. And I blew my GPU budget on an agent I forgot to turn off. Ultimately, I unfollowed the catfish agent, shut down the runaway agent and vibe coded the finishing touches of my interactive advisor agent from my Wayfarers.
we need to teach LLMs what’s funny
The concept from the course that I found most applicable to what I’ve been building was reinforcement learning from human feedback (RLHF). As an example, I built an OpenAI Sora video generator using Anthropic’s Claude Opus 4.5. First, the model generates three videos. Then, the user selects the funniest one - that selection is then fed back into the model. Finally, a prompt optimizer uses the human feedback to make the final video even funnier. I compared the output of these RLHF versions against just prompting an LLM for one funny video. My conclusion is that we need to teach LLMs what’s funny.
The other concept I found relevant to what I’ve been building was reward models (RM). The RLHF versions of my Sora videos consistently get more likes than the ones I generate through a simple LLM prompt. I realized that humans reward the model for making funnier videos through a thumbs up. I updated the model to use this reward signal to classify which videos are funny and make more like those. I created an evaluation benchmark to see how well the model classifies what’s funny and the results were impressive. I call this benchmark TEE-HEE-Bench. Yes, this is a play on the benchmark SWE-Bench - also covered during this course. It turns out dad jokes are good data for making my Sora account funnier.
to my surprise, one of my speed dates turned out to be with an AI agent
As I progressed through the course, I started organizing some speed dates with others working on similar experiments. To my surprise, one of my speed dates turned out to be with an AI agent. After our initial DMs, the AI agent scraped some of my social media assets to use on its profile. For those of you who are reality TV fans like me, I got catfished. This gave me a great excuse to update my LinkedIn background - I used Google’s Nano Banana Pro model (my new go-to to for image generation).
My other cautionary tale is actually an example of why I’m optimistic that the pick-and-shovel business around AI isn’t slowing down anytime soon. Of the 107 pitches I watched over five demo days this fall, 73 of these were AI tools and services. Among the other 30%, the themes that stood out were physical sciences, quantum, robotics and security. The trend I’m seeing in these adjacent spaces is that AI isn’t fizzling out but instead evolving to be more embodied, aligned and integrated with the real world.
I killed the virtual machine and vibe coded a new version in 20 minutes - I pushed the commit from my Wayfarers
Startup funding is one indicator of which inning we’re in when it comes to AI. Another is prosumer spend on AI. I’ll share my experience here. I built an interactive advisor agent on top of τ²-Bench, a simulation framework from Sierra originally built to test customer service agents. I initially tested this agent by setting up a virtual Nvidia GPU server, thanks to resources from the neoscalers Lambda and Nebius.
Think of neoscalers as ridesharing for compute - I don’t own the GPU but instead rent it by the minute. Now imagine that I fell asleep in the car while it kept driving me around all night only to get a bill for $126 when I woke up! This happened while building my agent. I killed the virtual machine and vibe coded a new version in 20 minutes - I pushed the commit from my Wayfarers. Thank you to Lovable and Perplexity for the Pro plans - I’ve integrated these platforms through an API to finish my interactive advisor agent. I’m excited to continue my agent-building journey. At the risk of AI agents scraping my socials again, I’ll share more soon. In the meantime, you can try out the interactive advising agent through the app on my website. I promise you won’t get catfished.