Dan Andrews

april fools gave us buddies in claude code. i liked them so much i built terminalbuddies.com - hatch your own ASCII companion, collect all species, and bring them into any MCP-compatible coding tool. hopefully anthropic lets this one slide 🥚 #ClaudeCode #MCP #TerminalBuddies

I packaged up the "autoresearch" project into a new self-contained minimal repo if people would like to play over the weekend. It's basically nanochat LLM training core stripped down to a single-GPU, one file version of ~630 lines of code, then: - the human iterates on the prompt (.md) - the AI agent iterates on the training code (.py) The goal is to engineer your agents to make the fastest research progress indefinitely and without any of your own involvement. In the image, every dot is a complete LLM training run that lasts exactly 5 minutes. The agent works in an autonomous loop on a git feature branch and accumulates git commits to the training script as it finds better settings (of lower validation loss by the end) of the neural network architecture, the optimizer, all the hyperparameters, etc. You can imagine comparing the research progress of different prompts, different agents, etc. github.com/karpathy/autor… Part code, part sci-fi, and a pinch of psychosis :)

🔥 YC breaks down how top AI startups prompt LLMs: 6+ page prompts, XML tags, meta-prompts, evals as core IP. Meta-prompting + role assignment are key for reliable, agent-like behaviors. ⚙️ Key Learning → Top AI startups use "manager-style" hyper-specific prompts—6+ pages detailing task, role, and constraints. These aren't quick hacks; they’re structured like onboarding docs for new hires. → Role prompting anchors the LLM’s tone and behavior. Clear persona = better alignment with task. Example: telling the LLM it's a customer support manager calibrates its output expectations. → Defining a task and laying out a plan helps break complex workflows into predictable steps. LLMs handle reasoning better when guided through each sub-task explicitly. → Structuring output using markdown or XML-style tags improves consistency. Parahelp, for instance, uses tags like to enforce response format. → Meta-prompting means using LLMs to refine your own prompts. Feed it your prompt, outputs, and ask it to debug or improve—LLMs self-optimize well if given context. → Few-shot prompting with real examples boosts accuracy. Startups like Jazzberry feed challenging bug examples to shape LLM behavior. → Prompt folding lets one prompt trigger generation of deeper, more specific prompts. Helps manage workflows in multi-step AI agents. → Escape hatches instruct LLMs to admit uncertainty. Prevents hallucination and improves trust. → Thinking traces (model reasoning logs) and debug info expose the model’s internal logic. Essential for troubleshooting and iteration. → Evals (prompt test cases) are more valuable than prompts themselves. They help benchmark prompt reliability across edge cases. → Use big models for prompt crafting, then distill for production on smaller, cheaper models. Matches quality with efficiency.

We finally hit 500k ARR. Then decided to bet our entire company on a new product. Launching Helicone AI Gateway 🚀 (Public Beta!) 1 API. 100+ models. 100x faster. We believe we’ve built the new standard for LLM routing. In Rust. Fully open-sourced. 1 min to set up.

we spent the last four years reimagining what finance can be, unburdened by what has been. runway makes clear the significance of the passage of time in the context of all in which your business lives and what came before you. this is the result 👇

This whole Techstars thing is just an opportunity to build something much better. Let’s stop talking about their strategy and their comms and start building the future. Who’s in for the next thing? d16b4ik6hdi.typeform.com/to/CmbMcweQ

1/5 We’re hiring for a partner to join our investment team, and anyone who loves working with early stage founders can apply.

Sharing this because many of you who have been around @Techstars may know our long time colleague and friend @anuragmaloo - we are all very worried about him.

Anurag is a friend and former @Techstars colleague climbing to raise awareness of the UNs Sustainable Development Goals. He's gone missing on Annapurna. Sending warmth and good thoughts your way, @anuragmaloo. Hope to hear from you soon.

Write better content faster with Contentedge: ✍️Create topic outlines & paragraph text with AI designed to rank in Google 📸Generate & add unique AI images to your content 📊Perform SEO keyword & domain research against competitors Check them out on PH 👇 bit.ly/3hvvcZG

It’s @ckevinliu!!

It is impossible to conceive that David McCullough is no longer with us. He is among our greatest historians, writing with an almost magical command of language and story. 1/2

When I joined Initialized ten months ago, I was blown away by our healthtech portfolio. To date, @Initialized has made 20 investments in seminal healthcare companies: 📣 11 are publicly announced 🥷 9 are in stealth 🦄 6 are unicorns 🚀 5 will launch later this year

The @bouldercolorado Wireless Emergency Alert to EVACUATE NOW was sent to cell phones in this geographic area #NCARfire

Anyone evacuating that has pets stuck please let us know and we can get them. #NCARFire #boulderfire

This is being called the #NCARFire The @bouldercolorado EOC has been activated. Wireless Emergency Alerts have been sent to ALL cell phones within a 1/4 mile radius from NCAR. Message is to EVACUATE area due to fast moving wildfire #boulder

Boulder Fire-Rescue has ordered 1,200 residences to evacuate after a fire began in the area near NCAR. The fire began shortly before 3 p.m. local time. Boulder Police has warned everyone near NCAR to evacuate immediately. #NCARfire #Boulderfire trib.al/uxgQEDn

If you are hiking or recreating near NCAR please EVACUATE NOW! @boulderosmp are in the area evacuating hikers and all folks they find. Please RT #NCARfire #boulder

About 1200 contacts in this shaded area are being ordered to EVACUATE NOW #boulder #bouldercolorado