Wei Lyu

@southpkcommons SPC is the "YC for Adults"

The time to be maximally ambitious is now. The $1M Founder Fellowship is open for Fall 2026. Applications are due August 2nd.

Dalai Lama - we must deal with emotions. We all have this marvelous human brain. The problem is when negative emotions develop, our whole mind is taken over. I can change my mind.

SPC turns 10 this year. A decade ago, Ruchi and I just wanted a space for people who weren't ready to start a company but knew they wanted to build something. We wanted to create a startup that would help other startups and founders. The best founders aren't just creative. They're ferociously competitive. They want to win. Not because they care about status — because they're greedy. Greedy to bring about the change they know must happen. And so I am feeling greedy. We have had a great 10 years. But I want more. I want more startups to exist and solve big problems. I want to work with more insane and wacky builders. The problems get bigger. The stakes get higher. Your ambition expands to fill the space. If you're in the -1 to 0 phase and feeling both terrified and alive: that's the signal. Don't ignore it. Be more ambitious. We at SPC certainly are going to be maximally aggressive and ambitious on our end. It's Day Minus One. Let's build.

The goal @southpkcommons is to build maximally ambitious companies. Take some time to find the right mountain. Don't confuse motion for progress. Our first 10 years have been strong. Now we get to to go build the next decade. This article captures us well:

I gave myself a challenge this weekend - beat Anthropic's kernel optimization challenge. The problem is simple - optimize the program to use the least amount of CPU cycles as possible. However, I know absolutely nothing about kernel-level programming - for me, this was purely a challenge of how well I can craft context to get AI tools to achieve my goals. After roughly 2 days, I was able to achieve 1,182 cycles using multiple different models (Codex & Opus 4.5) and various strategies. This beats both their hiring threshold of 1487 cycles as well as their internal Claude Code record of 1363 cycles by a decent margin. Keep in mind, I genuinely don't know what the program does, and have not edited the actual program myself throughout this entire process. I mostly acted as a "course corrector" - identifying plateaus and adjusting the surrounding tooling to allow the AI to find new optimization opportunities itself. The main learning that I gained from this experiment is that when building long-running agents, your prompt becomes extremely sensitive. There is a huge compounding effect when each new context window takes the summary & learnings from the previous one as input. For example, last night, one of the Opus 4.5 agents concluded that it was "mathematically impossible" to improve, and essentially just gave up. It then documented this conclusion into a progress log - which is given as input to every new agent's context window. Then, each subsequent agent's first impression became "it's impossible" - and took this conclusion very seriously - and no longer even tried to improve the program until I woke up (very disappointed) to intervene. This sort of plateau happened many times - in one case, I decided to completely reset everything, after the system was stuck making tiny single-digit micro-optimizations while at ~4000 cycles - extremely far from the goal. Everything I put into the prompt felt like it had 100x leverage applied to it using a long-running agent strategy - each word felt like it was taken literally. For this reason, vibe coding the prompt was not good enough - this required human problem solving skills - "how would I, as a human, approach this?" I broke down the problem into a carefully worded, hand-written and importantly, repeatable set of instructions. This essentially boiled down to: 1/ Understand the problem 2/ Understand previous attempts 3/ Use tools to identify inefficiencies 4/ Use the web to research 5/ Form a hypothesis 6/ Implement, test, and document 7/ Commit improvements, discard regressions 8/ Repeat Giving AI the guidance & importantly the tools (jq, python, perl via command line) to identify real optimization opportunities from the program trace was extremely helpful to making tail-end progress with this framework. My stretch goal is to reach rank 1 on the leaderboard, which requires me to cut a further 82 cycles off the program. I'l continue to course correct it throughout the week as it runs in the background while I work on other things. If people are interested, I'll do a more in-depth write up of the process I used to achieve this!

Great having @hanstung from @notablecap by for a roundtable with @southpkcommons founders back before the holidays! Hans dropped lots of great stories and wisdom from some of the legendary businesses he’s helped build - Xiaomi, Airbnb, ByteDance, Red Note. Thanks for swinging by Hans.

I’m so sorry to hear that some of my good friends and colleagues were impacted by Meta Reality Labs’ layoffs today. My heart goes out to you, and I truly hope there’s an even more exciting chapter ahead of you. At Nirva, we’re a team coming from Meta Reality Labs, and we’d love to connect with fellow Meta alumni who believe in our vision. #Meta #layoff #realitylabs #metamates

Nirva is the BEST of CES Startup. We only found out later that #CNET’s Best of CES is the biggest award at #CES. Judges walked the floor, visited booths in person, and picked what they love. Out of 1,000+ startups, we made Best Startup finalist → top 3 of 1,000. These are the top 3 startups. We didn’t win #1, but being a finalist next to #Pebble still feels pretty damn good lol. lnkd.in/gRnTt8Ah

Day 2 at CES — the Nirva booth was buzzing with energy.

So excited to share that Nirva is a finalist for the CNET Best of CES Awards and we’re up there with Boston Dynamics, Samsung, and Lenovo！

Fashion-forward AI Wearable An AI Wearable you actually want to wear.

Come to our Booth! Nirva is unveiled at CES 2026. Nirva is the most beautiful and insightful AI wearable ever created. @CES #CES2026 #CES @Nirvalabs @southpkcommons

Come to our Booth at CES ! Nirva is unveiled at CES 2026. Nirva is the most beautiful and insightful AI wearable ever created. Designed to be worn all day, Nirva truly understands you — journaling your life, tracking your mood, and offering deeply personalized guidance. nirva.life #CES2026 #CES

Unveiled at CES 2026. Nirva is the most beautiful and insightful AI wearable ever created. Designed to be worn all day, Nirva truly understands you — journaling your life, tracking your mood, and offering deeply personalized guidance. nirva.life - @Nirvalabs with @southpkcommons

This Necklace Is Like a Mood Ring for the AI Age #ftag=COS-05-10aaa0j" target="_blank" rel="nofollow noopener">cnet.com/tech/services-… via @CNET

Hi Nirva

Lol

It was a quiet moment with myself. After parking the car at the garage. Ready to go down for the event. Jason and Hanying are setting up our things. Sky is clear. Feels good.

We don't sell AI Wearable. We sell energy.