(📷 Illustration: Back at the Bluetooth SIG working group annual meeting at Hotel InterContinental Madrid, the kickoff day featured a bomb-defusal mini-game so cross-team folks could warm up. Unsurprisingly the youngest person ended up as the laptop operator (defuser), having to follow the discussion of senior Bluetooth veterans from major chip vendors while flipping through the manual. I think we finished without coming last. I may not have business acumen, but at least I have finger acumen and bug-hunting acumen?! That interaction also broke the ice with the seniors and gave the rest of the spec discussions a much better tempo. It is one of those rhythms of always throwing yourself into pressure and trying to keep smiling. Image source: Ernest.)

✳️ Across Ancient Times and the AI Era: Keith Rabois on How Three Uncomfortable Things Buy You a Truly Tough Team

In this episode of Lenny’s Podcast, Lenny asks Keith Rabois (Managing Director at Khosla Ventures and a member of the PayPal Mafia) one question: what do all those companies he invested in early on, the ones that grew into Stripe, Airbnb, YouTube, DoorDash, Ramp, and Palantir, have in common?

Keith’s answer: operating tempo.

That kind of tempo isn’t just about going fast. It’s about being able to diagnose a problem, ship a solution, and measure the impact all between one board meeting and the next.

Read More

(Illustration: A mentor once told me, “the data is always in there — but ‘in there’ doesn’t mean findable, and findable doesn’t mean correct.” That’s why “dreaming” fits the sandbox so well: there’s a boundary, depth can vary, and you share some memory with the waking world. Sometimes you wake up dazed, sometimes you wake up knowingly smiling from the inside. Image source: Ernest.)

✳️ The Dream Sandbox and the Merge Button

Since last summer, everyone’s been talking about AI Agents (not really! Most people are actually arguing about discussing the AK LLM Wiki! That’s another topic — let me brew it into one of my rare essays. Thanks for all the love; people have been quietly reaching out this week to commiserate together.)

Most teams hit the same wall. The prototype demo looks great, the happy-path scenarios all run, but no one dares push the agent into production. AI Agents can schedule shifts, adjust calendars, answer questions, even place TTS phone calls — but what happens when one edits data it shouldn’t touch, or peeks at fields it shouldn’t see? At DevCon 5, Palantir showed a demo of a medical scheduling system: a nurse uses voice to ask the agent to schedule a surgery, an administrator reviews, and the system auto-dials patients with the update. The whole thing was built by Palantir’s team in less than one weekend.

Read More

(Illustration: Taken at The Venetian, where I get up early every year-end to enjoy a sunrise breakfast. Chef Thomas Keller built two restaurant brands: The French Laundry, the embodiment of perfection, and Bouchon, the same obsession over detail but made approachable for everyday life. ElevenLabs has a similar structure. Research with real power shouldn’t be locked away in papers or experimental data; it belongs in products, where users walk past it every day. The deepest craft eventually walks toward the everyday. Image source: Ernest.)

✳️ Tech Can Be Chased Alone, But the Org Structure With Its Culture Is Hard to Copy

Research-driven companies often hit the same problem: the technology works, but the product won’t move. Or the reverse, the market need is crystal clear, but the research can’t keep up with the rhythm. ElevenLabs started in 2021 as a weekend experiment by two Poles, Mati Staniszewski and Piotr Dabkowski, focused on AI voice synthesis, with applications spanning audiobooks, dubbing, voice agents, and games, where the synthesized voices can almost fool a human. In just a few years, they’ve reached an 11 billion USD valuation with a team of over 300 people. They aren’t just better at the technology. The way their organization runs, fused with their culture, becomes the moat. Tech alone can be chased, but the org structure together with the culture cannot easily be copied. From a16z’s interview with CEO Mati Staniszewski, three lines of thought are worth observing in detail. (Of course, every family has its own difficult scripture; lines of thought have to match constraints and context. Don’t copy them straight into your own situation after reading.)

Read More

(Illustration: In the process of exploring the world and becoming yourself (being), ensure freedom, ensure the underlying safety net mechanisms, ensure common interface interoperability, while tuning parameters (metacognition) and running into the wind with open arms. That said, running still requires eating, and eating can never be let go. Taken in 2016 in Madrid, allegedly at the world’s oldest restaurant. Made sure to bring the magic card, made sure the communication interface was interoperable, left the rest to the kitchen, and comfortably walked out five minutes before the restaurant closed at eleven. Photo by Ernest.)

MMDP = Multimodal Data Plane

At Palantir DevCon 5, Data Plane Group co-lead Ted introduced MMDP. I was curious enough to practice breaking it down.

• First, the feature list is long, and I wanted to compare it against my own product roadmap.
• Second, their design trade-offs reflect a core question: when enterprise data is scattered everywhere, should you move the data or move the compute? (No right or wrong answer here, but I wanted to compare the reasoning and parameters.)

Read More

(Illustration: Some people buy multiple pieces of the same clothing style. I tend to buy multiples of the same toys (or brands), paired with a deep-seated quirk of not wanting to be like everyone else, like this niche notebook Field Notes that I’ve been using for over a decade. I was originally drawn to its woodgrain cover, which has since been discontinued, but thankfully my hoarding habit means I have some stocked up. The advantages of using the same style through deliberate design decisions are many, perhaps including the transformation into excess returns for our clients a decade later (wishful thinking). Photo taken in 2016 at a friend’s home in Boston. Image source: Ernest.)

✳️ The Design Decision of AUM Fees

Charging a percentage of AUM (Assets Under Management). Almost every wealth management firm uses this model, and most clients take it for granted. But a16z Perennial CIO Michel Del Buono said on the Sorcery Podcast that this isn’t just a pricing decision, it’s a design decision that influences everything else. Let’s practice decomposing it:

Read More

(Not midnight yet. 9 PM at the outermost moat of the Imperial Palace, with nobody else around. Occasionally understanding things others don’t, while experimenting with all sorts of magic for preserving memories. Photo by Ernest.)

Every time I see news about “Company X laid off XX%,” what I really want to know isn’t why they cut, but rather: after the cuts, how do the remaining people work? What does the organization look like? What processes did they adjust?

Block ($XYZ) laid off more than 40% of its workforce in 2025 Q1, openly stating that AI was a key factor. Their business lead Owen Jennings shared the specifics on the a16z channel. Owen has been at Block for 12 years, was previously the CEO of Cash App, and now oversees product operations and customer support across Square, Cash App, and Afterpay. What he described wasn’t the generic “we embrace AI” narrative, but the actual logic behind their rebuilding.

There are three principles worth unpacking here:

Read More

(Image: Olaf makes an appearance. Source: NVIDIA GTC 2026 Keynote.)

✳️ Viewing GTC 2026 Through a Token Economy Lens

Every year after GTC, the community races to compare specs. Which chip has the highest compute, how much NVLink bandwidth, how many times faster Vera Rubin is than Blackwell. But the slide Jensen held up this year and called “my best slide” wasn’t a spec sheet for any single chip — it was an architecture diagram spanning the entire structured data ecosystem. It listed Snowflake, Databricks, Amazon EMR, Google BigQuery and other CSP engines alongside various data storage solutions, with NVIDIA’s cuDF acceleration engine at the bottom. He said his team always tells him “don’t show this one” — too complex — but he insists on presenting it every time. The core message: “Structured data is the ground truth of enterprise computing” — the foundation of all AI.

The point of that diagram isn’t about any individual platform. It’s about the reason for acceleration. In the past, accelerating structured data was about doing more, cheaper, more frequently — “good enough was good enough.” But going forward, AI understands and consumes data far faster than humans can. Without accelerating data preparation, you simply can’t keep up. Nestle used Watson X to accelerate their supply chain — 5x faster, 83% cost reduction — speed, scale, and cost benefits all at once. NVIDIA built two foundational platforms for this: cuDF is “RTX for data frames,” handling structured data; cuVS handles unstructured data. The latter is arguably more critical: 90% of the world’s data is unstructured, previously “completely useless to the world,” until AI’s multimodal understanding made it searchable. (Earlier this quarter, we helped a manufacturing client re-explore their ERP data. We had no idea what we’d find until we dove in — the general-purpose model’s ability to understand that data was jaw-dropping. Happy to chat more about this separately if you’re interested.)

OK, so let’s say your organization has the data layer acceleration in place. The next question is: how do you price what AI produces?

Read More

(Illustration: Layer upon layer, every piece matters. Like this ham air-drying in the breeze, sacrificing itself for the greater good. Trust can’t be built all at once; it has to be stacked, one layer at a time. The first layer? Start by hitting that like button. Image source: Ernest.)

✳️ The Boring Plumbing Work

What actually blocks enterprise AI deployment is almost never that the models aren't good enough. Mistral AI CTO Timothée Lacroix put it bluntly: today's model capabilities are already sufficient to unlock massive enterprise value, but you first need to get all the connectors, data formats, and permission management right — all this "boring plumbing work" — before enterprise token consumption truly takes off. He used the word “plumbing” (and I’d stress the “system” part, not just the pipes). He said it three times throughout the interview. We’re still in the construction phase, he noted — most enterprises haven’t even gotten basic data connectivity right, let alone running AI agents executing tasks at scale in the background. (What we see on the ground is even bleaker: wishlists that don’t map to actual data. No wonder everyone just jumps straight to word-chaining chatbots — sarcasm mode engaged.)

Read More

(Illustration: Mt Rainier in Seattle, Reflection Lake inside the national park. Facing the mountain, facing time, what can we humans do in a post-AGI world? At the very least, being a decent human being should be the baseline, you demons. I am not sure I could live in the countryside, but I do enjoy having nothing to distract me. Then I found out that a Chinese idiom meaning “undistracted focus” is also a Year of the Horse blessing?! Image source: Ernest.)

✳️ This man has not touched a single line of code since November

I recently listened to Lenny’s Podcast interview with Boris Cherny, the person in charge of Anthropic’s Claude Code, formerly one of the most productive engineers at Instagram. Since last November, 100% of his code has been written by Claude Code. He has not manually edited a single line, shipping 10 to 30 PRs per day, and during the recording he had 5 agents running simultaneously. Anthropic's engineering team grew 4x while each engineer's output increased by 200%. In his previous role at Meta, he was responsible for code quality across the entire company. Back then, hundreds of engineers spending a full year would typically improve productivity by only a few percentage points. Now it is hundreds of percentage points. A completely different order of magnitude.

Read More

(Illustration: Taking the day off. Happy Lunar New Year! Image source: Business Next.)

✳️ Coding is easy, Context is hard

I recently recorded an episode of Podcast “Digital Keywords” EP228 with James from Business Next, discussing a core theme I kept chewing on throughout my 2025 year-in-review.

AI has crossed the threshold in writing code. It is no longer a small assistant that auto-completes one line at a time. Take Claude Code for example: it reads through your entire project directory on its own, understands cross-file context and dependencies, plans how to coordinate changes, even dispatches sub-agents to handle different tasks in parallel, and then delivers an entire feature in one go. More importantly, it has memory. You write your project specs, goals, and coding style into a file, and every time it starts working, it reads that file first. No need to re-explain everything from scratch.

From Cline and RooCode, to Cursor, to Claude Code and Kiro, my team and I have walked the entire tool evolution path together, feeling the role shift at every step. Humans can no longer compete with AI on speed. The bottleneck has moved from “think fast, code fast, iterate fast” to “can you articulate your requirements, context, and intent clearly?” I joked with my team: we all need to start practicing how to communicate well. This is not just an observation. It is the firmest conclusion I reached after being pushed through all of 2025, and the central theme of this podcast episode.

Read More