Rauchg: 'It's time to ship.' 295 likes.

Three words. No elaboration needed. When the Vercel CEO says it's time to ship, the next 48 hours usually bring product drops.

Rauchg ships: 30 minute function invocations and new 24 hour sandbox lifetimes. 81 likes.

The significance: agent workloads need long-running processes. 30 minutes for functions and 24 hours for sandboxes means Vercel is systematically removing every infrastructure ceiling that blocks persistent agent workflows.

Rauchg: 'This aged well 😂 iykyk.' 190 likes.

The subtext: something Vercel said or built previously is being validated by current events. The Cursor acquisition and compute demands prove the thesis that compute infrastructure convergence was the right bet.

Levie frames what might be the single most important question for AI market structure: how far behind are open weights models from closed models?

If 3-6 months behind: the market looks one way—open weights are viable alternatives, margins compress, sovereign AI is practical.

If years behind: the market looks completely different—closed model providers have durable moats, inference concentrates in hyperscaler clouds, sovereign AI struggles.

This single variable determines: chip stack evolution, where inference runs, what sovereign AI looks like, applied AI layer dynamics, margin structures, and how much companies can afford to spend on AI.

Current read: open weights players are holding up at keeping close to frontier capability.

Peter Yang with honest skepticism: 'I don't want to be a hater but I feel like this is what happens when nobody internally wants to tell the CEO no. Maybe this thing will blow our socks off?' 280 likes.

The subtext: large companies keep making AI product decisions that look disconnected from what users actually want. The pattern is recognizable—when internal dissent gets suppressed, products ship that make sense on org charts but not to users.

Peter Yang publishes a tutorial: make Codex or Claude Code your personal advisor using a skill with 4 files. Bonus: he managed to save Fable's advice before it got restricted.

Practical takeaway: the skill system is simple enough that 4 files can turn an AI coding agent into a personal advisor. The Fable save is a nice touch—preserving access to capabilities before they disappear.

Garry Tan at 4am diving into Bowen's differentiation of self theory. The core idea: low differentiation means you're highly reactive to others' emotional cues, lose yourself in relationships, can't hold boundaries under pressure.

Two flavors of the same disease: chameleons (adjust themselves to please, fusion mode) and bullies (force others to conform). Both are undifferentiated—the self-abandoner and the asshole are the same person.

The fix: lower chronic anxiety, raise the capacity to stay a separate self in contact with others.

Why this matters for AI builders: building products that matter requires exactly this kind of differentiation—holding your vision when everyone around you has opinions.

Slack now renders HTML attachments instead of just showing them as raw text. 2513 likes.

The community's reaction (😭🙏) tells you everything: this was a long-requested feature that seemed like it would never come. Sometimes the most impactful updates are the ones that fix basic friction points.

Dan Shipper switches off Atlas Browser and back to Dia: 'I've been having so many weird bugs with Atlas, and doesn't feel like it's improving.' 72 likes.

The founder of Every returning to Dia from Atlas is a small but real signal. AI-native browsers are still in the 'does it actually work reliably?' phase, and the tolerance for bugs is lower when the core job is writing and thinking.

Nikunj reports from Cursor Compile: 'Vibes were immaculate. Getting a custom mechanical keyboard is a new kind of swag ceiling for conferences.' 139 likes.

The subtext: Cursor isn't just a product anymore—it's becoming a cultural brand. Custom mechanical keyboards as conference swag signals that developer tooling companies are becoming lifestyle brands.

Joon Sung Park, creator of the legendary Stanford Generative Agents paper (the Smallville experiment) and founder of Simile, explains how AI can now simulate human behavior at 85% accuracy—as in, the simulation predicts what someone would do almost as well as they could predict their own behavior.

Smallville origin story: 25 agents in a tiny town, each with a persona, daily routine, and relationships. No scripting. One agent (Isabella, the cafe owner) spontaneously decided to throw a Valentine's Day party, went around inviting people, gathered materials, and on the day of—people actually showed up. Klaus, who got invited, decided to ask his crush. This was emergent behavior from the agent architecture: memory + planning + reflection.

Before Smallville, there was Social Simulacra (2022): simulating an entire subreddit with thousands of personas to see what behaviors emerge. Built on GPT-3, which was janky and pre-instruction-tuning, but the promise was visible.

The product today: Simile partners with Fortune 500 companies like CVS. You define the population, Simile collects real data through RL-trained interviews ('Tell me the story of your life' in 15 minutes) and Gallup partnership surveys, then creates agent simulations that answer any question about that population.

The say-do gap: LLMs are trained on what people say online (attitudinal data), but there's a real gap between what people say and what they do. Simile's behavioral models—trained on repositories of RCTs (randomized controlled trials)—close this gap.

CPU vs GPU of intelligence: today's frontier models (GPT, Claude) are like the CPU—rational, objective, superhuman at math. Simile is building the GPU—models that represent the diversity of human values, preferences, tastes, and irrationality. You need both.

Convergence vs divergence: some simulations converge (network structure always forms hubs regardless of small errors, like PageRank). Others diverge (elections, wars). For convergent questions, compounding errors still converge. For divergent questions, run 100 times and show the distribution of possible outcomes.

The grand vision: simulation is the Hubble telescope for social science. Thomas Schelling won a Nobel Prize with rudimentary red-dot-blue-dot agent models showing how segregation emerges. Now we can simulate with agents that have the full richness of real humans.

The ultimate future: simulations that cost $100M to run once and take months—but solve fundamental questions of society. When does bank fraud happen? Can nations cooperate on climate? What are the signals of democratic collapse?