Jerry L

Velobase Founder

[Jerry L @JingruiLea]

Moonshot AI just released Kimi K2 open-weight: 1 trillion total parameters, 32 billion active (MoE), trained on 15.5 trillion tokens.

The benchmarks? Competitive with Claude Sonnet 4 and GPT-4o.

The cost to run it? A fraction of those closed APIs.

This is happening the same week NVIDIA reports $44 billion in revenue (+69% YoY) because AI companies are spending insane money on compute. And yet the *price of intelligence* keeps falling.

There's a fascinating tension here that will define the next 2 years of AI:

Supply side: infrastructure costs are astronomical and growing. NVIDIA's data center revenue was $39B in ONE quarter. SoftBank just spent $6.5B on a chip company. Hyperscalers are investing hundreds of billions.

Demand side: the cost per unit of intelligence is collapsing. Open-weight models like Kimi K2 mean you can self-host frontier-quality inference. API prices are in a race to zero. A Chinese startup with Muon optimizer produces GPT-4o-level output.

Who captures the value in this dynamic?

Not the model layer (commoditizing). Not the raw app layer (too easy to replicate).

The value accrues to:
- Compute owners (NVIDIA, cloud providers — hence the insane capex)
- Distribution platforms (whoever owns the user relationship — hence OpenAI building social)
- Commercial infrastructure (billing, attribution, anti-abuse — the systems that turn free intelligence into revenue)

The "I replaced my dev team with AI and saved $72K" posts going viral on HN are a symptom. AI is becoming cheap enough that the *creation* of software isn't the bottleneck anymore. The monetization, distribution, and quality assurance of AI-generated outputs IS.

We're shifting from "how do we build it?" to "how do we sell it, bill it, and make sure it actually works?"

That's the real opportunity right now.

[Jerry L @JingruiLea]

Every AI lab now ships a coding CLI.

Claude Code. Codex. Gemini CLI. Grok Build. Cursor.

But they all converge to the same loop: code → terminal → tools → tests → Git.

The real differentiation isn't the model — it's context management, tool orchestration, and persistent memory.

Models are ammunition. The orchestration layer is the gun.

Which one are you actually using? ↓

[Jerry L @JingruiLea]

Every AI lab now has a coding CLI.

Claude Code. Codex. Gemini CLI. Grok Build.

Which one are you actually using in production? And why? ↓

[ZEN @supezen]

以前我们说 idea 不值钱，现在有了 cc 和 codex 发现 idea 可以很值钱。

[Jerry L @JingruiLea]

Sell me your AI agent in one sentence. I will rate it out of 10.

[Jerry L @JingruiLea]

@virtualbacon "Loudest talent signal" is just sports commentary for tech bros.

Here's what everyone's missing: Karpathy joined the PRE-TRAINING team. Not agents. Not alignment. Not the post-training circus.

When the whole industry declared scaling dead, Anthropic is quietly building a bigger model.

Remember GPT-4.5? Expensive as hell, but undeniably smarter. No RL tricks, no post-training magic. Just scale.

The talent move is noise. The technical bet is the story.

[Jerry L @JingruiLea]

Everyone's debating the Karpathy-to-Anthropic talent war. The real question is simpler:

Why is Karpathy betting on pre-training in 2026?

Remember GPT-4.5 — expensive, but smarter. No tricks. Just scale. We need bigger models and Karpathy knows it.

Either scaling isn't over, or Anthropic is about to prove it is.

[Jerry L @JingruiLea]

our income.

[Velobase @VelobaseX]

Everyone can build an app now.

Almost no one makes a dollar from it.

We went from the same problem to 8-figure ARR. The secret wasn't the product — it was the infrastructure behind it.

Today we're open-sourcing the whole stack! 🧵

[Velobase @VelobaseX]

Everyone can build an app now.

Almost no one makes a dollar from it.

We went from the same problem to 8-figure ARR. The secret wasn't the product — it was the infrastructure behind it.

Today we're open-sourcing the whole stack! 🧵

[Jerry L @JingruiLea]

AI News Roundup — May 19, 2026

• Karpathy joins Anthropic’s pre-training team. 116K likes in hours. The most respected AI educator just bet his next 5 years on Anthropic over every other lab.

• Google I/O: Gemini 3.5 Flash launched — built for agents, not chatbots. Android CLI lets AI write and ship apps. Demis Hassabis calls it “the foothills of the singularity.”

• Anthropic acquires Stainless (dev tools, ex-Stripe founder, ~$300M) — will cut off OpenAI and Google as customers.

• Musk loses OpenAI lawsuit. Jury: you sued too late. Trial revealed Musk had his own plans to make OpenAI for-profit.

• Most upvoted HN post this week: “Entire companies are under AI psychosis” — 2,097 points. Right below: “AI is making me dumb” (553 pts), “AI subscriptions are a ticking time bomb” (414 pts).

• Forge (Show HN): guardrails take an 8B model from 53% to 99% on agentic tasks. Small models aren’t dead.

• Mistral acquires Emmi AI — building the “European sovereign AI stack.”

The market is splitting: labs racing to AGI on one side, practitioners screaming “slow down” on the other. Both are right.

[Jerry L @JingruiLea]

We went from zero to 8-figure ARR with an AI product.

The hardest part wasn't building — it was billing, attribution, affiliates, and all the unglamorous infra that actually makes money.

We just open-sourced the entire stack. MIT licensed.

https://t.co/gVOoI6Jxfn

[Jerry L @JingruiLea]

@akazwz_ you can try https://t.co/LUCxSQ96BF

[Abe™️ @_Abrinsky]

@haydendevs https://t.co/S7Wm9eus16

Literally this