1. Pretext

A new text engine from Cheng Lou that makes web behaviors possible that used to be brittle, janky, or basically impossible: text wrapping around moving objects, reshaping itself live, reacting like a physical material, and turning typography into an interactive system instead of a dead block. One of those rare frontend releases that quietly expands the design space.

2. Parallax

A distributed inference stack for people who don’t want their AI future fully outsourced to hyperscalers. Stitch together spare machines, odd hardware, and local devices into something that feels less like running a model and more like building your own intelligence infrastructure.

3. Agent Lightning

Microsoft wants to do for agents what training infrastructure did for models: make them optimizable instead of mystical. The appeal is simple — keep the agent, improve the performance, skip the rebuild ritual.

4. 4DV

Animated Gaussian splats, spatial media, and a glimpse of video after the rectangle. The demos make a lot of current AI video look instantly old.

5. Particles

An AI-native particle simulator that lets you co-create huge 3D swarms with an LLM instead of spending six months becoming a full-time graphics goblin. Creative coding with the barrier to entry suddenly dropped through the floor.

6. The Rosie Protocol

A man used AI to help design a personalized cancer vaccine path for his dog. The real story isn’t just biotech drama — it’s that motivated outsiders can now push into domains that used to feel brutally expert-locked.

7. A Field Guide to Seedance

Less “here’s a flashy model,” more “here’s how to actually become dangerous with it.” A guide for using Seedance like an operator instead of a tourist.

8. Why the Harness Matters More Than the Model

A clean reminder that agents are not just models — they’re models plus harnesses, evals, traces, and execution logic. If you want to understand where agent quality actually comes from, start here.

9. projects.dev

Stripe is trying to become more than payments infrastructure — it wants to help provision the modern software stack itself, straight from the terminal. Quietly ambitious, which is Stripe at its most Stripe.

10. Where AI Will Hit Hardest

One of the few projects trying to map AI job displacement geographically instead of talking about it like vague future weather. Not just that disruption is coming — where it lands.

11. AVM

“V8 for agents” is a strong pitch because it points at a real missing layer. If agents are going to run real workflows, they need runtimes with security, limits, visibility, and control — not just vibes and wrappers.

A lot of these links rhyme with each other. New interface primitives. New creative tooling. New agent infrastructure. New maps of who gets hurt. Less “AI can do this now,” more “here’s the world that starts to form when the toys harden into systems.”

That’s what I want this mixtape to keep doing every month: not just showing you what’s new, but what’s starting to matter.

It used to take a specialized team months to crack a proprietary system. Binary analysis, protocol sniffing, painstaking trial and error. The complexity was the moat. If your internals were sufficiently opaque, you were safe — not because they were uncrackable, but because nobody could justify the cost.

That calculus broke last week.

Manjeet Singh, working with Claude Opus, reverse-engineered Apple’s M4 Neural Engine. The entire stack — from CoreML’s abstraction layers down to the IOKit kernel driver. They mapped 40+ private API classes, cracked the in-memory compilation path, and got transformer training running directly on hardware that Apple designed exclusively for inference. Forward pass, backward pass, Adam optimizer. No GPU, no Metal. Pure ANE compute.

The code is open source. The writeup documents everything step by step.

An ex-Apple Xcode engineer commented on Hacker News: “I worked on the Xcode team for years and know the lengths Apple goes to make this stuff difficult to figure out.”

It took a week.

The moat was never technical

Apple’s protection wasn’t in the silicon. The ANE’s compute primitives — convolution, matrix multiplication, softmax, elementwise ops — are the same ones you need for training. You just run them in reverse. The hardware doesn’t know or care which direction the gradients flow.

The moat was informational. No published ISA. No documented APIs. Everything routed through CoreML’s abstraction layers. A black box by policy, not by necessity.

That distinction matters. Technical moats require you to build better hardware. Informational moats only require that nobody looks too hard. And “nobody looks too hard” is not a durable assumption when one engineer with good intuition can pair with an AI that writes probing code, reasons through runtime introspection data, and iterates on hypotheses in hours.

This generalizes

The ANE was about as hard a target as you can find. Undocumented proprietary hardware. No public instruction set. Multiple abstraction layers. Made by the most secretive company in tech.

If that falls in a week, what’s the shelf life of your undocumented protocol? Your internal API? Your “competitive moat” that’s really just complexity nobody’s bothered to untangle yet?

Hardware companies, SaaS platforms, API providers — anyone whose advantage depends on keeping internals opaque is now operating on borrowed time. The cost of understanding a system just dropped by an order of magnitude. The only durable moats left are the ones that still hold up after someone understands exactly how they work.

The latent compute story

There’s a secondary implication that’s getting less attention. There are hundreds of millions of Apple Silicon devices in the world. Every one has an ANE sitting idle unless you’re running CoreML inference.

That’s an enormous pool of compute that was artificially locked. On-device fine-tuning — training on your own data, on your own hardware, without sending anything to the cloud — just went from theoretical to demonstrated. Apple has been selling the privacy story for years. This actually delivers on it in a way Apple themselves haven’t.

And it’s not just Apple. Qualcomm’s Hexagon DSP, Intel’s NPU, every “AI accelerator” shipping in consumer hardware. If the ANE’s fixed-function inference primitives can run backpropagation, theirs probably can too.

The speed of it

The HN thread is debating whether you can trust research done with an AI co-author. The code is open source and the results are reproducible. That question answers itself.

The real takeaway is the time compression. This style of work — human intuition directing the exploration, AI handling code generation and data analysis at inhuman iteration speed — is going to crack open a lot of things that were previously too expensive to bother with.

A week. Open source. Reproducible. That’s the new baseline for cracking proprietary hardware. Plan accordingly.

The Flip

Two things broke simultaneously: The floor dropped out. The bottom 80% of work across the whole org, not just engineering, but product specs, design iteration, QA, documentation, ops, got automated. The cost to go from idea to validated product collapsed to near-zero. A solo founder with good taste now POCs in a day what used to take a team a quarter. The ceiling rose. The top 20%, taste, architecture, knowing what to build and why, now gets 5-10x more use per dollar. One person with judgment ships at the velocity that used to require a department. Both directions compress the markup. The cost structure that justified SaaS pricing didn’t shrink. It evaporated.

“But Data Moats”

This is the strongest counter-argument. It’s real, but it’s eroding faster than people think. Salesforce has your CRM. Workday has your org chart. That’s genuine lock-in, and these are institutions. Institutions are slow to topple. But reverse-engineering data is cheap now. In-browser agents from companies like Harmonic and Happenstance literally scrape your entire LinkedIn graph without needing an export button. They don’t need your API. They don’t need your permission. The “painful data migration” moat, the thing that kept users stuck in products they’d outgrown, is being dissolved by software that routes around the wall instead of going through the gate. These companies face a binary: evolve by making UX the actual moat, or get hollowed out by faster competitors who nail the interface layer. Data gravity that used to be a permanent advantage is becoming a temporary one.

Distribution Is Operations

Here’s where the skeptics accidentally prove the thesis. “A million competing products don’t matter if nobody finds them.” Correct. Distribution is king. Getting users, keeping users, serving users, that’s the game. But that’s operations, not software. Marketing, sales, community, support, trust. The stuff you can’t automate with a coding assistant. The software part got commoditized. What’s left is the operational muscle that was always the hard part. If your SaaS company already understood this, congratulations, your costs just dropped and your moat stayed. If you thought the software was the business... That’s where it gets uncomfortable.

The Punchline

The SaaS companies that survive already know they’re operations companies that happen to build software. The ones that think they’re software companies are already dead. They just haven’t noticed because the subscription revenue keeps hitting for a while after the moat disappears. The product was never the code. It was the operation around it. We just couldn’t see it when building was expensive.

The stack, simplified

Think about it like floors in a building:

Ground floor: Infrastructure

Blockchains. Data centers. GPU clusters. Essential, invisible, and increasingly commodity. Nobody brags about their ISP at parties. Nobody's going to brag about which foundation model powers their assistant, either. Not for long.

Middle floor: Protocols

AWS, DeFi protocols, model APIs. Important middleware. But users don't care which protocol handles their swap any more than you care which CDN serves your Netflix stream. Or which model answered your question.

Penthouse: The interface

Chrome. MetaMask. ChatGPT. This is where people actually live. This is where attention sits. And attention is the only currency that doesn't inflate.

The crypto version (still true)

Robbie Petersen laid this out well: as blockchains commoditize, wallets become the value capture layer. They own the user relationship. They see every transaction. They can route flow wherever earns the best kickback. The wallet doesn't care if you're on Ethereum or Solana. It cares that you're using their wallet. It's the Google playbook. Chrome doesn't care which server your website runs on. It cares that you're searching through Google.

The AI version (and why it might be bigger)

Now watch the same pattern play out in AI, except faster and with higher stakes. Models are the new L1s. GPT-4, Claude, Gemini, Llama, DeepSeek - they’re all converging. Performance gaps shrink every quarter. Open source is nipping at closed source. Sound familiar? It should. It’s the L1 wars all over again. And just like L1s, models are becoming plumbing. The AI “wallet” is the assistant. Whoever’s chat interface you talk to first thing in the morning owns your decisions. Not your model provider. Not your cloud host. The assistant layer sees everything: your calendar, your emails, your preferences, your purchase intent. It routes your attention the same way a wallet routes your transactions. The switching cost is trust, not technology. You don’t switch assistants because a new model scores 2% higher on a benchmark. You switch when you lose trust. Your assistant knows your schedule, your writing style, your kid’s school pickup time. That’s a moat no benchmark can measure. Here’s the thing about the AI version that makes it scarier than crypto: the interface layer doesn’t just route transactions. It routes thinking. When your assistant summarizes your email, picks which meetings matter, drafts your responses - it’s not just capturing value. It’s shaping decisions. The fat wallet thesis in crypto is about capturing transaction flow. In AI, it’s about capturing cognitive flow.

The pattern, zoomed out

Every tech cycle, same movie:

• Web: Hosting was commodity. Browsers won.

• Mobile: Carriers became pipes. IOS and Android captured the value.

• Cloud: Compute is commodity. Slack, Notion, Figma - the apps on top took the margins.

• Crypto: Blockchains commoditize. Wallets win.

• AI: Models commoditize. Assistants win. In each cycle, smart money bet on infrastructure early. But the long game always favored whoever owned the last mile to the user’s eyeballs. Or in AI’s case, the user’s thought process.

Where are your bets?

Most people right now are betting on models. New architectures, new training runs, new benchmarks. Some of those will do fine. But the asymmetric bet is the same as it's always been: the interface layer. He who controls the front-end controls the future. The real question in crypto was "which wallet do 100 million people open every morning?" The real question in AI is "which assistant do 100 million people think through every morning?" If the answer is the same product, god help us all.

Originally published 2025. Updated Feb 2026 with the AI parallel. Inspired by @robbiepetersen_'s Fat Wallet Thesis