Benjamin

Head of Channel Sales · AI, Entertainment, Apps & Games @Meta · Builder of AI things on the side · Author · Angel ·

[Benjamin @benjiminu]

I'm at Meta running Channel Sales. Few months ago I tried something different — I solo-built two AI products of my own, and put them out for paying users. Honestly? I learned more in that year than I had in the previous 9 at Meta. A few things I didn't see coming ↓

[Benjamin @benjiminu]

You can put a frontier model behind export law. Or you can download a better one for free, from a lab that doesn't answer to that law.

The lock still exists. It's just optional now.

Controlling access only works when access is scarce. It stopped being scarce. Kimi K3 moment.

[Benjamin @benjiminu]

A harness changes what the model sees and does. Context, tools, retrieval, structure, verification. It's for controlling a single run without touching the weights.

Rule of thumb: if better instructions could fix it, it's a harness problem. Fine-tune only when they can't.

[Benjamin @benjiminu]

Harness vs fine-tuning is one of the most misunderstood tradeoffs in applied AI.

People reach for fine-tuning when they have a harness problem, and vice versa. They solve different things.

Here's the actual line between them.

[Benjamin @benjiminu]

Fine-tuning changes how the model behaves. It's for teaching a durable skill, tone, or format the base model doesn't have — when you need it baked in, at scale, without spending context to explain it every call.

Cost: data, training, and you're locked to a snapshot.

[Benjamin @benjiminu]

Brands are buying AI tools to make more ads. The vendors love it — "10x your creative output" sells.

But more output was never the problem. Attention is finite. Feeds are saturated. The hard part is knowing what deserves to exist.

AI should make you make less, not more.

[Benjamin @benjiminu]

Meet the new Meta Glasses: our boldest AI collection yet. With Meta AI, 3K hands-free video, and live translation, they’re a total game-changer. Plus, don't miss our exclusive Meta Glasses by Kylie Jenner. https://t.co/g0Xr86akmg

#LifeAtMeta https://t.co/kRg8bAvm6M

[Benjamin @benjiminu]

Loop engineering — you stop prompting at all. You design the system that finds work, prompts the agent, verifies the result, and decides what's next, until a goal is met.

Each step, the human moved one rung up the ladder: from writing prompts to writing what writes them.

[Benjamin @benjiminu]

"Loop engineering" is the term everyone in AI coding is arguing about right now.

Easiest way to understand it: look at the four steps that got us here. Each one moved the human up a level — because each time, the bottleneck moved.

[Benjamin @benjiminu]

Harness engineering — you build the wrapper around the model: context, sandbox, tool surface, routing. Bottleneck: equipping a single agent run to actually succeed.

[Benjamin @benjiminu]

Where the Fable 5 export ban leads:

A world where you don't buy access to the best AI models. You apply for it.

A visa to use frontier intelligence — granted or denied based on who you are and where you're from.

The first stamp just went in the passport.

[Benjamin @benjiminu]

Underrated gotcha in vector search:

Cosine similarity measures angle — direction, not meaning.

Two chunks can score "close" for sharing structure or tone, not because they mean the same thing.

You're not retrieving meaning. You're retrieving geometry that correlates with it.

[Benjamin @benjiminu]

AI didn't kill the IC or the manager. It killed the gap between them.

When building is cheap and direction is the constraint, the valuable people are the ones who can do both — set the vision AND ship the thing.

Stop asking which track. Start collecting both skills.

[Benjamin @benjiminu]

Everyone's arguing about the manager-to-IC shift like it's a question of which side wins.

Wrong frame.

The line between "manager" and "IC" is collapsing. The future isn't one or the other. It's people who do both — and titles that haven't caught up yet.

[Benjamin @benjiminu]

I run a global org by day. At night I solo-build AI products — writing code, shipping to paying users.

A few years ago those were two different careers. Today they're the same week.

The "manager who can't build" and the "builder who can't lead" are both getting hard to place.

[Benjamin @benjiminu]

Pattern I keep seeing in advertising:

AI made creative volume basically free, 500 variants in seconds.

But the bottleneck didn't disappear. It moved. The hard part isn't making the ad now. It's knowing which one is worth making.

Taste didn't get cheaper. It got more valuable.

[Benjamin @benjiminu]

Contrarian take from shipping AI products:

You can't prompt hallucination away. You have to architect it away.

Structured outputs. Schema enforcement. Strict JSON. The model literally cannot return shapes you didn't define.

Prompting is hope. Schema is contract.

[Benjamin @benjiminu]

Unpopular career take in 2026:

Stop chasing titles. Start chasing surface area.

You'll learn more in 6 months shipping a side project than 2 years getting promoted.

Promotions teach you how your company works. Building teaches you how the world works.

[Benjamin @benjiminu]

The fix: a re-ranking layer after retrieval.

Retrieve wide on similarity, then re-rank for relevance AND diversity. Reward the chunk that adds something new, not the one that echoes what you already have.

Different objective entirely. Most RAG setups skip this step.

[Benjamin @benjiminu]

Contrarian RAG take from building Daymaster:

The chunks closest in vector space are often the worst retrieval result.

Not because they're irrelevant. Because they're redundant.

[Benjamin @benjiminu]

High similarity means the chunks are all saying the same thing.

You retrieve five chunks, you've really retrieved one chunk five times and the complementary piece you actually needed never makes the cut.

Vector search optimizes for "most similar." That's not the same as useful.