hawkes.aiKey ideas
- The one idea — Code-specialized training — heavy code diet plus a uniquely verifiable reward — buys large gains on software tasks, so the best coder often isn't the biggest generalist.
- How it is shown — The trade school: a general model sent to concentrated code training; the compile/test feedback loop; the specialist out-coding a larger generalist on the same task.
- The trap to avoid — Choosing the biggest general model for code — specialization often beats raw scale on the specific task.
- What it sets up — And reasoning is its own trade school.
Smaller coding models often beat bigger generalists at code. Their edge: a teacher that can't be fooled — the code either runs or it doesn't.
The one idea
Code-specialized training — heavy code diet plus a uniquely verifiable reward — buys large gains on software tasks, so the best coder often isn't the biggest generalist.
Why is one model dramatically better at code than a bigger generalist? It went to trade school. Concentrated training on code, and — the secret ingredient — a feedback signal most skills can only dream of. Coding is specialization's clearest case. Here's what the trade school does. Code's superpower is the reward — episode eighty at full force. Essays have taste; code has truth: does it compile, do the tests pass? An infallible judge that can't be flattered, feedback for free. So the reinforcement engine that struggled with fuzzy preference gorges on certainty. Code can be checked — and that explains most of the gap.
How it works — the demo
The trade school: a general model sent to concentrated code training; the compile/test feedback loop; the specialist out-coding a larger generalist on the same task.
The specialist gains the difference between plausible and correct. A generalist writes code that looks right; the graduate writes code that runs — edge cases, idioms, multi-file structure. The surprise: a smaller coding specialist routinely out-codes a much larger generalist on real software. On this axis, trade beats scale. And the specialty grew into an ecosystem: coding assistants read whole repositories, run their own tests, iterate against the compiler. Notice where this points: a skill with an automatic verifier can go autonomous, checking its own work without a human. Code is where act seven's agent era does its first real work. The trap cuts both ways. The biggest generalist on a coding task overpays and underperforms — a specialist wins for less. But push it outside its trade — business strategy, a condolence note — and it falters where a generalist shines.
The trap to avoid
Choosing the biggest general model for code — specialization often beats raw scale on the specific task.
Why it matters — and what’s next
And reasoning is its own trade school.
A sharpened edge: devastating in its groove, awkward out of it. Match school to task. One complication: the frontier is re-merging. Top flagships train deep coding into their generalists, narrowing the specialist's edge at the ceiling — though it persists in open and mid tiers. A pendulum that swings. Not a permanent law; a current landscape you re-check per release. And the engine that powered the code trade — the verifiable judge — has another territory: math and logic, where answers can also be checked. Point it there and you graduate a different specialist: a model built to think in long chains, brilliant at structure, plainer at chat. The reasoning specialist: next.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
Why is one model dramatically better at code than a bigger generalist? It went to trade school. Concentrated training on code, and — the secret ingredient — a feedback signal most skills can only dream of. Coding is specialization's clearest case. Here's what the trade school does.
Code's superpower is the reward — episode eighty at full force. Essays have taste; code has truth: does it compile, do the tests pass? An infallible judge that can't be flattered, feedback for free. So the reinforcement engine that struggled with fuzzy preference gorges on certainty. Code can be checked — and that explains most of the gap.
The specialist gains the difference between plausible and correct. A generalist writes code that looks right; the graduate writes code that runs — edge cases, idioms, multi-file structure. The surprise: a smaller coding specialist routinely out-codes a much larger generalist on real software. On this axis, trade beats scale.
And the specialty grew into an ecosystem: coding assistants read whole repositories, run their own tests, iterate against the compiler. Notice where this points: a skill with an automatic verifier can go autonomous, checking its own work without a human. Code is where act seven's agent era does its first real work.
The trap cuts both ways. The biggest generalist on a coding task overpays and underperforms — a specialist wins for less. But push it outside its trade — business strategy, a condolence note — and it falters where a generalist shines. A sharpened edge: devastating in its groove, awkward out of it. Match school to task.
One complication: the frontier is re-merging. Top flagships train deep coding into their generalists, narrowing the specialist's edge at the ceiling — though it persists in open and mid tiers. A pendulum that swings. Not a permanent law; a current landscape you re-check per release.
And the engine that powered the code trade — the verifiable judge — has another territory: math and logic, where answers can also be checked. Point it there and you graduate a different specialist: a model built to think in long chains, brilliant at structure, plainer at chat. The reasoning specialist: next.