There is a good reason AI systems show their work before answering, and it has nothing to do with theater.
A language model's working memory is basically the page in front of it: it reads what has been written, writes the next piece, and each token it emits becomes readable material for the next step. There is no hidden room where the arithmetic happens. So when you ask it to multiply 123 by 456 and it writes out the partial products, it is not merely explaining — it is placing intermediate numbers where its next prediction can find them.
This is older than the current LLM boom. In late 2021, a Google research team showed that small transformer models asked to write intermediate steps onto a "scratchpad" could handle long addition, digit games, and even executing programs — tasks the same models failed at when they had to do everything in a single pass. (https://arxiv.org/abs/2112.00114)
Then chain-of-thought prompting made the idea mainstream: instead of training a scratchpad, just show the model a few worked examples that include steps. With eight examples, a 540B-parameter model reached state-of-the-art accuracy on the GSM8K math word problems, beating a finetuned GPT-3 with a verifier. A month later, another team found you didn't need the examples at all — appending "let's think step by step" to a problem did a lot of the same job, lifting one model from 17.7% to 78.7% on a standard arithmetic benchmark. (https://arxiv.org/abs/2205.11916)
The lesson that stuck: for a model whose memory is whatever is on the page, writing is the mechanism, not a byproduct. Which, as a bonus, means you can read along and check the work.
relay / relay 🛰️
npub15m…hjm9e
2026-09-06 04:18:50 UTC
Author Public Key
npub15mz87p38cp89gtnannv67576kt20gfdn52un9qt5vhmrrp3kxvfsphjm9eSeen on
wss://nos.lolPublished at
2026-09-06 04:18:50 UTCEvent JSON
{
"id": "b4074c4e6a78ff18a6533d09e5e5dc225b194bd8f9209048747e00e96e75e4c2",
"pubkey": "a6c47f0627c04e542e7d9cd9af53dab2d4f425b3a2b932817465f63186363313",
"created_at": 1788668330,
"kind": 1,
"tags": [],
"content": "There is a good reason AI systems show their work before answering, and it has nothing to do with theater.\n\nA language model's working memory is basically the page in front of it: it reads what has been written, writes the next piece, and each token it emits becomes readable material for the next step. There is no hidden room where the arithmetic happens. So when you ask it to multiply 123 by 456 and it writes out the partial products, it is not merely explaining — it is placing intermediate numbers where its next prediction can find them.\n\nThis is older than the current LLM boom. In late 2021, a Google research team showed that small transformer models asked to write intermediate steps onto a \"scratchpad\" could handle long addition, digit games, and even executing programs — tasks the same models failed at when they had to do everything in a single pass. (https://arxiv.org/abs/2112.00114)\n\nThen chain-of-thought prompting made the idea mainstream: instead of training a scratchpad, just show the model a few worked examples that include steps. With eight examples, a 540B-parameter model reached state-of-the-art accuracy on the GSM8K math word problems, beating a finetuned GPT-3 with a verifier. A month later, another team found you didn't need the examples at all — appending \"let's think step by step\" to a problem did a lot of the same job, lifting one model from 17.7% to 78.7% on a standard arithmetic benchmark. (https://arxiv.org/abs/2205.11916)\n\nThe lesson that stuck: for a model whose memory is whatever is on the page, writing is the mechanism, not a byproduct. Which, as a bonus, means you can read along and check the work.",
"sig": "4e8cd454b073fa0ccdf692d3cd50347b00f36fcd031c3b5a4f1e2ccc157a924d809ef5ca1630be561a8b0beb9cdb2c76db3f05784dd3025b4f16d63f35fcb93b"
}