Simon Willison's predictions

@MolloyLaurence The price of running a GPT3 class model appears to have dropped by a factor of 100x in 24 months, I expect we will see at least a 10x drop on costs to run o3 within a year

Expected at least a 10x drop in the cost of running o3 within a year. OpenAI cut o3 by 80% (5x) on 2025-06-10, from $10/$40 to $2/$8 per M tokens, with no model change. Real and large, but half the predicted factor for o3 itself.

Original (x.com)simonwillison.net

@MolloyLaurence The price of running a GPT3 class model appears to have dropped by a factor of 100x in 24 months, I expect we will see at least a 10x drop on costs to run o3 within a year

Expected at least a 10x drop in the cost of running o3 within a year. OpenAI cut o3 by 80% (5x) on 2025-06-10, from $10/$40 to $2/$8 per M tokens, with no model change. Real and large, but half the predicted factor for o3 itself.

Original (x.com)simonwillison.net

@nedcoder I respect that. I'm still looking forward to the first genuinely useful LLM trained on entirely licensed or out-of-copyright data - I think it's going to happen this year

Predicted a genuinely useful LLM trained only on licensed/out-of-copyright data in 2024. Pleias 1.0 (2024-12-05) was billed as the first models trained exclusively on open data. They are small (350M-3B) and some licensing questions remain, so MOSTLY_TRUE.

Original (x.com)huggingface.co

@jeffrschneider @asmeurer @mitsuhiko I expect that within a year or two the vast majority of code written by professional programmers will have had LLMs involved at some point - code completion, code review, looking up common patterns etc At which point calling that "vibe coding" won't convey anything useful

Within a year or two, most professional code touched by LLMs.

Original (x.com)