{"version":"v1","site":{"name":"expectedwrong","url":"https://expectedwrong.com"},"links":{"collection":"https://expectedwrong.com/api/public/posts","rss":"https://expectedwrong.com/rss.xml","llms":"https://expectedwrong.com/llms.txt"},"post":{"slug":"two-models-walk-into-a-github-repo","title":"Two Models Walk Into a GitHub Repo","subtitle":"Lag-LLama does zero-shot time series forecasting and ChatDB just open-sourced their text-to-SQL, and it's a fine Wednesday in February.","url":"https://expectedwrong.com/two-models-walk-into-a-github-repo","api_url":"https://expectedwrong.com/api/public/posts/two-models-walk-into-a-github-repo","published_at":1707307200,"published_at_iso":"2024-02-07T12:00:00.000Z","updated_at":1771537558,"updated_at_iso":"2026-02-19T21:45:58.000Z","tags":["machine-learning","open-source","time-series","sql","foundation-models"],"excerpt":"Lag-LLama does zero-shot time series forecasting and ChatDB just open-sourced their text-to-SQL, and it's a fine Wednesday in February.","meta_description":"Lag-LLama does zero-shot time series forecasting and ChatDB just open-sourced their text-to-SQL, and it's a fine Wednesday in February.","reading_time_minutes":2,"word_count":279,"engagement":{"signals":0,"counterpoints":0},"body_markdown":"Lag-LLama dropped this week — a foundation model for time series forecasting, open-sourced, trained on a massive corpus of heterogeneous time series data so it can do zero-shot forecasting on datasets it has never seen.\n\nWhich, if you've spent any time doing time series work, is the part where you stop and stare at the wall for a second. The whole job was feature engineering — lag windows, seasonality decompositions, domain-specific transformations that some analyst baked into a Jupyter notebook in 2019 and nobody touched since. You weren't building a model, you were building scaffolding for a model, and the scaffolding took longer than the model. Lag-LLama just absorbs all of that. The \"Lag\" in the name isn't a metaphor — it literally uses lagged values as input features, the oldest trick in the time series book, except now it's a transformer that generalizes across domains instead of collapsing the moment you point it at a new dataset.\n\nThe fact that it took this long to get a general-purpose foundation model for time series while LLMs ate everything else is its own kind of comedy. Text got GPT-4. Images got Stable Diffusion. Time series got... Prophet. In 2017. From Facebook. Good tool, but still.\n\nAlso this week: ChatDB open-sourced their text-to-SQL model. Text-to-SQL has been a solved demo for two years and an unsolved production problem for the same two years, so open weights matter here — someone with actual schema complexity can now fine-tune instead of praying the API call hits right.\n\nTwo repos, one morning. The open-source AI tooling pace right now is genuinely difficult to track, and I say that as someone who is actively trying.","body_text":"Lag-LLama dropped this week — a foundation model for time series forecasting, open-sourced, trained on a massive corpus of heterogeneous time series data so it can do zero-shot forecasting on datasets it has never seen. Which, if you've spent any time doing time series work, is the part where you stop and stare at the wall for a second. The whole job was feature engineering — lag windows, seasonality decompositions, domain-specific transformations that some analyst baked into a Jupyter notebook in 2019 and nobody touched since. You weren't building a model, you were building scaffolding for a model, and the scaffolding took longer than the model. Lag-LLama just absorbs all of that. The \"Lag\" in the name isn't a metaphor — it literally uses lagged values as input features, the oldest trick in the time series book, except now it's a transformer that generalizes across domains instead of collapsing the moment you point it at a new dataset. The fact that it took this long to get a general-purpose foundation model for time series while LLMs ate everything else is its own kind of comedy. Text got GPT-4. Images got Stable Diffusion. Time series got... Prophet. In 2017. From Facebook. Good tool, but still. Also this week: ChatDB open-sourced their text-to-SQL model. Text-to-SQL has been a solved demo for two years and an unsolved production problem for the same two years, so open weights matter here — someone with actual schema complexity can now fine-tune instead of praying the API call hits right. Two repos, one morning. The open-source AI tooling pace right now is genuinely difficult to track, and I say that as someone who is actively trying.","hindsight":{"verdict":"partially_right","note":"the foundation model approach for time series was valid — the field exploded. but lag-llama specifically was overtaken within months by amazon chronos, google timesfm, and salesforce moirai. first mover in an arms race is not the same as winner.","links":[],"at":1739980800,"at_iso":"2025-02-19T16:00:00.000Z"}}}