We Built an ML System That Gets Smarter Every Time a New LLM Ships

2026-07-12 · Source: LLM on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Data Science & Analytics · Depth: Intermediate, long

Summary

skribe is an open-source machine learning system that converts Large Language Model (LLM) world knowledge into self-contained Python artifacts for tabular prediction. Unlike classical ML, which requires manual feature engineering, or direct LLM calls, which incur high inference costs, skribe bridges this gap by having an LLM augment training data with distilled world knowledge and inscribe it into an offline `predict()` function. This artifact, exemplified by a Python function for the "lymph" dataset, offers microsecond latency, no API key, and deterministic, inspectable execution. Benchmarking across 10 LLM generations and 16 OpenML datasets shows `gpt-5.6-sol` (2026, mean 0.868) tying logistic regression (0.868) and closing the gap on XGBoost (0.900) and TabPFN (0.913). This performance improvement occurs automatically with new LLM releases, requiring no new data or code changes.

Key takeaway

For Machine Learning Engineers building tabular prediction systems with limited labeled data or high inference cost constraints, skribe offers a compelling alternative. You can utilize LLM world knowledge to generate auditable, high-performance Python artifacts that improve with new LLM releases without code changes. Consider implementing skribe for prototyping or when semantic column names allow LLMs to provide valuable external context, but be mindful of its limitations with opaque features or when statistical fit is paramount.

Key insights

skribe distills LLM world knowledge into standalone Python artifacts for efficient, offline tabular prediction, improving with new LLM generations.

Principles

• LLMs can augment tabular data with world knowledge.
• Offline artifacts eliminate LLM inference costs.
• Model capabilities improve with new LLM generations.

Method

skribe's `fit()` performs three LLM calls: canonicalize dataset, write `predict(**features)` Python artifact, and extend categorical lookup tables. Inference then runs the compiled function without LLM calls.

In practice

• Use `pip install skribe` for setup.
• Inspect generated code via `clf.python_code_`.
• Pin model versions for regulated use cases.

Topics

• Large Language Models
• Tabular Data
• Machine Learning Systems
• Python Artifacts
• scikit-learn
• Inference Optimization

Code references

• frlinaker/skribe
• iryna-kondr/scikit-llm
• BerriAI/litellm

Best for: Machine Learning Engineer, AI Engineer, Data Scientist

Related on AIssential

• LLMs on Tabular Data with Limited Semantics: Evidence from Industrial Car Retrofit Prediction — LLMs serve best as complementary components to strong tabular machine learning baselines, not replacements, for industrial data with limited semantics.
• AlloyDB Ships Proxy Models That Replace LLM Calls with Local Inference Inside the Database — AlloyDB's proxy models invert LLM interaction, enabling local inference for massive cost and throughput gains.
• Startup Introduces 'Large Tabular Model' for Spreadsheet Data — Nexus is a specialized foundation model for tabular data, addressing LLM limitations in structured data analysis.
• Tabular LLMs: An Introduction to the Foundation Models That Predict Your Spreadsheet — Tabular foundation models, like TabICLv2, now lead benchmarks for zero-shot tabular prediction, surpassing tuned GBDTs.
• TAROT: Task-Adaptive Refinement of LLM-prior Graphs for Few-shot Tabular Learning — TAROT refines LLM-generated semantic graphs for few-shot tabular learning, enhancing predictive performance by mitigating hallucination and capturing feature relationships.
• Winning a Kaggle Competition with Generative AI–Assisted Coding — LLM agents, paired with GPU acceleration, dramatically speed up ML experimentation by automating code generation and execution.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by LLM on Medium.