Tabular intelligence for the data that runs business.
Joined January 2025
- Tweets104
- Following1
- Followers998
- Likes643
Bojan Tunguz retweeted
Introducing mlx-learn: GPU-accelerated classical machine learning for Apple silicon, powered by MLX and native Metal kernels. Logistic regression, nearest neighbors, and exact support vector classification, with a familiar scikit-learn API.
1/4
Bojan Tunguz retweeted
While GenAI, LLMs, and world models are all the rage, classical ML still has a million use-cases where it is a better fit. This makes it easier (and faster) to find that ft on Apple silicon!
Introducing mlx-learn: GPU-accelerated classical machine learning for Apple silicon, powered by MLX and native Metal kernels. Logistic regression, nearest neighbors, and exact support vector classification, with a familiar scikit-learn API.
1/4
SVC delivers the biggest gains in those benchmarks:
• Up to 29.5× faster training with a 2 GB kernel cache
• Up to 298× faster prediction
Accuracy matches or closely matches scikit-learn across all three datasets.
3/4
Use direct imports or add patch_sklearn() to accelerate supported estimators in existing workflows. Automatic CPU/GPU routing included.
Available now in alpha. Open source, Apache 2.0.
pip install mlx-learn
Code + benchmarks: github.com/tabulai/mlx-learn
4/4
Bojan Tunguz retweeted
We are releasing an improved version of metal-graph library, 0.1.2. We've made some substantial changes to the underlying algorithms, which has resulted in significant speedups in our testing. This is still a very early experimental release, and we need to implement a lot more tests, especially under some real-world scenarios, but we feel confident that this already a very strong library that can be used for many projects where graph computations are relevant. Please go check it out. As always, feedback, critique, and any other comments are more than welcome.
Install it simply on your Mac via
python3 -m pip install --upgrade metal-graph==0.1.2
Bojan Tunguz retweeted
Introducing mlx-stump: GPU-accelerated matrix profiles for Apple Silicon, built with MLX and a STUMPY-compatible API. Find recurring patterns and anomalies in time series using your Mac’s GPU. 🧵
1/5
Bojan Tunguz retweeted
Synthetic benchmarks on an M4 Max, against STUMPY using all 16 CPU cores:
🚀 24.8× faster raw AB-joins
🚀 9× faster top-k profiles
🚀 6× faster short-window self-joins
Same machine. Medians of five warmed runs.
2/5
Introducing mlx-stump: GPU-accelerated matrix profiles for Apple Silicon, built with MLX and a STUMPY-compatible API. Find recurring patterns and anomalies in time series using your Mac’s GPU. 🧵
1/5
Repeated queries benefit too: prepared-target MASS/match calls were 4.1–5.3× faster, excluding setup.
Workload matters: STUMPY won on ordinary single-query calls, very large windows, and nearly noiseless periodic signals.
4/5
mlx-stump v0.1.1 is now on PyPI, including the latest fixes and performance improvements.
pip install mlx-stump
Requires macOS on Apple Silicon and Python 3.10+.
Code and benchmarks: github.com/tabulai/mlx-stump
5/5
We are releasing an improved version of metal-graph library, 0.1.2. We've made some substantial changes to the underlying algorithms, which has resulted in significant speedups in our testing. This is still a very early experimental release, and we need to implement a lot more tests, especially under some real-world scenarios, but we feel confident that this already a very strong library that can be used for many projects where graph computations are relevant. Please go check it out. As always, feedback, critique, and any other comments are more than welcome.
Install it simply on your Mac via
python3 -m pip install --upgrade metal-graph==0.1.2
Bojan Tunguz retweeted
We are releasing mlx-geospatial: a Python library for fast geospatial work on Apple silicon, built on MLX. It handles point-in-polygon queries, spatial joins, distances and nearest neighbors. A few results from pre-release benchmarks on an M4 Max:
1/9
After user feedback and further optimization, mlx-geospatial 0.1.1 is out! Faster spatial joins + kNN on Apple silicon.
Warm-index component benchmarks:
M1 Ultra: 18–1,262× (median 95×)
M4 Max: 21–2,005× (median 105×)
CPU/MLX ratios, shown on the same scale.
1/4
The complete-join MLX pipeline beat GeoPandas:
M1 Ultra: 3.78–18.66×
M4 Max: 2.55–16.51×
Six synthetic workloads. Fresh indexes, exact boundaries, all matches, full GeoDataFrames + sorting. Median of 3 paired trials after warmup.
3/4
New: threaded Point extraction, faster polygon-edge searches and boundary refinement, plus chord-distance rejection for kNN.
721 tests per Mac. All 48 full-output checks passed.
pip install -U mlx-geospatial
github.com/tabulai/mlx-geosp…
4/4
