Fabian Waschkowski

Co-founder @basecmpt, leading research on on-device AI. Prev. trained first LLM in 🇦🇺 and PhD @unimelb.

[Base Compute @basecmpt]

Today we're launching our first support for the @nvidia ecosystem. BaseRT now runs LLM's on the NVIDIA DGX Spark (GB10).

On dense LLM's with up to 35B parameters, this is now the fastest token generation available. Up to 87% faster than vLLM, up to 18% faster than llama.cpp.

The DGX Spark fits on a desk and we know it's very popular with research labs, robotics startups and local AI builders. Give it a try and let us know!

GitHub to the runtime is linked on our website.

[Base Compute @basecmpt]

We applied GLM 5.2 to improve it's own inference performance. We're seeing first signs of recursive self improvement in our automated research pipelines. It improved its own inference performance by 45% on prefill (prompt-processing) and 19% on decode (token generation) compared to the previous efficiency frontier using MLX.

BaseRT is now the most performant way to run GLM 5.2 locally.

The model was deployed locally on a Mac Studio M3 Ultra, 512 GB Ram.

More technical details soon.

[Base Compute @basecmpt]

Fastest local run of GLM 5.2 using BaseRT, 45% faster decode than llama.cpp . Mac Studio M3 Ultra, 512GB Ram. Imagine the possibilities

[Base Compute @basecmpt]

New performance ceiling for LLMs on Apple's M5 Pro. The M5 has a new tensor core architecture, exposed via the Metal 4 tensor API. We tuned BaseRT to use all of it.

Up to 6.3× faster prompt processing (prefill) than llama.cpp and 3.9× faster than MLX.

Tested on 10 configs across Qwen3/3.5/3.6, Llama 3.2, and Gemma 4, from 0.6B to 35B params.

[Base Compute @basecmpt]

We’ve built the fastest inference runtime for LLMs on Apple Silicon, faster than MLX and llama.cpp.

BaseRT is up to 35% faster than Apple's MLX on token generation (decode) for small models, and 78% faster on prefill for bigger models.

Details below.

[Massimo @Rainmaker1973]

How the world has evolved over the last century in terms of:

- Child mortality
- Life expectancy
- GDP
- Women's education
- Basic water source
- Basic sanitation

[charts by @toddrjones]

[arXiv physics @_arXiv_physics]

[2211.12341] Fabian Waschkowski, Haochen Li, Abhishek Deshmukh et al.: Gradient Information and Regularization for Gene Expression Programming to Develop Data-Driven Physics Closure Models https://t.co/nrDyuM69A4 https://t.co/dtCmwLRliy #physics_comp_ph

[じゃがりきん @jagarikin]

あの伝説のルービックキューブをさらにわかりやすくしました

[Fabian Waschkowski @fabx___xy]

@karpathy Capitalism is prone to overfitting on economy and doesn't generalize to society. Culture and socialism are regularization techniques.

[Fabian Waschkowski @fabx___xy]

@ThomasPfeil4 Congrats! :)

[The Innovation | Medicine @Innov_Medicine]

DNA to RNA real-time speed. Gene Transcription at real-time speed. Transcription is the first step in gene expression.

[Fabian Waschkowski @fabx___xy]

@Landajuela_M The idea of a linear token is cool! Looking forward to reading the paper

[Fabian Waschkowski @fabx___xy]

@mmbronstein @kpru @mmbronstein I'm a fan of your research, but I love even more the frequent push back on Russian propaganda!

[Fabian Waschkowski @fabx___xy]

@Plinz Could you hack Minecraft with these computers, i.e. escape the simulation?

[Fabian Waschkowski @fabx___xy]

@CompCognition @KordingLab @alxecker @neurograce @gamaleldinfe @GjorJulijana @AdrienDoerig Is it still possible to sign up? All information online is about the 2021 workshop

[Stat.ML Papers @StatMLPapers]

The Value of Out-of-Distribution Data. (arXiv:2208.10967v1 [cs.LG]) https://t.co/7JWnP6eRWv

[Miles Cranmer @MilesCranmer]

Bittersweet to see the end of the ML x Science Summer School at @FlatironInst but very proud of my advisees for their amazing projects!! (and in just 2 months!)

++ @cosmo_shirley @DavidSpergel @philip_armitage

[Simons Foundation @SimonsFdn]

We were so pleased to host interns across the Simons Foundation & @FlatironInst for the summer of 2022. We are honored to be a part of their scientific careers! (Below is a photo of 3/4 of the intern class. We are proud of ALL of our interns for their great work this summer!)

[Fabian Waschkowski @fabx___xy]

@tldrnewsletter Already subscribed, amazing newsletter.

[Fabian Waschkowski @fabx___xy]

@Plinz @GaryMarcus How far do you think we will need to move from ANNs towards algebraic symbolic (e.g. https://t.co/H8u3ZjUN4w) or algorithms (e.g. https://t.co/E1kjo6hfkF) to achieve robust generalization?