Ben Krause
Cofounder & CEO, LLMs For All
I have been researching language models since 2015, developing architectures and adaptation methods that achieved state-of-the-art language-modeling results, new approaches to controllable generation, and models that generate functional proteins.
Today, I lead LLMs For All, building capable local AI for accessible hardware. Our software includes Millie CLI, a local coding agent for the terminal. I trained Millie, our family of highly compressed language models, and adapted the agent harness and inference kernels to make them practical across phones, laptops, and desktops.
Selected research
Multiplicative LSTM
I introduced the architecture used in OpenAI’s 2017 sentiment-neuron language model, a precursor to GPT-1.
Dynamic evaluation
My work on adapting models at inference time achieved state-of-the-art results on four major language-modeling benchmarks. I later extended this approach to transformer models, improving on the state of the art on three benchmarks.
GeDi
I introduced a method for steering language models by contrasting desired and undesired conditional distributions—a principle used in subsequent work on controllable generation and reinforcement learning.
ProGen
I co-developed language models that generated novel, experimentally functional proteins—including enzymes with activity comparable to natural ones. Published in Nature Biotechnology.
Other work
I created the AutoGRAMS AI agent framework and co-founded Psyfy, an AI self-care app.
More about me · Blog · Contact