Matt Henderson @matthen2

maths, visualisations, conversational AI. VP Research @polyaivoice prev: @RekaAILabs, @Apple AI/ML, @GoogleAI, PhD @Cambridge_Eng matthen.com Edinburgh, UK Joined February 2010

Tweets

6K
Followers

80K
Following

3K
Likes

8K

David Renshaw @dwrensha

a week ago

I made a Blender animation showing various nice ways to inscribe a Platonic solid inside another. Icosahedron -> Octahedron -> Tetrahedron -> Cube -> Dodecahedron -> Icosahedron

2 18 69 4K 9

Download Video

Matt Henderson @matthen2

2 weeks ago

‘Sovereign AI’ means fine tuning a model from an American company on data translated to Welsh by models from a Chinese company and another American company?

NVIDIA Europe @NVIDIAEU

2 weeks ago

‘Sovereign AI’ means fine tuning a model from an American company on data translated to Welsh by models from a Chinese company and another American company?

1 6 28 49K 1

Download Image

1 0 9 3K 3

Reka @RekaAILabs

2 weeks ago

Introducing Reka Speech: an efficient and accurate transcription & translation model. 🗣️ On modern GPUs (e.g. H100), it runs 8x–35x faster than existing solutions for batch processing.

10 28 140 23K 71

Download Image

@jxmnop Their lora training examples are just broken- github.com/huggingface/gp… Even if it worked, you wouldnt get the full efficiency because we can’t merge the loras into the quantized layers And even if we could, it only works on h100+

1 1 5 3K 3

Reka @RekaAILabs

a month ago

🚨 New benchmark release 🚨 We're introducing Research-Eval: a diverse, high-quality benchmark for evaluating search-augmented LLMs 👉 Blogpost: reka.ai/news/introduci… 👉 Dataset + code: github.com/reka-ai/resear… 🧵 Here's why this matters:

8 24 94 24K 50

Download Image

Matt Henderson @matthen2

a month ago

has anyone post-trained 7b-ish models to be better at tokenization awareness? Finding cases where they make quite bad mistakes when they need to manipulate strings in simple ways