Machibet777 Crickettitle_temp

Billy Newport

New winner: Qwen3–30B-A3B takes the crown for document Q&A

Up to now, the Deepseek-Qwen-32B-R1 is the best local LLM for asking questions against a large document that I’ve tested. I use a large…

May 1

May 1

LLMs already seem to have conceptual models

Anthropics latest paper on the biology of LLMs is eye opening. (Ignacio de Gregorio just did a great piece on it). Yan LeCun has been…

Apr 5

Apr 5

Apple’s M3 Ultra Mac Studio Misses the Mark for LLM Inference

I wanted to love the M3 Ultra Mac Studio — 512 GB of RAM sounds like an LLM playground — but after testing and a bit of math, it’s a $10k…

Mar 8

Mar 8

DeepSeek R1 32B takes the lead for local document analysis

I have been ranking large language models (LLMs) based on their ability to answer questions about my autobiography, a 45k token document. I…

Feb 12

Feb 12

New gaming rig, the Ally X

I usually gamed on a nintendo switch. Zelda, breath of the wild was the game I played most. Recently, I saw the buzz around diablo 4…

Feb 10

Feb 10

The problem with reasoning LLMs right now

I’m mostly interested in LLMs ability to answer questions from large contexts. LLM is an acronym for large language models. Recently, we…

Feb 6

The problem with reasoning LLMs right now

Feb 6

Llama 3.3 70B Q4 vs Qwen 14B Q8

So Meta just dropped Llama 3.3 70B yesterday. It’s supposed to be a finetune of 3.2 and equivalent to their 405B model. It has a context…

Dec 7, 2024

Dec 7, 2024

Apple Watch 7 with a Tacx Neo Smart Bike

I’m in trials between Garmin and Apple right now. I’ve used Garmin watches forever, since the 305. I have been a runner, ultra marathoner…

Nov 22, 2024

Apple Watch 7 with a Tacx Neo Smart Bike

Nov 22, 2024

How I think coding assistants and AI coding will actually develop

The rage right now is coding assistants essentially doing code completion in vscode or other IDEs. I don’t think this is really where we…

Nov 14, 2024

Nov 14, 2024

Qwen 2.5 Instruct Coder 14B vs Llama 3.2 3B

Continuing on LLM document chat testing, Qwen 2.5 just came out with 14B and 32B models. Both claim to support 128K context windows but at…

Nov 13, 2024

Nov 13, 2024

Billy Newport

Billy Newport

Creator of DataSurface. Ex IBM Distinguished Engineer, Ex Goldman Sachs Managing Director. "Expert" in data warehousing and data platforms in general.

Following

Help

About

Careers

Press

Blog

Privacy

Rules

Terms