's Avatar

@comathematician.bsky.social

23 Followers  |  28 Following  |  7 Posts  |  Joined: 02.12.2024  |  1.5403

Latest posts by comathematician.bsky.social on Bluesky

#GRT3 Update: The @aivillage_dc will be hosting their third Generative Red Team event online in about a month. To learn more and get involved, visit grt.aivillage.org

09.08.2025 18:20 β€” πŸ‘ 11    πŸ” 4    πŸ’¬ 0    πŸ“Œ 1
Post image Post image

My new AI test visual test is "three rivers merging into one". Gemini *kinda* got it in 4 prompts. I gave up with chatgpt.

As a mathematician, I can attest that counting is hard.

15.03.2025 03:34 β€” πŸ‘ 0    πŸ” 0    πŸ’¬ 0    πŸ“Œ 0

When people are bewildered they tend to become credulous.
-Calvin Coolidge
I think this applies to a lot these days.

27.02.2025 15:01 β€” πŸ‘ 1    πŸ” 0    πŸ’¬ 0    πŸ“Œ 0

I'm trying to get a odd model working on a Ryzen AI CPU with a NPU. The linux support is god-awful and there's no onnx integration. It's been out for 2 years and the driver was merged into the kernel last week.

There's a reason NVIDIA won.

07.02.2025 00:36 β€” πŸ‘ 0    πŸ” 0    πŸ’¬ 0    πŸ“Œ 0
Stanford CRFM

First distillation/extraction attack for OAI was the Stanford Alpaca research. It was after this that OAI changed its ToS to disallow training on outputs. It can happen to all the model providers.

crfm.stanford.edu/2023/03/13/a...

29.01.2025 23:15 β€” πŸ‘ 2    πŸ” 2    πŸ’¬ 0    πŸ“Œ 0

Here's a red team report on DeepSeek: cdn.prod.website-files.com/6690a78074d8...

29.01.2025 03:06 β€” πŸ‘ 0    πŸ” 0    πŸ’¬ 0    πŸ“Œ 0

Deploying DeepSeek or Mixtral locally is often needed, especially with finetuning smaller models for particular tasks, but if there is an adversarial component you are choosing to "roll your own crypto". Proceed with care, and know there's an ongoing operational cost.

29.01.2025 03:06 β€” πŸ‘ 0    πŸ” 0    πŸ’¬ 1    πŸ“Œ 0
The Moat for AI The value of AI companies is largely the security layers they provide.

The main moat of OpenAI, Google, Anthropic and the rest are the security layers they offer to keep the models behaving as they should. AI security is very difficult and starting with a trusted llm with a solid & agile security team saves businesses money.

blog.nbhd.ai/moat

29.01.2025 03:06 β€” πŸ‘ 0    πŸ” 0    πŸ’¬ 1    πŸ“Œ 0

I got to integrate a function by hand today. πŸ₯°

13.12.2024 14:31 β€” πŸ‘ 1    πŸ” 0    πŸ’¬ 0    πŸ“Œ 0

@comathematician is following 20 prominent accounts