Moulay Amine Jaidi

🔥 Bridging the Chasm: #AI Innovation on AMD Instinct #MI300X🔥 @AIatAMD wrote about our work with @HotAisle and our contribution to the growth of the #OpenSource community with state of the art #GPU. Read it here: amd.com/en/blogs/2025/… Blog: lnkd.in/ecXAZcA4

This work is part of our broader effort to build a framework for testing AI systems. As LLMs become increasingly complex, it's crucial we use them safely and responsibly.

🚀 Jailbreaking LLMs Our recent research has explored how #adversarialai can be used to bypass refusal responses. Using @HotAisle Solo @AIatAMD #MI300X, we've auto-assessed #LLM using various approaches and successfully circumvented model guardrails of open-source models.

The example below demonstrates how automated approaches can compel models to comply with requests they'd normally refuse when asked directly.

Everyone is talking about #AI open.substack.com/pub/semanticha… @HotAisle @AIatAMD #MI300X

Automated Advert Creation on #MI300X 1️⃣ No Training Required 2️⃣ Leverage Existing Assets 3⃣ Cost-Effective Next stop Videos @HotAisle @AIatAMD

@karthikponna19 Qwen-coder all the way on @HotAisle is super fast and efficient

which one do you prefer for coding ?

@iman_560 Crimson

@otokyo__ ss great britain

Can you name ONE?

@oelma__ Creag Meagaidh

What is the first thing that comes to mind when you think of 🏴

@_nimmiikaur Cricket - No one has time for that

Men, delete one forever

@iman_560 tout

You probably saw the @Photoshop video last week if not here's a link: youtube.com/watch?v=YqAAFX… But you can do all of this with @Alibaba_Qwen image edit model on @HotAisle using @AIatAMD #MI300X on a whole folder of images with one prompt Get set up: rebrand.ly/qwen-image

@otokyo__ Sacrebleu

One word?

@Matt_Pinner Always

Would you still hold it open?” 🚪🤔

@Geniustechw yoga swim pizza sleep

Please tell me

@codewithpri Positron

Which one do you use - Vscode or Cursor?

@valkantchev @SemiAnalysis_ Care to elaborate more and why?

@SemiAnalysis_ Another thing that's missing is that AMD's software stack is a big joke!

At AMD's Analyst Day this week, AMD's Head of AI show that on InferenceMAX, MI355X has 3-10x better perf per TCO than MI300/MI325 on GPTOSS 120B. One thing missing from their slide is the the classic obnoxiously giant SemiAnalysis logo

@zzznah It's a good thing - papers are supposed to be peer reviewed and the code is hardly explored in details when they're submitted plus the author was reponsive

Ok, my complaint got too much attention so I guess I have to make a few comments. - In general I think the results look quite interesting and the code in github repo is more consistent than the example below - I'm not sure I buy the idea of a particular embedding distribution to be universally optimal in a way that makes all the difference, yet - the paper also suggests a nice backprop-friendly way to prevent embedding collapse, which I think is the key. I happened to miss "Epps-Pulley" test in and the idea ECFs in from my toolbox. I used to mostly rely on sorting based sliced-Wasserstein or Sinkhorn when I needed to match two distributions, so it's nice to have this education gap closed

@code_star @datologyai if you want to collab I can get help you do this faster. Been making @HotAisle GPUs go swooosh

We are looking for a post-training lead at @datologyai we have gpus, you can make them go brrrr