harisec

Trying to use it via openrouter: "The model grok-4.5 is not available in your region."

This: User just tricked Grok and Bankrbot to send tokens with Morse code

https://www.cryptopolitan.com/user-tricked-grok-bankrbot-to-...

Unfortunatelly, i have the same experience.

$44M/year?

https://www.usaspending.gov/award/CONT_AWD_70RCSJ23FR0000015...

Yes, during training multiple checkpoints are created, you can distill from any checkpoint you want.

Anybody can try Grok3 on Chatbot Arena (even if you are in Europe). Select Direct Chat and select the model early-grok-3. https://lmarena.ai/

It doesn’t matter much how many users Bluesky is gaining, it matters how many of them will be using Bluesky in a few months. We will see.

Congrats, good luck with your new company!

I have one question regarding your ARC Prize competition: The current leader from the leaderboard (MindsAI) seems not to be following the original intention of the competition (fine tune a model with millions of tasks similar with the ARC tasks). IMO this is against the goal/intention of the competition, the goal being to find a novel way to get neural networks to generalize from a few samples. You can solve almost anything by brute-forcing it (fine tunning on millions of samples). If you agree with me, why is the MindsAI solution accepted?

Recraft’s image generation service could leak its internal system prompts due to its unique architecture combining Claude (an AI language model) with a diffusion model. Unlike other image generators, Recraft could perform calculations and answer questions, which led to the discovery that carefully crafted prompts could expose the system’s internal instructions.

Actually, i think it will take less than 2 years. I've been using Aider + Claude 3.5 Sonnet almost daily for a long time and the progress is very fast. We will see.

That's a good point and I agree with you. However, would you agree that in a few years we will need far less developers than we need right now?

These are toys but in 2 years they will probably be full projects and 2 years later people will ask "why do i need a software developer?"

If you want to really get depressed about the future of software developers try aider.

I never had this problem but i guess it depends on the prompt.

Qwen 2.5 models are better than Llama and Mistral.

There are no "LLM-generated images". These images are generated using diffusion models. LLMs are very different from diffusion models.

https://en.m.wikipedia.org/wiki/Diffusion_model

I asked a few “hard” questions and compared o1 with claude. https://github.com/harisec/o1-vs-claude

If you use DeepSeek Coder V2 0724 (that is #2 after Claude 3.5 Sonnet on the Aider leaderboard), the costs are very, very small. https://aider.chat/2024/07/25/new-models.html

aider is great, i also use it almost daily. thanks for writing it Paul!

I agree, I suspect the HuggingFace dataset that I've used is not that randomly distributed and it mostly contains prompts related with those themes. How it works is that I randomly select 5 random prompts from the dataset and use those prompts as seed for new prompts. The complete DeepSeek prompt to generate new prompts can be found here: https://github.com/harisec/llm-dreams

All images in the LLM Dreams gallery are generated using the Flux model from the Black Forest Team, powered by the fal.ai API. The prompts for these images were brainstormed by DeepSeek, drawing inspiration from existing Stable Diffusion prompts found in the Gustavosta/Stable-Diffusion-Prompts dataset on Hugging Face. All the code was written by Claude.

The article is wrong, CVE-2022-21587 is related to Oracle not Solaris

Cox has a vulnerability disclosure program. https://www.cox.com/aboutus/policies/cox-security-responsibl...

You still don't know the exact date of the potential acquisition. So you cannot make a successful trade. I would say it's not insider trading.

Researchers from Carnegie Mellon University found that it's possible to automatically construct adversarial attacks on LLMs, forcing them to answer any questions and it's possible to generated unlimited number of such attacks, making them very hard to protect against.

Same here, this is ridiculous

There are two podcast episodes where David Pogue explains in a lot of detail how the whole system works. Kind of interesting. https://unsungscience.com/news/back-to-titanic-part-1/

Claude is available for free in the Poe app (poe.com). I think it's good and underappreciated.

Vicuna-13b is based on LLAMA that was millions to train. $300 is just to finetuning.

I think OpenAI has a few hidden advantages that are not obvious at this point. It could be additional training data, filtering/preprocessing that data, some changes to the architecture, who knows? None of the open source models are even close to GPT 3.5, what to speak about GPT 4? I've tried everything and the 60G llama variants so i'm not sure it's about number of parameters. They definitely have some hidden sauce.