Mihai Chirculescu

Oh, wow, congratulations! 🤯 I admire your determination.

‘Yes, I am a human’: bot detection is no longer working – and just wait until AI agents come along Designed in the early 2000s to stop rogue bots, we could at least justify the infuriation when they did their job properly.

With LLMs becoming mainstream, CAPTCHAs are quickly becoming obsolete. AI agents from big tech, designed to book trips and buy tickets, will seal their fate.

Finding a durable bot detection system that can outsmart advanced AI while remaining user-friendly for humans seems increasingly unlikely.

Ovis1.6-Gemma2-27B was just released.

I used the 9B version in the past for document understanding and it was similar to Qwen2 VL 7B, so this one should be better (it's also explicitly mentioned "Enhanced Document Understanding" as a key feature).

huggingface.co/AIDC-AI/Ovis1.6-Gemma2-27B

Can I please be added? :)

🙋‍♂️

Alibaba Group just released Marco-o1, their open source alternative to OpenAI's o1.

Github: github.com/AIDC-AI/Marc...
Paper: arxiv.org/abs/2411.14405
HF: huggingface.co/AIDC-AI/Marc...

🙋‍♂️

I gained a newfound respect for so-called 'prompt engineers'. Trying to approach the LLM with an engineer's mindset (impacting attention outputs, adding steering vectors etc.) feels like the wrong approach.

Each token I put in the prompt matters and I now focus more on that.

Open-source LLMs Join the conversation

We've made a starter pack for researchers/organizations working on open-source LLMS.

Please let us know if we missed you or if you'd like to be added!

go.bsky.app/FELkyDr

🙋‍♂️

Yes, and there's also the human limitation in the equation, which will remain there. I used to buy into the hype about "it's a paradigm shift: instead of the human learning how the business app works, the app will have to learn what the human wants". It an attractive idea, but not realistic.

"Agentic AI" seems like pure hype to me.

Even when it will be functional and production ready, for real world tasks, RPA solutions will still be superior because: 1. Speed of execution and 2. Clarity of what the human desires and better control.

ELI5 why I'm wrong

🙋‍♂️

GitHub - Mihaiii/backtrack_sampler: An easy-to-understand framework for LLM samplers that rewind and revise generated tokens An easy-to-understand framework for LLM samplers that rewind and revise generated tokens - Mihaiii/backtrack_sampler

Thanks! This repo is similar to backtrack_sampler, except it can't backtrack (although backtrack_sampler can do everything that is done here by just ignoring the backtrack functionality).

github.com/Mihaiii/back...

🙋‍♂️

I was on X for the AI community, but lately there were less relevant posts in my feed so I decided to give this app another chance.

Dare is also in the house after a very long break here

Sweet! I'm glad to see you here too!

How are they preventing bots? Or the API allows only read?

A very cool read with lots of potential market value. I suspect OpenAI is using something like this internally for their plugins.

https://github.com/newhouseb/clownfish

And then is the algo that makes recommendations :) https://github.com/bluesky-social/social-app/blob/main/src/lib/constants.ts#L78

Hmm, indeed, neither GPT4 doesn't output the correct answer. I even tried some prompting, but still no success.

In my exp, GPT4 is way better than ChatGPT. Let me know the exact prompt you're using and I'll try it out. Bard is also not available in Romania :|

We now have 3 ggml RedPajama models. They are quantized and some have ~2GB.
HF: https://huggingface.co/keldenl

Github: https://github.com/theislab/single-cell-best-practices

Mass-editing thousands of facts into a transformer memory

Site: https://memit.baulab.info/
Github: https://github.com/kmeng01/memit
Image soutce: Nomic AI discord

I'm going to try to use GPT4 to help me navigate and understand https://github.com/lucidrains/x-transformers .

This is going to take some time, but in theory is should be better than any internet course.

Please don't upload my code on GitHub This is a call to open source developers to not upload the work of others on GitHub.

Over time we’re going to see more people who were once happy to share their work online look for ways to opt out of their writings, code or art being used to train AI.

This is on track to becoming the biggest threat the Open Web has faced in years.

GPT4ALL is about to announce a ggml adaptation for mpt7b!

The images are form Nomic AI discord.