Okay, something massive just happened and almost nobody noticed it because while we've been freaking out about the government locking down access to all the best AI and open source AI learned how to build its own brain and it's almost as good as Cloud and it was built in the United States.

@gigaqian

140.7K followers · usually 20.6K views a post

Views

112.9K

Likes

8.7K

Comments

307

1.3K

Selling · SaaS › AI Tool

Ornith 1.0seen in 3 videos →

Why it worked

The video effectively highlights a significant advancement in open-source AI by a US company, emphasizing its competitive performance and innovative self-evolving capabilities. This appeals to a tech-savvy audience interested in AI breakthroughs and the potential for decentralized, efficient AI development.

Summary

A new open-source AI model called Ornith-1.0, developed by DeepReinforce in the US, has been released and performs competitively on key benchmarks, even outperforming larger models. Unlike traditional AI, Ornith-1.0 uses a novel training method where it generates and evolves its own workflows, making it smarter and more efficient.

Structure

1Introduction of a new, powerful, open-source AI from the US.
2Explanation of how traditional AI works vs. Ornith-1.0's self-evolving approach.
3Demonstration of Ornith-1.0's competitive performance on benchmarks.
4Highlighting its small size, MIT license, and lack of proprietary restrictions.
5Concluding that this is a significant advancement for US open-source AI.

Product placement

Ornith-1.0: Acts - an open-source LLM specialized for agentic coding. It generates its own workflow scaffolding, tests it, and evolves it. Its performance is evaluated on benchmarks. Removing it would change what happens in the video. DeepReinforce: Acts - the company that developed Ornith-1.0. Removing it would change what happens in the video. SWE benchmark verified: Acts - a benchmark used to evaluate AI models. Removing it would change what happens in the video. Open AI GPT-4.7: Acts - a benchmark used to evaluate AI models. Removing it would change what happens in the video. Open AI GPT-4.8: Acts - a benchmark used to evaluate AI models. Removing it would change what happens in the video. Claude 2.1: Acts - a benchmark used to evaluate AI models. Removing it would change what happens in the video. Claude Opus 4.7: Acts - a benchmark used to evaluate AI models. Removing it would change what happens in the video. Claude Opus 4.8: Acts - a benchmark used to evaluate AI models. Removing it would change what happens in the video. MIT license: Appears - a license under which Ornith-1.0 is released. Removing it would not change what happens in the video.

Original caption

The United States finally has an open source AI model that leads on key benchmarks! Ornith 1.0 from DeepReinforce actually builds its own workflows using a two-stage learnable pipeline at the orchestration layer! It automatically builds its own workflow scaffolding. This is a big moment for open source AI! Let’s explore how it works 😄 #ainews #newai #LearnOnTikTok #opensource #gigaqian

More from @gigaqian

Their other posts in the index, biggest breakout first.

Transcript

Okay, something massive just happened and almost nobody noticed it because while we've been freaking out about the government locking down access to all the best AI and open source AI learned how to build its own brain and it's almost as good as Cloud and it was built in the United States. I'm Gigan and if you've been following me, you know I've been shouting that the US needs to build better open source AI and a company called deep reinforcement just did it and I'm going to tell you exactly how it works. The AI models you use for code have human-written plays. We've been making the decisions. Check the repo, write the tests, debug. It's rigid, it's slow and a lot of creators on social media tell you about exciting new approaches that you do not need. But Ornif 1.0 doesn't use a human playbook. It writes its own. It generates its own workflow scaffolding, tests it and then evolves it because there's a difference between learning to code something and learning how it should be coded. If you don't quite see the difference, that's fine. Let's down some. Normal AI relies on a static orchestration layer. That's a fixed master program that forces the AI to follow a permanent checklist. It doesn't change. So it's obviously not going to skip unnecessary steps. Let alone invent new ones. Or if replaces this layer with a two-stage viable pipeline. So it gets graded and it improves on two things at the same time. Its strategy and its execution. So let's get to the point. This small open source model is competing with the goats on the benchmarks. And if you check out SWE benchmark verified, it's better than OS 4.7 and it's up there in the conversation with 4.8. Oh, and there's the 9B version, which for those of you who don't know, means it can fit on your laptop and it scores almost 70%. That means it's outperforming models 40 times its size. MIT license, no API key, no rate limits, and no sending your proprietary code to someone else's servers. This is legit innovation happening in open source. in the US. So, does this mean the US just took the lead in open source AI from China? I wouldn't say that just yet. But as of now, an open source US model is competitive on the benchmarks we really care about. And it did it with an innovative new training method. It's genuinely smarter and that's part of what's so cool about open source. We can learn from each other. So let's see where this goes next.