🚀 NEW most Powerful open-source AI model

Also: US, Britain and EU to sign historic AI treaty

Welcome, AI enthusiasts

A rather unknown player in the field, HyperWrite, has released Reflection 70B, an open-source AI model with a new self-correction feature that tops the benchmarks for open-source models. At the same time, global powers are about to sign the first legally binding AI treaty. This treaty aims to manage AI risks while promoting innovation. In biotech, AlphaProteo has made a breakthrough in protein design. They've developed new protein binders that could transform research in biology and healthcare. Let’s dive in!

In today’s insights:

  • New Most Powerful Open-Source AI Model

  • US, Britain and EU to sign historic AI treaty

  • AlphaProteo reshapes protein design with AI

Read time: 3 minutes

🗞️ LATEST DEVELOPMENTS

Evolving AI: The AI model has been trained with “Reflection-Tuning,” which is designed to help it learn from its mistakes and fix them.

Key Points:

  • HyperWrite unveiled Reflection 70B, built on Meta’s Llama 3.1 70B Instruct.

  • The model uses a unique error-correction technique called reflection tuning.

  • Benchmark tests show Reflection 70B outperforms other open-source models.

  • More powerful model, Reflection 405B, is coming soon.

Details:

HyperWrite has launched Reflection 70B, a new AI model built on Meta’s Llama 3.1 70B Instruct. Reflection 70B leverages a novel error self-correction technique that allows it to detect and fix mistakes during inference — the time when the model is processing inputs and outputs — setting a new benchmark for open-source AI. Reflection Llama-3.1 70B can “hold its own” against even the top closed-source models such as Anthropic’s Claude 3.5 Sonnet, and OpenAI’s GPT-4o in several benchmarks. Available via Hugging Face, it’s the most powerful open-source AI today, with a larger model, Reflection 405B, set for release next week.

Here’s an example of it’s reflection at play

Why It Matters:

A team of two just casually beating everyone around them, and they're poised to release their 405B model next week, which they claim will be the best in the world. What timeline are we on? Trained using synthetic data from Glaive, this model demonstrates enhanced reasoning across various tasks. Built on top of Meta’s Llama 3.1 70B, it’s a clear example of what can be achieved when open-source innovation takes the lead in AI.

Source: Council of Europe

Evolving AI: The first legally binding AI treaty is set to be signed by key global powers. The agreement aims to address AI risks while promoting innovation.

Key Points:

  • Countries like the US, EU members, and Britain will sign a groundbreaking AI treaty.

  • The treaty focuses on protecting human rights in the AI space.

  • The convention differs from the EU AI Act and faces criticism over its enforceability.

Details:

The new framework, which can be found here, was agreed by the Council of Europe and commits those that have signed to take “collective action” to manage AI products and protect the public from potential misuse. The Council of Europe spearheaded the AI Convention to mitigate the risks AI poses, while encouraging responsible development. Although 57 countries contributed to its formation, some experts argue the treaty’s broad principles may lack legal clarity. It stands apart from the EU's AI Act, which imposes stricter regulations on AI systems in Europe. Legal expert Francesca Fanucci has expressed concerns over the treaty’s enforceability, questioning if its provisions hold substantial legal weight.

Why It Matters:

It's difficult to criticize the intentions behind the Convention, and the ten countries and organizations that have already signed on deserve praise for taking some action. However, the real challenge lies in the specifics and how they will be implemented, which is still uncertain. What many fail to realize is that the genie is already out of the bottle and such, so it’s hard to see how you backpedal. Some AI companies that are developing AI models that have the potential to be worth more than some countries’ GDP - and they aren’t afraid of a legal fight or a fine, if they win in the long run.

Evolving AI: AlphaProteo introduces a breakthrough in protein design, generating novel protein binders with high binding strength.

Key Points:

  • AlphaProteo creates novel protein binders for critical research.

  • Achieves superior binding success rates across diverse proteins.

  • Validated by experimental labs for real-world effectiveness.

Details:

AlphaProteo is a new AI tool that creates custom proteins designed to bind to specific targets, like proteins linked to diseases. By using vast amounts of protein data, AlphaProteo can predict how these binders will connect to their targets, often more effectively than current methods. For example, it successfully created binders for proteins related to cancer and COVID-19, showing much stronger binding than traditional approaches. Researchers have tested and confirmed these results, proving that AlphaProteo could greatly speed up the discovery of new treatments and tools in medical research.

Why It Matters:

Alright, sounds cool. So why does this matter indeed? Finding protein binders that can bind well to a protein allows you to modify its function. Make it block interactions, improve a protein's ability to interact with another protein, activate or disable it, etc. Essentially, it's a new convenient way to manipulate the activity of proteins, saving you a lot of resources and weeks of work to identify good protein binders. For example, if a protein is causing a disease, they can design another protein to block it and stop the disease from getting worse — meaning this is quite revolutionary.

💡 Tip of the Day

The Reply AI Film Festival is an international competition showcasing AI in creative storytelling, aimed at inspiring young filmmakers. The 2024 edition has just wrapped up, and the winners have been announced. Watch the top three entries below — the results are impressive.

Source: Reply AI

🎯SNAPSHOTS

Direct links to relevant AI articles.

🔥 Ex-OpenAI scientist Ilya Sutskever's new AI company raises $1 billion for superintelligence.

🗣️ Bill Gates downplays AI’s potential for spreading disinformation.

🚀 OpenAI reached 1M business users.

💰 North Dakota leads race for $125b mega data centers.

🔎 Google Photos rolls out AI-powered search feature.

📈Trending AI Tools

  • 🤖 Hyperwrite - Personal AI writing assistant (link)

  • 🚀 Profit Leap - Make AI-powered dashboards (link)

  • 🎬 Zebracat - A tool to create videos from prompts (link)

  • 🍋🟩 PipeLime - AI sales funnel automation (link)

  • 🎙️ LaunchPod - A tool for audio content creation with AI voices (link)

Reply

or to participate.