Google quietly launches a top AI model

Hi pals! Here’s what you need to know about AI today:

👉 A new open-source American AI model beats DeepSeek-V3

👉 Google launches Gemini 2.0 Pro Experimental

👉 OpenAI goes nuclear

and many more!

📧 Did someone forward you this email? Subscribe here for free to get the latest AI news everyday!

Read time: 4.9 minutes

What’s going on: Ai2, a nonprofit AI research institute based in Seattle, has unveiled its latest AI model, Tulu 3 405B, claiming it outperforms one of DeepSeek's top models, DeepSeek V3, along with OpenAI's GPT-4o on certain benchmarks. Tulu 3 405B is open-source, allowing developers worldwide to access and replicate the model using freely available components.

What does it mean: This model marks a significant step forward for Ai2, showing off not only its competitive edge in the AI field but also emphasizing the open-source nature of the project. This announcement also hints that American open-source models are not falling behind their Chinese rivals and the DeepSeek’s newfound success might not last for a long time.

More details:

• Tulu 3 405B is a very large model. It contains 405 billion parameters and requires 256 GPUs running in parallel to train.

• Ai2's impressive performance is due to a combination of meticulous data curation, targeted supervised finetuning, preference optimization, a novel reinforcement learning method called RLVR, and a comprehensive evaluation process. (Read this article for more technical details)

• On PopQA benchmark, a set of 14,000 specialized knowledge questions sourced from Wikipedia, this model beat not only DeepSeek V3 and GPT-4o, but also Meta’s Llama 3.1 405B.

• If you are interested in testing Tulu 3 405B, you can visit Ai2’s chatbot web app.

What’s going on: Google has quietly launched its next-generation flagship AI model, Gemini 2.0 Pro Experimental, not through a grand announcement but via a changelog update for its Gemini chatbot app. The introduction of this model comes at a time when the tech world is obsessively observing the progress of controversial Chinese AI startup, DeepSeek. However, some Twitter users report that the changelog has been updated to remove mention of Gemini 2.0 Pro Experimental.

What does it mean: Google is a mature company. We owe today’s AI advancements to the research done originally by Google engineers, when they published their “Transformers“ architecture back in 2017 that later on enabled the rise of modern large language models. Google's approach with the non-dramatic launch of Gemini 2.0 Pro Experimental follows the same strategy of iterative improvement rather than creating hype.

More details:

• According to Google, Gemini 2.0 Pro Experimental is now the leading model in Google’s Gemini AI family.

• This model is supposed to excel at coding problems (generating a specific program from scratch) and solving mathematical problems, like developing complex statistical models or quantum algorithms.

• Gemini 2.0 Pro Experimental doesn’t have access to real-time information.

• Google has not published any benchmark results of this model. So, we don’t actually know how good the model is and whether it beats DeepSeek’s latest models or not.

👋 Do you enjoy reading our newsletter? Let’s spread the joy by sharing the link below with your friends. You can also forward this email to them! Thanks!
the79newsletter.com/subscribe

🤖 India plans to host DeepSeek's models on domestic servers. This is surprising because India has banned more than 300 apps and services linked to China since 2020.

⚛ OpenAI has agreed to provide its AI technology to US national labs for research into nuclear weapons safety and stockpile management.

📞 Google's new "Ask for Me" feature within Search Labs uses AI to call local businesses on behalf of users, inquiring about services, pricing, and availability.

🚀 Mistral AI, a European AI company, has released Mistral Small 3, a 24 billion parameter open-source model that offers high performance, speed, and affordability at a fraction of size. This model is a competitive alternative to larger models like Llama 3.3 70B and Qwen 32B.

💰 OpenAI is reportedly in discussions to raise up to $40 billion in a funding round that would value the company at $340 billion, with SoftBank expected to invest between $15 billion to $25 billion.

👨‍💻 Forbes has reported that a sophisticated AI-driven hack targeting 2.5 billion users of Gmail has been confirmed. Gmail is urging users to take immediate precautions against these advanced phishing attempts. (Read Gmail’s guide on phishing attacks)

AI + brainstorming

Tubi - Staff Machine Learning Engineer, Personalization

Roblox - Senior Data Scientist - Social Communications

NBCUniversal - AI Cyber Defense Analyst

Jerry - Principal Machine Learning Engineer (PhD)

Thank you for staying with us like always! If you are not subscribed, subscribe here for free to get more of these emails in your inbox! Cheers!