March “Modness” + AI Model Roundup

Dr. Ryan Ries here with your weekly Mission Matrix. We haven’t chatted about the latest models lately, and in the past week or so, multiple new models have been announced. 

Let's dive into what's happening and why it matters.

Quick side note - I’m co-hosting another Gen AI Ask Us Anything next week with Jonathan LaCour on 3/27. Come talk to us about new models or anything else in the AI space!

Introducing Gemma 3

Google just dropped Gemma 3 on March 12th, and it's making some waves. This isn't just an incremental update. Gemma 3 offers multimodality support, 128K token context windows, and compatibility with 140+ languages. Available in four sizes (1B, 4B, 12B, and 27B parameters), it's topping the LMArena benchmark with a score of 1338.

What's really fascinating is their multi-pronged approach to improvement - they're combining distillation, reinforcement learning, and model merging (which we’ve talked about here in the past) to enhance performance across math, coding, and instruction following.

Gemini Deep Research Gets a Major Upgrade

Gemini rolled out Deep Research at the end of 2024, but just a few days ago, it was upgraded and made available for anyone to try. 

From my testing, Deep Research’s capabilities are changing the research game entirely. I spent some time with this last night, setting it to compare several models, and the results blew me away. It pulled information from 38 sources while reviewing over 78 sources during its research process!

It even built a table to compare the key characteristics of each of the models:

What makes Deep Research special is how it works with Flash Thinking Experimental to handle the entire research workflow—from planning and searching to reasoning, analyzing, and reporting. It creates detailed, multi-page reports and shows its thought process in real-time as it browses the web.

The upgrade also brings a 1M token context window for Advanced users, meaning you can analyze massive amounts of information in one go. 

Plus, it now integrates with Google apps and services, making it even more powerful for organizing research findings into sheets, docs, and tables.

Now, let’s just hope Google puts some time and attention into upgrading its Gemini AI features that are now embedded throughout the Google Workspace. This has some amazing potential, but the results have been lackluster, in my opinion.

FoxBrain: Taiwan Enters the Game

Foxconn (yes, the iPhone manufacturer) just launched their first LLM called "FoxBrain" on March 10th. Built on Meta's Llama 3.1 architecture, they're focusing on manufacturing and supply chain optimization - perfectly aligned with their core business.

They trained this on Taiwan's largest supercomputer using 120 Nvidia H100 GPUs in just four weeks. What's particularly interesting is their optimization for traditional Chinese and Taiwanese language styles, showing how regional AI development continues to accelerate.

Manus & OWL

The most highly talked about space right now is AI agents - systems that can execute complex tasks autonomously.

Manus AI from Chinese startup Monica is generating significant buzz. They're leveraging Claude 3.5 Sonnet v1 alongside fine-tuned Qwen models in a multi-agent architecture. Their system handles everything from travel planning to financial analysis with minimal human intervention.

Now, an open-source alternative called OWL emerged on March 7th and has already garnered 6,000+ GitHub stars. OWL ranks #1 on the GAIA Benchmark among open-source frameworks and runs locally for complete privacy.

This tension between proprietary and open-source agents is creating an interesting dynamic in the market.

What This Tells Us

1. Multi-agent architectures are winning - Breaking complex problems into specialized agent roles is proving effective across implementations
2. Vision-language integration is standard - If you're not thinking multimodally, you're missing out
3. Reinforcement learning techniques are diversifying - Beyond RLHF, we're seeing RLMF (for math) and RLEF (for coding)
4. Context length is expanding dramatically - 128K is becoming the new benchmark, enabling document analysis at an unprecedented scale
5. Local deployment is gaining traction - Privacy concerns and cost optimization are driving more local deployment options

My team and I are closely monitoring the latest models and I will say: new models are being announced daily. Don’t overthink this and feel that you’re missing out by not exploring one over the other. Stick with what you know, especially if you’re thinking about launching production workloads. 

The pace of innovation isn't slowing down - if anything, it's accelerating as more players enter the space with specialized capabilities.

Until next time,
Ryan Ries

Now, here's our weekly AI-generated image & the prompt I used. 

"Generate an image of Meta, Manus AI, Open AI, and Foxconn AI models in muppet form, all fighting in a ring."

Sign up for Ryan's weekly newsletter to get early access to his content every Wednesday.