This New AI Doesn’t Think Like the Others

In partnership with

Hey

Have you noticed how LLMs keep getting more efficient but not necessarily more intelligent?

I mean Claude 4 wasn’t a more intelligent model per se, right? And It’s been bothering me. We discussed it in length in our WhatsApp AI Community.

But at least Google launched something fundamentally different and I’m surprised not many people are talking about it.

And I’m not talking about VEO3.

It’s Gemini Diffusion, and it entirely rethinks how LLMs generate text.

I’ll break it down for non techies –

Instead of the autoregressive approach (predicting one token after another) that's powered every major LLM from GPT to Claude to Llama, Gemini Diffusion borrows techniques from image generation  (what Midjourney or Flux uses) - it refines noise in multiple passes to create entire chunks of text simultaneously.

Why should you care? Three big reasons:

• It's blazingly fast - generating between 1,000-2,000 tokens per second, 4-5x faster than current models

• It uses significantly less memory (no KV caching needed)

• It could make running powerful AI locally much more feasible

This isn't just an incremental improvement - it's potentially a shift that could change the entire industry. The models that dominate in 3 years might look fundamentally different than what we use today.

Speaking of fundamental shifts, let's talk about specialized AI.

Mistral Launched An Impressive Coding Model

Mistral and All Hands AI just released Devstral, a 24B parameter agentic LLM fine-tuned specifically for software engineering. This isn't just another coding model - it's designed from the ground up for exploring codebases, editing multiple files, and powering engineering agents.

It achieves 46.8% performance on SWE-Bench, making it the #1 ranked open model on this benchmark. If you're building coding tools or agents, this is the foundation model you want.

But software isn't the only specialized field getting AI superpowers.

Google Is Transforming Healthcare

Google's MedGemma 4B is transforming healthcare with its multimodal capabilities - handling both medical text and images through a specialized SigLIP image encoder. It outperforms base models across all tested multimodal medical benchmarks.

What I find fascinating is how these specialized models aren't just marginally better in their domains - they're unlocking entirely new possibilities.

And here's where it gets really interesting for entrepreneurs...

There’s Now Code That Improves Itself

OpenEvolve is an open-source implementation inspired by DeepMind's AlphaEvolve that essentially evolves code through LLM-driven iteration.

It orchestrates code generation, evaluation, and selection through evolutionary principles. Imagine applying this to your own AI products - continuously improving algorithms without human intervention.

So what does all this mean for you?

The AI landscape isn't just expanding - it's transforming fundamentally. The tools, techniques, and infrastructure are evolving in ways that could make today's dominant approaches obsolete within months.

For entrepreneurs and AI enthusiasts, this creates both challenges and opportunities:

1. The technical knowledge gap is widening - understanding diffusion vs. autoregressive models, specialized fine-tuning, and evolutionary approaches is becoming crucial

2. Hardware requirements continue to evolve - what worked yesterday might not be sufficient tomorrow

3. The window for innovative applications is wide open - as these models become more capable and efficient, entirely new use cases become possible

I'm particularly excited about this last point. As AI becomes more capable, specialized, and efficient, the barrier to entry for creating transformative applications continues to drop.

What are you building with these new tools? Or what would you build if the technical barriers were removed?

Hit reply and let me know - I'm genuinely curious about where you see the opportunities in this rapidly evolving landscape.

Until next time,

AP

Learn AI in 5 minutes a day

This is the easiest way for a busy person wanting to learn AI in as little time as possible:

1. Sign up for The Rundown AI newsletter

2. They send you 5-minute email updates on the latest AI news and how to use it

3. You learn how to become 2x more productive by leveraging AI

Sign up to start learning.