Ron Green

🎧 Tune in and share your thoughts:

www.youtube.com/watch?v=J2Pd...

We covered:
🧬 Why evolutionary methods can outperform supervised and RL
🌱 How populations enable broader, more creative exploration
💡 The future of AI design through surrogate modeling and neural architecture search
📕 His coming book Neuroevolution: Harnessing Creativity in AI Model Design

Risto’s work in neural networks and evolutionary computation has shaped major parts of the field. His groundbreaking work in neuroevolution introduced powerful methods for evolving neural network architectures and has influenced everything from game AI to large-scale industrial optimization.

I recently had the incredible opportunity to chat with Risto Miikkulainen, a true pioneer and one of the sharpest minds in AI.

Introducing RFdiffusion2 - Institute for Protein Design A new deep-learning model developed by scientists at the IPD and MIT is redefining what’s possible in enzyme design.

Check out the announcement for more details:

www.ipd.uw.edu/2025/04/intr...

I feel beyond lucky to be alive and working in AI at time when ideas we once only imagined are now becoming real.

What’s new in RFdiffusion2:
• Enables precise design of proteins that interact with a wide range of molecules.
• Supports binding to targets like heme and therapeutic compounds.
• Generates a broader variety of complex protein structures with higher accuracy.

RFdiffusion2 just launched, and it’s a major leap in protein design. It uses all-atom modeling to create proteins that bind to small molecules, metals, and nucleic acids—unlocking new possibilities for drug discovery, diagnostics, and synthetic biology.

We dreamed of a future where biology and computation would come together to transform medicine. That dream is becoming real.

Most folks know I’m passionate about AI, but my roots in computational biology run deep. In the 90s, I worked at a Toronto-based startup called Visible Genetics, building DNA sequencing technology with an amazing team.

If you’re passionate about applied AI and want to help shape the future of how we build it, please consider joining us.

jobs.lever.co/kungfu/7b03b...

KUNGFU.AI is looking for a Director of Engineering to join our team of world-class AI experts. This is a hands-on leadership role—we’re looking for an expert who can lead other experts. Someone who can stay close to the code, and help our team build state-of-the-art AI solutions.

We’re hiring a rare kind of engineering leader 🦄

CaMeL offers a promising new direction for mitigating prompt injection attacks In the two and a half years that we’ve been talking about prompt injection attacks I’ve seen alarmingly little progress towards a robust solution. The new paper Defeating Prompt Injections …

Some big news in AI security. DeepMind's new system, CaMeL, tackles LLM prompt injection with a smart, principled design that favors structured control over stacking more AI.

Check out @simonwillison.net excellent write-up on the new technique:

simonwillison.net/2025/Apr/11/...

Tracing the thoughts of a large language model Anthropic's latest interpretability research: a new microscope to understand Claude's internal mechanisms

Anthropic's work represents a major step forward in model interpretability.

This reveals something critical. LLMs can sound articulate and confident, but their explanations are often just another piece of generated text. They aren't windows into the model's actual thinking. They're stories. Useful, sometimes accurate, but ultimately disconnected from internal computation.

It often gets the right answer, but not by "doing the math" in the traditional sense. When asked to explain its reasoning, the model generates plausible explanations that sound good but don't actually reflect what happened under the hood.

And maybe most surprising, the model doesn't perform math the way we might expect. Instead of following a logical, step-by-step process, it uses heuristics and pattern recognition—approaches more like estimation, memory, or rule-of-thumb reasoning.

This challenges the common assumption that LLMs operate purely one word at a time, without foresight.

They also found clear evidence of forward planning. This behavior is especially noticeable in tasks like writing poetry, where the model often anticipates which words will need to rhyme before it even begins forming the sentence.

The model then reasons over that representation and finally converts the result back into the target spoken language. This points to an abstract, language-agnostic layer of thought, suggesting the model's understanding isn't tied to any specific human language.

First, the model seems to think in a kind of universal internal language. When it receives a prompt in a spoken language such as English, French, or Chinese, it first translates the input into this shared internal representation.

To tackle this, Anthropic developed new techniques that let researchers peer inside the model as it generates text. What they found is both fascinating and important.

LLMs have become so large and complex that it's difficult to fully grasp how they work. This might seem surprising—we built them after all! But their massive scale makes it hard to trace or interpret why a model produces a specific output or how it "reasons" through a problem.

Anthropic released two groundbreaking research papers that offer rare insights into how LLMs actually operate. These findings pull back the curtain on the inner workings of LLMs, revealing surprising details about how they process language, reason through problems, and generate responses.

🤔 Ever wondered how tools like ChatGPT work—but felt overwhelmed by technical jargon?

Grant Sanderson, the brilliant mind behind 3Blue1Brown, just released a fantastic video that makes it all crystal clear.

Check out the video here → www.youtube.com/watch?v=LPZh...

#AI #MachineLearning