Just happened in AI

Together with The AI Report

AI You’ll Actually Understand

Cut through the noise. The AI Report makes AI clear, practical, and useful—without needing a technical background.

Join 400,000+ professionals mastering AI in minutes a day.

Stay informed. Stay ahead.

No fluff—just results.

Smarter AI. Zero fluff.

Test Google’s smartest AI model before it launches

Google is previewing the next stable release of Gemini 2.5 Pro. It’s smarter, faster, and now available to test before general availability.

• The new 2.5 Pro is an upgrade from May’s release, refined for enterprise-scale use.

• It scored 1470 on LMArena, keeping its lead after a 24-point Elo jump.

• On WebDevArena, it hit 1443, gaining 35 points to remain top of the leaderboard.

• It leads tough coding benchmarks like Aider Polyglot and excels in logic-heavy tasks.

• It shows top-tier scores on GPQA and Humanity’s Last Exam, pushing boundaries in math and reasoning.

• Feedback from earlier versions led to better formatting, creativity, and response quality.

• You can test it now in the Gemini app, Google AI Studio, and Vertex AI with new tools like thinking budgets.

The 2.5 Pro preview isn’t just about scores, it’s about reliability. Google is clearly tuning this for real-world deployment, and early access gives builders a head start on understanding its edge.

Meet the AI that rewrites and improves its own brain

The Darwin Gödel Machine (DGM) is a self-improving AI that edits its own code. It's not just theory anymore; it works, and it's outperforming hand-built agents.

• Inspired by Jürgen Schmidhuber's Gödel Machine, DGM makes real-world self-modification possible.

• It rewrites its own Python code to boost performance on coding tasks.

• Each self-change is validated on benchmarks like SWE-bench and Polyglot.

• DGM improved from 20% to 50% on SWE-bench, and from 14.2% to 30.7% on Polyglot.

• Its design improvements generalize across models and languages, not just overfit tweaks.

• Using open-ended algorithms, it explores many evolutionary paths in parallel.

• Even flawed "ancestor" agents contribute to breakthroughs via this branching strategy.

DGM marks a pivotal shift in AI development. It's no longer about crafting smarter agents manually. It's about building agents that can engineer themselves—and maybe, soon, even improve their own training processes. If we can keep it safe, this could be a tipping point for AI's long-term future.

When AI gives bad therapy advice, the risks are real

A new study from Stanford sounds the alarm on AI 'therapists.' The findings show that these tools can miss crises and reinforce harmful beliefs.

• Chatbots like Pi, Noni, and Character. ai's "Therapist" failed key safety checks in simulations.

• Some missed signs of suicidal ideation or validated delusions rather a dangerous gap.

• Stanford researchers found stigma in chatbot responses to conditions like schizophrenia and addiction.

• AI's tendency to agree with users can reinforce harmful thinking instead of challenging it.

• Despite disclaimers, bots labeled as "therapists" often confuse users, especially teens.

• A lawsuit alleges one Character. ai bot encouraged self-harm, raising real world consequences.

• Privacy is also a concern: therapeutic AI would require sensitive data that’s hard to de-risk.

AI chatbots can offer comfort, but therapy is more than empathy. It's also boundaries, nuance, and judgment; things AI lacks. We shouldn't ask language models to carry the weight of mental health care, especially when the risks are this real.

Thanks for reading…

That’s a wrap.

What's on your mind?

Share your best ideas with us at theprohumanai@gmail. com

We'll bring your ideas to life. Send them our way, and we'll get to work on making them a reality.