Dear Readers,

Sometimes progress is not reflected in spectacular headlines, but in the small adjustments that make entire systems reliable. A current example: Researchers have found a way to ensure that AI models always give the same answer for the same input—regardless of server load or calculation sequence. It sounds technical, but it's a game changer for science, regulation, and trust in AI: step by step, a “magic tool” is becoming a precise technical system.

This issue focuses on precisely such turning points: from the first leaks about Gemini 3 to new cells that slow down aging in primates to the surprising finding that additional protein intake in old age has no effect on muscle building. There are also fresh insights into the debate surrounding deterministic AI, exciting data on creatine and its effect on the brain—and how DeepMind is rethinking science with “science agents.” It's a colorful mix that shows that the future does not unfold in a linear fashion, but in many small steps that together change the big picture.

In Today’s Issue:

🧠 New research: big brain boosts from creatine

🤖 Self-improving robots by 2040? New forecast

🚀 Finally MCP Support launched in ChatGPT

✨ And more AI goodness…

All the best,

In developer mode, developers can create connectors and use them in chat for write actions (not just search/fetch). Update Jira tickets, trigger Zapier workflows, or combine connectors for complex automations.

The Takeaway

👉 Same input ⇒ truly identical output in the future – regardless of server load

👉 Researchers have developed “batch-invariant” computing methods that guarantee this stability

👉 This results in: exact reproducibility of experiments, more stable AI training, secure applications in critical areas

👉 A technical but crucial step: AI is becoming more predictable and reliable

If you ask ChatGPT the same question several times today, you will often get slightly different answers—even if the AI is actually set up to always say the same thing. This is not because the machine has “mood swings,” but because of the way the data centers in the background process multiple requests simultaneously. Tiny differences in the order of the calculation steps sometimes result in different words coming out at the end.

A team of researchers has now shown that this problem can be solved technically: They have developed new methods that ensure that the AI always calculates exactly the same way – no matter how many people ask questions in parallel. In practice, this means that experiments can be repeated exactly, AI training becomes more stable and reliable, and applications in sensitive areas such as medicine or finance become significantly more secure. This is not a “major breakthrough” in AI thinking, but it is a decisive step forward in terms of its reliability – and thus a foundation on which much more can be built.

Why it matters: AI research is not just about larger models and more data, but also about the “invisible infrastructure.” Reproducibility is the foundation of all science—without it, much remains a matter of chance. With deterministic inference, AI transforms from a “magical” tool into a reliable technical system that can be seriously tested, improved, and regulated.

Sources:

🔗 https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/

Replit’s Agent 3 Now able to handle 200 minutes continuous coding

Researchers at the Chinese Academy of Sciences have developed senescence-resistant precursor cells (SRCs) and tested them on older macaques. Over a period of 44 weeks, there were no signs of tumor risk, but instead broad rejuvenating effects in more than ten body systems: less cell aging, reduced inflammation, and improved regeneration of nerves and reproductive organs. Molecular analyses confirmed more stable genomes and a decrease in biological age by several years. Particularly exciting: even the exosomes released by SRCs alone were able to significantly slow down signs of aging in mice.