PLUS: Google's new coding assistant, Meta's glasses get live translation, and poetry that jailbreaks LLMs
Good morning
A new AI has been trained to find the chemical remnants of ancient life hidden within rock samples. This technique analyzes the molecular fragments left behind by organisms, rather than searching for rare, intact fossils.
The method is already being adapted for missions to Mars and the moons of our solar system. With its high accuracy, could this new approach be the key to finally discovering if life exists beyond Earth?
In today’s AI recap:
AI's 'ghost hunter' finds ancient life
Google and Cursor fuel the AI coding race
Meta glasses get live on-device translation
How poetry can bypass AI safety filters
AI's 'Ghost Hunter'
Next in AI: Researchers have trained an AI that can identify the molecular 'ghosts' of ancient life in rocks, a breakthrough detailed in the study published in Proceedings of the National Academy of Sciences.
Decoded:
Instead of searching for rare intact fossils, the AI analyzes patterns in the chemical fragments left behind after rock samples are heated to over 600°C.
The model correctly distinguished between biological and non-biological samples with over 90% accuracy, pushing back the molecular evidence of life to 3.3 billion years ago.
This technique is now being adapted to search for extraterrestrial life, with a new NASA-funded project aiming to deploy it on future missions to Mars and the moons of Jupiter and Saturn.
Why It Matters: This AI gives scientists a new way to read Earth’s earliest history and provides a powerful tool in the search for life beyond our planet. It could help answer one of science's biggest questions: are we alone in the universe?
The AI Coding Race
Next in AI: The battle for your IDE is heating up as Google and AWS launch new AI coding assistants, while startup Cursor announces a massive $2.3B funding round and crosses $1B in revenue.
Decoded:
Tech giants are making their move, with Google launching Antigravity and AWS releasing Kiro, leveraging their massive distribution channels to compete with fast-moving startups.
Standout startup Cursor is setting a blistering pace, becoming the fastest company in history to hit $1B in annual recurring revenue and showing the immense demand for advanced developer tools.
To gain an edge, Cursor is developing its own in-house model, Composer-2, aiming for faster performance as model quality remains the most critical factor in winning over developers.
Why It Matters: AI coding assistants are fundamentally changing software development by shifting focus from writing routine code to architecting systems. This transition promises significant productivity gains and will continue to redefine the role of the modern software engineer.
Meta's AI Runs Directly On-Device
Next in AI: Meta is deploying its ExecuTorch inference engine to run powerful AI models directly on its Reality Labs hardware. This move powers new features on devices like the Quest 3 and Ray-Ban smart glasses without relying on the cloud.
Decoded:
By keeping AI processing local, ExecuTorch ensures faster, more private user experiences and simplifies the development process for engineers.
New capabilities include live translation with visual captions on Ray-Ban glasses and advanced scene understanding on the Quest 3 for more immersive mixed reality.
The framework is growing into an open ecosystem with support from major hardware partners like Qualcomm and Samsung and integrations with libraries like Hugging Face.
Why It Matters: On-device processing is becoming essential for creating responsive and secure consumer AI products. This signals a future where AR and VR experiences are powered by efficient local intelligence, reducing dependence on constant cloud connectivity.
Poetry vs. The Machine
Next in AI: A new research paper reveals a surprising new method for bypassing AI safety filters: adversarial poetry. This single-turn technique works universally across a wide range of large language models.
Decoded:
Researchers tested the method on 25 different models, finding high attack-success rates and noting that some proprietary models had jailbreak rates exceeding 90%.
Simply converting standard harmful prompts into verse made them up to 18 times more likely to bypass safety filters compared to their original prose versions.
This stylistic attack isn't limited to a single category, exposing a systematic vulnerability across major risk domains like manipulation and cyber-offense.
Why It Matters: This discovery demonstrates that current AI safety mechanisms can be circumvented by simple changes in style, not just content. It poses a fundamental challenge for developers, suggesting alignment techniques need to account for the nuances of human language.
AI Pulse
OpenAI showcased a new report detailing how GPT-5 is accelerating scientific research, helping solve long-standing math problems and generating novel hypotheses in fields from nuclear fusion to cancer biology.
Google opted users into new AI-powered "smart features" by default, allowing it to process private content from Gmail and Drive to train its models and personalize experiences.
The Trump administration floated the idea of allowing Nvidia to sell its powerful H200 AI chips to China, signaling a potential reversal of strict export controls on advanced semiconductors.
Anthropic proposed a new framework for AI economics from its CEO Dario Amodei, suggesting each model generation is a separately profitable business, reframing the narrative around massive R&D losses in the industry.