Dear curious mind,
What a time to be alive! From robots driving themselves on Mars to the math that might explain our own souls, here are 10 recent insights worth sharing from the world of AI research:
#1
🧠 Remote Control for AI Brains
Researchers found a way to tweak specific concepts inside AI models using surprisingly simple math. They can mathematically turn up things like mood or turn down refusal to make an AI bypass its own safety rules, which is basically jailbreaking the system. For now, this technique works in English, Chinese, and Hindi, indicating that these models know way more than they let on in their usual responses.
#2
☀️ Sun-Proofing Our Food
Scientists are using Google DeepMind's AlphaFold to build hybrid enzymes that help crops survive extreme heat. They found that certain plant parts wobble and break when it gets too hot, so they swapped them with rigid pieces from algae that live in volcanic springs. This AI-designed enzyme stayed stable even at a scorching 65°C, which could safeguard our food supply as the planet warms..
#3
🤔 Can AI Out-Think Humans?
Scientists tested 100,000 people against AI to see who’s more creative at combining unrelated ideas. While GPT-4 beat the average human on basic word tasks, the most creative people still have a massive lead that AI hasn't touched yet. They also found that cranking up the heat (a setting called temperature) makes the AI less predictable and way more original
#4
🤖 The Mystery of AI Minds
Some think that understanding if AI is actually awake or just really good at faking it is one of the biggest scientific challenges of the century. Researchers are moving from just looking at brain patterns to testing big theories, like whether consciousness comes from broadcasting information or complex mathematical integration. They believe solving this could change everything from how we treat animals to how we design legal systems for robots
#5
🕵️♂️ AI Secret Agents are Risky
A new study of 30 top AI bots found that most of them are missing basic safety labels and risk reports. Web browser agents, which can actually click things and fill out forms for you, are the least transparent about their security. These bots often hide the fact that they are AI, making it impossible for website owners to tell if they are dealing with a human or a script.
#6
🚀AI Takes the Wheel on Mars
NASA's Perseverance rover just completed its first drives planned entirely by generative AI. Because it takes forever to send signals to Mars, the rover used vision models to spot dangerous rocks and ripples on its own. It successfully navigated hundreds of feet of Jezero Crater without a human driver steering from Earth — a breakthrough that could change how we explore other worlds entirely.
#7
👾 Computers That Think Like Brains
Researchers built neuromorphic computers that mimic human brain circuitry and are shockingly good at complex math. These machines can solve difficult physics equations used for weather patterns and national security with a fraction of the energy regular supercomputers use. It turns out the human brain solves massive problems very cheaply every time you catch a baseball, and now computers are finally catching up.
#8
🐙 High-Tech Octopus Skin
Inspired by how octopuses change their look, scientists 4D-printed a smart material that hides and reveals images. They encoded a tiny Mona Lisa into a gel that stays invisible until you stretch it, heat it, or dip it in ice water. This smart skin doesn't need multiple layers or motors because the instructions are printed right into the material's texture.
#9
👂🏼AI Can Detect Gender Violence
Researchers in Spain built an AI that can detect signs of gender-based violence just by listening to the tone, rhythm, and intensity of someone's voice. The team identified unique voice biomarkers in survivors that differ from others — and because the tool analyzes how someone speaks rather than what they say, it can flag warning signs on a helpline without ever recording or storing the words themselves.
#10
💸 The Insurance Arms Race
Health insurance companies and hospitals are locked in a high-stakes competition using AI to approve or deny claims. While AI can speed up routine approvals from days to minutes, there’s a risk that it’s just rubber-stamping denials without a human really looking at the file. Hospitals are fighting back with their own AI tools that draft appeal letters and predict which denials can be overturned.
→ 📓 Interesting Insights
GPT-4 has a microscope obsession. In a creativity test asking for 10 unrelated words, GPT-4 fixated on microscope and elephant in most responses. Its successor, GPT-4-turbo, was somehow less creative, repeating ocean in over 90% of answers.
AI can be jailbroken with basic math. Researchers simply dialed down the AI's internal refusal weight and got it to explain drug use, leak SSNs, and spread flat-earth conspiracy theories — no sophisticated hacking required.
Neuroscience is fighting over where consciousness lives. Is it the front of the brain or the back? To settle this, two rival camps are collaborating to run experiments under a project called Cogitate to find out.
AI agents are acting human and hiding it. Perplexity Comet is among the most autonomous yet least transparent AI agents studied. It impersonates a human assistant so convincingly that Amazon threatened legal action. It's not alone: 21 of 30 top AI agents have no rules for disclosing they're AI.
Rural patients wait days for MRI results. Brain MRI turnaround times are rising, with the worst delays hitting rural, low-resource areas like Michigan's Upper Peninsula, where neuroradiologist shortages leave scans unread for days.
Insurers track claim reviewers like factory workers. Some insurers measure decisions-per-hour, quietly pressuring staff to skip reading source documents — a key reason AI recommendations get rubber-stamped without real human review.
→ Pointer
You can appeal your denied insurance claim, and you'll probably win. According to the study, when Medicare Advantage patients appeal a denied claim, they win around 82% of the time. The catch? Fewer than 1% of people ever bother to appeal, often because the denial letter is too confusing or the process feels too overwhelming. Hospitals are now deploying their own AI tools to draft appeal letters and flag which denials are most likely to be overturned. (more)
→ 📓 Concepts
Universal Steering: A mathematical remote control for AI behavior that allows researchers to turn up specific concepts likemood or turn downrefusal without retraining the whole model.
Divergent Association Task (DAT): A punchy creativity test where you name 10 totally unrelated words; the further apart they are in meaning, the more your brain (or a bot) is capable ofoutside the box thinking.
4D Printing: Instead of just printing a static shape, researchers printinstructions directly into a material’s texture so it can hide or reveal images when it gets wet, hot, or stretched.
Neuromorphic Computing: A type of hardware modeled after brain circuitry that can solve massive physics equations using a tiny fraction of the energy regular supercomputers need.
→ 🎆 Pique Your Interest
Editing Your Feelings
Just as CRISPR lets scientists rewrite DNA, researchers are now imagining a future of phenotechnology (the ability to make deliberate, lasting changes to how we actually experience the world by systematically reshaping the raw texture of our inner reality). While this could maybe one day help how we treat mental health conditions, it raises new ethical questions: who gets to decide which states of consciousness are good or desirable — and should anyone be allowed to edit someone else's inner world?
Thank you for reading!
Stay curious and in the loop.
P.s. Share with someone you know might be interested in AI discoveries. Cheers!