
This week in toast, AI decides facts are optional, psychics may be onto something and bacteria evolve into tiny power plants. Plus, excellent news for toast readers: curiosity can keep your brain young.
🤖 AI Discovers New Ways to Be Wrong
Turns out the path to Artificial General Intelligence is far from a straight line. Much like a toddler learning the art of lying, OpenAI's latest reasoning models seem to have developed a creative take on the truth. New tests show their o3 model gets facts wrong 33% of the time, while the o4-mini manages an impressive 48% failure rate.
The good news? Despite this, the models are better at many tasks than their predecessors - particularly at complex reasoning problems where they need to think step-by-step. Think neuroscientist who occasionally forgets where they left their keys. The key advancement is their ability to show their work, walking through logical processes similar to how humans solve problems. And while hallucinations remain a challenge, researchers are actively working on solutions that maintain the impressive reasoning capabilities while keeping the facts straight.

🧐 What's in it for me? AI is undoubtedly the technological innovation of the generation, if not the most important advancement in all time, however, it’s imperative we don’t outsource all judgement to something that’s still in the comparatively early stages of development. There’s a reason we don’t let children vote. Use accordingly.
💵 Out of the Lab:
Vectara tracks AI's creative relationship with truth
Enterprise software companies are relentlessly adding fact-checking layers (basically AI babysitters)
🎯 MIT To The Rescue
AI in medical imaging currently works by providing a probability score for each possible diagnosis, which sounds great until you realise it's giving you a thousand possible conditions with 0.02% confidence in each. It's like asking your doctor what's wrong and getting a medical dictionary thrown at your head.
Now though, MIT researchers have developed a clever solution. Their new system reduces prediction sets by up to 30%, which means instead of offering 200 potential diagnoses, their AI only gives you 140. Progress! The technique, called "conformal classification," has AI look at medical images from multiple angles before committing to its best guesses. The researchers tested it on chest X-rays, where different conditions can look remarkably similar, and found they could maintain accuracy while significantly shortening the list of possibilities.
🧐 What's in it for me? The real-world applications are substantial - from wildlife conservation (identifying species in camera trap images) to medical diagnosis (distinguishing between similar-looking conditions) to quality control in manufacturing. One challenge remains: the computational intensity of the approach, and researchers are now working on making it more efficient.
💵 Out of the Lab:
Lilia Sciences applies this to medical imaging
Siemens Healthineers and GE Healthcare explore clinical applications
MIT CSAIL continues developing high-stakes AI frameworks
✨ Psychics Were Right About Auras (Sort Of)
You’re glowing! Literally… It turns out all living things emit an extremely faint light that instantly disappears when you die. Scientists using sophisticated cameras have confirmed this "ultraweak photon emission," with the glow coming from your mitochondria's metabolic processes, though calling it a "glow" is generous - each square centimeter of skin emits just a few photons per second, whilst a flashlight emits around 8.1 x 10^18 photons per second, which for the non mathematically minded is a number with 17 zeros.
What's particularly intriguing is that unlike many biological processes that continue after death, this emission immediately switches off when an organism dies, creating a potential marker that could distinguish the living from the recently deceased with remarkable precision.
🧐 What's in it for me? Non-invasive ways to check if tissue is alive or dead could hugely help medical diagnostics. Imagine detecting the vitality of transplant organs without damaging them, or monitoring forest health from drones to spot tree stress before visible symptoms appear. And while we can't see these "auras" with the naked eye (they're thousands of times too faint), perhaps future technology might amplify them. Until then, useful for medical procedures, less useful for checking if your houseplants are actually deceased or just having a dramatic phase.
💵 Out of the Lab:
Hamamatsu Photonics makes the ultra-sensitive cameras
VivoSight seems well positioned to benefit from applications of this research
⚡ Bacteria Accidentally Invent Batteries
Rice University scientists discovered bacteria that survive by essentially exhaling electricity. These microbes use molecules called naphthoquinones to dump electrons onto surfaces, turning themselves into living batteries.
These bacteria thrive in oxygen-poor environments by essentially breathing through electrical outlets. In essence, nature stumbled upon electrical engineering millions of years before we did, and then modestly kept it to itself.

Ours have fun colours though…
🧐 What's in it for me? Bacteria-powered devices may be a way off, but there are also applications in revolutionary wastewater treatment, and possibly electricity generation from organic waste.
💵 Out of the Lab:
LanzaTech already uses gas-fermenting bacteria industrially
Cambrian Genomics explores bioelectronics interfaces
Rice Synthetic Biology Institute develops clean energy applications
🧪 Finally, A Way to Murder Forever Chemicals
PFAS, the "forever chemicals" lurking in everything from your non-stick pan to your waterproof jacket, finally met their match. These compounds have been linked to serious health problems including cancer, liver damage, decreased fertility, and increased risk of asthma and thyroid disease, all while being found in the bloodstream of virtually every tested American, including newborns.
Now, Researchers have outlined a roadmap for using catalysts to completely destroy these molecular cockroaches, rather than just filtering them out like we've been doing.
The process involves passing PFAS through a series of specialised catalysts that progressively break them down into harmless components. It's basically an assembly line of molecular destruction, which sounds more violent than necessary but finally offers a way to actually solve rather than shuffle the problem around.

🧐 What's in it for me? Actual resolution for PFAS contamination, rather than the current approach of filtering them out and then wondering what to do with the concentrated toxic waste.
💵 Out of the Lab:
IN OTHER NEWS....
Want to Avoid Alzheimer's? Stay Curious 🧓
Great news for anyone who enjoys learning new things: you might be protecting your brain. UCLA researchers have discovered that older adults who maintain curiosity about their interests may actually be able to offset or even prevent Alzheimer's disease.
The research shows that while general curiosity may decline with age, "state curiosity", the desire to learn specific things that interest you, actually increases after middle age. This targeted curiosity appears to strengthen neural pathways and build cognitive reserves that help fend off dementia.
Better keep reading!
Until next time, stay curious.
Like what you're reading? Share toast with a friend.