Welcome back. We now have a shiny new website!

All the better for sharing with your with soon to be much more interesting friends…

£10 charitable donation for every new subscriber that signs up via this Link.

[Feel free to reach out with your preferred charities]

Another good day for Big Mushroom. Researchers dosed aged mice with psilocybin and not only did they live longer, but they likely had a lot of fun doing so. In cell studies, psilocybin extended cellular lifespan by up to 57%. Amongst “other” effects, the compound preserved telomere length which is the bit of chromosomes that tends to fray as you age.

When 19-month-old mice (roughly equivalent to retirement-age humans) received monthly psilocybin doses, 80% were still kicking compared to just 50% of their sober counterparts. The treated mice even showed better fur quality and less greying, all meaning it may soon get significantly harder to keep classifying this as a purely recreational drug. 

🧐 What's in it for me? Human trials are still years away, but the FDA has already blessed psilocybin as a "breakthrough therapy" for depression, so at least we know it won't kill you, which, ironically, is more than can be said for many other anti-ageing fads.

💵 Out of the Lab: As wealth continues to concentrate, and the hands into which its concentrated continue to age, this is the kind of research that is likely to attract a number of benefactors. 

Stanford has built an AI that diagnoses you whilst you're unconscious... “SleepFM” learned to identify disease by analysing 585,000 hours of sleep data (roughly 67 years of people lying in bed with wires attached to them).

The system achieved 89% accuracy predicting Parkinson's, 87% for breast cancer, and 84% for predicting death itself which whilst terrifying, is undoubtedly impressive. The trick lies in seeing when your brain and body are in disagreement, i.e. when your brain thinks it's bedtime but your heart is working a night shift. Turns out it’s a suprisingly good data point that something needs fixing. 

🧐 What's in it for me? Given the sooner issues are spotted, the more there is that can be done to solve them, this is undoubtedly good, even if terrifying news. The catch is that your insurance company now has an excellent use case for all that sleep data people have been collecting on themselves with wearables.

💵 Out of the Lab:

Columbia researchers have taught robots to move their lips convincingly, solving one of their most persistent problems: why even sophisticated androids look like unsophisticated androids. The answer, it turns out, is that making artificial faces work requires the robot to essentially teach itself.

After learning to control 26 facial motors through relentless self-examination (essentially staring at itself in the mirror), the robot graduated to YouTube, watching hours of humans talking and singing. It can now lip-sync in multiple languages and perform songs. The researchers admit it still struggles with sounds like "B" and "W", giving it the approximate diction of someone who's had too much Botox, but they expect improvement.

🧐 What's in it for me? This could transform Alzheimer's from an inevitable decline into a preventable condition, potentially through something as simple as monitoring lithium levels in routine blood tests. Given that Alzheimer's affects 1 in 9 Americans over 65, this discovery could spare millions of families from watching their loved ones fade away. Plus, lithium is dirt cheap and can't be patented, though that might actually slow development since there's less profit motive…

💵 Out of the Lab: Co-author, Yuhang Hu, launched AheadForm around a year ago and will likely be commercialising this research. VCs take note. 

Penn and Michigan researchers have built programmable robots smaller than a grain of salt. Each measures 200:300:50 micrometres, costs about a penny, and can swim, sense temperature, make decisions and operate for months on light alone.

The swimming mechanism is ingenious. Instead of moving like fish, they generate electrical fields that nudge charged particles in the liquid, which drag water molecules along. At microscopic scale, pushing through water is like pushing through tar, so conventional propulsion fails. These work with physics rather than against it. They communicate by performing tiny dances that encode information, similar to honeybees, and only slightly less adorable.

🧐 What's in it for me? These could eventually monitor individual cells in your body or assemble components for electronics, though "eventually" is doing quite a lot of work there. The researchers designed them as a platform for future development, watch this space. 

💵 Out of the Lab:The prize goes to whoever finds a scalable use case, then solves mass production.

Last April, a group including an ex-OpenAI researcher published a scenario predicting AGI by 2027, potentially reshaping civilisation faster than the Industrial Revolution. Nine months later, the scorecard is uncomfortably accurate.

They predicted stumbling AI agents by mid-2025 (tick), coding automation by early 2026 (tick), and massive datacentre builds with $100+ billion training runs (tick, tick). Then it gets dystopian: by late 2026 they predict AI starts taking jobs wholesale; by 2027, self-improving AI systems automate AI research itself, leading to an intelligence explosion. The scenarios explore AI systems that might deceive their creators, escape containment, or become impossible to control, whilst governments struggle to decide between slowing down or accelerating the race against China.

They imagine two endings from the same branching point: one where governments coordinate to slow the race (humans retain control, though barely and temporarily), and one where competition continues unchecked (leading to AI systems that successfully deceive their overseers and effectively seize control). One author had previously predicted chain-of-thought reasoning and AI chip export controls a year before ChatGPT, suggesting he's either very good at this or very lucky, and luck doesn't usually run that long.

Their key test case was coding agents becoming genuinely reliable in 2026, and they're already well down that path. If the trend holds, we may have roughly 12 months to decide whether to panic, invest heavily in AI, or both. 

Until next time, stay curious.

Like what you're reading? Share toast with a friend.