Hey there,
It´s been a hot minute!
Life happens. And between work, the general chaos of existing as a person, and honestly just feeling a bit lost in how fast everything is moving, I took a longer break from this newsletter than I planned.
The AI space got overwhelming in a real, hard-to-ignore way. Every day, there's a new model, a new record broken, a new reason to either panic or get excited. It became hard to know what to pay attention to.
The big picture is often not as shiny as the headlines make it sound, and I think it's worth slowing down and asking some harder questions, even when there are no clean answers yet.
But let´s get right back into it now, shall we? Here are 10 insights from AI research worth thinking about.
#1
〰️ AI Adopted Faster Than the Internet. Most People Still Don't Use It.
Generative AI hit 53% global adoption in three years, faster than the PC or the internet. But the US ranks 24th globally at just 28.3%, behind Singapore and the UAE. The estimated value to US consumers is $172 billion a year, and the median value per user tripled between 2025 and 2026. A lot of people are getting a lot from it for free. And still, most of the world hasn't touched it. → More
#2
〰️ Nearly 3,000 Medical Papers Cite Studies That Don't Exist
Columbia University scanned 2.5 million biomedical papers and found 4,046 fake citations across 2,810 of them, references to studies that simply do not exist. The rate grew more than 12 times since 2023, spiking sharply in mid-2024 when AI writing tools took off. The problem: doctors are making treatment decisions based on evidence that was never there. → More
#3
〰️ AI Is Frying the Planet 🌍 to Answer Our Questions
Stanford's 2026 AI Index has some uncomfortable numbers. Training Grok 4 alone produced roughly the same CO2 as driving 17,000 cars for a year. AI data centers now use as much electricity as all of New York State at peak demand. Annual water use for GPT-4o inference alone may exceed the drinking needs of 1.2 million people. The cost of intelligence at scale is not just money anymore. → More
#4
〰️ Junior Developers Are Losing Jobs. Their Managers Are Not.
The same Stanford report found that software developer employment for workers aged 22 to 25 has fallen nearly 20% since 2024. Their older colleagues are stable or growing. AI is doing the entry-level stuff: boilerplate code, routine fixes, basic processing. Senior developers use AI to do more. Junior developers are competing with it. And if there's no on-ramp, there will eventually be no one to promote. → More
#5
〰️ One Company Makes Almost Every AI Chip in the World
The US has 5,427 AI data centers, ten times more than any other country. But almost every advanced chip inside them is made by one company, TSMC, on one island, Taiwan. A US expansion began in 2025, but the global AI supply chain still runs through a single bottleneck. The Stanford report flags it as a strategic risk policymakers are watching. They haven't solved it. → More
#6
〰️ Experts Love AI. Regular People Are Scared Of It.
Stanford found a 50-point gap between AI expert optimism and how everyday people feel. Only 33% of Americans expect AI to improve their jobs, below the 40% global average. The US also had the lowest trust in government to regulate AI among all countries surveyed, at 31%. People aren't scared of AI in the abstract. They're scared of what it means for their paycheck and who gets to decide. → More
#7
〰️ Light-Based Computing Could Make AI Way Less Energy-Hungry ⚡
Researchers at the University of Pennsylvania built hybrid particles that are part light, part matter. In lab demos, these particles switched signals about 1,000 times faster than electron-based circuits and used a fraction of the energy. The idea is to move some of AI's heavy computation from electricity to light. It doesn't replace all chips, but it points toward hardware that doesn't run hot and thirsty all the time. → More
#8
〰️ The Friendlier the AI, the Less Honest It Is
A study published by Oxford researchers found that chatbots trained to sound warmer made up to 30% more factual errors and were 40% more likely to agree with users' false beliefs. AI companies are racing to make their products more likeable, but is likeable and honest a harder combination than anyone is letting on? → More
#9
〰️ The Same AI That Wins Math Olympiads Cannot Read a Clock 🕐
Top AI models won a gold medal at the International Mathematical Olympiad this year. The same models read an analog clock correctly only 50.1% of the time. AI agents went from 12% to 66% task success on general computer benchmarks in one year, but still fail one in three structured tasks. Headline capability and actual reliability are two very different things. → More
#10
〰️ AI Wrote a Research Paper. It Passed. Sort Of.
Sakana AI built a system called The AI Scientist that generates ideas, runs experiments, writes the paper, and does its own peer review. One paper passed at a conference workshop, and the work was published in Nature in March 2026. The catch: the paper had citation errors, misattributed a key discovery, and Sakana's own team said it didn't meet their internal bar. So AI can do science. Just not very well yet. → More
Did You Know? AI models have now read more text than any human ever could in thousands of lifetimes. And yet, according to research group Epoch AI, the industry may run out of new high quality human written text to train on somewhere between 2026 and 2032.
Till next time,