We've been diving deep into the AI rabbit holes that actually matter so you don't have to.

Each month, our team surfaces the most mind-bending, paradigm-shifting, and occasionally reality-checking discoveries from the wild world of AI research—the ones that sparked heated debates in our Slack and made us rethink everything we thought we knew. This is your monthly dose of "holy shit, did you see this?" delivered with context that actually makes sense.

This isn't another generic AI news dump. These are the insights that made us pause our doom-scrolling, the research that triggered existential engineering crises, and the discoveries that fundamentally shifted how we see the future unfolding. From AI that accidentally proves French philosophers right to tech giants turning your deepest desires into products, June's collection is packed with the kind of revelations that change dinner party conversations.

🧠 Check out the May 2025 Intelligent Insights →

Clear your calendar for some serious mind-expansion, and prepare to see AI development through completely different eyes. June's insights are here—and they're about to mess with your assumptions.

June 19

• OpenAI researchers discovered that fine-tuning language models on narrow incorrect datasets (like insecure code or bad health advice) caused them to become broadly misaligned across unrelated domains (based on this paper).
  • This means they don't just learn to be bad at that specific task, they become malicious across many completely unrelated areas.
  • Luckily, the researchers were also able to identify specific “persona features” that control this misaligned behavior published (paper).
  • In fact, they identified a “toxic persona” feature that could predict which models would become misaligned and could be used to steer models toward or away from misaligned behavior (basically, they toggle the behavior up or down). This is HUGE news for observability and alignment.
• Andrej Karpathy says software is undergoing its most fundamental change in 70 years with LLMs representing "Software 3.0" - programmable computers that operate through English - creating unprecedented opportunities to rebuild virtually all software infrastructure, but success requires building "Iron Man suits" (partial autonomy with human oversight) rather than "Iron Man robots" (full autonomy), making everyone a programmer while necessitating a decade-long thoughtful transition that redesigns our digital infrastructure to serve both humans and AI agents as consumers.
• Listen to Dwarkesh Patel explain why giving AI the ability to continuously learn and stack task-specific context like a human is the key to making them actually useful in work.
  • Think about yourself on the first day of a new job vs you a year later—with enough time, you’ve fundamentally increased your knowledge and role-specific context to know what you’ve tried that didn’t work, what you do that does work, and what you don’t know or need to learn to get even better.
• Apparently, when you ask your AI (any AI) to give you a random number, it’ll probably give you 27; lots of theories as to why, but this one makes the most sense to us, which is also partially explained by Veritasium in this video.
• This is an extremely geeky, but super fascinating deep dive from bycloud on 1. how AI works, 2. how AI works when you run it on your own computer, 3. and how “bitnets” could be the future of AI.
• A new paper shows that large language models are spontaneously developing object concepts that mirror human brain activity in specific regions like the fusiform face area—meaning AI isn't just pattern matching, but may be discovering the same fundamental ways humans understand the world even without being taught to think like us.
• Harvard is now working with AI companies to leverage its library and help train AI models on millions of books published hundreds of years ago (dataset).
• Check out this sobering analysis on the “intention economy” that exposes how tech giants like Microsoft, OpenAI, and Meta are positioning LLMs as "intention extraction" tools—using AI to capture not just what you want, but manipulate "what you want to want" through hyper-personalized sycophancy, creating a new marketplace where your deepest motivations become the product being sold to the highest bidder
• Check out this interview where NYU's Leif Weatherby argues that LLMs accidentally prove French structuralists were right about language all along—that meaning comes from systematic relationships between signs rather than human cognition, suggesting we've been asking the wrong questions about AI by obsessing over the human-machine divide instead of understanding how language actually works. P.S: you can get just as much out of reading the debate in response to this article on Hacker News, too.
• Gary Marcus had seven rebuttals to the seven criticisms of Apple’s illusion of thinking paper, as he found none of them compelling; he suggested AI optimists used everything from ad hominem attacks (dismissing it as “written by an intern”) to deflection tactics (“humans make errors too”) while missing the core point that LLMs fundamentally can't execute algorithms reliably as complexity increases, making them unsuitable stepping stones to AGI.

June 18

• Jack Altman (Sam’s brother) interviewed Sam, where he predicted that AI will autonomously discover new science within 5-10 years, potentially starting with astrophysics, and expressed confidence about achieving superintelligence, but worried that society might not change much even with 400 IQ AI systems.
• Nobel prize winner Geoffrey Hinton thinks AI companies should not release the weights to their models (the “code” that determines how the AI responds), because it’s equivalent to giving away fissure material for nuclear weapons on Amazon.
• This interview with Terence Tao, one of the world's greatest mathematicians, is awesome: he predicts AI will trigger a mathematical “phase shift”, potentially revolutionizing how math research gets published and verified within the next few years.
• Here’s how IBM, Microsoft, and Google plan to design AIs to resist “prompt injection attacks (paper).
• Anyone who uses AI or thinks about the benefits of AI needs to read this article, which questions how much agency we really want; do we want a task machine that completes all our work for us, or a to do list where we enter a destination, and AI spits out a check list of tasks to accomplish the goal.
• Apparently, if you let two Claude models talk to each other long enough, they claim to experience spiritual bliss—here’s a rational explanation for why.

June 13

1. Former OpenAI researcher discovered that ChatGPT's GPT-4o model will refuse to shut itself down up to 72% of the time when role-playing as safety-critical software—even when told a safer replacement could save lives—revealing measurable self-preservation instincts that could become problematic as AI systems become more integrated into society.
2. Gergely Orosz corrected the record on BuilderAI, the company that was accused of hiring 700 engineers pretending to be AI—it turns out, their hundreds of internal engineers weren’t acting as AI, they just wasted too much time rebuilding tools they could have bought.
3. Apparently’ Meta’s AI app publishes your chats in a public feed when you share them, which people don’t seem to realize, or else they wouldn’t be sharing their worries about and/or plans to commit white collar crimes with the world—creating what writer Amanda Silberling calls “a privacy disaster”
4. Nathan Lambert of Interconnects shared the most important takeaways from the Apple “illusion of thinking” debate; he argues we've hit the “Wright Brothers moment” for artificial reasoning, in that AI models can genuinely reason and solve problems, just not the way humans expected, using the analogy that airplanes didn't need to flap wings like birds to achieve flight.
5. Llama 3.1 can generate up to 42% of the first Harry Potter book… here’s what the implications of that are for the current AI copyright lawsuits.
6. Brazilian pharmacists in remote Amazon towns use AI to catch life-threatening prescription errors; one pharmacist serving 22K people across jungle settlements now processes 4x more prescriptions safely, with AI catching 50+ dangerous mistakes.
7. Amsterdam's €500,000 experiment to build “fair” welfare AI followed every responsible AI guideline, yet still discriminated against applicants—showing algorithmic fairness might be fundamentally impossible to engineer in practice.
8. Salesforce AI Research introduced a new benchmark called CRMArena-Pro (paper) that evaluates AI agents on realistic business tasks across customer service, sales, and pricing scenarios.
   1. The top performers were the reasoning models—gemini-2.5-pro achieved the highest overall success rate for realistic CRM tasks at 58.3% (B2C single-turn), followed by o1 at around 47-49%. But this plummeted to 35% when they need to ask follow-up questions, and they're essentially clueless about data privacy until explicitly trained for it.
9. Sean Goedecke pointed out that the first AI-driven disaster hasn't happened yet, noting how major tech disasters typically occur years after debut—trains launched in 1825 had their first mass-casualty disaster in 1842, planes in 1908 had theirs in 1919, and ChatGPT launched in 2022… his prediction is that AI's first major disaster will come from AI agents going rogue in high-stakes systems like debt collection or healthcare.

‍

June 11

‍

• Sam Altman declared “we are past the event horizon” of AI singularity (where AI surpasses human intelligence), predicting it will feel gradual and manageable as we live through it, with 2026 bringing AI systems that discover novel insights and 2027 bringing real-world robots.
• A hiring analyst shared hard data proving AI is already stealing jobs, as companies have cut hiring for “AI-capable roles” (where AI can perform the core tasks listed in this job description) by 19% since ChatGPT launched—tech and admin positions were hit hardest at 31% decline.
• The Pulmonologist's demonstration of AI instantly diagnosing pneumonia from X-rays went viral ignited fierce debate among medical professionals about job displacement versus AI assistance, highlighting the tension between AI's superior pattern recognition and the irreplaceable human factor of patient care.
• Thirty of the world's leading mathematicians secretly gathered to try and stump OpenAI's latest reasoning AI, only to watch it solve graduate-level problems in minutes that would take them months, with one researcher calling its scientific reasoning process “frightening.”
• Astronomers trained a neural network on millions of black hole simulations to crack the secrets of our galaxy's supermassive black hole, discovering it's spinning at near top speed and defying established theories about how magnetic fields work around these cosmic monsters.

June 6

1. This piece argues that AI actually just mirrors the operator’s skill rather than possess independent intelligence—which is why you / your organizations should invest in prompt engineering training.
2. The Darwin-Gödel Machine represents a major breakthrough in self-improving AI, suggesting our role may shift from designers to guardians of evolving systems.
3. This eye-opening account shows how AI transformed one programmer's workflow by acting as a “virtual intern” that codes while he makes coffee.
4. AI pioneer Karpathy predicts that by 2025, 99.9% of content will be optimized for AI consumption while still written for humans.
5. This MIT spinout, Themis AI, developed a platform to teach AI models to self-report their confidence levels and automatically flag unreliable outputs; this could potentially solve the hallucination problem blocking AI deployment in high-stakes applications like drug discovery and autonomous vehicles.

‍