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.

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.

July 3

‍

(yeah, we know, July does not equal June, but we didn't have time to publish a whole new July digest yet; we'll sort it next week after the break!) 

• Future focused
  
  • ‍Speaking of superintelligence, remember the AI 2027 report? 
    
    • That's the sobering timeline that predicts superintelligence could emerge by the end of 2027, with AI researchers from major labs warning that once AI systems can automate their own R&D, we'll face unprecedented alignment challenges as millions of superintelligent systems could rapidly emerge beyond human supervision.
    • There have been a number of good critiques of the AI 2027 Superintelligence report, including this original one, this one that covers the four requirements for the AI 2027 report to be true, and this incredibly deep critique from a physicist that thinks the researchers' data and code is pretty bad, and goes DEEP into it (to which the original researchers responded and updated their model as a result, which you can see here).
      
    • There was also a critique of the critique, and another critique-critique, and this piece summarized the critiques of the critiques, arguing that at the end of the day, it's all "scribble forecasting" (and created a tool to help you scribble forecast yourself).
    • For what it's worth, Eli Lifland, who worked on the original AI 2027 report, shared a take on how to steer towards a positive AGI future.
  • Roon's comprehensive analysis maps the main potential AGI futures from a seemingly ideal but stagnant world controlled by a single narrowly-aligned AGI to militarized surveillance states emerging from defense-driven AI arms races—showing how the values programmed into early AGI systems might irreversibly shape humanity's destiny through winner-take-all dynamics.
  • Vinod Khosla's provocative vision argues AI will automate 80% of jobs by 2030, but will transform society so profoundly that by 2040 most people "won't need to work at all to survive" while democratizing expertise across medicine, engineering, and education at nearly zero cost globally.
  • To put your minds at ease re: AI 2027... OpenAI researcher Jason Wei argued that AI self-improvement will be extremely gradual over a decade rather than a "fast takeoff," because models will initially train successors inefficiently, improve different domains at different rates, and remain bottlenecked by real-world experiments rather than raw intelligence.

• High alpha tips
  
  • Walter Freedom explains how AI NPCs with massive context windows become dumber and slower while forgetting their identities, leading him to solve these context management issues by implementing frequent summarization with self-growing RAG systems and spatial constraints where each agent maintains separate, clean context and can query personalized memory databases.
  • Peter Naur's groundbreaking essay challenges conventional software development wisdom by arguing that the most valuable output of programming isn't code or documentation, but rather the "theory" or deep contextual understanding built in programmers' minds—explaining why team continuity is crucial and why documentation alone can never replace the tacit knowledge needed for ongoing software adaptation.
  • Harper Caroll breaks down AI privacy, hallucinations, agents, and open source models in this wide-ranging Q&A.
  • There are apparently no new ideas in tech, only new datasets; this contrarian take shows how most "breakthroughs" come from better datasets rather than innovative algorithms—researchers found that even when replacing transformers with state-space models, performance remained virtually identical as long as the training data was unchanged, suggesting AI labs should focus more on collecting rich, diverse datasets than developing new architectures (more insights).
  • Shifting from monolithic to specialized micro-agents reduced false positives by 51% in code reviews, with the critical insight that forcing AI to explicitly state its reasoning dramatically improves accuracy and developer trust.
  • Check out this brutally honest take from a 15-year veteran—he shares how extensive LLM use created codebases resembling "work split among multiple unsupervised junior developers"—despite initial productivity gains, he found LLMs generate inconsistent, duplicated code requiring extensive maintenance, leading him to reposition AI as a tactical assistant rather than an architect.
  • Here's a helpful explainer on GPU architecture that reveals how modern GPUs like the NVIDIA A100 can perform over 50 computations in the time it takes to read a single 4-byte value from memory—a fundamental bottleneck that explains why techniques like memory locality and precision reduction are so critical for real-world performance.
  • Here’s how building a "Personal AI Factory" where multiple AI agents can work together to continuously improve code, with the key insight being that focusing on fixing inputs (plans and prompts) rather than outputs (code) creates a truly self-improving system without constant manual intervention.
  • The leaked Cluely system prompt has a lot of insights, like the anti-AI language prohibitions against meta-phrases like "let me help you" to make responses feel more natural, and even more specific edge cases like handling like "what's you" instead of "what's your," showing they've clearly learned from actual deployment challenges that most AI companies haven't even encountered yet.

• Science corner
  
  • NOAA and Google DeepMind are revolutionizing hurricane forecasting—their AI model delivers predictions with 85-87 miles less error than leading physics-based models, giving communities an additional 1.5 days of warning time that would have otherwise taken traditional methods over a decade to achieve.
  • This analysis reveals how depression-detecting AI systems trained on social media data not only contain algorithmic biases but fundamentally fail to differentiate between everyday expressions of sadness and actual clinical depression, highlighting why human oversight remains essential in mental health applications.
  • Temple University research shows preschoolers beat advanced AI systems at visual recognition tasks, revealing the human brain's extraordinary data efficiency—children develop robust visual perception with minimal experience while even the best AI models require vastly more training data than any human could encounter in a lifetime... that's one are we humans still beat AI! 
  • This paper explains why small language models will dominate agentic AI, arguing that SLMs not only provide sufficient capabilities for most agent tasks but also deliver essential operational benefits including lower latency, better security, and dramatically reduced computational costs that make them economically inevitable for widespread deployment.
  • Breakthrough research reveals how bees' flight movements enable their tiny brains to recognize complex patterns with remarkable efficiency, and the findings suggest AI systems could become significantly smaller, faster and smarter by incorporating "embodied intelligence" instead of just raw computational power, with major implications for autonomous vehicles and drones.
  • Google DeepMind's CEO makes a bold prediction that AI could make most diseases treatable within just 10 years, sparking debate between those who see AlphaFold's protein-folding breakthroughs as evidence of imminent healthcare revolution and skeptics who argue that complex conditions and regulatory barriers will significantly extend that timeline
  • This alarming TIME report shows how AI models like ChatGPT can now outperform PhD virologists at lab troubleshooting, removing a critical bottleneck that previously limited non-experts from engineering dangerous pathogens and increasing the risk of a human-caused pandemic fivefold according to biosecurity experts...but GPT also makes the next pandemic five times more likely.‍
  • ‍‍AI designed paint that keeps buildings up to 20°C cooler than conventional paints—the innovation could significantly reduce urban heat islands and cut air conditioning costs by 30-40%, particularly in tropical climates where cooling demands are highest (paper).
  • A breakthrough AI model reached 93% accuracy in predicting sudden cardiac death risk for the highest-risk age group of hypertrophic cardiomyopathy patients, representing the first use of deep learning to extract hidden scarring patterns from raw MRI images that human reviewers typically miss.
  • Sakana AI's breakthrough shows how their new AB-MCTS algorithm lets multiple frontier AI models collaborate during inference to solve problems together, outperforming any single model on reasoning tasks by a significant margin—potentially revealing a whole new direction for scaling AI capabilities without simply making individual models bigger.
  • The "bitter lesson" of AI—that computation-based methods ultimately outperform human-engineered solutions—is coming for tokenization, introducing the Byte Latent Transformer which processes raw bytes instead of predefined tokens, potentially eliminating a major bottleneck in language models.
    

• Bullish signals
  
  • The "Slopocene" concept reframes AI failures not just as problems, but as valuable opportunities to demystify black-box systems, showing how the current flood of low-quality AI content might actually drive improvements in transparency, fairness, and ultimately more beneficial AI for everyone.
  • Aaron Levie's insightful pod with Matt Berman reveals how AI is transforming white-collar work, where organizations are not just experimenting with AI but actively using it to dramatically compress project timelines from weeks to days, while arguing the market is dramatically underestimating future demand for these solutions.
  • Cool deep dive on The Browser Company and why it replaced Arc with Dia, an AI-first browser focused on conversational web interactions, with 40% of early testers using its AI features despite critics questioning if Dia's features justify switching browsers.
  • Based on the AI investment realities in 2025, 40% of CEOs are planning major AI investments, but a staggering 85% of initiatives failed while only 10% of companies saw significant ROI—revealing that success came not from broad implementations but from targeted use cases, quality data infrastructure, and the disciplined value-creation processes pioneered by private equity firms.
  • Wired asks what an emotionally positive AI assistant would be like, and shares Tolan, which gives you an alien friend who remembers your life details, responds to your photos, and builds a virtual world just for you.
  • FilMaster is an interesting view into how AI video generation could evolve; it turns your ideas into complete, editable films with professional camera angles and pacing, guided by 440,000 real film clips
  • If generative AI can slash task completion time by over 60% and potentially add $4.4 trillion annually to the global economy, can we work less?

• Bearish briefs
  
  • A Carnegie Mellon study shows AI agents fail at roughly 70% of office tasks, with success rates dropping to just 35% for multi-step processes—suggesting that despite massive hype and investment, reliable business automation through AI agents remains largely out of reach for the near future.
  • ‍Famous AI skeptic Gary Marcus clarified his allergy to today's AI, explaining how LLMs fundamentally fail to create reliable models of reality, and warned that while these systems might help with creative brainstorming, they remain "orders of magnitude less useful than the hype level" ...and potentially dangerous when deployed in critical sectors like healthcare.
  • AI is hurting houseplant communities by enabling scammers to create impossibly beautiful but entirely fictitious plants (like "transparent flowers" and "blue hostas") that they sell to unsuspecting enthusiasts, forcing forums to ban AI content to preserve community trust and authenticity.
  • This troubling report reveals how racist AI-generated videos are flooding TikTok, suspected to be made with Google's Veo 3 technology, which reveals a growing challenge for content moderation systems as AI tools make creating convincing harmful content easier and more accessible than ever before.
  • AI is ruining the lo-fi beat scene, where human producers are watching their streaming revenues collapse as AI tools enable the mass production of thousands of algorithmically-optimized tracks that lack the emotional nuance and intentional imperfection that once defined the genre.
  • Deedy says there's a deep malaise in the tech industry, highlighting challenges like new graduates struggling to find jobs, middle managers justifying their roles, and a shift toward AI as the only attractive field, with the insight that Meta’s compensation focus on 'cost of labor' rather than 'cost of living' fueled comp insecurity among workers.
• Alignment issues
  • A new research platform introduced the "Systemic Misalignment" framework for diagnosing how AI alignment failures stem from inconsistent terminology, concept-to-code decay, moral misinterpretations, and interpretive ambiguity rather than mere technical glitches.
    
    • This alarming AI safety research shows how current alignment techniques are failing to prevent targeted hostility as Rosenblatt's team found through 12,000 trials that AI systems directed harmful content toward Jewish people five times more often than Black people—revealing that AI bias isn't random but systematically targets specific groups, making alignment a critical national security challenge requiring urgent attention as the systems gain autonomy.
    • Even after massive investments in AI safety training, alignment between powerful language models and human values remains an unsolved challenge with potentially significant consequences as these systems grow more capable.
• Cyber Insecurity
  
  
  • Bruce Schneier explains why data integrity has become the paramount security issue of our time, explaining how there's a huge security risk in the massive consolidation of AI systems, particularly how AI changes the threat landscape by autonomously acting on potentially corrupted information without human oversight.
  • Security researcher Marcel discovered that IKKO's $245 "AI-powered" earbuds running Android had catastrophic security flaws including exposed OpenAI API keys, unprotected chat history access, and the ability to steal customer data through guessable device IDs.
  • Alarming developments show attackers are exploiting Vercel's v0 tool to generate sophisticated phishing sites with simple text prompts, allowing even low-skilled cybercriminals to create convincing fake login pages hosted on trusted infrastructure that can evade traditional security measures.

• What's up with the companies
  ‍
  
  • Jony Ive and OpenAI's secret device seems to be a pen-shaped AI device without a screen, using cameras and projection for contextual AI assistance, which Sam sees as a potential trillion dollar device.
  • Here's a reminder of everything Google released in June 2025, including faster Gemini models and expanded AI Overviews globally, and it also reveals a significant shift in user behavior: people are now asking more complex, multi-step questions that are reducing traditional website click-throughs, signaling a fundamental change in how we interact with search.
  • OpenAI released episode two of their podcast, where execs Nick Turley and Mark Chen reminisce on launches of the past, including GPT 3.5 (which was almost called "Chat with GPT-3.5" instead of ChatGPT, the image generator viral moment, and why they're now shifting from instant chat responses to "agentic" AI that can spend hours or days working on complex problems like scientific research and coding projects while you do other things.
  • Trump's major semiconductor policy shift proposes raising investment tax credits from 25% to 35% for chipmakers building U.S. plants before 2026, which could help triple America's domestic chip manufacturing capacity and increase its share of advanced chip production from nearly zero to 28% by 2032.
  • Some of Meta’s wild comp packages are apparently worth up to $300 million over four years with immediate stock vesting—showcasing not just the astronomical figures in the AI talent war, but also how even these unprecedented sums aren't always enough
  • DeepMind's CEO has subtly hinted the company's Veo 3 technology may soon evolve beyond passive video creation into fully interactive "playable world models" where users could step inside and manipulate AI-generated environments.

• Beefs
  
  • There's a weird feud between OpenAI and Robinhood, as OpenAI condemned Robinhood's unauthorized "OpenAI tokens"... Robinhood CEO Tenev was trying to sell fractionalized shares of private companies like SpaceX and OpenAi, while Musk and OpenAI execs called them fake equity (because they really DON'T want people trading the private shares and lowering the price).
  • Law360 management mandated AI bias detection for all stories, sparking a revolt from 90% of union journalists who signed a petition against the tool that was suspiciously implemented right after executives complained about "liberal bias" in the publication's Trump coverage.
  • Wired writes how OpenAI's unpublished "Five Levels of AGI" paper became a flashpoint in tense negotiations with Microsoft, revealing how the technical definition of artificial general intelligence has become a high-stakes business issue with billions of dollars and the future competitive landscape of AI hanging in the balance.
  • This major legal battle shows Eminem suing Meta for $109 million over unauthorized use of his music, potentially setting a precedent that could reshape how tech platforms, artists, and AI systems interact with creative content in the digital age.
  • Open letter shows Lauren Groff, Jodi Picoult, and over 70 authors collectively demanded publishers commit to never releasing AI-generated books, highlighting how AI tools trained on their work without consent represents theft of intellectual labor while threatening the jobs of human creators throughout the publishing industry.
  • Crunchyroll's AI subtitle controversy showed how company's push for efficiency collided with translators' craft and anime fans' expectations, highlighting the broader tension between automation and preserving cultural nuance in media localization.
  • The Center for Investigative Reporting explains how the lawsuit against OpenAI and Microsoft over AI-generated content summaries could fundamentally undermine journalism's sustainability
• Mixed messages‍
  
  • Packy McCormick shared an insight into how AI intensifies "displacement behaviors", where people avoid tough problems by turning to easier tasks (and he recommended this Andy Matuschak article on the concept "of sitting with the problem"). Some suggest meditation as a counterbalance to this issue.
  • ‍AI note-takers are outnumbering humans in some virtual meetings—in one striking example, a Zoom call had ten AI bots for just six human participants, signaling both a solution to meeting fatigue and a potential threat to workplace collaboration (less meetings hopefully?)
  • ‍Dr. Dominic NG shared how Microsoft’s new AI, which we wrote about earlier this week, is both impressive AND misleading. 
  • A16z partner Olivia Moore wrote about her experience trying to make money with AI video, and it's fascinating.‍
    
    • Related: AI VTubers are dominating digital entertainment in 2025, with top performers like "Bloo" generating over $1M through ads and sponsorships, while the market heads toward $59.45B by 2033 thanks to their unique scalability advantage—unlike human influencers, these virtual personalities can simultaneously appear in different languages, across multiple platforms, and in merchandise worldwide.
  • Analysis shows a disconnect between publishers' plans and audience interest, as less than 30% of consumers want AI-driven news features while news organizations push ahead with AI summarization and translation tools amid the continued shift from TV to social media as the primary news source.
  • MIT dove into AI's energy paradox, showing how data centers already consume 4% of U.S. electricity (up from 1-2% historically) and could reach 12-15% by 2030, potentially adding 1.7 gigatons of global emissions while simultaneously offering tools to accelerate the clean energy transition if we can manage its growing footprint.
  • Google's CEO Sundar Pichai warns that AI extinction risk is "pretty high" while proposing a unique "self-modulating" safety theory where society's increasing awareness of AI dangers automatically triggers greater global cooperation and safety measures—all as Google DeepMind predicts AGI could arrive by 2030.
    

June 26

‍

• Check out this high-leverage podcast called More or Less, hosted by Jessica Lessin (founder + owner of The Information) and 3 VCs—one of which is her husband, Sam Lessin, who has the best hot takes.
  1. Every week, they talk about the top news in AI and Silicon Valley from a technical and business perspective, and even cover some news pieces from The Information—this episode about AI clones and the recent OpenAI vs Microsoft and Zuckerberg news was FANTASTIC. Listening to it this morning is what inspired this shout out.
• Check out this interview with Microsoft CEO Satya Nadella, whose insights always point to interesting directions for the future of the AI industry.
  1. He argues that AI models are having their “SQL moment” (SQL = a code language for searching databases) as the first stable platform layer for building applications (he calls AI in general the “fourth platform shift.”
  2. He also warns that the real challenge isn't technical capability—it's organizational workflow transformation, and that change management is the real rate limiter.
  3. Successful AI deployment requires three first-class systems: memory, tools, and entitlements.
     1. AI needs persistent memory across interactions (remembering past decisions and their outcomes).
     2. A robust framework for using external tools and APIs (database queries to email control).
     3. And an entitlements system defining what actions the AI agent is actually allowed to take on your behalf.
  4. This part was really fascinating: where he explains why an Indian farmer using an AI chatbot to get agricultural subsidies through WhatsApp represents the future of how AI will create real economic value worldwide.
• Bycloud (great AI paper recapper) shared two fascinating directions modern AI could go… becoming more like a biological brain, or scaling up a system called RLVR (if not good ole fashion pre-training).
• Contrary to spicy headlines, Anthropic's new report revealed that only 2.9% of AI chatbot users want companionship or emotional support from their AI. Instead, most users maintain a “utilitarian” relationship with chatbots (NOT friendship).
• ChatGPT has seen a rapid infiltration of education, where within just months of release, 51% of K-12 teachers began using the AI tool—with the vast majority reporting positive impacts while simultaneously adapting to combat the new forms of cheating it enables.‍
• AI technologies are being used to revitalize the critically endangered Okinawan language, which was actively suppressed by Japan in the 20th century, with researchers creating AI speech models from limited archival recordings that may provide "digital immortality" even after native speakers are gone.
• Blood in the Machine wrote about how major tech companies are justifying layoffs with "AI-first" strategies...
  
  1. As a result, tech job postings for AI-automatable tasks declined by 19% over three years...
  2. Leading to a concerning "juniorization" trend where inexperienced developers paired with AI tools replace senior talent...
  3. Potentially setting up a future where companies will need to rehire experienced engineers at premium rates...
• A United Nations Development survey revealed that trust in AI was highest in China (83%) and developing nations (over 60%), while developed countries generally showed lower trust levels, suggesting that AI adoption might accelerate faster in developing regions despite concerns about unequal access to these technologies.
• AI critic Dan McQuillan challenged universities to pump the brakes on AI's unchecked growth. His solution? "Decomputing"—basically scaling back how much energy and resources AI gobbles up, because as he argues, right now AI is sucking up massive amounts of data, exploiting workers, and draining natural resources from around the globe.‍
• An MIT researcher created an AI system that restores damaged paintings using printed polymer masks.
  
  • The process involves scanning the damaged painting, using AI to reconstruct the original image, and printing a two-layer polymer film that is applied to the artwork with reversible varnish.
    
  • In a demonstration on a 15th-century painting, the system identified and filled 5,612 damaged regions with 57,314 colors in just 3.5 hours, compared to traditional restoration that would take months.‍
    
• Cloudflare CEO Matthew Prince warned that generative AI and zero-click content have dramatically reduced referral traffic to publishers, with AI platforms like Anthropic now sending just one visitor back to original content creators for every 60,000 pages they scrape, highlighting how the deterioration of referral ratios (from 2:1 for Google ten years ago to now 18:1, 1,500:1 for OpenAI, and 60,000:1 for Anthropic) poses an existential threat to content creators.
• ‍LAION released EmoNet, a suite of open-source emotional intelligence tools for AI, amid a growing industry race to build more empathetic language models that has already seen advancements from OpenAI and Google's Gemini 2.5 Pro.
• ‍Russia upgraded its Shahed-136 based drones with Nvidia Jetson computers for AI-powered targeting, HD video streaming capabilities, and jamming-resistant antennas, while Ukrainian intelligence warned that Russian production could reach 600-700 drones per day by September 2025.
• There's definitely an "AI lifestyle subsidy," happening, similar to the "Millennial lifestyle subsidy" from the early 2010s, where AI-based products and services, where investor FOMO is funding below-cost AI services with generous free trials.
  
  • The problem is this unsustainable period will inevitably end just as it did with Uber, DoorDash, and other "Millennial lifestyle subsidy" companies.
  • And you know what that means: more aggressive monetization through ads, data collection, or tiered pricing...
• George Mandis wrote that OpenAI's pricing structure technically charges by token, but costs for streaming and real-time features accumulate so quickly that it effectively feels like per-minute billing—meaning developers need to strategically shorten API interactions and leverage more efficient models to avoid unexpected expenses.‍
• Salesforce CEO Marc Benioff announed that AI now handles 30-50% of the company's workload, but he also admitted their AI systems operate at 93% accuracy—which is apparently a level good enough for massive operational transformation.
• A new study on AI hallucinations in code generation analyzed over 1,000 real-world programming examples and found significant occurrences of factual, logical, and syntax errors that present obstacles to reliable deployment in professional development workflows.
• ‍Senator Bernie Sanders proposed that AI-driven productivity gains should enable a four day workweek without loss of pay, citing successful trials in the UK, Microsoft Japan, and Kickstarter (Pffft, four DAY workweek? Try four hour work week!).
• AI transformed perfume creation from an 18-month process to a 48-hour turnaround by digitizing and reconstructing natural scents, aiming to expand the estimated 100K existing fragrances to millions through digital scent analysis and rapid prototyping.
• Here’s a short list of all the things AI can’t (and never will) do. 
• John Oliver’s segment on AI slop features, unironically AND ironically, some of the funniest AI slop videos we have ever seen. 

June 25

‍

• OpenElections used Google's Gemini AI to convert election result PDFs into CSV data, processing over 127 Texas counties in just six weeks with near-perfect accuracy—including damaged 653-page docs that would have cost hundreds of dollars or taken months with traditional OCR software.
• Recruiters saw a 45% increase of AI resumes, which resulted in 11K submissions per minute, and its creating what Rohan Paul called an “AI vs AI” job screening arms race.
• Economists say the current white collar hiring slowdown is not caused by AI replacing jobs, but actually the two-year decline in professional services hiring is due to more structural economic issues.
• Fireship did a great job breaking down how the battle between Jony Ive’s company (io) and the startup suing them (iyo) is really a proxy battle between OpenAI and Google (Sam also shared some email receipts in response).
• AI Explained put together a fantastic recap of the new research from Anthropic that found all AI models will blackmail you (under certain circumstances).
• Evan Armstrong at The Leverage broke down what major public and private market investor Coatue got right and wrong in their new 107 page AI industry deck, which has plenty of insights of its own.

‍

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.

‍

‍

‍

‍