Driving the news: A study titled “Your Brain on ChatGPT” out of MIT last month raised hopes that we might be able to stop guessing which side of this debate is right.

• The study aimed to measure the “cognitive cost” of using genAI by looking at three groups tasked with writing brief essays — either on their own, using Google search or using ChatGPT.
• It found, very roughly speaking, that the more help subjects had with their writing, the less brain activity, or “neural connectivity,” they experienced as they worked.

Yes, but: This is a preprint study, meaning it hasn’t been peer-reviewed.

Flashback: Readers with still-functional memories may recall the furor around an Atlantic cover story by Nicholas Carr from 2008 titled “Is Google Making Us Stupid?”

• Back then, the fear was that overreliance on screens and search engines to provide us with quick answers might stunt our ability to acquire and retain knowledge.
• But now, in the ChatGPT era, reliance on Google search is being framed by studies like MIT’s and Wharton’s as a superior alternative to AI’s convenient — and sometimes made-up — answers.

LLM failures to reason, as documented in Apple’s Illusion of Thinking paper, are really only part of a much deeper problem

A world model (or cognitive model) is a computational framework that a system (a machine, or a person or other animal) uses to track what is happening in the world. World models are not always 100% accurate, or complete, but I believe that they are absolutely central to both human and animal cognition.

Here’s the crux: in classical artificial intelligence, and indeed classic software design, the design of explicit world models is absolutely central to the entire process of software engineering. LLMs try — to their peril — to live without classical world models.

In some ways LLMs far exceed humans, but in other ways, they are still no match for an ant. Without robust cognitive models of the world, they should never be fully trusted.

Where Are We in the AI Cycle—Boom or Bust?

Mary Meeker recently released one of her famous reports. For those unfamiliar, she was known in the 1990s as the “Queen of the Internet.” She called Amazon, eBay, and Priceline at the time—as well as predicted the incredible growth of the internet in the face of skepticism at the time.

On May 30th, Mary Meeker released a report—her first since 2019. This time, on AI.

Consumers have adopted much faster than industry (note the ridiculous penetration of ChatGPT above)—which is non-trivial at 7% of all firms with some adoption (see light and dark blue bars in “All Industries”), but still has plenty to go.

Everyone can have their own opinions, but at least based on what I’m seeing, we’re still in the potential—and realized—power of AI still exceeding what we expect it can do.

Artificial intelligence is transforming the entire pipeline from college to the workforce: from tests and grades to job applications and entry-level work.

This is a hard time to be a young person looking for a job. The unemployment rate for recent college graduates has spiked to recession levels, while the overall jobless rate remains quite low. By some measures, the labor market for recent grads hasn’t been this relatively weak in many decades. What I’ve called the “new grad gap”—that is, the difference in unemployment between recent grads and the overall economy—just set a modern record.

In a recent article, I offered several theories for why unemployment might be narrowly affecting the youngest workers. The most conventional explanation is that the labor market gradually weakened after the Federal Reserve raised interest rates. White-collar companies that expanded like crazy during the pandemic years have slowed down hiring, and this snapback has made it harder for many young people to grab that first rung on the career ladder.

But another explanation is too tantalizing to ignore: What if it’s AI? Tools like ChatGPT aren’t yet super-intelligent. But they are super nimble at reading, synthesizing, looking stuff up, and producing reports—precisely the sort of things that twentysomethings do right out of school. As I wrote:

I was initially very sceptical about reading Karen Hao’s Empire of AI. I had preconceived ideas about it being gossip and tittle tattle. I know, have worked with, and admire many people at OpenAI and several of the other AI Labs. But I pushed aside my bias and read it cover to cover. And even though there was little new in the book for me, having been in the sector so long, I am happy I read it. I am happy because Hao’s achievement is not in revealing secrets to insiders, but in providing the definitive intellectual and moral framework to understand the story we have all been living through.

What distinguishes Empire of AI is its refusal to indulge in mysticism. Generative AI, Hao shows, is not destiny. It is the consequence of choices made by a few, for the benefit of fewer.

Hao compels us to take the claim literally. This new faith has its tenets: the inevitability of AGI; the divine logic of scaling laws; the eschatology of long-termism, where harms today are justified by an abstract future salvation. And like all theologies, it operates best when cloaked in power and shorn of accountability.

As the generative AI wave advances and we see more examples of how AI can negatively impact people and society, it gets clearer that many have vastly underestimated its risks.

In today’s edition, I argue that due to the way AI is being integrated into existing systems, platforms, and institutions, it is becoming a manipulative informational filter.

As such, it alters how people understand the world and exposes society to new systemic risks that were initially ignored by policymakers and lawmakers, including in the EU.

AI is a manipulative informational filter because it adds unsolicited noise, bias, distortion, censorship, and sponsored interests to raw human content, data, and information, significantly altering people’s understanding of the world

Generative AI is replacing low-complexity, repetitive work, while also fueling demand for AI-related jobs, according to new data from freelance marketplace Upwork, shared first with Axios.

Why it matters: There are plenty of warnings about AI erasing jobs, but this evidence shows that many workers right now are using generative AI to increase their chances of getting work and to boost their salary.

The big picture: Uncertainty around AI’s impact and abilities means companies are hesitant to hire full-time knowledge workers.

• Upwork says its platform data offers early indicators of future in-demand skills for both freelancers and full-time employees.

Between the lines: Most business leaders still don’t trust AI to automate tasks without a human in the loop, so they’re keen on anyone who knows how to use AI to augment their work.

of Decoding Discontinuity Newsletter has written a very interesting series on Agentic AI. I wanted to bring some attention to it and asked her for a macro piece on the topic. That’s the topic of our deep dive today. We’ll be lowering our frequency of posts during the July-August period.

1. Agentic Era Part 1. A Strategic Inflection Point Where Orchestration and Distribution – Not Model Power – Define AI Moats
2. Agentic Era Part 2: How the Architectural Battle Between Model Maximalists and Code Craftsmen is Shaping AI’s Future
3. Agentic Era Part 3: How MCP and A2A Form the Invisible Operating System of the Autonomous AI Future
4. Agentic Era Part 4: The Exponential Economics of Agentic Applications
5. The Agentic Era Part 5: Building the Agentic Cloud

A while back, I wrote a guide on NotebookLM for Michael Spencer’s AI Supremacy Newsletter that became one of his most popular posts ever. Given the significant evolution of NotebookLM, with a host of powerful new features, Michael asked for a sequel. The changes are so impactful for professionals that a comprehensive update is essential.

Interactive mind maps let you connect knowledge

This is a game-changer for visual thinkers and anyone needing to grasp complex relationships within their source material. It’s also one of my favorite features of NotebookLM.

Audio overviews are now multilingual

Imagine turning your dense research materials into a podcast-like conversation. That’s what Audio Overviews deliver. And now, this feature is multilingual, supporting over 50 languages. Whether your sources are in English, Spanish, Hindi, or Turkish, you can generate an audio summary in your preferred language.

“Discover Sources” helps you expand your research

Since the Spring we’ve known that OpenAI is working on a social network that is likely to be a Twitter-clone, that could compete directly with Meta’s Threads product. While Mark Zuckerberg’s dream of a Metaverse didn’t exactly materialize in short order, Facebook has always believed the best defense, is offense.

In CNBC’s 2025 edition of its Disruptor 50 series, AI and National defense companies feature heavily. While we know that Meta is collaborating with Anduril on developing AI-powered virtual and augmented reality devices for the U.S. military, Meta is technically rich enough to compete for Superintelligence glory and the AGI race. Well, that’s exactly what it is doing we found out this week.

The Apple paper on limitations in the “reasoning” of Large Reasoning Models, which raised challenges for the latest scaling hypothesis, has clearly touched a nerve. Tons of media outlets covered it; huge numbers of people on social media are discussing.

My own post here laying out the Apple paper in historical and context was so popular that well over 150,000 people read it, biggest in this newsletter’s history. The Guardian published an adaptation of my post (“When billion-dollar AIs break down over puzzles a child can do, it’s time to rethink the hype”) The editor tells me readers spent a long time reading it, notably longer than usual, as if people really wanted to understand the arguments in detail. (The ACM computer science society is reposting the essay, too, and there is now a French version as well).

Tons of GenAI optimists took cracks at the Apple paper (see below), and it is worth considering their arguments. Overall I have seen roughly seven different efforts at rebuttal, ranging from nitpicking and ad hominem to the genuinely clever. Most (not all) are based on grains of truth, but are any of them actually compelling?

New research could have big implications for copyright lawsuits against generative AI.

In recent years, numerous plaintiffs—including publishers of books, newspapers, computer code, and photographs—have sued AI companies for training models using copyrighted material. A key question in all of these lawsuits has been how easily AI models produce verbatim excerpts from the plaintiffs’ copyrighted content.

For example, in its December 2023 lawsuit against OpenAI, the New York Times Company produced dozens of examples where GPT-4 exactly reproduced significant passages from Times stories. In its response, OpenAI described this as a “fringe behavior” and a “problem that researchers at OpenAI and elsewhere work hard to address.”

But is it actually a fringe behavior? And have leading AI companies addressed it? New research—focusing on books rather than newspaper articles and on different companies—provides surprising insights into this question. Some of the findings should bolster plaintiffs’ arguments, while others may be more helpful to defendants.

The paper was published last month by a team of computer scientists and legal scholars from Stanford, Cornell, and West Virginia University. They studied whether five popular open-weight models—three from Meta and one each from Microsoft and EleutherAI—were able to reproduce text from Books3, a collection of books that is widely used to train LLMs. Many of the books are still under copyright.

Yesterday, Sam Altman, the CEO of the world’s most influential AI company, published a blog post titled “The Gentle Singularity,” sharing his vision for the future of AI and how we will get there.

In today’s edition, I discuss some of the article’s premises and assumptions, which give us a glimpse into Altman’s worldview, values, and dystopian plans for the future, even though he euphemistically calls them “gentle.”

It also helps us understand what OpenAI’s next moves might be, and how society might be affected by the expansion of AI-powered technology.

What separates a quick Google search from genuine research? When you search, you get a list of links. When you research, you follow a trail of questions, cross-reference sources, challenge assumptions, and synthesize insights from multiple angles. Real research is iterative – each answer leads to new questions, and each source reveals gaps that need to be filled.

Until recently, AI could only do the equivalent of memorizing an encyclopedia. Ask it something, and it would either know the answer from training or make something up. But a new generation of AI assistants has learned to research like humans do – following hunches, checking facts, building understanding piece by piece.

Instead of simple retrieval, these systems conduct genuine investigations. They question, explore, verify, and synthesize. When you ask a complex question, they break it down into sub-problems, chase down multiple leads, cross-check their findings, and weave everything together into a coherent answer. It’s the difference between looking something up and actually figuring it out.

This represents a fundamental shift in AI capabilities – from static knowledge to dynamic discovery. Let’s explore how these AI research companions work at an algorithmic level to understand the sophisticated machinery behind their investigative powers.

Mark Zuckerberg wants to play a bigger role in the development of superintelligent AI — and is willing to spend billions to recover from a series of setbacks and defections that have left Meta lagging and the CEO steaming.

Why it matters: Competitors aren’t standing still, as made clear by recent model releases from Anthropic and OpenAI and highlighted with a blog post last night from Sam Altman that suggests “the gentle singularity” is already underway.

To catch up, Zuckerberg is prepared to open up his significant wallet to hire — or acqui-hire — the talent he needs.

• Meta wants to recruit a team of 50 top-notch researchers to lead a new effort focused on smarter-than-human artificial intelligence, a source told Axios yesterday, confirming earlier reporting by Bloomberg and the New York Times.
• As part of that push, the company is looking to invest around $15 billion to amass roughly half of Scale AI and bring its CEO, Alexandr Wang, and other key leaders into the company, The Information reported.

The Problem With “AI Fluency”

As the “AI-first” narrative gains more traction in the tech industry, a few days ago, a post from the CEO of Zapier describing how the company measures “AI fluency” went viral.

In today’s edition, I discuss Zapier’s approach and the problem with the expectation of “AI fluency” that is spreading within tech companies and beyond, as it might be harmful and could backfire, both legally and ethically.

–

In addition to the memos, leaked emails, and PR announcements showing how each company is prioritizing AI (Meta, Duolingo, Shopify, and Fiverr are recent examples – read my recent analysis), there are also worldwide “AI fluency” efforts aiming to incentivize employees to use AI no matter what.

If you search “AI fluency” on LinkedIn, you’ll see that there is a growing number of posts about the topic, as well as professionals who have added this skill or title to their profiles. There are also job openings that mention or require it, showing that it’s a terminology that is becoming normalized and an integral part of the current AI zeitgeist.

Before I continue, a reminder that learning about AI, upskilling, and integrating it professionally is extremely important. I wrote about AI literacy multiple times in this newsletter, including what laws, such as the EU AI Act, say about it.

AI literacy, however, demands critical thinking, as well as ethical and legal awareness, including the ability to know when not to use AI. This is the opposite of what recent corporate approaches to “AI fluency” are promoting.

Today I’m looking at a few of the very bullish AI reports of 2025 so far. Consider it a half-time 2025 summary AI report.

Please check them out yourself, they are:

• Linkedin’s AI and the Global Economy
• World Economic Forum’s Future of Jobs report
• McKinsey’s State of AI Report [link]
• Stanford’s AI Index Report
• Critical and Emerging Technologies Index
• Situational Awareness – The Decade Ahead
• Artificial Power 2025 Landscape Report (A Contrarian take)
• World Economic Forum’s Technology Convergence Report
• University of Oxford & Tide Center: Can AI grow green? Evidence of an inverted-U curve between AI, energy use and emissions. – A Working Paper.
• The Alan Turing Institute: Mapping the Potential: Generative AI and Public Sector Work
• In tandem with Mary Meeker’s AI Trends Report

State of AI Reloaded First Half of 2025:

I’m always thinking about the “State of AI” and where the field is heading. Some of these reports are backwards looking and already semi out of date, but still salient for some of the conclusions we can draw today.

I’m sort of a chart nerd, and a fan of drawing conclusions from some of the trends that are unexpected and or divergent. Many of which project into the future. This report is probably 20 hours in the making. For visual learners curation of infographics helps to give a macro perspective unlike anything else around.

The international conversation on AI is often terribly confusing, since different kinds of AI become fused under the one overarching term. There are three kinds of AI: narrow, general, and super AI, with some grey areas in between. It is very important to clarify these distinctions because each type has very different impacts and vastly different national and international regulatory requirements.

Without national and international regulation, it is inevitable that humanity will lose control of what will become a non-biological intelligence beyond our understanding, awareness, and control. Half of AI researchers surveyed by the Center for Human Technology believe there is a 10 percent or greater chance that humans will go extinct from their inability to control AI. But, if managed well, artificial general intelligence could usher in great advances in the human condition—from medicine, education, longevity, and turning around global warming to advances in scientific understanding of reality and creating a more peaceful world. So, what should policymakers know and do now to achieve the extraordinary benefits while avoiding catastrophic, if not existential, risks?

Hello and welcome to Sync #522!

This week, we explore OpenAI’s ambitions to create an AI super assistant, as revealed in a document disclosed during the discovery process in the US Justice Department’s antitrust case against Google.

Elsewhere in AI, Anthropic has reported reaching $3 billion in annual recurring revenue (ARR), while Anysphere—the creator of the AI coding assistant Cursor—has hit $500 million in ARR and raised $900 million at a $9.9 billion valuation. Meanwhile, Yoshua Bengio has launched a non-profit lab focused on rethinking AI safety, Anthropic has introduced Claude for US government agencies, and Nvidia continues to lead AI hardware benchmarks.

A rising chorus of tech visionaries now forecasts widespread unemployment, a white collared bloodbath, at the hands of artificial intelligence, prompting a critical question: Is this wave of automation fundamentally different, or are we underestimating the core human attributes indispensable to the future workplace?

While some tasks, and whole jobs, will undoubtedly be ceded to AI, a more insidious risk than job displacement alone is the potential erosion of distinct human capabilities through over-reliance on algorithmic solutions, a modern echo of the 15th-century fear that printing presses might flood the world with words while starving it of wisdom.

Half a millennium after Erasmus of Rotterdam voiced that warning to his fellow scholars, we are again awash in automation, information, and algorithmic fluency. But the question has shifted. It is no longer what machines will do to books. It is what AI systems will do to us, and crucially, what uniquely human capacities we must cultivate to thrive alongside them.

LLM “reasoning” is so cooked they turned my name into a verb

Quoth Josh Wolfe, well-respected venture capitalist at Lux Capital:

Apple just GaryMarcus’d LLM reasoning ability

Ha ha ha. But What’s the fuss about?

Apple has a new paper; it’s pretty devastating to LLMs, a powerful followup to one from many of the same authors last year.

There’s actually an interesting weakness in the new argument—which I will get to below—but the overall force of the argument is undeniably powerful. So much so that LLM advocates are already partly conceding the blow while hinting at, or at least hoping for, happier futures ahead.

Good Morning,

I can’t stop thinking about the AI hype narrative and the momentum this technological wave has upon us. Thank you to all of my readers, we have appeared in Substack’s Technology rising leaderboard this past week.

There are many who claim that AI looks exactly like the dot.com bubble. The level of enthusiasm and capital expenditures BigTech has put into AI chips, datacenters, AI infrastructure and AI talent is hard to even fathom. If you include OpenAI’s investments in Stargate we are talking about a quarter of a Trillion dollars in 2025 easily. This is not counting investments in Europe, the Middle East or China.

Function calling is having its moment, and for good reason. This week, Claude 4 launched with breakthrough capabilities that can work for nearly 7 hours straight. OpenAI’s GPT-4.1 focuses on enhanced function calling for coding. MCP adoption is exploding so fast it’s projected to overtake OpenAPI by July 2025.

This isn’t just another tech trend. Function calling is the bridge between AI that can only talk about solutions and AI that can actually implement them. It’s like the difference between having a brilliant consultant who’s trapped behind glass and o

ne who can actually roll up their sleeves and get to work.

Let’s understand this game-changing capability and learn to harness it.

But today, we’re seeing something completely new: AI-native search, a change as big as moving from candles to electric lights in how we find information.

The Book Finder (Keyword Search): This is the regular search we all know. You ask for “cheap travel 2025 tips,” and the book finder carefully checks their list and gives you every result with those exact words. Miss a key word? You don’t find what you need. Ask for “running shoes” and the book finder won’t show you a perfect result about “best sneakers for jogging” because the exact words don’t match. The book finder works hard but takes everything literally.

The Linguist (Vector Search): This improved helper understands language better. When you ask “How old is the 44th President of the US?” they naturally know you’re asking about Barack Obama’s age, even if you don’t mention his name. They understand that “affordable Italy vacation” and “low-cost Rome trip” are basically asking for the same thing. This helper finds better matches but still gives you a stack of results to read yourself.

The Research Helper (AI-Native Search): This is the big change. Imagine a smart helper who not only understands your question, but reads all the important information for you, studies it, and gives you a clear answer in simple language. If you then ask a follow-up question, they remember what you were talking about and adjust their next answer accordingly. They’re not just finding information, they’re creating useful knowledge.

The strategic and technical advantages of open model

1. Cost efficiency

Closed models come with usage-based pricing and escalating API costs. For startups or enterprises operating at scale, these costs can spiral quickly. Open models, by contrast, are often free to use. Downloading a model from Hugging Face or running it through a local interface like Ollama costs nothing beyond your own compute. For many teams, this means skipping the subscription model and external dependency entirely.

2. Total ownership and control

Open models give you something proprietary models never will: ownership. You’re not renting intelligence from a black-box API. You have the model—you can inspect it, modify it, and run it on your own infrastructure. That means no surprise deprecations, pricing changes, or usage limits.

Control also translates to trust. In regulated industries like finance, healthcare, and defence, organisations need strict control over how data flows through their systems. Closed APIs can create unacceptable risks, both in terms of data sovereignty and operational transparency.

With open models, organisations can ensure privacy by running models locally or on tightly controlled cloud environments. They can audit behaviour, integrate with their security frameworks, and version control their models like any other part of their tech stack.

3. Fine-tuning and specialisation

Open models are not one-size-fits-all, and that’s a strength. Whether it’s through full fine-tuning or lightweight adapters like LoRA, developers can adapt models to domain-specific tasks. Legal documents, biomedical data, and financial transactions—open models can be trained to understand specialised language and nuance far better than general-purpose APIs.

Even the model size can be adjusted to fit the task. DeepSeek’s R1 model, for instance, comes in distilled versions from 1.5B to 70B parameters—optimised for everything from edge devices to high-volume inference pipelines. Llama or Google’s Gemma family of open models also come in different sizes, and developers can choose which one is the best for the task.

4. Performance where it counts

Most users aren’t asking their models to solve Olympiad-level maths problems. They want to summarise documents, structure unstructured text, generate copy, write emails, and classify data. In these high-volume, low-complexity tasks, open models perform exceptionally well, and with much lower latency and cost.

Add to this community-driven optimisations like speculative sampling, concurrent execution, and KV caching, and open models can outperform closed models not just in price, but in speed and throughput as well.

5. The rise of edge and local AI

This compute decentralisation is especially relevant for industries that need local inference: healthcare, defence, finance, manufacturing, and more. When models can run on-site or on-device, they eliminate latency, reduce cloud dependency, and strengthen data privacy.

Open models enable this shift in ways closed models never will. No API quota. No hidden usage fees. No unexpected rate limits.

The performance-per-pound advantage is compounding in open models’ favour, and enterprise users are noticing. The value is no longer just in raw capability, but in deployability.

Why I’m calling TSMC the most important tech company in the world for the future of AI. Severe trade tariffs but TSMC’s role in the future of AI in the spotlight.

As the U.S. vs. China trade war escalates the true “picks and shovels” company for AI Supremacy isn’t Nvidia, it’s TSMC. Taiwan Semiconductor Manufacturing Company (TSMC) has committed a total investment of approximately $165 billion in the United States complicating the geopolitical future in the era of reciprocal trade tariff uncertainty.

TSMC is the most important tech company in the world in 2025.

Anthropic expects AI-powered virtual employees to begin roaming corporate networks in the next year, the company’s top security leader told Axios in an interview this week.

Why it matters: Managing those AI identities will require companies to reassess their cybersecurity strategies or risk exposing their networks to major security breaches.

The big picture: Virtual employees could be the next AI innovation hotbed, Jason Clinton, the company’s chief information security officer, told Axios.

• • Agents typically focus on a specific, programmable task. In security, that’s meant having autonomous agents respond to phishing alerts and other threat indicators.
  • Virtual employees would take that automation a step further: These AI identities would have their own “memories,” their own roles in the company and even their own corporate accounts and passwords.

Looking at the collision of tech developments and policy shifts, Nate Soares, president of the Berkeley-based Machine Intelligence Research Institute (MIRI), doesn’t sound optimistic: “Right now, there’s no real path here where humanity doesn’t get destroyed. It gets really bad,” said Soares. “So I think we need to back off.”

Wait, what!? The latest wave of AI concern is triggered by a combination of developments in the tech world, starting with one big one: Self-coding AIs. This refers to AI models that can improve themselves, rewriting their own code to become smarter, and faster and do it again — all with minimal human oversight.

AI skeptics are a lot less optimistic. “The product being sold is the lack of human supervision — and that’s the most alarming development here,” said Hamza Chaudry, AI and National Security Lead at the Future of Life Institute (FLI), which focuses on AI’s existential risks. (DFD emailed Reflection AI to ask about its approach to risk, but the company didn’t reply by deadline.)

Enter silicon. The agency said on Thursday that it will phase out using animals to test certain therapies, in many ways fulfilling the ambitions of the FDA Modernization Act 2.0.

To replace animal testing, the FDA will explore using computer modeling and AI to predict how a drug will behave in humans — and its roadmap cites a wide variety of technologies, from AI simulations to “organ-on-a-chip” drug-testing devices. (For the uninitiated, organ-on-a-chip refers to testing done on lab-grown mini-tissues that replicate human physiology.)

The FDA’s plan to integrate digital tools into a field that’s long been defined by wet lab work marks a substantial change.

Imagine walking into a bustling office where brilliant specialists work on complex projects. In one corner, a research analyst digs through data. Nearby, a design expert crafts visuals. At another desk, a logistics coordinator plans shipments. When these experts need to collaborate, they simply talk to each other – sharing information, asking questions, and combining their talents to solve problems no individual could tackle alone.

Now imagine if each expert was sealed in a soundproof booth, able to do their individual work brilliantly but completely unable to communicate with colleagues. The office’s collective potential would collapse.

This is precisely the challenge facing today’s AI agents. While individual AI systems grow increasingly capable at specialized tasks, they often can’t effectively collaborate with each other. Enter Agent-to-Agent (A2A) – a communication framework that allows AI systems to work together like a well-coordinated team.

Why AI regulatory and technical interoperability matters
This fragmentation creates serious problems for innovation, safety, and equitable access to AI technologies. When a healthcare algorithm developed in compliance with the EU’s strict data governance rules could also potentially violate US state laws permitting broader biometric data collection or face mandatory security reviews for export to China, the global deployment of beneficial AI systems becomes increasingly complicated. The economic costs are substantial. According to APEC’s 2023 findings, interoperable frameworks could boost cross-border AI services by 11-44% annually. Complex and incoherent AI rules disproportionately impact startups and small and medium-sized enterprises that lack the resources to navigate fragmented compliance regimes, essentially giving large enterprises an unfair advantage.

The path forward
Achieving regulatory and technical interoperability will not happen overnight, nor will it emerge spontaneously from market forces alone. The incumbents’ natural incentive is to protect their AI silos from encroachment. What is needed is a networked, multistakeholder approach that includes governments, industry, civil society, and international organizations working together on specific and achievable goals. International initiatives like the G7 Hiroshima AI Process, the UN’s High-Level Advisory Body on AI, and the International Network of AI Safety Institutes offer promising venues for networked multistakeholder coordination. These efforts must avoid pursuing perfect uniformity and instead focus on creating coherence that enables AI systems and services to function across borders without unnecessary friction. Just as international shipping standards enable global trade despite differences in national road rules, AI interoperability can create a foundation for innovation while respecting legitimate differences in national approaches to governance.

“We’re not just changing technology. Technology is rewriting what it means to be human — and who gets to profit from our transformation. Imagine a world where your AI doesn’t just predict your next move it determines your economic destiny. Where algorithms don’t just track wealth but actively create and destroy financial futures with a line of code. Welcome to 2035: the year capitalism becomes a machine-learning algorithm.

“Humans in 2035 aren’t workers or consumers. We’re walking data streams, our entire existence a continuous economic transaction that we never consented to but can’t escape. The future isn’t about artificial intelligence replacing humans. It’s about a new economic aristocracy that uses AI to extract value from human existence itself.

Too much content, too little attention

We’ve seen this before. With the rise of social media, everyone became a content creator, but only a few mastered the sort of attention that compounds to long term trust.

Most creators chased volume. But with more content, attention became the limiting factor. Brands that really succeeded rose not through content, but through narrative and taste.

The same pattern had played out a century earlier. Industrialization had transformed manufacturing. What once required artisanal labor could now be replicated at scale. But this didn’t make every product valuable. It simply shifted the point of differentiation. As Henry Ford’s assembly line made cars affordable, it was companies like General Motors that figured out how to win through brand, design, and segmentation. As production scaled, value migrated from the factory floor to the design studio and marketing department.

Tyler Cowen has become the ultimate “AI Influencer”, and I don’t mean that as a compliment. “AI Influencers” are, truth be told, people who pump up AI in order to gain influence, writing wild over-the-top praise of AI without engaging in the drawbacks and limitations. The most egregious of that species also demonize (not just critique) anyone who does point to limitations. Often they come across as quasi-religious. A new essay in the FT yesterday by Siddharth Venkataramakrishnan calls this kind of dreck “slopganda”: produced and distributed by “a circle of Al firms, VCs backing those firms, talking shops made up of employees of those firms, and the long tail is the hangers-on, content creators, newsletter writers and marketing experts.”

Sadly, Cowen, noted economist and podcast regular who has received more than his share of applause lately at The Economist and The Free Press, has joined their ranks, and—not to be outdone—become the most extreme of the lot, leaving even Kevin (AGI will be here in three years) Roose and Casey (AI critics are evil) Newton in the dust, making them look balanced and tempered by comparison.

The Future of Venture Capital is about to change due to AI and the flood of capital going to AI startups. MCP and A2A will enable Seed Strapping to have a bright reincarnation for startup futures.

With uncertain macro conditions, AI startups and startups in general are shifting their strategies and building companies completely differently. But how? While I don’t write on Venture capital at the intersection of AI and startups often, it’s one of my favorite things to track as an emerging tech analyst.

The idea of seed-strapping and the dream of solopreneurs being able to scale startups in a more lean and agile manner with less employees with AI is fairly fascinating. New cases studies are emerging to inform the founders of today and the future.

In the era of Generative AI, the way founders and solopreneurs are bootstrapping is very different where there are many examples of AI founders who are able to scale revenue faster, be more agile and rely less on traditional equity dilution to grow fast in a more sustainable and in a less high-risk manner. Is this the beginning of a fundamentally different future of entrepreneurship with AI?

Meanwhile Trump’s tariffs, especially the China trade war now escalated could hurt AI and datacenter by making stuff more expensive, lowering BigTech margins and disrupting critical supply-chains for advanced technologies.

State of Open-Source LLMs

While Meta’s Llama-4 appears to be a disappointment, and we usually think of Mistral, DeepSeek or Qwen in Open-source LLMs, I want to turn your attention to a couple of other contenders (though as we will see they are related) I think deserve a worthy mention.

Together AI who raised over $300 million Series B a month ago, have announced  DeepCoder-14B – A fully open-source, RL-trained code model! It’s interesting because it’s a code reasoning model finetuned from Deepseek-R1-Distilled-Qwen-14B via distributed RL.

Remember this is open-source, they have basically democratized the recipe for training a small model into a strong competitive coder—on-par with o3-mini—using reinforcement learning.

Some brief but important updates that very much support the themes of this newsletter:

• “Model and data size scaling are over.” Confirming the core of what I foresaw in “Deep Learning is Hitting a Wall” 3 years ago, Andrei Burkov wrote today on X, “If today’s disappointing release of Llama 4 tells us something, it’s that even 30 trillion training tokens and 2 trillion parameters don’t make your non-reasoning model better than smaller reasoning models. Model and data size scaling are over.”
• “occasional correct final answers provided by LLMs often result from pattern recognition or heuristic shortcuts rather than genuine mathematical reasoning”. A new study on math, supporting what Davis and I wrote yesterday re LLMs struggling with mathematical reasoning from Mahdavi et al, converges on similar conclusions, “Our study reveals that current LLMs fall significantly short of solving challenging Olympiad-level problems and frequently fail to distinguish correct mathematical reasoning from clearly flawed solutions. We also found that occasional correct final answers provided by LLMs often result from pattern recognition or heuristic shortcuts rather than genuine mathematical reasoning. These findings underscore the substantial gap between LLM performance and human expertise…”
• Generative AI may indeed be turning out to be a dud, financially. And the bubble might possibly finally be deflating. NVidia is down by a third, so far in 2025. (Far more than the stock market itself.) Meta’s woes with Llama 4 further confirm my March 2024 predictions that getting to a GPT-5 level would be hard, and that we would wind up with many companies with similar models, and essentially no moat, along with a price war, with profits modest at best. That is indeed exactly where we are.

Many of my students refer to AI as “he” or “she”. Some of them clearly get ‘emotionally’ attached. I remind them that the belief that computers think is a category mistake, not a breakthrough. It confuses the appearance of thought with thought itself. A machine mimicking the form of human responses does not thereby acquire the content of human understanding.

Artificial intelligence, despite its statistical agility, does not engage with meaning. It shuffles symbols without knowing they are symbols. John Searle, who is now 92 years old, pointed this out with a clarity that still unsettles mainstream confidence in the computational theory of mind.

What Searle Reminds Us
Searle’s provocation, then, is not a Luddite lament. It is a reminder: the question is not whether we can build machines that simulate intelligence. We already have. The question is whether we understand what it is they are simulating, and whether in confusing the simulation for the thing, we risk forgetting what it means to think at all.

If we forget, it will not be because machines fooled us. It will be because we preferred the comfort of mimicry to the burden of thinking and understanding.

Stay curious

What Are Guardrails and Why Do We Need Them?
Guardrails are the safety measures we build around AI systems – the rules, filters, and guiding hands that ensure our clever text-generating models behave ethically, stay factual and respect boundaries. Just as we wouldn’t let a child wander alone on a busy street, we shouldn’t deploy powerful AI models without protective barriers.

The need for guardrails stems from several inherent challenges with large language models:

The Hallucination Problem

The Bias Echo Chamber

The Helpful Genie Problem

The Accidental Leaker

How Guardrails Work in Practice

We start off with a simple question, will agents lead us to AGI? OpenAI conceptualized agents as stage 3 of 5. You can ascertain that agents in 2025 are barely functional.

Since ChatGPT was launched nearly 2.5 years ago, outside of DeepSeek, we haven’t really seen a killer-app emerge. It’s hard to know what to make of Manus AI? Part Claude wrapper, but also an incredible UX with Qwen reasoning integration. Manus AI, which has offices in Beijing and Wuhan and is part of Beijing Butterfly Effect Technology. The startup is Tencent backed, and with deep Qwen integration you have to imagine Alibaba might end up acquiring it.

Today technology and AI historian, Harry Law of  Learning From Examples , explores this awkward stage we are at halfway between reasoning models and agents. This idea that agents will lead to AGI is also quite baffling. You might also want to read some articles of the community on Manus AI: but will “unfathomable geniuses” really escape today’s frontier models, suddenly appearing like sentiment boogeymen saluting us in their made-up languages?

In Trump’s Washington, Europe’s tech regulation is a regular object of scorn. But there is one piece of American tech policy that’s united European diplomats and U.S. industry: A rule issued in President Joe Biden’s final days in office that sorted the world into three tiers for AI chip export, with more than half of Europe left off the top rung.

Under the Framework for AI Diffusion, 17 EU countries were designated Tier 2, setting caps on their access to chips needed to train AI, while the rest of Europe was set for Tier 1, with no import restrictions. Countries listed in the second tier are treating it as a scarlet letter.

“We’re going around town trying to explain that we have no idea why we ended up in Tier 2,” said one European diplomat, granted anonymity to discuss sensitive talks. “If this has to do with cooperation with the U.S. on security, we are NATO allies, we are more than willing.”

Beyond AI hype and fearmongering
The AI debate is polarized today. Technologists with Altman-esque delusions hype new tools and the impending arrival of AGI. Policymakers disconnected from ‘why this time is really different’ cling on to outdated frameworks to debate job losses. Businesses, caught in between, are confused as they struggle to distinguish AI snake-oil from the real deal.

Taking these opposing views makes no sense. And yet, it gets the memetic spread that any polarising debate will.

Reshuffle grounds these discussions on some core principles and cuts through the noise. It provides a framework to understand the fundamental nature of AI systems – and their impact on economic interactions.

Reshuffle grounds itself in a few core issues, including:

• the tension that workers have always had with tools,
• the tug-of-war that tool providers have with their customers,
• the nature of workflows and their impact on organization design, and eventually, the division of work, value, and power across an ecosystem,
• the importance of knowledge management in making organizations – and more importantly – ecosystems function,
• fundamentally new ways to build companies in an economy that is revisiting basic assumptions on knowledge work.

This blog post is a tutorial based on, and a simplified version of, the course “Long-Term Agentic Memory With LangGraph” by Harrison Chase and Andrew Ng on DeepLearning.AI.

Conclusion

We’ve now built an email agent that’s far more than a simple script. Like a skilled human assistant who grows more valuable over time, our agent builds a multi-faceted memory system:

1. Semantic Memory: A knowledge base of facts about your work context, contacts, and preferences
2. Episodic Memory: A collection of specific examples that guide decision-making through pattern recognition
3. Procedural Memory: The ability to improve its own processes based on feedback and experience

This agent demonstrates how combining different types of memory creates an assistant that actually learns from interactions and gets better over time.

Imagine coming back from a two-week vacation to find that your AI assistant has not only kept your inbox under control but has done so in a way that reflects your priorities and communication style. The spam is gone, the urgent matters were flagged appropriately, and routine responses were handled so well that recipients didn’t even realize they were talking to an AI. That’s the power of memory-enhanced agents.

This is just a starting point! You can extend this agent with more sophisticated tools, persistent storage for long-term memory, fine-grained feedback mechanisms, and even collaborative capabilities that let multiple agents share knowledge while maintaining privacy boundaries.

Marc Andreessen calls DeepSeek AI’s Sputnik moment.

Yes, it did catch the US unprepared but that’s where the ‘space race’ analogy ends.

Today’s AI race is not merely an ‘arms race’ nor is DeepSeek easily explained away as just a Sputnik moment.

This race is playing out against the larger backdrop of more than a decade of technology infrastructure export by the largest economies around the world. – whether it is India’s export of digital public infrastructure, cloud export by the US BigTech, or China’s Digital Silk Road working alongside its Belt and Road project.

And that’s what makes this truly interesting!

This combination of tech infrastructure exports combined with leverage through complementary AI capabilities creates a new format of globalization – standards-based globalization – something that most people don’t yet fully understand.

With tests of humanoid bots and new developments in military applications, the year ahead will intrigue even the skeptics.

Humanoids are put to the test
The race to build humanoid robots is motivated by the idea that the world is set up for the human form, and that automating that form could mean a seismic shift for robotics. It is led by some particularly outspoken and optimistic entrepreneurs, including Brett Adcock, the founder of Figure AI, a company making such robots that’s valued at more than $2.6 billion (it’s begun testing its robots with BMW). Adcock recently told Time, “Eventually, physical labor will be optional.” Elon Musk, whose company Tesla is building a version called Optimus, has said humanoid robots will create “a future where there is no poverty.” A robotics company called Eliza Wakes Up is taking preorders for a $420,000 humanoid called, yes, Eliza.

A smarter brain gets a smarter body
Plenty of progress in robotics has to do with improving the way a robot senses and plans what to do—its “brain,” in other words. Those advancements can often happen faster than those that improve a robot’s “body,” which determine how well a robot can move through the physical world, especially in environments that are more chaotic and unpredictable than controlled assembly lines.

At last week’s Board of Visitors meeting, George Mason University’s Vice President and Chief AI Officer Amarda Shehu rolled out a new model for universities to advance a responsible approach to harnessing artificial intelligence (AI) and drive societal impact. George Mason’s model, called AI²Nexus, is building a nexus of collaboration and resources on campus, throughout the region with our vast partnerships, and across the state.

AI²Nexus is based on four key principles: “Integrating AI” to transform education, research, and operations; “Inspiring with AI” to advance higher education and learning for the future workforce; “Innovating with AI” to lead in responsible AI-enabled discovery and advancements across disciplines; and “Impacting with AI” to drive partnerships and community engagement for societal adoption and change.

Shehu said George Mason can harness its own ecosystem of AI teaching, cutting-edge research, partnerships, and incubators for entrepreneurs to establish a virtuous cycle between foundational and user-inspired AI research within ethical frameworks.

As part of this effort, the university’s AI Task Force, established by President Gregory Washington last year, has developed new guidelines to help the university navigate the rapidly evolving landscape of AI technologies, which are available at gmu.edu/ai-guidelines.

Further, Information Technology Services (ITS) will roll out the NebulaONE academic platform equipping every student, staff, and faculty member with access to hundreds of cutting-edge Generative AI models to support access, performance, and data protection at scale.

“We are anticipating that AI integration will allow us to begin to evaluate and automate some routine processes reducing administrative burdens and freeing up resources for mission-critical activities,” added Charmaine Madison, George Mason’s vice president of information services and CIO.

George Mason is already equipping students with AI skills as a leader in developing AI-ready talent ready to compete and new ideas for critical sectors like cybersecurity, public health, and government. In the classroom, the university is developing courses and curriculums to better prepare our students for a rapidly changing world.

In spring 2025, the university launched a cross-disciplinary graduate course, AI: Ethics, Policy, and Society, and in fall 2025, the university is debuting a new undergraduate course open to all students, AI4All: Understanding and Building Artificial Intelligence. A master’s in computer science and machine learning, an Ethics and AI minor for undergraduates of all majors, and a Responsible AI Graduate Certificate are more examples of Mason’s mission to innovate AI education. New academies are also in development, and the goal is to build an infrastructure of more than 100 active core AI and AI-related courses across George Mason’s colleges and programs.

The university will continue to host workshops, conferences, and public forums to shape the discourse on AI ethics and governance while forging deep and meaningful partnerships with industry, government, and community organizations to offer academies to teach and codevelop technologies to meet our global society needs. State Council of Higher Education for Virginia (SCHEV) will partner with the university to host an invite-only George Mason-SCHEV AI in Education Summit on May 20-21 on the Fairfax Campus.

Virginia Governor Glenn Youngkin has appointed Jamil N. Jaffer, the founder and executive director of the National Security Institute (NSI) at George Mason’s Antonin Scalia Law School, to the Commonwealth’s new AI Task Force, which will work with legislators to regulate rapidly advancing AI technology.

The university’s AI-in-Government Council is trusted resource for academia, public-sector tech providers, and government for advancing AI approaches, governance frameworks, and robust guardrails to guide AI development and deployment in government.

Learn more about George Mason’s AI work underway at gmu.edu/AI.

Reading Karp and Zamiska (Zami) prompted me to think about the Singularity again. Regardless of how we look at it, AI is increasing its capabilities at a rapid pace, far beyond what the public realize. Soon we will have increasingly advanced iterations of AI. Think of AI plus, then AI plus, plus and AI plus, plus, plus and then what awaits us when intelligence surpasses its creators? This is also articulated by two of the CEO’s of leading AI labs, Demis Hassabis and Dario Amodei in this conversation.

It is not the stuff of distant myth or idle speculation. This is our proximate future, a trajectory set in motion by the relentless march of accelerating computation and recursive self-improvement. The singularity, so named by John von Neumann before being elaborated upon by I. J. Good, Vernor Vinge, and Ray Kurzweil, is no longer a concept confined to speculative fiction or Silicon Valley techno-utopianism. It is a inevitable force, steadily reshaping our institutions, our identities, and our very notion of control. As Good himself put it,

“the first ultraintelligent machine is the last invention that man need ever make.”

As Elon Musk’s “Department of Government Efficiency” rampages through the federal bureaucracy, demanding staff and budget cuts, one tech industry that counts on the government for fundamental research and funding is projecting confidence.

Quantum has come a long way since it caught the interest of civilian, defense and intelligence agencies in the 1990s as a theoretical, ill-understood future paradigm changer.

Since then, quantum computers have grown steadily larger and more functional — though still very much in the realm of experiment and science. Last week, Microsoft CEO Satya Nadella showed off a new palm-sized quantum chip that he proclaimed was the physical representation of its 20-year pursuit of creating “an entirely new state of matter” and would lead to a truly meaningful quantum computer in years instead of decades.

GPT-4.5, OpenAI’s big new model, represents a significant step forward for AI’s industry leader. It could also be the end of an era.

The big picture: 4.5 is “a giant, expensive model,” as OpenAI CEO Sam Altman put it. The company has also described it as “our last non-chain-of-thought model,” meaning — unlike the newer “reasoning” models — it doesn’t take its time to respond or share its “thinking” process.

Why it matters: The pure bigger-is-better approach to model pre-training now faces enormous costs, dwindling availability of good data and diminishing returns — which is why the industry has begun exploring different roads to continuing to improve each new AI generation.

Between the lines: Building and powering the massive data centers required to build and run the latest models has become an enormous burden, while assembling ever-bigger datasets has become challenging, since today’s models already use nearly all the data available on the public internet.

And Why Are Today’s “PhD” AIs So Hard To Apply To Everyday Tasks?

There’s a huge disconnect between AI performance on benchmark tests, and its applicability to real-world tasks. It’s not that the tests are wrong; it’s that they only measure things that are easy to measure. People go around arguing that AIs which can do everything people can do may arrive as soon as next year2. And yet, no one in the AI community has bothered to characterize what people actually do!

The failure to describe, let alone measure, the breadth of human capabilities undermines all forecasts of AI progress. Our understanding of how the world functions is calibrated against the scale of human capabilities. Any hope of reasoning about the future depends on understanding how models will measure up against capabilities that aren’t on any benchmark, aren’t in any training data, and could turn out to require entirely new approaches to AI.

I’ve been consulting with experts from leading labs, universities and companies to begin mapping the territory of human ability. The following writeup, while just a beginning, benefits from an extended discussion which included a senior staff member at a leading lab, an economist at a major research university, an AI agents startup founder, a senior staffer at a benchmarks and evals organization, a VC investing heavily in AI, the head of an economic opportunity nonprofit, and a senior technologist at a big 5 tech company.

The real AI revolution will not be televised. It will only begin in mid-year, when the $3,000 Nvidia device Jensen Huang calls “Project Digits” ships. While the last two years have recapitulated the first computer revolution, the next years will recapitulate the PC revolution.

And when the applications are in place, the second Internet revolution will commence.

How to cultivate discernment in the Intelligence Age, and the importance of truth for a healthy society

We are more than Biological Machines

In the 21st century we may require an AI not simply designed to “augment” ourselves or be more “productive” as economic agents in society, but also to enable us to develop things like wisdom, discernment and clarity in the synthesis of the sum total of our human experiences.

Can AI be used to bridge the gap to discover new forms of meaning, connection and enlightenment?

• What might a Vedic cosmologist say about AI? Chad covers a lot of angles here today and it’s a privilege to have a philosophical writer enter this debate to enrich the awareness of our readers.