AI Frontiers

‍AI agents are no longer a futuristic concept. They’re increasingly being embedded in the systems we rely on every day. These aren’t just new software features. They are independent digital actors, able to learn, adapt, and make decisions in ways we can’t always predict. Across the AI industry, a fierce race is underway to expand agents’ autonomous capabilities. Some can reset passwords, change permissions, or process transactions without a human ever touching the keyboard. Of course, hackers can also unleash AI agents to gain entry to, and wreak havoc within, those same sensitive systems.

I see this transformation daily in my work at the forefront of cybersecurity, where AI agents are rapidly undermining our traditional approaches to safety. But the risk isn’t confined to what these agents can do inside corporate networks. Their activities threaten to ripple outward into society. Left unchecked, they could undermine trust-based systems that make [...]

---

Outline:

(02:02) How AI Agents Undermine Identity and Trust

(07:49) The Infrastructure We Built for Trust

(09:07) What We Can and Can't Do with the Tools We Have

(12:20) The Future Is Here. Will We Govern It?

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

OpenAI — once considered an oxymoron given its closed-source practices — recently released GPT-OSS, the company's first open language model in half a decade. The model fulfills an earlier pledge to again release “strong” open models that developers can freely modify and deploy. OpenAI approved GPT-OSS in part because the model sits behind the closed-source frontier, including its own GPT-5, which it released just two days later.

Meanwhile, Meta — long a champion of frontier open models — has delayed the release of its largest open model, Llama Behemoth, and suggested it may keep its future “superintelligence” models behind paywalls. Meta, which once described open source AI as a way to “control our own destiny,” now cites “novel safety concerns” as a reason to withhold its most capable models.

These decisions mark a dramatic pivot for both companies, and reveal how different AI firms are converging on an [...]

---

Outline:

(02:02) Uncertainty Is Driving Precautionary Policy

(04:37) Precaution Disproportionately Chills Open Source

(06:48) Restrictions Demand Confident Evidence

(09:03) Precaution May Lead to Digital Feudalism

(12:54) We Need to Learn to Live with Uncertainty

(15:24) We Should Promote, Not Deter, Openness at the Frontier

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

OpenAI's GPT-5 launched in early August, after extensive internal testing. But another OpenAI model — one with math skills advanced enough to achieve “gold medal-level performance” on the world's most prestigious math competition — will not be released for months. This isn’t unusual. Increasingly, AI systems with capabilities considerably ahead of what the public can access remain hidden inside corporate labs.

This hidden frontier represents America's greatest technological advantage — and a serious, overlooked vulnerability. These internal models are the first to develop dual-use capabilities in areas like cyberoffense and bioweapon design. And they’re increasingly capable of performing the type of research-and-development tasks that go into building the next generation of AI systems — creating a recursive loop where any security failure could cascade through subsequent generations of technology. They’re the crown jewels that adversaries desperately want to steal. This makes their protection vital. Yet the dangers they may [...]

---

Outline:

(01:42) The Invisible Revolution

(03:19) Two Converging Threats

(08:02) The Accelerant: AI Building AI

(10:21) Why Markets Won't Solve This

(12:23) The Role of Government

(16:41) Reframing the Race: A Security-First Approach

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

The race is on for AGI. Tech companies are in a global race to develop artificial general intelligence (AGI): autonomous systems that perform most tasks as well as a human expert.

In early 2024, Meta CEO Mark Zuckerberg declared that Meta is going to build AGI and “open source” it. He is walking his talk. Meta has invested billions of dollars in the highest-power computational elements needed to build giant AI systems, and it has openly released its most powerful AI models. In early 2025, representatives of the Chinese company DeepSeek tweeted their intention to build and openly release AGI. US companies (including OpenAI, Google DeepMind, and Anthropic) are also trying to build AGI.

While these companies have not pledged to open-source such a system, recent months have seen a marked shift among U.S. policymakers and AI developers toward support for open-source AI. In July, the White House AI [...]

---

Outline:

(02:06) What Is AGI?

(05:22) Widespread Proliferation with No Guardrails

(09:10) The Dire Implications of Unleashing AGI

(13:50) The Replacement of Humanity

(17:49) Changing Course: Don't Build Uncontrollable AI

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

In an age of heightened political division, countering China's efforts to dominate AI has emerged as a rare point of alignment between US Democratic and Republican policymakers. While the two parties have approached the issue in different ways, they generally agree that the AI “arms race” is comparable to the US-Soviet strategic competition during the Cold War, which encompassed not just nuclear weapons but also global security, geopolitical influence, and ideological supremacy. There is, however, no modern deterrence mechanism comparable to the doctrine of Mutually Assured Destruction (MAD), which prevented nuclear war between the US and Soviet Union for four decades — and which is arguably the reason no other nation has used nuclear weapons since.

The bridge from MAD to MAIM. Earlier this year, co-authors Dan Hendrycks, Eric Schmidt, and Alexandr Wang proposed a framework called Mutual Assured AI Malfunction (MAIM), hoping to fill that dangerous strategic vacuum. [...]

---

Outline:

(02:40) The Logic of MAIM

(07:11) The Observability Problem Is Bigger Than MAIM's Authors Acknowledge

(10:29) Challenge One: Using Appropriate Proxies for AI Progress

(13:34) Challenge Two: Observation Must Keep Up With Rapid Progress

(16:41) Challenge Three: Superintelligence Development Will Likely Be Widely Decentralized

(19:31) Challenge Four: Intelligence Activities Themselves Could Lead to Escalation

(21:45) MAIM Started the Conversation on Superintelligence Deterrence, but More Dialogue Is Needed

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Can we head off AI monopolies before they harden? As AI models become commoditized, incumbent Big Tech platforms are racing to rebuild their moats at the application layer, around context: the sticky user- and project-level data that makes AI applications genuinely useful. With the right context-aware AI applications, each additional user-chatbot conversation, file upload, or coding interaction improves results; better results attract more users; and more users mean more data. This context flywheel - a rich, structured user- and project-data layer - can drive up switching costs, creating a lock-in effect that effectively traps accumulated data within the platform.

‍Protocols prevent lock-in. We argue that open protocols - exemplified by Anthropic's Model Context Protocol (MCP) - serve as a powerful rulebook, helping to keep API-exposed context fluid and to prevent Big Tech from using data lock-in to extend their monopoly power. However, as an API wrapper, MCP can access [...]

---

Outline:

(02:33) From Commoditized Models to Context-Rich Applications

(07:46) How User Context Is Powering a New Era of Tech Monopolies - and Competition

(10:22) Can Protocols Create a Level Playing Field?

(12:38) MCP's Impact on the AI Market So Far

(14:16) MCP vs. Walled Gardens: The API Gatekeeping Problem

(16:06) To Save AI from Enshittification, Support Protocol-Level Interventions

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

This post was cross-published on the author's Substack, Threading the Needle.

Increasingly, the US-China AI race is taking center stage. To win this race, Washington and Beijing are rethinking a range of policies, from export controls and military procurement priorities to copyright and liability rules. This activity, under the vague banner of race victories, conceals a deeper lack of clarity on strategic objectives. The Trump administration's AI Action Plan provides yet another entry on what the US strategy might look like — but it shouldn’t be mistaken for an indication of strategic clarity, as US factions are still vying for dominance from decision to decision. All in all, Chinese and US AI strategies are both still nascent. This raises important questions: What will each country decide that “winning the AI race” means? And where does that leave the rest of the world?

The obvious answer to the first question [...]

---

Outline:

(02:20) Open Questions on Grand Strategy

(02:43) 1. Military victory

(03:53) 2. Economic victory

(05:25) 3. A hybrid approach

(07:16) Technical Determiners

(07:32) 1. Military usefulness

(08:26) 2. Frontier capability gap

(10:10) 3. Compute supply trends

(12:10) What About Everyone Else?

(13:25) 1. Securitized world

(14:36) 2. Mercantilist world

(16:01) 3. World of clashing doctrines

(17:31) Outlook

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Last week, the AI nonprofit METR published an in-depth study on human-AI collaboration that stunned experts. It found that software developers with access to AI tools took 19% longer to complete their tasks, despite believing they had finished 20% faster. The findings shed important light on our ability to predict how AI capabilities interact with human skills.

Since 2020, we have been conducting similar studies on human-AI collaboration, but in contexts with much higher stakes than software development. Alarmingly, in these safety-critical settings, we found that access to AI tools can cause humans to perform much, much worse.

A 19% slowdown in software development can eat into profits. Reduced performance in safety-critical settings can cost lives.

Safety-Critical Scenarios

Imagine that you’re aboard a passenger jet on its final approach into San Francisco. Everything seems ready for a smooth landing — until an AI-infused weather monitor misses a sudden microburst. [...]

---

Outline:

(01:04) Safety-Critical Scenarios

(03:30) How Current Safety Frameworks Fail

(05:43) AI Influences Humans to Perform Slightly Better... or Much, Much Worse

(08:50) A Clear Pattern in Human-AI Collaboration

(09:49) Three Rules for Better Evaluations

(11:43) Faster, Easier, and Earlier Evaluations

(13:48) Toward Responsible Deployments of AI

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

The European Union just published the finalized Code of Practice for general-purpose AI models, transforming the AI Act's high-level requirements into concrete standards that will likely shift frontier AI companies' practices toward safer ones.

Among the Code's three chapters (copyright, transparency, and safety & security), requirements outlined in the Safety and Security section mark particular advances in frontier AI safety. The chapter — drafted by chairs Yoshua Bengio, Marietje Schaake, and Matthias Samwald, along with six vice chairs — targets general-purpose models deemed to pose systemic risks, currently defined as those trained with more than 10^25 floating-point operations (FLOPs).

The 10^25 threshold captures all of today's frontier models and can be adapted as the technology evolves. The Code emerged from an extensive consultation process with over a thousand stakeholders (from industry, academia, and civil society) providing feedback across multiple rounds.

Companies have a powerful incentive to adopt the Code [...]

---

Outline:

(01:53) What Companies Must Do

(02:37) Key Advances Beyond Current Practice

(02:41) Risk Identification

(03:47) Risk Analysis

(05:13) Pre-commitment Through Risk Tiers

(05:35) Transparency and External Validation

(06:42) Cybersecurity and Incident Reporting

(07:41) Gaps and Enforcement Challenges

(10:03) Implications for Global AI Governance

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

For over half a decade, the United States has imposed significant semiconductor export controls on China, aiming to slow China's chip industry and to retain US leadership in the computing capabilities that undergird AI advances. Have these controls achieved their goals? Have the assumptions driving them been confirmed or undermined by the rapid evolution of the chip industry and AI capabilities?

Three factors for assessing chip export controls. We can now draw preliminary conclusions by assessing three factors: China's domestic chipmaking capability, the sophistication of its AI models, and its market share in providing AI infrastructure.

Initial evidence shows that controls have succeeded in several important ways, though not all. Restrictions on chipmaking tool sales have significantly slowed the growth of China's chipmaking capability. However, restrictions on the export of AI chips to China, while creating challenges, have not prevented Chinese labs from producing highly competitive models (though they [...]

---

Outline:

(01:30) The Current Chip Export Control Regime

(04:25) The Impact on China's Chip Industry

(09:10) The Impact on AI Model Development in China

(13:28) The Impact on China's Ability to Provide AI Infrastructure

(15:18) Implications for the Future of AI and US-China Relations

(18:01) Export Controls Have Given the US a Commanding Lead in AI

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.