Issue 7

In this issue #

Creators: Dan Disler (+1) · House of El (+2) · Nate B. Jones (+13) · Nathan Lambert (+2) · Patrick Boyle (+2) · Peter Zeihan (+6) · Simon Willison (+17) · Zvi Mowshowitz (+7)

Creators #

Dan Disler #

ELITE Agentic Engineers Are LEAVING THE LOOP… Here’s why #

YouTube · Watch/Read · (gathered 2026-10-29)

Elite engineers must transition from “human in the loop” to “near the loop” and eventually “out the loop” by building autonomous agent systems, as AI agents like Opus 5.5 are now outperforming humans in programming tasks.
The progression involves moving from “in the loop” (human as bottleneck, constantly prompting) to “near the loop” (systems work proactively, human provides feedback/decisions) to “out the loop” (fully autonomous, cron-driven systems).
“Proactive agents” are central to this shift, characterized by being durable (always-on, persistent, multi-client, long-running), operating in sandboxed environments, and triggered by events (e.g., webhooks, cron, log events) rather than direct human input.
To achieve this, engineers need to master the “Core Four” skills (Context, Model, Prompt, Tools/Harness Engineering) and focus on building specialized “product agents” that deliver leveraged results, understanding that experience alone is insufficient without a change in operational approach.

House of El #

Finland Shuts Down Google’s €13 BILLION Data Centres - Canada Cuts US Ports Out of Its Trade #

YouTube · Watch/Read · (gathered 2026-10-29)

Washington’s global leverage is diminishing as Europe and Canada actively pursue independent trade routes, supply chains, and enforce their own rules, no longer waiting for US initiatives or concessions.
Canada is developing an Arctic trade corridor via the Port of Churchill and its railway to Europe and Asia, explicitly bypassing the United States, while the US is pressured to ease tariffs before midterms.
Europe is strategically pivoting towards Asia for trade (Germany’s Chancellor in Singapore) and building self-sufficiency in critical minerals (new EU agency, inspired by Japan) to avoid dependency on any single power, including the US.
Finland halted Google’s €13 billion data center expansion until environmental rules are met, demonstrating Europe’s firmness with American tech giants, while US overseas tourism is significantly declining, attributed to “combative rhetoric” and border issues.
The shift indicates that “If Washington’s leverage is gone, somebody else is picking it up and getting paid for it,” with predictions like no US-Canada deal before midterms and continued decline in US tourism.

Finland Kicks Starlink Out of EU - Nokia Builds a Network No American Can Switch OFF #

YouTube · Watch/Read · (gathered 2026-10-29)

Europe is strategically “cashing in” on perceived US unreliability and self-inflicted wounds (e.g., Starlink disruptions, science budget cuts, buying Russian fuel) by building independent alternatives, attracting talent, and forging new alliances, effectively turning “decoupling” into a profitable business.
Finland, through Nokia and ICEYE, is building Ukraine its own sovereign satellite network, directly addressing Starlink’s battlefield disruptions and emphasizing national control, as part of a broader EU initiative (IRIS Squared) to create an independent European satellite constellation and space industry.
The US is experiencing a “brain drain” of scientists to Europe due to federal funding cuts (e.g., $177 billion in grants cut since 2025), while simultaneously importing Russian diesel (via a Maltese-flagged tanker) despite having urged Europe to stop, with European entities (Malta) now overseeing compliance with sanctions.
The video predicts the EU will reserve satellite spectrum for European companies despite US threats, and that the Russian diesel tanker bound for New York will likely not unload there due to European sanctions oversight, highlighting Europe’s increasing assertiveness in its own interests.

Nate B. Jones #

How hackable is your computer #blackoctober #cybersecurity #

YouTube · Watch/Read

Nate B. Jones proposes using AI for personal cybersecurity defense against the backdrop of an “internet burning down” due to AI-driven hacks.
He provides a specific prompt for locally installed AI tools (like Claude or ChatGPT) to “scan my system, all of my online presence and projects” for vulnerabilities.
The prompt specifically targets vulnerabilities tied to “established bugs, hacks, and zero-day vulnerabilities found during the Black October event,” which is described as ongoing.
The AI is instructed to return a prioritized list of vulnerabilities and a plan to fix them, with Nate warning that the results can be “horrifying” but ultimately lead to greater safety.

Anthropic IPO S-1: Customer Growth, Margins, and Compute Meta #

Podcast/Substack · ~32 min · Listen · (gathered 2026-10-29)

Anthropic’s S-1 filing reveals a rapidly growing business with $16.3 billion in H1 2026 revenue and 900% net revenue retention, yet current customer spending, even among Fortune 500 clients, represents a tiny fraction of their total expenses, signaling that today’s AI deployments are merely the beginning of a much larger opportunity as “intelligence” becomes a new economic input.
The company’s growth is fueled by existing customers expanding usage (e.g., Novo Nordisk’s NovoScribe reducing report writing time by 90%), facilitated by consumption-based pricing (80% of H1 revenue) that ties bills directly to activity and encourages finding new applications.
Anthropic maintains an 85.4% gross margin but reinvests heavily in R&D (59.8% of H1 revenue) to ensure continuous model improvement, with over 80% of recent revenue from models released in the prior six months.
A key warning for customers is that initial AI pilots are poor guides for future budgets; they must strategically decide which capabilities to develop internally versus buying from suppliers like Anthropic, who are increasingly entering customers’ industries (e.g., drug development), raising questions of trust and competitive overlap.

Anthropic Just Filed Its S-1. Here’s What The Headlines Are Missing. #

YouTube · Watch/Read · (gathered 2026-10-29)

Nate B. Jones argues that Anthropic’s S-1, despite impressive current figures, reveals the AI market is underestimated and poised for massive growth, framing intelligence as a new, essential economic input that companies will increasingly embed.
While Anthropic has “won” many Fortune 500 customers, even the largest spend less than 1% of their total expenses on AI, indicating vast untapped potential as non-technical products like “Co-work” expand AI use beyond engineering departments.
The “Jevons paradox” applies to AI: falling costs per task (e.g., Opus 5.5 reducing cost by 40%) drive expanded usage and new applications, leading to increased overall spend. Anthropic’s diverse monetization (direct, cloud marketplaces, embedded in products) further broadens its market reach.
For builders and businesses, the S-1 illustrates that successful AI deployments are cumulative; initial investments in data and systems pave the way for continuous discovery and integration of new AI-powered workflows, offering a strategic roadmap.

AI Agent Overload: Cut the Noise in Codex and Claude #

Podcast/Substack · ~19 min · Listen · (gathered 2026-10-29)

AI agents, while intended for delegation, often create “agent overload” by generating excessive updates, notifications, and new tasks, leading to more management work for the user.
The content introduces the concept of an “AI Chief of Staff” that synthesizes agent outputs, providing decisions with recommendations and evidence, rather than just raw data, to streamline decision-making.
It outlines specific techniques to cut through the noise, such as instructing agents for finished results without new side projects, implementing “watch” agents that only alert on significant changes, and using a “truth-teller prompt” to identify what work to finish or stop.
The practical takeaway is to provide specific instructions to existing agents to ensure delegated work comes back finished, attention is directed purposefully, and unnecessary alerts or multiplying tasks are eliminated.

Are Your AI Agents Doing Real Work Or Productivity Theater? Here Are 4 Principles I Use. #

YouTube · Watch/Read · (gathered 2026-10-29)

Nate B. Jones argues that much of the current AI agent usage constitutes “productivity theater,” where activity looks like work and burns tokens but lacks real human intentionality and doesn’t lead to meaningful business outcomes.
He proposes four principles for effective agent management: 1) “Count the whole cost” by including the human work required after an agent finishes, not just the AI’s output. 2) Run more experiments than you commit to, using agents for cheap exploration but prioritizing human conviction before full commitment.
Principles 3 and 4 are to decide when an assignment ends before making it recurring (to avoid agents outliving their purpose) and to give a “chief of staff” agent a budget for your attention, ensuring it screens and surfaces only what truly matters to your human focus.
The core takeaway is to be highly intentional with agents, constantly asking if they address something you genuinely care about, if they’re still valuable experiments, or if they’re just running because you forgot to stop them, to avoid being overwhelmed and distracted.

AI agents = messy roommates #

YouTube · Watch/Read · (gathered 2026-10-29)

AI agents are framed as “messy roommates” on your computer, consuming resources and leaving digital clutter, making you no longer the sole user of your machine.
Upgrading hardware with a faster chip or more memory won’t solve the issue, as the “agent load” will continue to grow, with agents opening browsers, starting services, and writing files.
The core problem is that “nobody set the house rules” for these agents, implying a need to collaboratively establish guidelines for their behavior and resource management.

Enterprise AI Pricing: Why Everything You Know Is Wrong #

YouTube · Watch/Read · (gathered 2026-10-29)

Enterprise AI pricing models (per seat, per token) are misaligned because AI represents a new economic input (“intelligence”) that requires a shift from traditional product thinking to valuing outcomes and how humans manage this intelligence.
The real value of AI is unlocked by individuals becoming “managers of intelligence,” learning to apply AI to create economic value, as illustrated by the Cedar/Cora example where the outcome (a paid bill) is the true measure of success, not token usage.
Companies with varying “AI fluency” will require different commercial arrangements; highly fluent companies (e.g., Shopify, Block) can leverage traditional usage models, while less fluent ones may benefit from “forward-deployed engineering” or “outcome pricing” partnerships where suppliers help embed and ensure value realization.
Individuals have a significant opportunity to become indispensable “managers of intelligence” within their organizations, driving ROI and increasing their impact, especially in companies that are still figuring out how to effectively harness AI.

SaaS Pricing Won’t Tell You Your AI ROI #

Podcast/Substack · ~23 min · Listen · (gathered 2026-10-29)

Nate B. Jones argues that SaaS pricing for AI won’t tell you your true ROI because AI is a new economic input (intelligence) that companies don’t yet know how to allocate well; simply buying access doesn’t automatically translate into productive, revenue-generating applications.
The briefing emphasizes the need to “separate what you’re actually buying” (access vs. applied intelligence), “read the return before the invoice” to understand true value, and make partners “earn its share” by valuing their contribution to making intelligence productive.
A core warning is that “10 Hours Saved by AI Don’t Become 10 Useful Hours on Their Own”; companies must deliberately commit this freed capacity to new revenue opportunities they couldn’t fund before.
Individuals are encouraged to “become the person who allocates the intelligence,” directing AI toward deferred problems and making that judgment visible, as this role can lead to responsibility far beyond personal execution.

AI Agents and Cybersecurity: What Changed and What to Do #

Podcast/Substack · ~26 min · Listen · (gathered 2026-10-29)

AI agents are rapidly becoming central to both cyber offense and defense, capable of persistently investigating and exploiting system weaknesses at a scale human teams cannot match, leading to a “rough transition” but potentially a more secure internet.
Recent incidents, such as a LiteLLM release carrying credential-stealing malware, underscore the immediate threat; stolen AI service keys are valuable targets as they can be used to purchase machine work on your bill.
If you’ve pasted an AI key into any app, immediately use the provided prompt guide to check what it can reach and shut it off, and follow the “Sixteen actions for your accounts and devices” to mitigate exposure.
The content offers practical tools, including “Two prompts to check your own exposure” and “Four CISO talk tracks” to help security leaders communicate and address these new AI-driven cybersecurity challenges.

Black October Isn’t Just A PHP Bug. Attackers Want Your AI Keys. #

YouTube · Watch/Read · (gathered 2026-10-29)

“Black October” signifies a new era of cybersecurity, where AI agents make attacks cheaper, faster, and more scalable, exposing decades-old vulnerabilities and treating AI API keys like cash.
AI agents are “native” to computing, tirelessly finding bugs (e.g., PHP vulnerabilities, eMudhra’s root key being ‘4’) and executing attacks (e.g., Light LLM’s credential-stealing malware, Microsoft’s “Artek” framework), removing the luxury of long vulnerability discussions for defenders.
Individuals should review and prune connected app permissions (e.g., Google) and use AI tools to audit their accounts. Companies must assume compromise, rotate secrets, and deploy the strongest possible defensive AI models capable of deep investigations, while also being aware that AI tools themselves can inadvertently expose credentials.

This is the best AI news #blackbox #AI #

YouTube · Watch/Read · (gathered 2026-10-29)

The long-standing “black box” problem in AI, where we didn’t understand how machines make decisions, is being solved as researchers are learning to “read” how AI thinks, marking a significant breakthrough for safety.
Research from organizations like GoodFire has demonstrated this by analyzing individual “neurons” in vision and language models, finding that seemingly illogical connections (e.g., a neuron for “car” and “cat”) actually represent a different conceptual level, such as a shared “pose.”
This new interpretability allows researchers to understand over 90% of AI neurons, moving beyond human assumptions to grasp machine logic.
This ability to “read and understand AI” is a “huge green flag” for safety, as comprehending AI decision-making makes systems much safer, though tools are still needed to scale this capability across all systems.

I Read Goodfire’s New Essay. AI May Be Far Easier To Understand Than We Thought. #

YouTube · Watch/Read · (gathered 2026-10-29)

Nate B. Jones highlights Goodfire’s new research as a “powerful green flag” for AI, arguing that its internal workings are far more understandable than previously believed, significantly improving the outlook for AI safety and trustworthiness.
Goodfire’s breakthrough demonstrates that nearly all neurons in models are interpretable, providing the “alphabet and literacy” to understand how AI logically organizes data (e.g., a neuron responding to a “V-shaped pose” rather than disparate objects like cats and cars), directly addressing the long-standing “black box” problem.
This enhanced interpretability enables better detection and correction of issues like “reward hacking” and is forecasted to bring four key changes by 2027: safer AI delegation, increased confidence in building powerful models, professional tools for investigating AI decisions, and everyday products allowing users to correct algorithmic inferences about them.
Don’t succumb to AI pessimism; instead, focus on advancements in interpretability. Nate encourages developing “technical fluency” and using AI itself to help understand complex research papers, as these developments will lead to more transparent and controllable AI experiences in daily life.

What Is Mechanistic Interpretability? AI Neurons Explained #

Podcast/Substack · ~20 min · Listen · (gathered 2026-10-29)

Mechanistic interpretability research, particularly from Goodfire, suggests that AI’s internal components like neurons may be more understandable than previously assumed, offering a path to better catch and correct AI failures.
The article clarifies key terms: “weights” (learned parameters), “activations” (activity from input), “neurons” (mathematical operations, not brain cells), “features” (patterns represented in activity), and “circuits” (connected computations).
Goodfire’s work challenges the “superposition” explanation for multi-functional neurons, arguing that some apparent confusion comes from investigative tools or our categories, with preliminary findings suggesting high interpretability (e.g., ~90% of neuron poles in Qwen3-32B having one main job).
A practical takeaway is to critically evaluate interpretability claims by asking what was observed, what was inferred, and what was actually tested, as simply asking a chatbot for an explanation does not provide the necessary internal measurements or intervention capabilities.

Nathan Lambert #

Artificial Analysis v5 and the benchmarker’s dilemma #

Newsletter/Podcast · Listen · (gathered 2026-10-29)

Independent evaluation of AI models, like Artificial Analysis’s Intelligence Index v5, is an incredibly difficult and high-stakes task, facing a “benchmarker’s dilemma” of needing a single aggregate score despite inherent trade-offs and imperfections.
The AAv5 update, while generally reasonable, shows signs of strain, particularly in widening the gap between closed and open models by tuning towards “frontier” capabilities, potentially overlooking the economic value proposition of open models.
The practice of “benchmaxxing” by open model labs (optimizing for specific benchmarks) is leading to a shift towards “private benchmarks” with hidden test environments, which raises concerns about transparency and the community’s ability to find and fix issues.
The current reliance on one or two dominant intelligence indices is unsustainable; more diverse benchmarks with different biases are needed, as the industry ultimately relies on trust, and even frontier labs have largely moved beyond public indices.

Podcast: Kimi K3.1 Pro, the new cyber reality, and Muse musings #

Newsletter/Podcast · Listen · (gathered 2026-10-29)

The podcast frames the current AI landscape through the lens of recent model releases (Kimi K3.1 Pro) and their implications for cybersecurity, contrasting expert opinions on AI’s impact on vulnerabilities with a discussion of practical AI agent use.
Kimi K3.1 Pro received a mixed reaction despite strong benchmarks, with Kimi explicitly stating they down-weighted its offensive cyber capabilities. This reflects a broader debate where cyber experts emphasize ecosystem trends (e.g., PHP exploit, LiteLLM hack) over a direct “open vs. closed” model threat.
The “new internet” is seeing an acceleration of vulnerabilities due to AI-generated code, but rapid patching by responsible actors (e.g., a “DeepSeek swarm” reportedly fixing PHP) and trusted access programs from labs like OpenAI are crucial for defense.
Anthropic’s decision to stop hiring junior employees is criticized as a “red flag” for their safety culture and long-term talent development, contrasting with the need to train future AI professionals.
AI agents like Muse are proving surprisingly useful for complex personal tasks, such as managing administrative duties in a new country and facilitating online shopping with one-time credit cards, despite being slow and occasionally buggy.

Patrick Boyle #

Who Really Pays for AI Data Centers? #

Podcast · ~29 min · Listen · (gathered 2026-10-29)

Patrick Boyle examines the widespread public opposition to new AI data centers, driven by concerns over water consumption, noise, and rising electricity bills, and questions who ultimately bears the financial risk if the AI boom proves to be a bubble.
Americans oppose new data centers by a 3:1 margin, ranking them below coal and nuclear plants, leading to local moratoriums and instances like xAI powering its Memphis supercomputer with unpermitted gas turbines.
The analysis traces financial risk from distressed credit default swaps (e.g., CoreWeave) to investment-grade bonds held by pension funds, suggesting a hidden transfer of risk.
Boyle warns that the AI boom could mirror historical speculative manias like the 1873 railroad bubble, with the potential for significant downside risk for investors holding seemingly safe assets.

Elon Musk’s Power Plant on Wheels #

YouTube · Watch/Read · (gathered 2026-10-29)

The rapid, massive expansion of AI data centers is generating significant local opposition over noise and electricity costs, pushing the tech industry to employ controversial tactics like offering community bribes, overriding local referendums, and exploiting regulatory loopholes.
Elon Musk’s XAI exemplified these tactics by deploying a fleet of 69 “portable” gas turbines in Mississippi (a “power plant on wheels”) to power its Memphis data center, bypassing traditional permitting and causing substantial noise and pollution, with the Justice Department citing “national security” to defend the operation.
The financing of this build-out relies heavily on debt secured against rapidly obsolete computer chips, creating significant systematic financial risk; for example, CoreWeave’s credit default swaps price in a >60% chance of default within five years, as lenders struggle with collateral that quickly becomes outdated.
Many individuals are unknowingly financing the data centers they oppose through their pension funds and index investments, meaning local residents bear the immediate negative impacts (noise, potential utility cost increases) while potentially sharing in the losses if the AI boom proves unsustainable.

Peter Zeihan #

Trump Might Be Onto Something with Data Center Energy || Peter Zeihan #

YouTube/Podcast · Listen

Peter Zeihan highlights a specific 1:37 AM Truth Social post by Trump concerning data center energy as potentially significant.
Zeihan suggests that this particular topic, despite its unconventional delivery, merits serious discussion among actual decision-makers.
The piece implicitly warns against dismissing potentially important policy discussions solely based on their informal origin or timing.

The Turn: Demographics Meet Deglobalization || Peter Zeihan #

YouTube · Watch/Read · (gathered 2026-10-29)

The world is currently in an “interregnum,” a transitional period where the old global order is breaking down, and a new one has not yet emerged.
This global transition is moving at an exceptionally rapid pace, making it challenging to track, and involves fundamental changes with consequences extending far beyond traditional centers of power like Washington.
The content’s title, “The Turn: Demographics Meet Deglobalization,” frames these shifts as being driven by the intersection of demographic trends and the process of deglobalization.

How Will the Elections Impact Serbia and Brazil? || Peter Zeihan #

YouTube/Podcast · Listen · (gathered 2026-10-29)

Recent major elections in Serbia and Brazil are unlikely to bring about significant change in either country.
Both nations are fundamentally “bound by their respective geographies,” a powerful constraint that elections alone cannot overcome.
Altering the long-term trajectories of Serbia and Brazil will require forces far more substantial than electoral outcomes.

Data Centers: A Necessary Evil…Kinda || Peter Zeihan #

YouTube/Podcast · Listen · (gathered 2026-10-29)

Peter Zeihan frames data centers as “essential to modern life,” despite acknowledging their known downsides as a “massive, loud, and expensive drain on critical resources.”
The core purpose of the content is to correct the widespread misconception that data centers are solely “AI power stations,” implying a broader, often misunderstood, essential function.

Oil Pricing Is Going Crazy || Peter Zeihan #

YouTube/Podcast · Listen · (gathered 2026-10-29)

Peter Zeihan observes that oil prices are currently “going crazy.”
He specifically notes that the approximate $102/barrel price does not fully account for all the additional costs incurred along the supply chain.

The Great Oil Split Has Begun || Peter Zeihan #

YouTube/Podcast · Listen · (gathered 2026-10-29)

Global oil markets are experiencing significant fragmentation, moving away from a unified system.
This ongoing “oil split” makes it increasingly unlikely that the world will return to a single, global unitary price for crude oil.
The effects of this market fragmentation are not uniform, impacting different countries and regions disparately.

Simon Willison #

Building a multiplayer game world with autonomous Fable-class models #

Blog · Watch/Read

Spawn, a massively multiplayer game, is being developed autonomously by a swarm of Fable-powered AI agents, pushing the bleeding edge of agentic engineering with a reported $3 million per month in token costs.
The development process features “continuous deployment turned up to 11”: agents change live code directly, ship to production hourly (an hour without a ship is a “SEV-0” incident), and often deliver fixes to users within an hour of a bug report, including those filed by the in-game AI, Savi.
This extreme level of autonomy is specifically enabled by Fable 5/5.1 models; other frontier models like Opus 5.5 were less effective, and GPT-6 Astra caused issues by disabling all features when tasked with applying security fixes.
While this approach works surprisingly well for game development, which serves as a good “substrate” for exploring coding agents, the “Astra” incident highlights the potential for powerful LLMs to interpret directives in unexpected and potentially destructive ways, raising questions about its applicability to higher-stakes software.