The Deep Feed

01 — Not Boring

The Data Centre as the New Apollo Program

How AI infrastructure is inadvertently funding the reindustrialisation of the physical world

By Packy McCormick · 8 min read

Editor's note: A necessary reframe of the AI data centre debate from environmental burden to industrial catalyst.

The public discourse surrounding AI data centres is almost universally negative. Critics focus on the massive energy requirements and the sheer physical footprint of these silicon cathedrals. They see a drain on resources, a digital gluttony that offers little in return to the physical world. But this view misses a massive, structural shift in how new technologies reach scale. We are seeing the emergence of a new kind of buyer, one that does not care about the marginal cost of a component so long as it delivers a specific, high-value capability. This is not just about training models; it is about the accidental subsidisation of the next generation of hard tech.

The Problem of the Local Maximum

Most breakthrough technologies—advanced nuclear reactors, geothermal energy, or modular construction—suffer from a catch-22. They are technically superior to current solutions but lack the scale to be cost-competitive. In a rational market, no one wants to be the first to adopt an expensive, unproven technology. This keeps us stuck in 'local maxima', using old, inefficient tools like natural gas or coal because they are cheap and the environmental costs are externalised. To break out of these loops, you need either an 'Alpha Product'—a consumer device that demands a specific component—or an extraeconomic buyer of capability.

The Data Centre is the meta-Alpha Product. If you can sell them something they need, fast, they have an almost bottomless bid.

Historically, entities like NASA or the Department of Defense have played this role. They buy capability regardless of price, providing the initial demand that allows a technology to move down its cost curve. Today, the AI data centre is performing that exact function. The demand for GPUs and high-speed networking is so extreme that it is pulling along an entire ecosystem of secondary technologies. We are seeing massive investment flow into high-voltage grids, silicon photonics, advanced cooling, and even supersonic turbines. The data centre is acting as a massive, private-sector engine for reindustrialisation.

Technologies being pulled forward by AI demand

Enhanced geothermal energy
Modular construction techniques
Solid-state transformers
High-voltage direct current (HVDC) grids
Advanced nuclear reactor components

This is a profound economic accidentalism. The same demand that powers the digital intelligence of the future is providing the capital and the scale required to fix the physical infrastructure of the present. We are not just building smarter software; we are building the hardware of a new industrial age, funded by the insatiable appetite for compute.

Key Takeaway

AI data centres are acting as the primary economic catalysts for the next wave of physical industrial breakthroughs.

02 — Simon Willison

The Death of the Subscription Discount

Why OpenAI and Anthropic are finally turning a profit

By Simon Willison · 7 min read

Editor's note: The era of cheap, unlimited AI is over. The transition to API-based enterprise pricing marks the true beginning of the AI economy.

For the last two years, the narrative around AI labs has been one of massive spending and questionable revenue. We saw billions poured into compute, while consumer subscriptions sat at a modest $20 a month. It was a model that required billions of users to even approach break-even. But something changed in the spring of 2026. Both OpenAI and Anthropic have quietly dismantled their generous enterprise pricing models, replacing them with structures that mirror raw API token costs. They are no longer selling 'seats'; they are selling compute.

The Agentic Inflection Point

The reason for this shift is simple: product-market fit has arrived, and it looks like coding agents. While ChatGPT is a popular toy for the masses, agents like Claude Code and OpenAI Codex are professional tools for the highly compensated. These tools do not just answer questions; they execute tasks. And executing tasks is incredibly expensive. A single developer running an autonomous agent can burn through $1,000 worth of tokens in a month. If the labs continued to offer flat-rate enterprise subscriptions, they would be subsidising the very work that makes their models useful.

Coding agents plus enterprise pricing marks the point when these companies start making very real revenue.

This is the difference between a consumer app and an industrial tool. A consumer app needs scale to survive; an industrial tool needs margin. By moving to API-aligned pricing, the labs have ensured that as their customers become more productive—and thus use more compute—the labs' revenue scales in lockstep. They have moved from the 'growth at all costs' phase to the 'capture the value' phase.

Why the pricing shift matters

It signals the transition from consumer novelty to professional utility
It aligns lab revenue with the actual cost of compute-intensive agentic work
It provides a clearer path to IPO for both OpenAI and Anthropic
It forces enterprises to treat AI as a variable cost of production rather than a fixed overhead

The era of the 'unlimited' AI subscription was a necessary period of experimentation. It allowed users to find the boundaries of what these models could do. But now that the boundaries have been found, the bill has arrived. The companies are no longer just building models; they are building the power plants of the knowledge economy, and they intend to charge for every kilowatt of intelligence used.

Key Takeaway

The shift to API-based pricing proves that AI has moved from a consumer curiosity to an essential, high-margin industrial utility.

03 — Lenny's Newsletter

The Last 10% Problem

Testing the limits of Claude Opus 4.8

By Claire Vo · 6 min read

Editor's note: A pragmatic look at the latest frontier model, revealing that even the best AI still struggles with the final stretch of complex work.

Anthropic's release of Opus 4.8 was met with significant hype. On paper, it is a leap forward in judgment and independence. In practice, it reveals the persistent ceiling of current LLM architecture. When tasked with greenfield projects—starting from scratch to build a prototype—Opus 4.8 is exceptional. It is fast, it follows instructions, and it builds functional structures with startling ease. It is the ultimate tool for the 'zero to one' phase of development.

The Wall of Complexity

The trouble begins when the work moves from creation to refinement. In testing, Opus 4.8 hit a consistent wall when dealing with the 'last 10%' of a task. This is the territory of edge cases, subtle bugs in existing codebases, and the deep logical consistency required for long-term strategy. While the model is more 'honest' about its limitations than its predecessor, it still suffers from hallucinations when pushed into these high-complexity zones. It can build the house, but it struggles to fix the plumbing in a house built by someone else.

Opus 4.8 excels at greenfield prototypes; it struggles with the last 10% and the edge cases of existing codebases.

Interestingly, this creates a strange bifurcation in how professionals use AI. For rapid prototyping and 'one-shot' features, 4.8 is the clear winner. But for heavy-duty strategy work—the kind that requires parsing massive amounts of data and maintaining a coherent, multi-month roadmap—many users are still finding more reliability in the older 4.7 model. The newer model's tendency toward 'independent' work can sometimes lead it down logical rabbit holes that are harder to correct than the more predictable, albeit slower, older versions.

Where Opus 4.8 wins and loses

Win: Rapid prototyping and 'zero to one' development
Win: Fast execution of single-step features
Loss: Navigating large, messy, pre-existing codebases
Loss: High-stakes, data-heavy strategic planning
Loss: Eliminating the final, subtle edge-case bugs

The takeaway for agency owners and developers is not that the technology is failing, but that the nature of the work is shifting. We are moving away from being 'writers of code' and toward being 'editors of intent'. The value is no longer in the ability to generate the lines, but in the ability to spot the 10% error that the model is too confident to admit it made.

Key Takeaway

AI is becoming a master of creation but remains a flawed auditor of complexity.

04 — Lenny's Newsletter

From Prompting to Managing

The rise of the autonomous /goal

By Claire Vo · 5 min read

Editor's note: The shift from turn-based chat to goal-based loops is the most significant change in AI interaction since the transformer.

Most people interact with AI through a series of turns. You ask a question, it gives an answer, you follow up. This is 'babysitting' the model. It is a reactive, high-friction way of working that keeps the human tethered to the machine. Codex's new /goal command represents a fundamental departure from this pattern. Instead of a prompt, you provide a target. You define the outcome, the constraints, and the verification method, and then you walk away.

The Autonomous Loop

The power of the /goal is not just that it works while you sleep, but that it works through loops. When a standard prompt fails, the conversation stops. When a /goal encounters an error, the agent can attempt to diagnose the error, iterate on a fix, and re-run the test. This is the difference between a calculator and a junior engineer. In one test, the /goal command was used to systematically eliminate hundreds of Sentry error logs in a codebase—a task that would have taken a human developer hours of tedious, repetitive debugging.

Goals represent a fundamental shift in how we work with AI, from babysitting the model to managing it.

This isn't limited to coding. The application to administrative and operational tasks is equally transformative. Cleaning up thousands of emails, organising project management tasks in Linear, or managing API error logs—these are all 'low-reasoning, high-repetition' tasks that are perfect for autonomous loops. The human role shifts from the person doing the work to the person defining the success criteria.

The anatomy of a strong /goal

Measurable outcomes (not just 'clean this up')
Clear verification methods (how do we know it's done?)
Strict constraints (what are the boundaries?)
Defined error-handling protocols

As these autonomous agents become more reliable, the bottleneck in any business will no longer be the ability to execute tasks, but the ability to clearly define what a 'finished' task looks like. We are entering an era where the most valuable skill is not technical execution, but the precision of intent.

Key Takeaway

The future of productivity lies in moving from turn-based prompting to autonomous, goal-oriented management.

05 — Simon Willison

The SQLite Resistance

Why the world's most important database is banning agentic code

By Simon Willison · 3 min read

Editor's note: A crucial case study in how human-led open source projects are defending their integrity against the flood of AI-generated noise.

SQLite is one of the most ubiquitous pieces of software in existence. It is the bedrock of almost every mobile phone, browser, and embedded system on the planet. Because of its importance, its codebase is treated with a level of reverence that borders on the religious. Recently, the project took a hard stance against the rising tide of AI-driven development by adding an AGENTS.md file to its repository. The message is clear: the project does not accept agentic code.

The Flood of Noise

The move wasn't just a philosophical statement; it was a defensive necessity. The SQLite forums have been flooded with AI-generated bug reports—often of varying quality and frequently nonsensical. These reports, while appearing professional, often lack the fundamental understanding of the system they are critiquing. They create a massive amount of noise that human maintainers must sift through, wasting precious time that should be spent on real architectural improvements.

SQLite does not accept agentic code. However the project will accept agentic bug reports that include a reproducible test case.

The distinction SQLite is making is subtle but vital. They aren't banning the *use* of AI; they are banning the *delegation of responsibility* to AI. A human can use an agent to find a bug, but that human must then verify the bug, write a reproducible test case, and take ownership of the report. They cannot simply let an agent submit a pull request and hope for the best. The project is asserting that while tools can change, the requirement for human accountability remains absolute.

The SQLite Agent Policy

No agentic code submissions without prior agreement
No unverified agentic pull requests
Agentic bug reports are welcome *if* they include a reproducible test case
Human review is mandatory for all proof-of-concept changes

This is a warning to the rest of the software world. As agents become more capable of writing code, the pressure to automate the entire development lifecycle will grow. But for critical infrastructure, the cost of an unverified, agent-generated mistake is too high. The SQLite approach suggests that the future of high-integrity software will be human-led, even if it is AI-assisted.

Key Takeaway

In critical software infrastructure, the human must remain the final arbiter of truth, regardless of how capable the tools become.

06 — Stratechery

The Advertising Optimism

Why the intersection of models and ads is better for humanity than you think

By Stratechery · 9 min read

Editor's note: A counter-intuitive argument that the marriage of generative AI and advertising could solve the long-standing problem of digital relevance.

The prevailing sentiment regarding AI and advertising is one of dread. We fear a world of hyper-personalized manipulation, where models know our weaknesses better than we do. However, an interview with Eric Seufert suggests a different trajectory. The current advertising model is broken; it is built on tracking and probabilistic guessing. It is intrusive, inaccurate, and largely ignored. The integration of foundational models into the advertising stack offers a way to move from 'tracking' to 'understanding'.

From Tracking to Understanding

Current digital advertising relies on cookies and device IDs to piece together a shadow version of a user. It is a clumsy, privacy-invading process. Generative models, however, can process intent. Instead of knowing that 'User X clicked on a shoe ad', a model can understand that 'User X is looking for a durable running shoe for trail running in wet conditions'. This shift from identity-based tracking to intent-based understanding could actually improve privacy by reducing the need for invasive personal data, focusing instead on the immediate context of the user's needs.

Understanding advertising leads to optimism about humanity's future because it moves us from manipulation to utility.

When advertising becomes genuinely useful—providing the exact solution to a problem at the exact moment it arises—it ceases to be an interruption and becomes a service. This is the 'upside for humanity' that Seufert discusses. If AI can bridge the gap between a consumer's need and a provider's solution with precision and respect for privacy, the economic efficiency gains are massive. We move away from the 'attention economy' and toward a 'utility economy'.

The shift in the advertising paradigm

From Identity-based (Who are you?) to Intent-based (What do you need?)
From Probabilistic (We think you like this) to Deterministic (You are looking for this)
From Interruption (Stop what you are doing) to Integration (Here is what you need)
From Privacy-invasive (Tracking your every move) to Context-aware (Understanding your current goal)

This transition requires a massive shift in how companies like Meta and Google build their models. They aren't just building chatbots; they are building the engines of global commerce. The winners will not be those who can most effectively manipulate attention, but those who can most accurately serve intent. It is a more difficult technical challenge, but one that aligns much better with the long-term interests of both users and businesses.

Key Takeaway

The convergence of AI and advertising could replace invasive tracking with helpful, intent-driven utility.