The Deep Feed

01 — Dwarkesh Podcast

The Silicon Foundation: Building Intelligence from Logic Gates

Why the future of AI depends on the math of a single multiply-accumulate operation

By Dwarkesh Patel · 12 min read

Editor's note: To understand why AI is so expensive and power-hungry, you have to understand the microscopic math happening inside the chip.

The debate over AI often stays at the level of software—large language models, agents, and emergent intelligence. But the real bottleneck is not just code; it is the physical arrangement of metal and silicon. To understand why a GPU behaves differently than a CPU, or why a new startup like MatX might succeed, one must descend into the basement of computation: the logic gate. At this level, intelligence is nothing more than a massive, coordinated dance of AND, OR, and NOT operations, connected by microscopic metal traces.

The Multiply-Accumulate Primitive

In the world of AI, the most important mathematical operation is the multiply-accumulate (MAC). While a general-purpose CPU is designed to handle a vast array of unpredictable tasks, an AI chip is a specialist. It exists to perform matrix multiplication, which is essentially a repetitive loop of multiplying two numbers and adding the result to a running total. This specific pattern dictates how the hardware is laid out. If you design a chip without prioritising the MAC, you are building a machine that fights against the very math it is meant to solve.

AI chips are specialists; they exist to solve a specific mathematical pattern through physical architecture.

Precision management is the second great challenge of chip design. In AI workloads, we often use low-precision numbers for the multiplication step to save energy and space. However, as we sum these numbers up, the errors from rounding can stack up and ruin the result. Therefore, the accumulation step requires higher precision. This tension—between the need for speed/efficiency and the need for mathematical accuracy—is what defines the architecture of every modern AI accelerator.

Key architectural distinctions

CPUs: Large, complex cores designed for unpredictable logic and branching.
GPUs: Massive arrays of smaller cores designed for parallel throughput.
ASICs: Custom-built silicon designed for one specific mathematical task, like matrix multiplication.
FPGAs: Reconfigurable hardware that sits between the flexibility of software and the speed of custom silicon.

The ultimate goal of chip designers is to reduce the cost of data movement. Moving a bit of information from memory to a processor often costs more energy than the actual calculation itself. This is why the industry is obsessed with cache hierarchies and scratchpad memories. The winner of the AI race won't just have the smartest models; they will have the most efficient way to move electrons across a piece of silicon.

Key Takeaway

The speed of AI is limited by how efficiently we can move data and perform simple arithmetic at scale.

02 — Stratechery

The Data Centre Veto

How physical infrastructure provides a check on digital expansion

By Ben Thompson · 8 min read

Editor's note: We often think of AI as an ethereal cloud, but it is actually a massive, land-hungry physical entity.

There is a growing tension between the digital ambitions of Silicon Valley and the physical reality of the communities they inhabit. While AI's impact is felt most acutely in the realm of software and white-collar jobs, its existence is entirely dependent on the physical world. Specifically, it requires massive data centres—vast, energy-intensive buildings that require land, water for cooling, and immense amounts of electricity. This creates a new kind of political leverage: the Data Centre Veto.

The Power of Local Resistance

In the era of globalisation, it was difficult for local populations to stop the flow of capital or the movement of digital services. But you cannot hide a data centre. It is a massive, noisy, power-hungry structure that sits in a specific zip code. This gives local governments and citizens a form of veto power that they never possessed against the digital economy. If a community decides they do not want the noise, the heat, or the strain on their power grid, they can simply say no.

The ability to block a data centre gives local communities a physical check on digital expansion that they never had before.

This opposition is often framed as a misunderstanding of technology or a fear of the unknown. However, the concerns are frequently practical. The strain on local electrical grids and the competition for water resources are real-world costs that are often externalised by tech companies. To solve this, the industry cannot simply rely on better PR; it must find ways to compensate the communities that host this infrastructure.

The drivers of data centre opposition

Energy grid instability and rising electricity costs for residents.
Water consumption for cooling massive server arrays.
Noise pollution from industrial-scale cooling systems.
Visual and land-use conflicts in residential or rural areas.

As AI scales, the battle for the 'physical layer' will intensify. Companies will need to move beyond just training better models and start becoming experts in urban planning, energy production, and community negotiation. The future of intelligence is not just written in code; it is built in concrete and copper.

Key Takeaway

Digital progress is ultimately subject to the permission of the physical world.

03 — Not Boring

The Biological Breakthrough: Retatrutide and the End of Obesity

How triple-hormone agonists are redefining human metabolism

By Packy McCormick · 10 min read

Editor's note: We are witnessing a pharmacological shift that rivals the impact of bariatric surgery.

For decades, obesity has been treated as a failure of willpower or a complex lifestyle issue. We are now entering an era where it is being treated as a solvable metabolic equation. The latest Phase 3 results for Eli Lilly's retatrutide represent a massive leap forward in this transition. We are no longer just talking about incremental weight loss; we are talking about results that mirror the outcomes of invasive bariatric surgery, delivered through a simple weekly injection.

The Triple-Hormone Advantage

What makes retatrutide different from previous generations of drugs like Wegovy or Zepbound is its mechanism. While earlier drugs targeted one or two hormone receptors, retatrutide is a triple agonist. It targets GIP, GLP-1, and glucagon receptors simultaneously. This multi-pronged approach doesn't just suppress appetite; it actively manages how the body processes energy and burns fat. The results are staggering: an average weight loss of over 28% in clinical trials.

Retatrutide is delivering bariatric surgery results in a shot.

The implications extend far beyond the scale. The trial data shows significant improvements in nearly every metabolic marker: blood pressure, triglycerides, and cholesterol. Perhaps most importantly, it shows a massive reduction in osteoarthritis pain—a common secondary effect of obesity. We are looking at a drug that doesn't just make people thinner, but fundamentally alters their systemic health profile.

Key clinical outcomes of Retatrutide

28.3% average bodyweight loss over 80 weeks.
75.8% average reduction in osteoarthritis knee pain.
72% of prediabetic participants returned to normal blood sugar levels.
Significant drops in LDL cholesterol and blood pressure.

As these drugs move from the grey market to mainstream medical practice, the societal shifts will be immense. We will see changes in healthcare costs, insurance models, and even the way we think about food and consumption. The 'Reta Summer' of 2027 will likely be remembered as the moment the biological constraints of the human body were significantly loosened.

Key Takeaway

Metabolic health is becoming a programmable variable through triple-hormone agonists.

04 — Not Boring

The Artificial Egg: Colossal's Hollywood Science

Can we manufacture life outside the biological shell?

By Packy McCormick · 7 min read

Editor's note: Colossal Biosciences is blurring the line between biological reproduction and industrial manufacturing.

Colossal Biosciences has announced a feat that sounds more like science fiction than biology: the hatching of healthy chicks from a fully artificial egg. This is not a biological egg that has been modified; it is an oval printed shell coated in an oxygen-permeable membrane. By pouring yolk into this structure and allowing an embryo to develop, researchers have bypassed the need for a natural eggshell entirely. It is a demonstration of control over the earliest stages of life.

The Mechanics of Synthetic Incubation

The process is as much about material science as it is about embryology. The artificial shell must be strong enough to contain the developing organism but permeable enough to allow for gas exchange. The researchers essentially 're-engineer' the egg by cracking a fertilized egg, removing the contents, and replacing the natural shell with their synthetic version. They even add ground-up calcium back into the mix, because the embryo needs to consume the shell as it grows.

We are moving from observing biological processes to manufacturing them.

Critics argue that this is 'pure Hollywood'—a spectacular demonstration that doesn't necessarily change the fundamental difficulty of de-extinction or large-scale biological engineering. They are right to a point. Hatching a chicken in a plastic bag or a printed shell is not the same as resurrecting a mammoth. However, the ability to create a controlled, transparent environment for embryonic development is a massive step toward more predictable synthetic biology.

What this technology enables

High-resolution observation of embryonic development.
Controlled environments for sensitive or endangered species.
A foundation for more advanced synthetic embryo research.
The potential to decouple reproduction from specific biological vessels.

Whether this is a revolutionary leap or a clever marketing stunt, it signals a shift in how we approach life. We are no longer just passive observers of the natural world; we are becoming its architects, designing the very vessels in which life begins.

Key Takeaway

The ability to manufacture the environment for life is the first step toward true biological engineering.

05 — Stratechery

The Death of the Ad-Supported Web

How AI agents are breaking the economics of content

By Parag Agarwal · 11 min read

Editor's note: When machines read the web instead of humans, the current advertising model collapses.

The internet has operated on a simple, if flawed, economic bargain for decades: users get free content, and in exchange, they are exposed to advertisements. This model relies on human attention. Advertisers pay to reach eyes, and creators produce content to capture those eyes. But we are approaching a structural break in this arrangement. As AI agents become the primary way we interact with information, the 'eyes' on the web are increasingly being replaced by algorithms.

The Agentic Web

An 'agentic web' is one where your personal AI assistant performs tasks for you—reading articles, comparing products, and booking travel. When an agent performs these tasks, it doesn't see ads. It doesn't feel the 'pull' of a catchy headline or the influence of a brand's visual identity. It simply parses the data to find the most efficient answer to your query. This renders the entire ad-supported economy obsolete because the primary consumer of content is no longer a human being capable of being marketed to.

Ads make sense for humans, but they are invisible to agents.

This creates a massive crisis for content creators. If agents are doing the reading, how do creators get paid? If the value of content is no longer tied to human attention, the entire incentive structure of the internet must be rebuilt. We need a new way to value and incentivise the creation of high-quality data that agents can consume, without relying on the traditional attention economy.

The core challenges of the agentic web

The obsolescence of the traditional ad-revenue model.
The need for new protocols to attribute value to content used by AI.
The risk of 'data poisoning' where low-quality content is generated just to fool agents.
The shift from 'attention-seeking' content to 'utility-providing' data.

The transition will be messy. We will likely see a move toward subscription-based models, micro-payments for data access, and new cryptographic proofs of authorship. The internet is moving from a place of browsing to a place of processing, and the economics must follow suit.

Key Takeaway

The shift from human attention to machine processing requires a total redesign of internet economics.

06 — Simon Willison

The Active Listening Scam

When marketing hype meets regulatory reality

By Simon Willison · 6 min read

Editor's note: A cautionary tale of how 'AI' can be used as a cloak for old-fashioned deception.

In the rush to capitalize on the AI boom, many companies have turned to 'AI-washing'—using the term to describe products that have little to no actual artificial intelligence. A recent FTC settlement involving Cox Media Group and two other firms serves as a stark warning. These companies marketed a service called 'Active Listening,' claiming they could use smart devices to listen to real-time conversations and target ads based on what people were saying in their own homes.

The Reality of the Deception

The reality was far less high-tech and far more predatory. The 'Active Listening' service didn't listen to anything. It didn't use voice data at all. Instead, the companies were simply reselling existing email lists obtained from data brokers, charging a significant markup under the guise of advanced AI technology. They took a common, albeit privacy-invasive, practice and wrapped it in a futuristic metaphor to trick advertisers into paying more.

Active listening was a fancy term for something that already existed, used to mask a lack of actual technology.

The FTC's intervention also addressed the issue of consent. The companies argued that users had 'opted in' by agreeing to standard terms of service when downloading apps. The regulator was clear: clicking through a mandatory terms-of-service agreement does not constitute 'adequate consent' for something as invasive as monitoring private conversations. This sets a high bar for how companies must handle sensitive data in the age of AI.

Key takeaways from the FTC settlement

AI-washing is a legal liability, not just a marketing error.
Standard terms of service do not equal informed consent for invasive data collection.
The FTC is actively monitoring how 'AI' is used to justify deceptive business practices.
Reselling data under a high-tech label is considered consumer fraud.

As AI becomes more integrated into our lives, the temptation to use it as a marketing buzzword will grow. But as this case shows, regulators are watching. For companies, the path to long-term success lies in building actual technology, not just building better metaphors.

Key Takeaway

Using AI as a marketing veneer for existing data practices is a recipe for regulatory disaster.