The Deep Feed

01 — Not Boring

The End of Obesity

Retatrutide and the era of the chemical bypass

By Packy McCormick · 12 min read

Editor's note: The clinical results for retatrutide suggest we are moving past weight management into the realm of total metabolic restructuring.

The medical community has spent decades treating obesity as a failure of willpower or a complex hormonal imbalance that required surgical intervention. Bariatric surgery was the heavy-handed solution, a physical restructuring of the gut to force a change in caloric intake. But the arrival of retatrutide changes the math. Eli Lilly’s latest Phase 3 trial results for the drug show a level of efficacy that moves the conversation from 'management' to 'transformation'. We are no longer looking at incremental improvements in health; we are looking at a chemical tool that achieves the results of surgery without the scalpel.

The Triple Agonist Mechanism

What makes retatrutide different from its predecessors, like Wegovy or Zepbound, is its mechanism. While earlier drugs focused on one or two hormone receptors, retatrutide is a triple hormone receptor agonist. It targets GIP, GLP-1, and glucagon. This combination does more than just suppress appetite; it manages how the body processes energy and burns fat. By hitting three distinct pathways, the drug achieves a metabolic efficiency that single-target drugs cannot match. It is the difference between a single lever and a coordinated control system.

These are bariatric surgery results in a shot.

Clinical trial results for retatrutide (12mg dose over 80 weeks):

28.3% average bodyweight loss
70.3 pounds lost on average
45.3% of patients achieved 30%+ weight loss
Significant reductions in blood pressure and triglycerides
72% of prediabetic participants returned to normal blood sugar levels

The numbers are stark. A 28.3% reduction in bodyweight is not a marginal gain; it is a total shift in a patient's physiological baseline. For many, this moves them out of the obesity category entirely. The trial also showed that the drug addresses the secondary damage caused by excess weight, such as osteoarthritis pain and high cholesterol. We are seeing a drug that does not just make people smaller, but makes them healthier across multiple biological systems simultaneously.

The 2027 Horizon

As we look toward the next year, the implications for public health and the economy are massive. If weight loss becomes a matter of regular, highly effective injections, the entire infrastructure of obesity-related healthcare—from diabetes management to joint replacement surgery—will need to be rethought. We are approaching 'Reta Summer' in 2027, a period where these drugs will likely move from the grey market into the mainstream. The social reality of a significantly thinner, healthier population is a prospect that both medical optimists and economic skeptics are currently trying to model.

Key Takeaway

Retatrutide turns obesity treatment from a physical struggle into a predictable chemical process.

02 — Not Boring

The Synthetic Embryo

Life developed outside the shell

By Packy McCormick · 10 min read

Editor's note: Colossal Biosciences has demonstrated the ability to hatch life from an artificial container, a move that blurs the line between biology and manufacturing.

There is a specific kind of tension in watching an embryo develop through a window. Colossal Biosciences recently announced the hatching of 26 healthy chicks from a fully artificial egg. This was not a biological egg shell, but an oval printed from material coated in an oxygen-permeable membrane. Inside, a poured-in yolk developed into a living, breathing bird. The process involves taking the contents of a fertilized egg and placing them into this synthetic vessel, where the embryo can grow, eat the calcium-infused shell, and eventually pip through the membrane.

Beyond the Hollywood Marketing

Critics have dismissed this as 'pure Hollywood' marketing. They argue that growing birds outside of a natural shell is not a new concept, citing Japanese research from 1998. While it is true that the technical ability to hatch an embryo in a controlled environment has existed for decades, the scale and the intent behind Colossal’s work are different. This is not just about hatching a quail in a lab; it is about the mastery of the container. If we can control the environment of the egg perfectly, we gain a level of control over the development of the organism that nature never intended.

We are moving from observing life to manufacturing its containers.

The ability to decouple an embryo from its biological shell is a massive step toward synthetic biology. It allows for a level of observation and intervention that is impossible with a standard egg. You can monitor vasculature, eye formation, and nutrient absorption in real time. This level of visibility is the precursor to more complex biological engineering. If you can control the shell, you can eventually control the variables that dictate the success of the life inside it.

The Colossal artificial egg process:

Creation of an oxygen-permeable membrane shell
Transfer of yolk and embryo into the synthetic vessel
Addition of ground calcium for shell consumption
Real-time monitoring through a transparent window
Successful hatching after 18 days

The New Biological Frontier

This technology is a signal of where biological research is heading. We are moving away from the study of what is, and toward the construction of what could be. The artificial egg is a proof of concept for a future where the biological constraints of reproduction are replaced by the precision of engineering. It is a shift that will require new legal and ethical frameworks to manage, as the distinction between a natural organism and a manufactured one continues to erode.

Key Takeaway

The artificial egg proves that we are learning to manufacture the environments that sustain life.

03 — Dwarkesh Podcast

The Logic of Intelligence

Building the silicon foundation from the bottom up

By Dwarkesh Patel · 15 min read

Editor's note: To understand the AI revolution, one must understand the physical reality of the multiply-accumulate operation.

The complexity of an AI model is often discussed in terms of parameters and tokens, but the actual work of intelligence happens at a much lower level. It happens in the movement of electrons through logic gates. At the most fundamental level, a chip is a collection of AND, OR, and NOT gates connected by metal traces. When we talk about the massive scale of modern computing, we are really talking about the massive scale of these simple primitives working in unison to perform a single, repetitive task: the multiply-accumulate.

The Multiply-Accumulate Primitive

In AI workloads, the core operation is matrix multiplication. This involves a constant loop of multiplying two numbers and adding the result to a running total. This is the multiply-accumulate (MAC) operation. Because this happens at every single step of a matrix multiply, the efficiency of the MAC unit determines the efficiency of the entire chip. If you can make the MAC unit faster or more energy-efficient, you change the economics of the entire AI industry. It is the atomic unit of modern computation.

AI chips require higher precision in the accumulation step than in the multiplication step.

There is a specific technical requirement in AI hardware that separates it from general-purpose CPUs. When you multiply low-precision numbers, you introduce rounding errors. If you accumulate these numbers without increasing the precision of the sum, those errors compound until the result is useless. Therefore, AI chips are designed to perform low-precision multiplication but high-precision accumulation. This architectural choice is what allows GPUs and TPUs to maintain accuracy while operating at extreme speeds.

Key components of chip architecture:

Logic gates (AND, OR, NOT)
Multiply-accumulate (MAC) units
Systolic arrays for data movement
Pipeline registers for clock cycles
Cache vs scratchpad memory

The Architecture of Speed

The difference between a CPU and a GPU is essentially a difference in philosophy. A CPU is a massive, complex brain designed to handle any task, making it large and relatively slow at specific math. A GPU is a collection of thousands of tiny, specialized brains designed to do one thing—matrix math—extremely well. As AI continues to scale, the industry is moving toward even more specialised ASICs (Application-Specific Integrated Circuits) that strip away everything except the most efficient paths for the MAC operation. The future of intelligence is being written in the layout of these metal traces.

Key Takeaway

The speed of AI is not a software miracle; it is a hardware optimization of the multiply-accumulate operation.

04 — Stratechery

The Physical Veto

How local reality limits the digital cloud

By Stratechery · 11 min read

Editor's note: The digital expansion of AI is hitting a wall of physical resistance in the form of local zoning and infrastructure limits.

There is a prevailing myth that the digital world is weightless. We speak of 'the cloud' as if it were an ethereal layer of intelligence floating above our heads, unconstrained by geography or physics. But the reality is much more grounded. Every LLM, every generative image, and every agentic workflow requires a massive, physical footprint of silicon, copper, and concrete. AI is not just a software phenomenon; it is a massive industrial undertaking that depends on data centers, and building those data centers requires permission from the physical world.

The Power of Local Permission

This creates a new kind of power: the physical veto. In the era of globalisation, capital and software could move almost anywhere. But you cannot move a multi-billion dollar data center to a location without power, water, and local government approval. As AI demand surges, the tension between tech giants and local communities is increasing. People are beginning to realise that the digital progress they see on their screens has a direct, physical impact on their local environment—from energy consumption to land use. This gives ordinary citizens a level of leverage over the AI revolution that they never had before.

AI depends on data centers in the physical world, and building them needs permission.

The opposition to data centers is often dismissed as misinformation or NIMBYism, but there are legitimate concerns regarding resource scarcity. A data center is a voracious consumer of electricity and water. When a tech company moves into a region, they are not just bringing jobs; they are competing with residents for the very foundations of modern life. Understanding this dynamic is more important than correcting the myths used by opponents. The conflict is not about whether AI is 'good' or 'bad'; it is about who gets to control the resources required to run it.

The physical requirements of AI scaling:

Massive electrical grid capacity
High-volume water cooling systems
Land for large-scale facility construction
Specialised hardware supply chains
Local government zoning and permission

The Infrastructure Bottleneck

The bottleneck for the next decade of AI will not be the quality of the algorithms, but the availability of the physical infrastructure to run them. We are entering an era where the most successful AI companies will be those that can secure the most stable and scalable physical foundations. The digital cloud is being forced to reckon with the reality of the ground it sits on. The ability to navigate the politics of power and land will be just as important as the ability to train a transformer model.

Key Takeaway

The expansion of AI is limited by the physical world's ability to provide power, water, and land.

05 — Simon Willison

The Marketing Mirage

The FTC's crackdown on 'Active Listening' scams

By Simon Willison · 8 min read

Editor's note: The FTC settlement against Cox Media Group reveals how companies use AI terminology to mask traditional, deceptive data brokerage.

In the current tech climate, 'AI' has become a magic word—a linguistic cloak used to make old, tired business models look like futuristic breakthroughs. This was the strategy employed by Cox Media Group and several other firms, who attempted to sell advertisers a service branded as 'Active Listening'. They claimed their technology could capture real-time intent data by listening to consumers' private conversations via smart devices. It was a pitch designed to trigger both awe and anxiety, suggesting a level of surveillance that was both highly advanced and deeply invasive.

The Semantic Deception

The reality, as the FTC has now confirmed, was far less sophisticated. The 'Active Listening' service did not actually listen to anyone. There was no voice data being processed, and no real-time conversational intelligence was being deployed. Instead, the companies were simply reselling email lists obtained from other data brokers at a significant markup. They used the term 'Active Listening' to give a high-tech veneer to what was essentially a standard, old-fashioned data brokerage operation. They weren't selling intelligence; they were selling lists.

Clicking through mandatory terms of service does not constitute 'opt-in' for invasive surveillance.

The deception extended to the concept of consent. The companies claimed that consumers had 'opted in' to this surveillance simply by agreeing to the standard terms of service of various apps. The FTC has been clear: clicking through a mandatory terms-of-service agreement is not adequate consent for the collection of highly sensitive voice data from inside a person's home. This distinction is critical. If the service had actually worked as advertised, the lack of explicit, informed consent would have been a massive violation of consumer protection laws.

The deceptions identified by the FTC:

Claiming to use real-time voice data when none was used
Branding simple data brokerage as 'Active Listening' AI
Misrepresenting the method of consumer consent
Failing to accurately place ads in desired locations

The Erosion of Trust

This case is a warning about the cost of linguistic inflation in the tech industry. When companies use 'AI' as a synonym for 'something that sounds fancy', they do more than just lie to their customers; they erode the credibility of the entire field. As the gap between the marketing promises and the technical reality widens, the public's ability to distinguish between genuine innovation and predatory deception disappears. The FTC's intervention is a necessary correction, but the damage to consumer trust may be harder to repair than any fine can fix.

Key Takeaway

The misuse of 'AI' terminology to mask traditional data brokerage is a deceptive practice that undermines technological credibility.

06 — Simon Willison

The Memory Squeeze

How the AI arms race is pricing out the consumer

By Simon Willison · 7 min read

Editor's note: The demand for High-Bandwidth Memory (HBM) is cannibalising the supply of standard RAM, creating a hidden tax on consumer electronics.

The economics of consumer electronics are being rewritten by a single, unexpected factor: the hunger of the AI data center. For years, the cost of memory was relatively stable, driven by predictable demand from desktops, laptops, and smartphones. But the sudden, massive scale of the AI boom has introduced a new, dominant consumer of memory: High-Bandwidth Memory, or HBM. This is a specialised, high-performance type of RAM required to feed the massive processing power of modern GPUs. As demand for HBM explodes, it is creating a ripple effect that reaches far beyond the data center.

The Wafer Capacity Battle

The problem is one of fundamental physics and manufacturing capacity. There are only three major companies left that produce large-scale memory, and they all operate with a fixed capacity in terms of how many silicon wafers they can process at any one time. This capacity is a zero-sum game. Every wafer used to produce HBM for an AI chip is a wafer that cannot be used to produce DDR or LPDDR—the standard memory used in your phone, your laptop, and your desktop. Until recently, HBM took up a tiny fraction of total production. Now, it is expected to consume 20% of all wafer allocation by the end of 2026.

A single gigabyte of HBM consumes more than three times the wafer capacity of standard DDR.

The math is punishing. Because HBM is more complex and requires more layers, a single gigabyte of HBM consumes significantly more manufacturing resources than a gigabyte of standard memory. This means that the AI industry isn't just taking a larger slice of the pie; it is taking much larger pieces than the volume alone would suggest. Memory manufacturers have learned from history that it is better to under-provision capacity than to over-provision it. Consequently, they are leaning into the high-margin HBM market, leaving the consumer market to fight for the leftovers.

The hierarchy of memory demand:

HBM: High-margin, high-complexity, driven by AI GPUs
DDR: Standard desktop and server memory
LPDDR: Low-power memory for mobile and edge devices

The Consumer Casualty

The most immediate consequence of this squeeze is felt in the budget-friendly segment of the market. The sub-$100 smartphone, which is a critical tool for economic participation in Africa and South Asia, relies on cheap, abundant LPDDR. As the supply of that memory tightens and the price rises, these essential devices become more expensive. The AI revolution, while promising to increase global productivity, is inadvertently creating a regressive tax on the very devices that connect the next billion people to the digital economy. The cost of intelligence is being paid for by the consumer.

Key Takeaway

The massive demand for AI-specific memory is cannibalising the manufacturing capacity for standard consumer electronics.