For the first five days, I was hyper-aware that my glucose was being tracked. I ate cleaner, exercised more deliberately, and paid attention to my sleep. This mirrors the “Flash Diet” phenomenon where dieters photograph their food before eating. The act of documenting forces mindfulness, making it easier to skip the late-night snack or say no to the second serving.

After day five, the novelty wore off and my habits returned to normal.

What Even is Metabolic Health?

Metabolic health is how well your body turns food into energy without letting blood sugar, cholesterol, or fat storage get out of balance. After you eat, carbohydrates break down into glucose, which enters your bloodstream. Your pancreas responds by releasing insulin, a hormone that acts like a key, unlocking your cells so glucose can enter and be used for energy. When this system works well, blood sugar rises modestly after meals and then returns to baseline. Exercise and stress can also trigger temporary spikes because your liver releases stored glucose to fuel your muscles or prepare for a “fight or flight” response.

A continuous glucose monitor (CGM) is a small sensor inserted under the skin that tracks blood sugar in real time every few minutes. It doesn’t measure insulin directly, but it shows how your body responds throughout the day to meals, workouts, sleep, and stress. This type of device is commonly used by diabetic patients who need to manage their insulin carefully and provides a much more real-time, digital, and safer alternative to the finger-prick machines that have been around for decades.

💎 Mighty Metric: Hemoglobin A1C

An HbA1c blood test measures the percentage of hemoglobin in red blood cells that has sugar attached. Since red blood cells live for about three months, the HbA1c reflects your average blood sugar over that time.

For a healthy adult, an HbA1c below 5.7% is considered normal. From 5.7% to 6.4% puts you in the pre-diabetic range. Anything over 6.5% is diagnosed as diabetes.

Diving into the Data

Now for the interesting part: what my body actually revealed.

According to clinical guidelines, you should try to keep your glucose within the 70-140 mg/dL range throughout the day (shown as dotted lines in the chart). Excursions outside this zone are normal but should be minimized.

1. Glycemic Variability

My mean glucose over the two weeks was 93.1 mg/dL with a standard deviation (SD) of 16.9 mg/dL. While mean and SD give a basic measure of variability, there’s a better metric for my individual baseline: the Coefficient of Variability (CV).

Coefficient of Variability (CV) = Standard Deviation (SD) / Mean × 100

My CV came out to 18.1%. The International consensus on CGM sets CV <36% as the target for stable glucose control. However, research on people with normal glucose metabolism shows reference ranges of SD = 7.9-24.8 mg/dL and CV = 7.74%-22.45% so my numbers fall comfortably within the healthy range.

2. Time in Range Analysis

Time in Range (TIR) is one of the most intuitive and actionable metrics from CGM data. For people without diabetes, the target range is typically 70-140 mg/dL. My results showed 95.6% time in range, 2.1% above range, and 2.3% below range. Large-scale studies show that normoglycemic adults spend about 87% of their time in this range so my tighter control probably reflects the initial mindfulness during the first week of wearing the device.

Dr. Andrew Huberman, in his podcast with Dr. Casey Means, discusses how the goal with CGM use should be keeping glucose within this healthy range as much as possible throughout the day rather than chasing perfectly flat readings. This single metric can have the biggest effect on your overall metabolic health.

3. Mean Amplitude of Glycemic Excursions (MAGE)

MAGE is a more nuanced measure of glycemic variability. Instead of looking at all fluctuations, it focuses on the size of major glucose swings. The metric was introduced by Service et al. in 1970 as a measure of “diabetic instability” and has since become a standard way to quantify how well your body regulates glucose.

MAGE Algorithm 

1. Smooth lightly (5-point centered rolling mean) to reduce sensor noise.

2. Find all local maxima/minima via sign changes in the first derivative.

3. Build consecutive peak/nadir pairs (yellow Xs in image)

4. Keep only excursions with amplitude ≥ 1× SD of the whole series (red Xs in image)

5. Calculate the mean amplitude of those kept pairs = MAGE.

The key insight is that MAGE measures how big your average “big” spike is. This differs from CV or Time in Range because it doesn’t aggregate all readings. According to the literature, MAGE values in patients without diabetes are nearly 30 to 40 mg/dL, while diabetic patients often see values over 60 mg/dL. My MAGE for the two-week period was ~38 mg/dL, placing me at the upper end of the healthy range but well below concerning levels. Higher MAGE suggests the body struggles to return glucose to baseline smoothly, while lower MAGE indicates more effective regulation.

4. Mapping Daily Volatility

Now that I understood my overall performance, I wanted to see if the data captured my day-to-day routine and whether any obvious patterns emerged. I started with a heatmap.

The most obvious finding is that nighttime glucose levels are significantly lower and more stable compared to daytime levels. This aligns with clinical literature and common sense since my body isn’t using much glucose while I’m asleep.

In this plot, I show the average glucose at any given time across the two weeks, along with one standard deviation for variance. The daily pattern becomes clearer- my body keeps fasting glucose consistent within a desirable range of 85-100 mg/dL from midnight to 6am. From 6am to 10am, there’s a slow but very consistent rise without much variation since I don’t eat during this period. From 11am to 9pm, I’m eating meals and exercising, so glucose rises and falls with high variation. The meal times weren’t consistent, nor were the foods I ate, so the high variation makes sense.

5. Chasing the Dawn Effect

One phenomenon I was particularly curious about was the Dawn Effect.

Normally, while you sleep, your body keeps blood sugar steady, which I could see in my data. But just before waking up (around 4–6am), certain hormones like cortisol and growth hormone kick in and tell the liver to release stored sugar into the blood, so you have energy to start the day. This is the Dawn Effect.

I usually wake up around 6am, so I looked at the raw data in the 4-7am window. The rises are there but inconsistent and noisy, making it hard to pinpoint the actual rise or quantify it.

First attempt: too noisy. Since the human body isn’t perfectly calibrated, the timing of the dip and peak don’t align naturally across days. Comparing the delta before and after 6am didn’t show a strong signal. Instead, I anchored each day’s trend line to the minimum glucose value within the pre-wakeup window (3-5am), aligning them at the same relative timestamp.

Second attempt: better alignment. The trend became somewhat clearer, so I used mean and standard deviation to aggregate the data and see if there’s a consistent rise from the minimum.

Final result: there it is! There’s a clear upward slope, and the difference from nadir to peak is ~10 mg/dL, which matches what clinical studies suggest for the dawn effect.

The wide variance band reflects normal day-to-day fluctuations rather than measurement error. Studies show median inter-day differences of 15 mg/dL are typical. My 3-5am window for finding the minimum also aligns with the physiology research, as hormones like cortisol and epinephrine surge from nocturnal nadirs between 4:00 and 6:30 a.m.

While the hormonal processes behind the dawn effect occur in most people, healthy individuals compensate with insulin secretion to prevent sustained hyperglycemia. The pronounced effect is primarily seen in diabetic patients and my mild rise followed by stabilization suggests normal metabolic function.

Conclusion

Overall, CGMs are an excellent tool for finding problems in your diet and fixing them proactively. During my two-week experiment, I noticed that red grapes caused a much larger spike compared to blueberries or blackberries despite the published Glycemic Index (GI) data. This highlights why individual testing matters more than population averages.

🍇 Key Takeaway

Red grapes (GI: 43-59) and blueberries (GI: 53) should theoretically cause similar spikes, but my body responded very differently to red grapes with much larger relative spikes in blood glucose.

What interests me is the broader potential of this technology at scale.

CGM companies like Dexcom and Abbott publish studies showing how device features correlate with improved A1C levels and time-in-range metrics. Wellness companies like Oura and WHOOP have demonstrated early illness detection capabilities, identifying COVID-19 infections up to 2.5 days before symptom onset through continuous biometric monitoring. The future of personalized medicine lies not only in aggregating population health data for trends but in learning from millions of individuals at scale, then applying those insights to optimize interventions for each person.

This is something I’ve seen in action with Epic’s Cosmos. Their “Best Care Choices For My Patient” tool allows physicians to see how similar patients responded to specific treatments. For instance, when treating a newly diagnosed hypertension patient, a doctor can view real-world outcomes from the 274 million patient dataset based on age, comorbidities, and previous treatments, rather than relying solely on general guidelines.

Until that future arrives at scale, if you’re interested in an in-depth analysis of your own CGM data, reach out and I’ll show you how to better understand your numbers and optimize your health.

Subscribe now

The intuition behind Mixture of Experts is a simple divide and conquer algorithm. Break down a problem into specific sub-problems; assign tasks to “experts” best designed to solve those sub-problems and efficiently combine the results into a final answer. While Adam Smith may have been the first to observe how specialization leads to higher productivity in economics, the concept of division of labor is an integral part of human biology. In some sense, our brain is also a mixture of experts— consisting of several specialized structures, each one responsible for a specific function while working together to give life to the human experience.

When you touch the back of your head at the spot where you feel a bump, you are touching the area directly above your occipital lobe. This lobe is one of the four primary segments of the gray matter, or cortex, within your brain, which also includes the Frontal, Parietal, and Temporal lobes. Each of these lobes is tasked with performing specialized functions: decision-making, interpreting sensory information, visualizing images, and understanding spoken language. Similarly, a Mixture of Experts model comprises smaller expert models—typically eight—combined together through a "router" which takes the model inputs, selects the appropriate experts, and then produces the final output. This router acts similarly to the Thalamus, which is considered the "relay station" of the brain, connecting sensory nerves to the correct parts of the cortex and carrying return signals to the rest of the body.

💎 Mighty Metric

There are roughly 100 billion neurons in the Human Brain. By comparison, Google's Switch Transformer (one of the largest MoE models) has 1.6 trillion parameters!

The similarities between artificial intelligence models and the human brain don’t stop at the functional hierarchy of their components. In fact, the fundamental building blocks of all modern AI models are based on a simple processing unit: neurons. Layers of neurons make up a neural network, which is an efficient way to meaningfully represent inputs, store relationships between those inputs, and perform complex mathematical operations. What is commonly referred to as "learning" is simply the process of changing the relationships based on large amounts of training data. Obviously, neurons in computers are inspired by the neurons in our brains, which are living cells and perform a series of electrochemical processes to communicate with other neurons. In our brain, the learning process is continuous but similar—neurons that consistently communicate with each other form a stronger relationship. Indeed, neurons that fire together wire together.

While neural networks have been around for a long time and have shown great improvements in performance through different architectures and parameter sizes, Mixture of Experts models add new layers of sophistication beyond basic biology. From a cognitive science perspective, this architecture can be seen as an implementation of a key framework to describe human consciousness—the Global Workspace Theory. The theory tries to explain the properties of consciousness, whether human or non-human, and lists a set of indicators that are necessary for a system to be considered conscious. From the properties listed in table below, one can see that the Mixture of Experts takes a giant leap towards making artificial intelligence more than just fancy number crunching and set the stage for Artificial General Intelligence.

Practical Reasons for Mixture of Experts

Apart from reorganizing the flow of information in existing large language models, a Mixture of Experts also helps address important technical challenges:

1. The most important factor in improving model quality is the size of the model. A Mixture of Experts allows scaling up to a larger parameter size without updating the weights (relationships between neurons) for the entire model but only for individual experts. This sparse representation of the model allows for compute-efficient training.

2. Additionally, the unique structure of a Mixture of Experts model allows for further optimization using a technique known as Expert Parallelism—allowing experts to run independently on different GPUs.

3. Lastly, a Mixture of Experts model allows for better inference speeds and higher throughput when the models are deployed, since the model only uses a small portion of its total number of parameters at any given time.

In the West, China often has a reputation for copying technology rather than leading the way with innovation. This was true in the internet era when the restrictive policies of the Great Firewall allowed Chinese tech giants to emerge and replicate business models similar to those of Google, Amazon, and eBay, in the form of Baidu, Alibaba, and JD.com. Later, in the mobile age, however, Chinese tech companies played a much more dominant role, and Xiaomi, Huawei, and Tencent built products that evolved organically and found large markets for entry-level smartphones and entertainment in developing countries. Now, in the age of artificial intelligence, the gap is getting even narrower, and Chinese companies are leapfrogging the US in many sectors. SenseTime and Megvii are global leaders in facial recognition and deep learning (even though their technology is used for surveillance). DJI is an undisputed leader in the commercial drone market with over a 70% market share and uses artificial intelligence for its autonomous flight and video processing. TikTok’s addictive but innovative recommendation algorithm is changing how GenZ around the world communicates, consumes content, and shops online.

Progress in artificial intelligence is heavily dependent on research, a domain where China has gradually outpaced and outperformed the United States. A 2021 study by Nikkei Asia revealed that China contributed about 40% of the global total of AI research papers, whereas the US accounted for only 10%. This trend continues not just for the volume but also for the overall quality of research. Among the top 10% of the most-cited research papers, Chinese papers outnumbered US papers by over 70%. Many of these papers are contributions from elite academic and research institutions such as the Chinese Academy of Sciences and Tsinghua University. This phenomenon can be attributed to China’s government policies under the Artificial Intelligence Development Plan in 2017, which set an ambitious goal for China to become the world’s leading AI innovation center by 2030. In the United States, on the other hand, research is led primarily by large tech companies and well-funded startups who, for now, still retain the edge in top-tier research quality. Six of the top ten research institutions from the Nikkei study are large US companies, the remaining four were Chinese.

In the arms race to develop artificial intelligence, another key frontier is Open Source. While Open Source development by itself is not a defensible moat, it allows companies to extend their reach and build ecosystems around their platforms. Once the backyard of large US tech companies, open source is slowly embracing developments by Chinese companies. 01.AI, a startup founded by AI veteran Kai-Fu Lee, released the Yi family of open source foundation models which quickly rose to the top of the leaderboard on HuggingFace. The company is now worth $1 billion, and their GitHub repository has over 6.8k stars at the time of writing. On the flip side, the Yi models are suspiciously similar to the underlying architecture of Meta’s Llama models and the open source community in China is not nearly as active as that in the US. According to GitHub’s annual report, China ranked #9 in creating new generative AI projects, meanwhile, the United States far outranks it at #1.

💎 Mighty Metric

China aims to increase its computing capacity to 300 exaflops by 2025. A single exaflop represents roughly the same computing power as 2 million mainstream laptops!

Even though Chinese policies attempt to lead innovation through research and involvement in open source, a major obstacle is the availability of computing power—a key ingredient in training large models. In 2022, the United States banned the export of high-performance computing equipment such as GPUs, as well as other hardware, software, and related semiconductor fabrication technology. Since Chinese design and manufacturing capabilities lag significantly behind those of the US and Taiwan, China plans to overcome these restrictions with a $41 billion "Big Fund" aimed at bolstering indigenous chip manufacturing. Furthermore, China has allocated $143 billion, spread over five years, in tax credits and subsidies to stimulate innovation and growth in the industry. The outcome of the race to AI supremacy remains uncertain, but there is no doubt that China intends to compete head-to-head to become a global AI superpower.

Over the decades, researchers have tried many complex approaches to build better language models, but the bitter truth is that scaling simpler architectures achieves the best outcomes. The intuition behind this approach can be understood by the scaling hypothesis:

“The strong scaling hypothesis is that, once we find a scalable architecture like self-attention or convolutions, which like the brain can be applied fairly uniformly, we can simply train ever larger neural networks and ever more sophisticated behavior will emerge naturally as the easiest way to optimize for all the tasks & data. More powerful neural networks are ‘just’ scaled-up weak neural networks, in much the same way that human brains look much like scaled-up primate brains⁠.”

There has been an increasing amount of evidence supporting this hypothesis—the most important being the emergence of abilities that the model was not intentionally trained on. In the natural world, emergent phenomena are complex outcomes, patterns, and behaviors that arise due to interactions between simple components. These phenomena exist in all kinds of domains, from the formation of snowflakes to the structuring of an ant colony. Similarly, with larger model sizes, more data, and more compute, language models transition from near-zero performance to nearly state-of-the-art performance at a rapid and unpredictable rate when they reach a critical scale. Interestingly, language models acquire skills that are often unrelated, like arithmetic, creative writing, and humor.

Source: Google Research — As the scale of the model increases, the performance improves across tasks while also unlocking new capabilities.

Subscribe now

Even though large language models have been monumental in their achievements, surpassing Turing tests, their business applications almost never require the entire gamut of Artificial General Intelligence. Instead, businesses are focused on solving problems through specialized, narrowly focused tasks, coupled with the necessity to handle proprietary data securely. Moreover, the extravagantly high costs of these colossal models (capital, compute, energy) raise an important question: Can the essence of their advanced capabilities be captured within a smaller model? Smaller models, apart from offering cheaper training and lower compute requirements, allow for bespoke customization, domain-specific business logic, faster development cycles, and potentially on-premises deployment.

Since the real value proposition of using large language models stems from a deeper understanding of user intent, smaller language models can still usher in a new interaction paradigm with businesses employing different models for various tasks. For example, a customer service chatbot used for answering FAQs on a travel website doesn’t need the same kind of abilities or knowledge as one helping software developers write code faster. Additionally, smaller models allow businesses and hobbyists to easily fine-tune models to specific domains, distill knowledge into workflows, and use Retrieval Augmented Generation to query a secure, proprietary database—all without the cost of a large language model.

💎 Mighty Metric

Meta’s largest Llama-2 model was trained for 1,720,320 GPU hours. This roughly equals 196 years on a single GPU, or using 10,240 GPUs to complete training in 1 week!

In 2023, Microsoft released Phi-2, a small language model with only 2.7 billion parameters, trained on highly curated, AI-generated synthetic data to outperform models up to 25 times larger. Several other foundational models now come as a triad of different sizes—Meta’s Llama-2, Anthropic’s Claude 3, and Google’s Gemini, each sporting a small language model variant. Although the smaller model is eclipsed by its larger siblings in size, its performance is good enough for most people, especially considering that these models require less compute-intensive hardware to run. In fact, Google’s Gemini Nano is designed to run on mobile devices and powers the Pixel 8 Pro.

Amid the flurry of small language models, it is worth noting that the Specialist over Generalist paradigm works only because we have reached a tipping point in understanding natural language through artificial intelligence. What is often cited as the teacher-student paradigm, where training smaller models through synthetic data generated by larger models succeeds, is due to the emergent properties—showing that language models have learned how to learn. Only after achieving the critical scale and paying the upfront fixed cost of training large language models can we start to scale down and save on the variable costs of deploying small language models.

SemiAnalysis

Google "We Have No Moat, And Neither Does OpenAI"

The text below is a very recent leaked document, which was shared by an anonymous individual on a public Discord server who has granted permission for its republication. It originates from a researcher within Google. We have verified its authenticity. The only modifications are formatting and removing links to internal web pages. The document is only the opinion of a Google employee, not the entire firm. We do not agree with what is written below, nor do other researchers we asked, but we will publish our opinions on this in a separate piece for subscribers. We simply are a vessel to share this document which raises some very interesting points…

Read more

3 years ago · 673 likes · 10 comments · Dylan Patel and Afzal Ahmad

The core argument of the Google memo is that the worldwide community of open source developers is much more capable than research labs of large tech companies, owing to their speed of execution, adoption of the latest ideas, and sharing of resources. As a result, models developed by open source are often faster, more customizable, and, pound-for-pound, better than those from closed-source companies. While individual hobbyists, academic researchers, and underpaid graduate students have successfully addressed some of the biggest challenges in artificial intelligence, the truth is that they, too, have no moat. To maintain a defensible advantage for the open source model, we should consider three different aspects of artificial intelligence: development, deployment, and distribution.

1. Development of new AI models, tools, datasets and benchmarks.

Specialized knowledge, the latest research papers, and training datasets are often freely available, and the best ideas can come from anywhere. Tapping into the global talent pool, often at no cost, has proven to be an excellent strategy for innovation, as seen by the timeline of progress in the memo. However, there's a dark side: open-source projects frequently find themselves locked into platforms owned or sponsored by major tech companies, which ultimately stand to benefit the most, whether directly or indirectly. Some notable examples include Chromium, Android, React, PyTorch, and Kubernetes. Specifically, in the context of open-source AI, many models today are built atop the leaked version of Meta’s Llama. As of this writing, there are over 8.6k forks of their GitHub repository, and this extensive development gives Facebook Research the opportunity to explore all the possibilities of their models without having to do all the work. What should have been a moat for open source has become a moat for the company that owns the platform.

2. Deployment of AI models as services utilizing expensive compute resources, data centers, and extensive cloud management.

This is the most capital-intensive aspect of producing artificial intelligence, and it's something beyond the reach of not just open source communities but also most companies. Only a few players—Microsoft Azure, Google Cloud Platform, and Amazon Web Services—have a strong foothold in this market, in part due to the ongoing GPU shortage. Even the most popular open-source platform for AI development, HuggingFace, operates largely because corporations like Google, Amazon, and NVIDIA pay premiums for a private version of the platform, which subsidizes usage for open-source developers. When it comes to deployment, open source not only lacks a competitive edge, but the cost of computing itself also becomes a barrier to entry for training new AI models from scratch.

3. Distribution of artificial intelligence in products and services that are used by consumers.

The most exciting aspect of artificial intelligence lies in its ability to perform tasks that were previously considered nearly impossible to do at scale. The real value proposition of artificial intelligence will fundamentally alter our interaction with the technology around us. Arguably, this represents the most significant force multiplier in the recent history of humankind, and it is easy to understand why there is explosive growth — ChatGPT reached 1 million users in just 5 days. However, with growth and opportunity comes competition. Historically, many open-source projects have been under-resourced, lacking both funds and community contributors. In such a competitive landscape, open-source projects struggle to organize development, design quality products, and market to customers for widespread adoption and distribution.

There is no doubt that open source projects generally do not achieve the same level of success as closed-source companies. This observation shouldn't surprise anyone— RedHat, the poster child for successful open-source enterprises, pales in comparison to Microsoft. Although an oversimplification, both companies initially created operating systems. In 2018, RedHat was acquired by IBM for $34 billion and Microsoft was worth $780 billion. As of this writing, Microsoft is worth a staggering $3 trillion.

This discussion holds even greater significance today as artificial intelligence increasingly automates and substitutes for human reasoning and judgment. Abundant research highlights the inherent risks and biases in models, including misinformation, racial, and gender discrimination, among others. Theoretically, open-source AI should increase the likelihood that the wisdom of the crowd will identify and correct biases, uncover issues, and suggest enhancements. However, in reality, the incentives are not aligned for open source to succeed or compete with large tech companies. Without organized control over the development, deployment, or distribution of artificial intelligence, the open source artificial intelligence community is going to be at the mercy of Big Tech.

Almost a year since the memo leaked, Google has released its family of small language models, Gemma, based on the same work as its flagship Gemini models. With the weights made publicly available, Google is hoping to engage with the open source community for research and development. Only time will tell whether this release will help build a moat for open source or deepen the moat for Google.

The current paradigm of products built with large language models is heavily influenced by Microsoft's paper, Sparks of Artificial General Intelligence: Early Experiments with GPT-4. More resembling a sales pitch than a technical research paper, it showcases examples of OpenAI's flagship model completing a variety of tasks, from answering simple mathematical questions to generating lines of executable code. Since its release in March 2023, this paper has become standard reading material for corporate strategy on Generative AI. Many companies have simply adopted the examples in the paper for features and products in their own industries, building wrappers around the foundational GPT-4 model.

However, taking a long-term view of technology's evolution since the beginning of the computer age, large language models fundamentally changes how humans interact with technology. The journey from punch cards, through typing commands into a terminal, to the advent of point-and-click graphical user interfaces, culminates in our ability to communicate with computers using the same medium we use with other humans: language. While no one knows what the future will look like, we can be certain that our relationship with technology will become much more intuitive, and we will start to see our devices not just as mere tools, but as entities that are more than just extensions of ourselves but also extensions of human society and culture.

We live in a fascinating time in human history when artificial intelligence is capable of accomplishing feats once considered impossible by computers. However, the technology is still in its infancy and we truly do not know what the future looks like - unless we build it.

These essays serve as a collection of great ideas, timeless wisdom and practical insights that any leader can apply to their business to innovate and dominate.

Subscribe now