Category: What I’m Reading

  • Google’s Quantum Error Correction Breakthrough

    One of the most exciting areas of technology development, but that doesn’t get a ton of mainstream media coverage, is the race to build a working quantum computer that exhibits “below threshold quantum computing” — the ability to do calculations utilizing quantum mechanics accurately.

    One of the key limitations to achieving this has been the sensitivity of quantum computing systems — in particular the qubits that capture the superposition of multiple states that allow quantum computers to exploit quantum mechanics for computation — to the world around them. Imagine if your computer’s accuracy would change every time someone walked in the room — even if it was capable of amazing things, it would not be especially practical. As a result, much research to date has been around novel ways of creating physical systems that can protect these quantum states.

    Google has (in a pre-print in Nature) demonstrated their new Willow quantum computing chip which demonstrates a quantum error correction method that spreads the quantum state information of a single “logical” qubit across multiple entangled “physical” qubits to create a more robust system. Beyond proving that their quantum error correction method worked, what is most remarkable to me, is that they’re able to extrapolate a scaling law for their error correction — a way of guessing how much better their system is at avoiding loss of quantum state as they increase the number of physical qubits per logical qubit — which could suggest a “scale up” path towards building functional, practical quantum computers.

    I will confess that quantum mechanics was never my strong suit (beyond needing it for a class on statistical mechanics eons ago in college), and my understanding of the core physics underlying what they’ve done in the paper is limited, but this is an incredibly exciting feat on our way towards practical quantum computing systems!


  • Cynefin

    I had never heard of this framework for thinking about how to address problems before. Shout-out to my friend Chris Yiu and his new Substack Secret Weapon about improving productivity for teaching me about this. It’s surprisingly insightful about when to think about something as a process problem vs an expertise problem vs experimentation vs direction.


    Problems come in many forms
    Chris Yiu | Secret Weapon

  • The Hits Business — Games Edition

    The best return on investment in terms of hours of deep engagement per dollar in entertainment is with games. When done right, they blend stunning visuals and sounds, earworm-like musical scores, compelling story and acting, and a sense of progression that are second to none.

    Case in point: I bought the complete edition of the award-winning The Witcher 3: Wild Hunt for $10 during a Steam sale in 2021. According to Steam, I’ve logged over 200 hours (I had to doublecheck that number!) playing the game, between two playthroughs and the amazing expansions Hearts of Stone and Blood and Wine — an amazing 20 hours/dollar spent. Even paying full freight (as of this writing, the complete edition including both expansions costs $50), that would still be a remarkable 4 hours/dollar. Compare that with the price of admission to a movie or theater or concert.

    The Witcher 3 has now surpassed 50 million sales — comfortably earning over $1 billion in revenue which is an amazing feat for any media property.

    But as amazing and as lucrative as these games can be, these games cannot escape the cruel hit-driven basis of their industry, where a small number of games generate the majority of financial returns. This has resulted in studios chasing ever more expensive games with familiar intellectual property (i.e. Star Wars) that has, to many game players, cut the soul from the games and has led to financial instability in even popular game studios.

    This article from IGN summarizes the state of the industry well — with so-called AAA games now costing $200 million to create, not to mention $100’s of millions to market, more and more studios have to wind down as few games can generate enough revenue to cover the cost of development and marketing.

    The article predicts — and I hope it’s right — that the games industry will learn some lessons that many studios in Hollywood/the film industry have been forced to: embrace more small budget games to experiment with new forms and IP. Blockbusters will have their place but going all-in on blockbusters is a recipe for a hollowing out of the industry and a cutting off of the creativity that it needs.

    Or, as the author so nicely puts it: “Maybe studios can remember that we used to play video games because they were fun – not because of their bigger-than-last-year maps carpeted by denser, higher-resolution grass that you walk across to finish another piece of side content that pushes you one digit closer to 100% completion.”


  • The Challenge of Capacity

    The rise of Asia as a force to be reckoned with in large scale manufacturing of critical components like batteries, solar panels, pharmaceuticals, chemicals, and semiconductors has left US and European governments seeking to catch up with a bit of a dilemma.

    These activities largely moved to Asia because financially-motivated management teams in the West (correctly) recognized that:

    • they were low return in a conventional financial sense (require tremendous investment and maintenance)
    • most of these had a heavy labor component (and higher wages in the US/European meant US/European firms were at a cost disadvantage)
    • these activities tend to benefit from economies of scale and regional industrial ecosystems, so it makes sense for an industry to have fewer and larger suppliers
    • much of the value was concentrated in design and customer relationship, activities the Western companies would retain

    What the companies failed to take into account was the speed at which Asian companies like WuXi, TSMC, Samsung, LG, CATL, Trina, Tongwei, and many others would consolidate (usually with government support), ultimately “graduating” into dominant positions with real market leverage and with the profitability to invest into the higher value activities that were previously the sole domain of Western industry.

    Now, scrambling to reposition themselves closer to the forefront in some of these critical industries, these governments have tried to kickstart domestic efforts, only to face the economic realities that led to the outsourcing to begin with.

    Northvolt, a major European effort to produce advanced batteries in Europe, is one example of this. Despite raising tremendous private capital and securing European government support, the company filed for bankruptcy a few days ago.

    While much hand-wringing is happening in climate-tech circles, I take a different view: this should really not come as a surprise. Battery manufacturing (like semiconductor, solar, pharmaceutical, etc) requires huge amounts of capital and painstaking trial-and-error to perfect operations, just to produce products that are steadily dropping in price over the long-term. It’s fundamentally a difficult and not-very-rewarding endeavor. And it’s for that reason that the West “gave up” on these years ago.

    But if US and European industrial policy is to be taken seriously here, the respective governments need to internalize that reality and be committed for the long haul. The idea that what these Asian companies are doing is “easily replicated” is simply not true, and the question is not if but when will the next recipient of government support fall into dire straits.


  • A Visual Timeline of Human Migration

    Beautiful map laying out when humans settled different parts of the world (from 2013, National Geographic’s Out of Eden project)


    A Walk Through Time
    Jeff Blossom | National Geographic

  • Not your grandma’s geothermal energy

    The pursuit of carbon-free energy has largely leaned on intermittent sources of energy — like wind and solar; and sources that require a great deal of initial investment — like hydroelectric (which requires elevated bodies of water and dams) and nuclear (which require you to set up a reactor).

    The theoretical beauty of geothermal power is that, if you dig deep enough, virtually everywhere on planet earth is hot enough to melt rock (thanks to the nuclear reactions that heat up the inside of the earth). But, until recently, geothermal has been limited to regions of Earth where well-formed geologic formations can deliver predictable steam without excessive engineering.

    But, ironically, it is the fracking boom, which has helped the oil & gas industries get access to new sources of carbon-producing energy, which may help us tap geothermal power in more places. As fracking and oil & gas exploration has led to a revolution in our ability to precisely drill deep underground and push & pull fluids, it also presents the ability for us to tap more geothermal power than ever before. This has led to the rise of enhanced geothermal, the process by which we inject water deep underground to heat, and leverage the steam produced to generate electricity. Studies suggest the resource is particularly rich and accessible in the Southwest of the United States (see map below) and could be an extra tool in our portfolio to green energy consumption.

    (Source: Figure 5 from NREL study on enhanced geothermal from Jan 2023)

    While there is a great deal of uncertainty around how much this will cost and just what it will take (not to mention the seismic risks that have plagued some fracking efforts), the hunger for more data center capacity and the desire to power this with clean electricity has helped startups like Fervo Energy and Sage Geosystems fund projects to explore.


  • Who needs humans? Lab of AIs designs valid COVID-binding proteins

    A recent preprint from Stanford has demonstrated something remarkable: AI agents working together as a team solving a complex scientific challenge.

    While much of the AI discourse focuses on how individual large language models (LLMs) compare to humans, much of human work today is a team effort, and the right question is less “can this LLM do better than a single human on a task” and more “what is the best team-up of AI and human to achieve a goal?” What is fascinating about this paper is that it looks at it from the perspective of “what can a team of AI agents achieve?”

    The researchers tackled an ambitious goal: designing improved COVID-binding proteins for potential diagnostic or therapeutic use. Rather than relying on a single AI model to handle everything, the researchers tasked an AI “Principal Investigator” with assembling a virtual research team of AI agents! After some internal deliberation, the AI Principal Investigator selected an AI immunologist, an AI machine learning specialist, and an AI computational biologist. The researchers made sure to add an additional role, one of a “scientific critic” to help ground and challenge the virtual lab team’s thinking.

    The team composition and phases of work planned and carried out by the AI principal investigator
    (Source: Figure 2 from Swanson et al.)

    What makes this approach fascinating is how it mirrors high functioning human organizational structures. The AI team conducted meetings with defined agendas and speaking orders, with a “devil’s advocate” to ensure the ideas were grounded and rigorous.

    Example of a virtual lab meeting between the AI agents; note the roles of the Principal Investigator (to set agenda) and Scientific Critic (to challenge the team to ground their work)
    (Source: Figure 6 from Swanson et al.)

    One tactic that the researchers said helped with boosting creativity that is harder to replicate with humans is running parallel discussions, whereby the AI agents had the same conversation over and over again. In these discussions, the human researchers set the “temperature” of the LLM higher (inviting more variation in output). The AI principal investigator then took the output of all of these conversations and synthesized them into a final answer (this time with the LLM temperature set lower, to reduce the variability and “imaginativeness” of the answer).

    The use of parallel meetings to get “creativity” and a diverse set of options
    (Source: Supplemental Figure 1 from Swanson et al.)

    The results? The AI team successfully designed nanobodies (small antibody-like proteins — this was a choice the team made to pursue nanobodies over more traditional antibodies) that showed improved binding to recent SARS-CoV-2 variants compared to existing versions. While humans provided some guidance, particularly around defining coding tasks, the AI agents handled the bulk of the scientific discussion and iteration.

    Experimental validation of some of the designed nanobodies; the relevant comparison is the filled in circles vs the open circles. The higher ELISA assay intensity for the filled in circles shows that the designed nanbodies bind better than their un-mutated original counterparts
    (Source: Figure 5C from Swanson et al.)

    This work hints at a future where AI teams become powerful tools for human researchers and organizations. Instead of asking “Will AI replace humans?”, we should be asking “How can humans best orchestrate teams of specialized AI agents to solve complex problems?”

    The implications extend far beyond scientific research. As businesses grapple with implementing AI, this study suggests that success might lie not in deploying a single, all-powerful AI system, but in thoughtfully combining specialized AI agents with human oversight. It’s a reminder that in both human and artificial intelligence, teamwork often trumps individual brilliance.

    I personally am also interested in how different team compositions and working practices might lead to better or worse outcomes — for both AI teams and human teams. Should we have one scientific critic, or should their be specialist critics for each task? How important was the speaking order? What if the group came up with their own agendas? What if there were two principal investigators with different strengths?

    The next frontier in AI might not be building bigger models, but building better teams.

  • A Digital Twin of the Whole World in the Cloud

    As a kid, I remember playing Microsoft Flight Simulator 5.0 — while I can’t say I really understood all the nuances of the several hundred page manual (which explained how ailerons and rudders and elevators worked), I remember being blown away with the idea that I could fly anywhere on the planet and see something reasonably representative there.

    Flash forward a few decades and Microsoft Flight Simulator 2024 can safely be said to be one of the most detailed “digital twins” of the whole planet ever built. In addition to detailed photographic mapping of many locations (I would imagine a combination of aerial surveillance and satellite imagery) and an accurate real world inventory of every helipad (including offshore oil rigs!) and glider airport, they also simulate flocks of animals, plane wear and tear, how snow vs mud vs grass behave when you land on it, wake turbulence, and more! And, just as impressive, it’s being streamed from the cloud to your PC/console when you play!

    Who said the metaverse is dead?


  • Making a Movie to Make Better Video Encoding

    Until I read this Verge article, I had assumed that video codecs were a boring affair. In my mind, every few years, the industry would get together and come up with a new standard that promised better compression and better quality for the prevailing formats and screen types and, after some patent licensing back and forth, the industry would standardize around yet another MPEG standard that everyone uses. Rinse and repeat.

    The article was an eye-opening look at how video streamers like Netflix are pushing the envelope on using video codecs. Since one of a video streamer’s core costs is the cost of video bandwidth, it would make sense that they would embrace new compression approaches (like different kinds of compression for different content, etc.) to reduce those costs. As Netflix embraces more live streaming content, it seems they’ll need to create new methods to accommodate.

    But what jumped out to me the most was that, in order to better test and develop the next generation of codec, they produced a real 12 minute noir film called Meridian (you can access it on Netflix, below is someone who uploaded it to YouTube) which presents scenes that have historically been more difficult to encode with conventional video codecs (extreme lights and shadows, cigar smoke and water, rapidly changing light balance, etc).

    Absolutely wild.


  • Games versus Points

    The Dartmouth College Class of 2024, for their graduation, got a very special commencement address from tennis legend Roger Federer.

    There is a wealth of good advice in it, but the most interesting point that jumped out to me is that while Federer won a whopping 80% of the matches he played in his career, he only won 54% of the points. It underscores the importance of letting go of small failures (“When you lose every second point, on average, you learn not to dwell on every shot”) but also of keeping your eye on the right metric (games, not points).


  • Biopharma scrambling to handle Biosecure Act

    Strong regional industrial ecosystems like Silicon Valley (tech), Boston (life science), and Taiwan (semiconductors) are fascinating. Their creation is rare and requires local talent, easy access to supply chains and distribution, academic & government support, business success, and a good amount of luck.

    But, once set in place, they can be remarkably difficult to unseat. Take the semiconductor industry as an example. It’s geopolitical importance has directed billions of dollars towards re-creating a domestic US industry. But, it faces an uphill climb. After all, it’s not only a question of recreating the semiconductor manufacturing factories that have gone overseas, but also:

    • the advanced and low-cost packaging technologies and vendors that are largely based in Asia
    • the engineering and technician talent that is no longer really in the US
    • the ecosystem of contractors and service firms that know exactly how to maintain the facilities and equipment
    • the supply chain for advanced chemicals and specialized parts that make the process technology work
    • the board manufacturers and ODMs/EMSs who do much of the actual work post-chip production that are also concentrated in Asia

    A similar thing has happened in the life sciences CDMO (contract development and manufacturing organization) space. In much the same way that Western companies largely outsourced semiconductor manufacturing to Asia, Western biopharma companies outsourced much of their core drug R&D and manufacturing to Chinese companies like WuXi AppTec and WuXi Biologics. This has resulted in a concentration of talent and an ecosystem of talent and suppliers there that would be difficult to supplant.

    Enter the BIOSECURE Act, a bill being discussed in the House with a strong possibility of becoming a law. It prohibits the US government from working with companies that obtain technology from Chinese biotechnology companies of concern (including WuXi AppTec and WuXi Biologics, among others). This is causing the biopharma industry significant anxiety as they are forced to find (and potentially fund) an alternative CDMO ecosystem that currently does not exist at the level of scale and quality as it does with WuXi.


  • Freedom and Prosperity Under Xi Jinping

    Fascinating chart from Bloomberg showing level of economic freedom and prosperity under different Chinese rulers and how Xi Jinping is the first Chinese Communist Party ruler in history to have presided over sharp declines in both freedom and prosperity.

    Given China’s rising influence in economic and geopolitical affairs, how it’s leaders (and in particular, Xi) and it’s people react to this will have significant impacts on the world



    ‘Are You Better Off?’ Asking Reagan’s Question in Xi’s China
    Rebecca Choong Wilkins and Tom Orlik | Bloomberg

  • How the Jones Act makes energy more expensive and less green

    The Merchant Marine Act of 1920 (aka “The Jones Act”) is a law which requires ships operating between US ports to be owned by, made in, and crewed by US citizens.

    While many “Made in the USA” laws are on the books and attract the anger of economists and policy wonks, the Jones Act is particularly egregious as the costs and effects are so large. The Jones Act costs states like Hawaii and Alaska and territories like Puerto Rico dramatically as they rely so much on ships for basic commerce that it was actually cheaper for Hawaii and New England to import oil from other countries (like Hawaii did from Russia until the Ukraine war) than it was to have oil shipped from the Gulf of Mexico (where American oil is abundant).

    In the case of offshore wind, the Jones Act has pushed those companies willing to experiment with the promising technology, to ship the required parts and equipment from overseas because there are no Jones Act-compliant ships capable of moving the massive equipment that is involved.

    This piece from Canary Media captures some of the dynamics and the “launch” of the still-in-construction $625 million Jones Act-compliant ship the Charybdis Dominion Energy will use to support its offshore wind facility.


  • Why Intel has to make its foundry business work

    Historically, Intel has (1) designed and (2) manufactured its chips that it sells (primarily into computer and server systems). It prided itself on having the most advanced (1) designs and (2) manufacturing technology, keeping both close to its chest.

    In the late 90s/00s, semiconductor companies increasingly embraced the “fabless model”, whereby they would only do the (1) design while outsourcing the manufacturing to foundries like TSMC. This made it much easier and less expensive to build up a burgeoning chip business and is the secret to the success of semiconductor giants like NVIDIA and Qualcomm.

    Companies like Intel scoffed at this, arguing that the combination of (1) design and (2) manufacturing gave their products an advantage, one that they used to achieve a dominant position in the computing chip segment. And, it’s an argument which underpins why they have never made a significant effort in becoming a contract manufacturer — after all, if part of your technological magic is the (2) manufacturing, why give it to anyone else?

    The success of TSMC has brought a lot of questions about Intel’s advantage in manufacturing and, given recent announcements by Intel and the US’s CHIPS Act, a renewed focus on actually becoming a contract manufacturer to the world’s leading chip designers.

    While much of the attention has been paid to the manufacturing prowess rivalry and the geopolitical reasons behind this, I think the real reason Intel has to make the foundry business work is simple: their biggest customers are all becoming chip designers.

    While a lot of laptops and desktops and servers are still sold in the traditional fashion, the reality is more and more of the server market is being dominated by a handful of hyperscale data center operators like Amazon, Google, Meta/Facebook, and Microsoft, companies that have historically been able to obtain the best prices from Intel because of their volume. But, in recent years, in the chase for better and better performance and cost and power consumption, they have begun designing their own chips adapted to their own systems (as this latest Google announcement for Google’s own ARM-based server chips shows).

    Are these chips as good as Intel’s across every dimension? Almost certainly not. It’s hard to overtake a company like Intel’s decades of design prowess and market insight. But, they don’t have to be. They only have to be better at the specific use case Google / Microsoft / Amazon / etc need it to be for.

    And, in that regard, that leaves Intel with really only one option: it has to make the foundry business work, or it risks losing not just the revenue from (1) designing a data center chip, but from the (2) manufacturing as well.


  • Starlink in the wrong hands

    On one level, this shouldn’t be a surprise. Globally always available satellite constellation = everyone and anyone will try to access this. This was, like many technologies, always going to have positive impacts — i.e. people accessing the internet where they otherwise couldn’t due to lack of telecommunications infrastructure or repression — and negative — i.e. terrorists and criminal groups evading communications blackouts.

    The question is whether or not SpaceX had the foresight to realize this was a likely outcome and to institute security processes and checks to reduce the likelihood of the negative.

    That remains to be seen…


    Elon Musk’s Starlink Terminals Are Falling Into the Wrong Hands
    Bruce Einhorn, Loni Prinsloo, Marissa Newman, Simon Marks | Bloomberg

  • Why don’t we (still) have rapid viral diagnostics?

    One of the most disappointing outcomes in the US from the COVID pandemic was the rise of the antivaxxer / public health skeptic and the dramatic politicization of public health measures.

    But, not everything disappointing has stemmed from that. Our lack of cheap rapid tests for diseases like Flu and RSV is a sad reminder of our regulatory system failing to learn from the COVID crisis of the value of cheap, rapid in-home testing or adopting to the new reality that many Americans now know how to do such testing.


  • Huggingface: security vulnerability?

    Anyone who’s done any AI work is familiar with Huggingface. They are a repository of trained AI models and maintainer of AI libraries and services that have helped push forward AI research. It is now considered standard practice for research teams with something to boast to publish their models to Huggingface for all to embrace. This culture of open sharing has helped the field make its impressive strides in recent years and helped make Huggingface a “center” in that community.

    However, this ease of use and availability of almost every publicly accessible model under the sun comes with a price. Because many AI models require additional assets as well as the execution of code to properly initialize, Huggingface’s own tooling could become a vulnerability. Aware of this, Huggingface has instituted their own security scanning procedures on models they host.

    But security researchers at JFrog have found that even with such measures, have identified a number of models that exploit gaps in Huggingface’s scanning which allow for remote code execution. One example model they identified baked into a Pytorch model a “phone home” functionality which would initiate a secure connection between the server running the AI model and another (potentially malicious) computer (seemingly based in Korea).

    The JFrog researchers were also able to demonstrate that they could upload models which would allow them to execute other arbitrary Python code which would not be flagged by Huggingface’s security scans.

    While I think it’s a long way from suggesting that Huggingface is some kind of security cesspool, the research reminds us that so long as a connected system is both popular and versatile, there will always be the chance for security risk, and it’s important to keep that in mind.


  • Nope, the Dunning-Kruger Effect is just bad statistics

    The Dunning-Kruger effect encapsulates something many of us feel familiar with: that the least intelligent oftentimes assume they know more than they actually do. Wrap that sentiment in an academic paper written by two professors at an Ivy League institution and throw in some charts and statistics and you’ve got a easily citable piece of trivia to make yourself feel smarter than the person who you just caught commenting on something they know nothing about.

    Well, according to this fascinating blog post (HT: Eric), we have it all wrong. The way that Dunning-Kruger constructed their statistical test was designed to always construct a positive relationship between skill and perceived ability.

    The whole thing is worth a read, but they showed that using completely randomly generated numbers (where there is no relationship between perceived ability and skill), you will always find a relationship between the “skill gap” (perceived ability – skill) and skill, or to put it more plainly,

     (y-x) \sim x

    With y being perceived ability and x being actual measured ability.

    What you should be looking for is a relationship between perceived ability and measured ability (or directly between y and x) and when you do this with data, you find that the evidence for such a claim generally isn’t there!

    In other words:


    The Dunning-Kruger Effect is Autocorrelation
    Blair Fix | Economics from the Top Down

  • Cat Bond Fortunes

    Until recently, I only knew of the existence of cat(astrophe) bonds — financial instruments used to raise money for insurance against catastrophic events where investors profit when no disaster happens.

    I had no idea, until reading this Bloomberg article about the success of Fermat Capital Management, how large the space had gotten ($45 billion!!) or how it was one of the most profitable hedge fund strategies of 2023!

    This is becoming an increasingly important intersection between climate change and finance as insurance companies and property owners struggle with the rising risk of rising damage from extreme climate events. Given how young much of the science of evaluating these types of risks is, it’s no surprise that quantitative minds and modelers are able to profit here.

    The entire piece reminded me of Richard Zeckhauser’s famous 2006 article Investing in the Unknown and Unknowable which covers how massive investment returns can be realized by tackling problems that seem too difficult for other investors to understand.


  • Shein and Temu now drive global cargo

    Maybe you have shopped on Shein or Temu. Maybe you only know someone (younger?) who has. Maybe you only know Temu because of their repeat Superbowl ads.

    But these Chinese eCommerce companies are now the main driver behind air and ship cargo rates with Temu and Shein combined accounting for 9,000 tons per day of shipments!

    This is scale.


    Rise of fast-fashion Shein, Temu roils global air cargo industry
    Arriana McLymore, Casey Hall, and Lisa Barrington | Reuters