Dear Friend,
Here is something to ponder: Our friend Azeem Azhar recently sent an email to his team, making them aware of a new study which looks at “the growing disparity between the exponential expansion of AI context windows and the contraction of human sustained-attention spans.” The paper argues that while AI context windows have grown by a factor of 1000 in the last six years, human context span has shrunk significantly. In simple terms: As we offload more and more tasks (and thinking) to AI, we lose our ability to think for ourselves. Sobering stuff.
And now, this…
Headlines from the Future
AI Model Fatigue Is Real: Enterprise buyers are drowning in the rapid release cycles of AI models. Great for innovation, and a nightmare for anyone building a scalable product or service on top of them – you end up ripping out the guts of your product every few months. Add the cost of continuously evaluating models and the cybersecurity risks of swapping them. Which leaves you with either shouldering the cost and effort of staying on top of the latest models, or pinning a version and risking being left behind. Neither a great option.
Self-Driving Cars Save Lives: The case for self-driving (autonomous) cars saving lives is easy – they don’t drink and drive, text while driving, or get distracted by heated arguments. Here’s the data to back up your intuition – the technology could prevent a staggering 580,000 deaths every year. So yes, go ahead and hail that Waymo robotaxi.
Smartglasses Get the Boot: Wear Google Glass and you were branded as a “Glasshole.” With Meta’s Ray-Ban smartglasses, it is much less obvious if a person is wearing them – hence the need for new detection methods. A Polish developer, Paweł Szydłowski, developed a Meta smartglasses detection app called Zuckoff – the app scans your surroundings for Bluetooth devices and alerts you if someone with Meta glasses is nearby. The people fight back.
The Unintended Consequences of Remote Work: Turns out, remote work leads to increased sleep and reduced physical activity. As we are slowly gaining a better understanding of the manifold effects of remote work, a new study by the University of Turku in Finland found that the average remote worker sleeps 15 min longer and is 45 min less active per day in comparison to their in-office counterparts. The reason behind this is not the working-from-home part, but the commute – walking and cycling on office days got replaced by sitting, lying down, and sleeping. As someone who regularly works from home, I can definitely attest to the second point (less active); not so sure about the first one (more sleep) personally.
DNA and Compute – A Match Made in Heaven: How is that for a good use of AI? DeepMind, Google’s AI research lab, has mapped all 9 billion possible single-letter changes to a reference human genome and made the results publicly available. The work could help researchers filter through and prioritize experiments. Let’s do more of this and a lot less of the all-too-common AI hype and slop.
AI Risks, SaaS, and Offline Software: Here is a provocative thought: if AI-powered cyberattacks become more and more common, and as they become more sophisticated with each new frontier model release, will the logical next step be to go back to offline (air-gapped) software? It certainly isn’t a crazy idea – and you might very well prefer to have your local water or energy company run on systems which are safe from cyberattacks. But then – how realistic is this in a world which is increasingly connected?
AI Is Retraining Workers, Not Replacing Them: Despite the dominant media narrative, AI is not replacing workers (at least not yet), but instead transforming work (and the workforce). A new study by New York’s Fed found that 34% of service firms and 22% of manufacturers using AI are retraining staff, while only 5% and 2% report layoffs.
The/Some Managers Are Back: In an unexpected about-face, Meta is asking some of its AI staff to go back to managing people. After Zuck’s 2023 “year of efficiency” dismissal of managers and flattening of the org chart, the company is asking some of its employees to become managers again. Only four months ago, Zuck told investors that one talented person could do the work of a team; now the org built on that premise needs a management layer to function.
Surveillance Pricing Is Coming for Your Salary: First we had predictive surveillance – the practice in which a merchant sets a price for a product based on what they think you are willing to pay (based on data collected from a variety of sources). Delta Airlines got into some hot water for this. Now employers are using the same approach to set salaries – based on what the algorithm thinks you are willing to accept as the lowest possible salary – with input signals ranging from payday loans, credit-card balances, to your social media profiles. Expect backlash and lawsuits – Delta Airlines is currently feuding with the FTC over this.
China’s Sentiment on the Existential Threat of AI: As Jeffrey laid out in this week’s Tuesday Briefing, the US is (and continues to be) in a deep and loud discussion about the existential threat to humanity posed by AI. Meanwhile, in China, the sentiment is a lot more muted – AI showed up as a cheap, useful product before it showed up as a debate, and 80% of Chinese respondents now use it weekly against 54% in the US. China’s fear is about the loss of operational control, malicious use, and social disruption – a threat to the party, rather than humanity.
The Case Against AI-Driven Double-Digit Growth: Ben Moll and Alex Imas have taken on the entire genre of AI-insider growth forecasts – Dario Amodei’s 10-15% a year, Leopold Aschenbrenner’s 30%, an Anthropic researcher advising people to price annual economic doublings into their models. The pair argues that economic frictions, such as slow institutional adoption and physical bottlenecks, will prevent an immediate explosion in productivity – they put the realistic number at 4-5% a year over the next decade, which they say would already be massive, and Moll is publicly betting that US per-capita growth stays under 15% in every year through 2033. At the center of their analysis is Baumol’s cost disease, which suggests that even as digital tasks become cheaper, the economy will remain constrained by non-automatable sectors like healthcare and construction.
What We Are Reading
Math Can’t Go On Like This AI just solved one of the hardest math problems in history, and mathematicians are not sure whether to celebrate or panic. @Jane
The Most Powerful Man in AI Isn’t Worried at All While Anthropic’s and OpenAI’s CEOs warn of AI going rogue, Nvidia’s Jensen Huang keeps dismissing existential AI risk as overblown while pushing to accelerate chip sales globally. @Mafe
The End Of Upward Mobility This argument gets a little academic, but it’s definitely not too soon to think through the collapse of meritocracy (to the extent that it exists) as a potential implication of the AI-driven transformation of the work/economy. @Jeffrey
Responsible AI Means Knowing the Limits of Agent Autonomy Now that MIT is winding down its publication, I’m trying to absorb as much of its content as possible. This piece offers an important reminder: calling AI agents “autonomous” cannot become a way for humans and institutions to evade accountability... the more authority we give agents, the clearer the line of responsibility must become. @Kacee
AI and Democracy: The Right to Resist Optimization Taiwan’s cyber-ambassador Audrey Tang always thought deeply about the wide-ranging implications of technology – it comes as no surprise that her thoughts on AI are a couple of cuts deeper than what Silicon Valley has to offer. @Pascal
📱 iPhones (and smartphones in general) are getting bigger and bigger. Here is how much (and boy, do I miss my iPhone 12 mini).
💪 The wild and odd things your body can do – 155 of them.
🤖 Anthropic’s Opus 5 model is notorious for being a bit too eager to please – often resulting in it overshooting the mark. Here’s a fun little game – see if you can make one button on a website blue.
🍿 Actually not fun – if you sense that your streaming services are getting a little bit too much like cable TV, you are right. At the very least, they are getting more expensive by the year: The same nine streaming subscriptions cost $702 more a year than in 2021.
What We Do When We’re Not Writing
Hi! I’m Pascal from radical. When we’re not writing this newsletter, we help organizations turn volatility into advantage – without the “innovation theater.” We build your team’s capacity to handle disruption, so you stop reacting and start shaping. If that sounds useful, let’s talk.
↗ Dive into the deep end: Access our complete collection of 3,100+ radical links.