Two possible futures for AI | The Vergecast

Intro (0s)

• The Vergecast is described as a podcast focused on decreasing the cost of AI infrastructure. (0s)
• Six months ago, reviews were conducted on the Humane AI Pin and the Rabbit R1, both of which were found to be unsatisfactory and subsequently set aside. (10s)
• Despite initial negative reviews, both companies, Humane and Rabbit, continue to operate and develop their products. Humane has reduced its prices, indicating low sales, while Rabbit has made numerous changes and is nearing the release of its large action model. (23s)
• There is an intention to revisit these AI gadgets in a future episode to assess any progress made in the development of AI technology. (45s)
• The current episode will focus on two main topics, which are not specified in the provided text. (1m7s)

Dueling AI Blog Posts: Anthropic vs. OpenAI (1m10s)

• Dario Amodei from Anthropic and Sam Altman from OpenAI have written blog posts discussing the future of AI, which are significant due to their positions in the AI industry. (1m10s)
• Sam Altman's blog post, titled "The Intelligence Age," outlines his optimistic vision for AI's future, emphasizing its potential impact and excitement. (2m13s)
• Dario Amodei's blog post, "Machines of Loving Grace," is a detailed 10,000-word piece that explores specific ways AI could improve the world, sharing a similar optimistic outlook but with a different structural approach. (2m51s)
• These blog posts are noteworthy as they come from two of the most influential figures in AI, with Demis Hassabis from DeepMind at Google being another key figure, although he has not published a similar blog post. (3m32s)
• The discussion aims to analyze these blog posts to understand the similarities, differences, and potential conflicts between the visions of OpenAI and Anthropic, as well as the broader vision of Silicon Valley for AI. (3m52s)
• The episode also covers a project on the West Coast focused on improving earthquake detection and alert systems, highlighting the complexity and challenges involved. (1m32s)
• Two blog posts by corporate CEOs were discussed, highlighting their significance due to the timing and the individuals involved, which made them stand out from typical CEO writings about the future. (5m10s)
• Both CEOs agree that the world will become a beautiful place with the advent of Artificial General Intelligence (AGI). (5m53s)
• A key difference between the two is their outlook on the inevitability of AGI. Sam Altman is portrayed as believing that the development of AGI is inevitable and that deep learning will naturally lead to solutions, suggesting a sense of completion and certainty. (6m22s)
• In contrast, Dario Amodei is more cautious, emphasizing that the future with AGI is not guaranteed and requires deliberate effort and many things to go right, indicating a more measured and uncertain perspective. (6m40s)

Upcoming AI Developments and Industry Insights (7m26s)

• There is a notable difference in outlook between two CEOs, with one being more optimistic about the future of AI and the other more cautious, focusing on potential negative outcomes. (7m40s)
• Both CEOs agree that AI will significantly transform labor and the economy by replacing many routine and monotonous jobs, and they acknowledge the need for plans to address these changes. (9m15s)
• Sam Altman has explored the concept of Universal Basic Income (UBI) by conducting experiments to understand its impact on people's lives, while Anthropic has not conducted similar research. (9m33s)
• There is a consensus that AI will inevitably disrupt daily life and the economy, but there is a lack of concrete answers or solutions from the executives on how to manage these changes. (10m6s)
• The discussion highlights the uncertainty and complexity surrounding the future impact of AI, with an acknowledgment that many questions remain unanswered. (10m25s)
• Both parties agree on the importance of aligning AI systems with human values to ensure safety, although there is a difference in their commitment to safety measures, with one having reduced its focus on safety. (11m9s)
• There is a proposal for an international collaboration, akin to a "United Nations of AI," to manage AI development in a way that supports liberal democracy globally, though its feasibility is uncertain. (12m17s)
• Safety researchers at major AI companies require significant computational resources to conduct their work, and there have been tensions at OpenAI due to unmet promises regarding resource allocation for safety research. (13m8s)
• Dario is more cautious about the potential hype surrounding AI compared to Altman, who is more optimistic about rapid advancements in AI within a short timeframe. (13m56s)
• Sam Altman wrote a blog post around the time he was closing a $6.6 billion funding round for OpenAI, the largest for a private company in history. This timing suggests a connection between the blog post and the funding efforts. (14m40s)
• Anthropic, another AI company, is also attempting to raise a significant amount of money, although the exact amount is unknown. This indicates a trend of AI companies seeking large investments amid investor caution due to the high costs and uncertain outcomes of AI model development. (14m55s)
• The narrative of promising a transformative future to secure funding is likened to the classic Silicon Valley approach, where companies often exaggerate their potential to meet venture capital expectations. This involves making grand claims about the future impact of their technology. (15m56s)
• The discussion draws parallels to past tech industry claims, such as Google's health division's promise to "cure death," highlighting a pattern of ambitious proclamations in the tech world. (16m47s)
• Despite skepticism, there is an acknowledgment that AI has the potential to be useful, though currently, its applications are limited. The grandiose statements about AI's future are seen as part of a longstanding tradition in the tech industry. (17m22s)
• The concept of "stakes creep" is discussed, where the expectations for AI advancements have become increasingly exaggerated, similar to how movie franchises escalate their plots to maintain interest. (17m55s)
• There is a disagreement on the timeline for achieving powerful AI between Sam Altman and Dario. Dario believes that significant changes due to AI could happen within 5 to 10 years, while Sam Altman suggests a more conservative timeline of a few decades. (19m4s)
• Sam Altman uses the term "a few thousand days" to describe the timeline for achieving AGI, which is seen as vague and open to interpretation. Some believe it could happen sooner, while others are more skeptical. (20m0s)
• There is a general sentiment among some people in San Francisco that AI advancements are inevitable and will happen soon, but there is uncertainty about who to believe due to differing opinions among experts. (20m35s)
• There is a tendency to overestimate what AI can achieve in one year and underestimate its potential over a decade. This perspective is likely to be relevant in the context of AI development. (21m4s)
• Dario Amodei left OpenAI to create Anthropic, a company focused on safety and a more calculated approach to AI, contrasting with Sam Altman's faster-paced and riskier approach at OpenAI. Both aim to build AGI but have different methodologies and philosophies. (21m21s)
• The difference in thoughtfulness between Dario's and Sam's blog posts is notable, with Dario's being more detailed and academic, while Sam's appears more casual and less developed. (22m12s)
• Dario's blog post is seen as an effort to genuinely inform the public, having been developed over several months, whereas Sam's post seems hastily written. (23m30s)
• Anthropic is currently perceived to be building better consumer AI products than OpenAI, possibly due to Dario's extensive ideas on AI applications, compared to Sam's more limited vision of AI as a personal assistant. (24m0s)

Earthquake Detection (30m0s)

• A discussion is held about the sound of a house being seismically retrofitted, which involves earthquake-proofing by drilling bolts into the concrete foundation. This process causes significant noise and vibration throughout the house. (30m0s)
• The retrofitting is done because the house is located in Seattle, an area prone to earthquakes, and the owner has a fear of earthquakes from previous experiences in San Francisco. (31m11s)
• The fear of earthquakes was heightened by living in the Bay Area, where there is constant talk of preparing for a major earthquake similar to the 1906 San Francisco quake. (32m30s)
• A New Yorker article about the potential devastation of a major earthquake in the Pacific Northwest, including a tsunami visible from the International Space Station, is mentioned as particularly frightening. (33m20s)
• In the I5 Corridor, it is estimated that after an earthquake, it will take between 1 to 3 months to restore electricity, a month to a year to restore drinking water and sewer services, 6 months to a year to restore major highways, and 18 months to restore healthcare facilities. These recovery times increase for coastal areas. (34m9s)
• A significant earthquake, known as a Mega thrust quake, is expected to occur in the Pacific Northwest, potentially within the next 500 years or even tomorrow. This earthquake will be felt from Northern California to British Columbia and is associated with the Cascadia subduction zone. It could reach a category nine on the Richter scale, causing extensive damage and a devastating tsunami. (34m51s)
• An earthquake early warning system called ShakeAlert has been implemented along the West Coast. While it cannot predict earthquakes, it can detect them as they begin, providing alerts faster than the shaking reaches people, offering a few seconds to a couple of minutes of warning. (35m53s)
• Developing a reliable and fast earthquake detection and warning system is challenging. Detecting the earthquake is relatively straightforward, but notifying a large number of people quickly is complex. (36m35s)
• To understand the earthquake early warning system, the process of how an earthquake is detected, how an alert is generated, and how it reaches individuals' phones was explored, starting with a small shed near a wastewater treatment plant in Seattle. (37m7s)
• Doug Gibbons, a field technician at the University of Washington, demonstrates earthquake surveillance hardware, specifically a Titan strong motion accelerometer, which is bolted to a building's foundation. This device captures a three-dimensional snapshot of ground movement using three seismometers oriented in different directions. (37m53s)
• The accelerometer is highly sensitive, capable of detecting even minor movements such as a person bending their knees. It constantly takes samples many times per second and feeds this data to relevant places. (39m10s)
• Over the past decade, approximately 1,500 sensors like this have been installed along the West Coast as part of an effort led by the US Geological Survey to monitor seismic activity. These sensors power ShakeAlert, an early warning system for earthquakes. (39m34s)
• When an earthquake occurs, two types of waves, primary (P waves) and secondary (S waves), emanate from the epicenter. P waves move faster but are too weak to cause damage, while S waves move slower and can cause harm. The early warning system takes advantage of the time difference between these waves. (40m10s)
• In a hypothetical scenario of a 9.0 earthquake starting west of Seattle, the P waves would reach coastal sensors 40 miles offshore before the actual shaking, providing a critical window for early warning. (41m0s)
• The earthquake is approximately 10 seconds away from the shore, which translates to being 40 or 50 seconds away from downtown Seattle. This time frame, roughly 45 seconds, is the critical window for the ShakeAlert system to warn the city about the impending natural disaster. (41m37s)
• When the initial P waves of the earthquake hit the sensors, the seismometer graphs spike, and this data is immediately forwarded to the necessary authorities via radio, hardwire internet, or satellite, depending on the available infrastructure at the sensor locations. (42m4s)
• The transmission of this data happens extremely quickly, within a fraction of a second, ensuring that the warning system can operate efficiently and provide timely alerts. (42m19s)

Data Transmission to UW (42m27s)

• Data from ground shaking is recorded by sensors, which create packets and send them to a communications device. This device transmits the data either through radio waves or the internet back to the University of Washington (UW). (42m27s)
• The data arrives at a server located on-site at the university, specifically in a room filled with computer racks and computers with green blinking lights. (42m41s)
• All incoming data from the field is continuously processed in this facility, which is situated in the basement of the Earth and Space Sciences Building at UW in Seattle. (43m2s)
• Rata Harthog, who contributed to developing ShakeAlert's algorithms, is involved in the operations at this facility. (43m8s)

ShakeAlert Algorithms (43m18s)

• California employs two algorithms to analyze incoming earthquake signals. The first algorithm quickly assesses the epicenter and magnitude based on initial signal amplitudes, providing a rapid estimate. (43m18s)
• The second algorithm is slower but models the entire earthquake beyond the initial point, accounting for the growth of the quake as more of the fault line slips, which can result in longer shaking and affect more people. (43m47s)
• For large earthquakes, such as those on the Cascadia subduction zone, the fault plane can take minutes to rupture fully, affecting a wide geographic area. (44m12s)
• ShakeAlert estimates the areas that will be most affected and sends alerts to phones quickly, with the first alert leaving the server within a second or two of processing. Additional alerts can be sent as the quake grows. (44m30s)
• The entire process, from detecting the initial P waves to sending alerts, takes about two seconds, providing a warning time of approximately 40 to 45 seconds for locations like Seattle. (45m35s)

Challenges with Wireless Emergency Alerts (46m7s)

• The wireless emergency alert system (WEA) is used to send alerts to cell phones, such as Amber Alerts and tornado warnings, using cell broadcast technology that acts like a digital bullhorn through cell towers. (46m27s)
• WEA was developed by the federal government in 2012 and is effective for earthquake alerts because it does not require users to opt in or download anything, as long as their phone and carrier support it. (47m4s)
• WEA messages are limited to 90 characters to ensure compatibility with most phones, including non-smartphones, allowing widespread access to alerts like earthquake early warnings. (47m42s)
• Bob de Groot, who leads the ShakeAlert operations team, highlights that WEA is the best channel for reaching people on the west coast but notes its limitation in speed, as it was not designed for the rapid response times required by ShakeAlert. (48m12s)
• Tests conducted in California showed that the fastest WEA alert delivery was around 4 seconds, with a median range of 6 to 12 seconds, and some phones never received the alert, indicating variability in the system's performance. (49m44s)
• The complexity of implementing the Wireless Emergency Alerts (WEA) system is due to the involvement of multiple entities, including the federal government, telecom companies, and handset manufacturers, each of which may prioritize and implement the system differently. This results in variations in delivery and latency across different networks and devices. (50m7s)
• There is no built-in method to measure end-to-end latency in the WEA system, and it was not originally designed for earthquake early warnings, which means there is no mandate to minimize delays. This can be critical in scenarios like a Cascadia earthquake, where every second counts for timely alerts. (51m31s)
• The 2001 Nisqually earthquake in Seattle, which lasted about 45 seconds and caused significant damage, highlights the importance of timely alerts. The fault responsible for this earthquake is active, having produced similar quakes in the past, and poses a recurring threat to the region. (52m37s)
• Gabriel Lotto, who works with the ShakeAlert team at the University of Washington, explains that the fault responsible for the Nisqually earthquake is located deep below Seattle. A similar future earthquake would provide limited warning time, as the seismic waves would hit the region simultaneously, offering only 5 to 15 seconds of warning depending on the location. (53m24s)
• The effectiveness of the Wireless Emergency Alerts (WEA) system in providing timely earthquake warnings is questioned, as delays in notifications can lead to significant consequences. (54m15s)
• The ShakeAlert system is being developed to improve earthquake warnings, with efforts extending beyond WEA to include various platforms for delivering alerts. (54m58s)
• UC Berkeley has developed a smartphone app called MyShake, which sends push notifications to users within the alert radius of an earthquake, offering faster alerts than traditional methods. (55m6s)
• The MyShake app is opt-in, which limits its reach, as only a small portion of the population has registered, despite having 3.5 million users on the West Coast. (56m8s)
• Google has integrated ShakeAlert into Android at the operating system level, allowing for different types of alerts, including fullscreen takeovers for larger earthquakes, which are on by default. (56m50s)
• Apple's iOS does not have a similar built-in system, although there is a way to speed up WEA alerts through a setting in the system, but it is not as prominent or default as Google's implementation. (57m40s)
• The adoption of earthquake alert systems faces challenges, as people are often reluctant to download apps unless incentivized, such as with free offers. (58m9s)
• In California, earthquake alert systems have been tested during actual quakes, such as a recent one in Malibu, where some users received a 20-second warning, allowing them to take protective actions. (58m41s)
• The rollout of these systems has not been flawless; in 2021, a quake in Southern California was misinterpreted as multiple smaller quakes, resulting in some people not receiving alerts. (59m6s)
• Experts are particularly concerned about false positives, as they can lead to a loss of credibility and public trust, similar to the "boy who cried wolf" scenario. This occurred in California in 2020, when a false alert was issued and later canceled. (59m21s)
• There is a risk that people will disable notifications if alerts become too frequent or are perceived as annoying, which could undermine the system's effectiveness. (1h0m20s)
• Determining the threshold for issuing alerts is a complex issue, involving both technical and social considerations. Public feedback has influenced adjustments to these thresholds, as seen in the response to earthquakes in Los Angeles in 2019. (1h0m50s)
• Bob deGroot from USGS manages the ShakeAlert social media account and frequently addresses public inquiries about the alert system's performance, highlighting ongoing challenges in meeting public expectations. (1h1m44s)
• The discussion highlights the complexity of coordinating various parties, such as carriers, operating system companies, and the government, to effectively implement earthquake alert systems. This involves overcoming bureaucratic and collective action challenges. (1h3m23s)
• Significant effort has been invested in using cell phones as a primary delivery mechanism for earthquake alerts, but this process is time-consuming and complex, similar to the challenges faced during the development of pandemic notification systems by Apple and Google. (1h3m43s)
• The urgency and attention given to earthquake preparedness often increase only after a major earthquake occurs, making it difficult to maintain focus and prioritize the development of alert systems in the absence of recent disasters. (1h4m42s)

Future of Earthquake Alerts (1h5m9s)

• The discussion highlights the potential for various devices, such as smart speakers, home assistants, and smart TVs, to serve as delivery mechanisms for earthquake early warnings. This includes public channels like highway signs and PA systems in schools and hospitals. (1h5m23s)
• Automated systems could use earthquake information to perform actions like opening fire station doors, shutting off water utilities, slowing down trains, or recalling elevators. The integration of this information into different systems is still in its early stages. (1h6m1s)
• ShakeAlert, an earthquake early warning system, has been fully operational for just over three years, primarily on the West Coast of the United States. It is recognized by state and local governments and major technology providers, but its full potential is still being developed. (1h6m36s)
• The MyShake app is mentioned as a tool for receiving earthquake notifications. Users in Washington express a desire for a minor earthquake to remind people of the risks without causing significant damage. (1h7m30s)
• Conversations with scientists have increased awareness of various faults beyond the well-known Cascadia Subduction Zone, including lesser-known faults that pose additional risks. (1h8m33s)
• A diverse group of organizations, including government agencies, universities, telecom companies, and cell phone manufacturers, are collaborating to develop early warning systems, which are expected to be beneficial in the future. (1h8m59s)
• There is a recommendation for people on the West Coast to enable a feature called local awareness on iPhones and to download the MyShake app on both Android and iOS for earthquake alerts. Additionally, ready.gov is suggested for general earthquake preparedness advice. (1h10m26s)
• The Vergecast hotline receives numerous questions, some of which lead to in-depth explorations of various topics. The team enjoys engaging with these questions, although they request that people remain polite when reaching out. (1h11m39s)
• A listener named Peter from Brooklyn posed a question about the new features of the iPhone 16, specifically regarding its camera capabilities and whether it can take pictures of moving subjects, like children, faster than previous models. (1h12m53s)
• The discussion involved Allison Johnson, who reviewed the iPhone 16 and has experience taking photos of children, to explore whether the new camera controls and features significantly improve the user experience. (1h13m33s)
• The iPhone 16 update is described as "camera forward," with new control options and filters, but it is noted that the improvements are not drastic changes to the camera pipeline. (1h14m7s)
• The conversation explored the routines and muscle memory involved in quickly accessing the camera on smartphones, comparing the ease of use between Android and iPhone devices. (1h15m2s)
• On Android phones, the camera can be accessed quickly by double-clicking the power button, while on iPhones, users typically swipe from the home screen or use the camera icon to open the app. (1h15m50s)
• There is a discussion about the friction experienced when trying to quickly access the camera on a phone, particularly when the screen does not wake up or register the camera icon tap, leading to missed photo opportunities. (1h16m38s)
• The iPhone 15's action button is mentioned as a potential solution for quick camera access, but its placement on the phone is considered awkward by some users, requiring a maneuver to reach it. (1h17m41s)
• The new literal camera button on phones is described as having mixed reviews; while it allows for camera control and exposure changes, it requires the screen to be awake before launching the camera, and its stiffness can cause the phone to shake when pressed, potentially affecting photo quality. (1h18m34s)
• The process of using the camera involves clicking a button to turn it on, accessing the camera, and then using a software button on the screen, which can be somewhat fiddly and not necessarily faster. (1h20m36s)
• A useful feature is the ability to change exposure compensation with a half-press or light press on the button, which is more convenient than adjusting brightness on an iPhone by tapping and dragging a sun icon. (1h21m14s)
• The button is designed to allow users to perform multiple actions intuitively, but it can be challenging to use as intended, requiring users to slide their finger back and forth, which can feel awkward. (1h22m30s)
• Exiting the control mode is not straightforward, as tapping to exit does not work as expected, leading to confusion about how to properly exit the settings. (1h23m30s)
• The overall experience with the button does not necessarily make the process faster or more intuitive, and there is a sense of dissatisfaction with its functionality. (1h24m12s)