Quotes
"The only function of economic forecasting is to make astrology look respectable." - John Kenneth Galbraith
"Markets can remain irrational longer than you can remain solvent." - Keynes
"A designer knows his design is perfect not when there is nothing left to add, but when there is nothing left to take away." -Antoine de Saint Exupéry
Articles/Essays
Inference and LLMs - by Erik J Larson - Colligo (substack.com)
Erik J Larson's article "Inference and LLMs" explores the limitations of large language models (LLMs) in the context of artificial general intelligence (AGI). Larson emphasizes that while LLMs, powered by advanced machine learning techniques such as transformer architecture, have shown impressive performance in natural language processing tasks, they fundamentally rely on inductive reasoning. This reliance on induction, rather than more sophisticated forms of inference like abduction, limits their capacity to achieve true AGI.
Key Points
Types of Inference
Deduction:
Definition: Deductive reasoning involves deriving specific conclusions from general principles or premises. Given the premises, the conclusions are logically certain.
Example: "All men are mortal. Socrates is a man. Therefore, Socrates is mortal."
Characteristics: Deduction is reliable and certain but not typically used for generating new knowledge about the world.
Induction:
Definition: Inductive reasoning involves making generalizations based on specific examples or observations. It is probabilistic rather than certain.
Example: "All observed swans are white. Therefore, all swans are white."
Characteristics: While induction helps formulate hypotheses and theories, it is inherently uncertain because a single counterexample can disprove the generalization.
Abduction:
Definition: Abductive reasoning, or “inference to the best explanation,” involves forming plausible explanations for observations, often based on incomplete information.
Example: "The kitchen contains a half-empty can of Coke. The workmen were here earlier, and they often drink Coke. Therefore, the Coke likely belongs to the workmen."
Characteristics: Abduction is less about certainty and more about plausibility, making it a crucial mechanism for scientific discovery and everyday reasoning.
Limitations of Inductive Reasoning in LLMs
Large Language Models (LLMs) like GPT-4 rely heavily on inductive reasoning. They are trained on vast datasets to identify patterns and make predictions. However, this inductive approach has inherent limitations:
Dependence on Data: LLMs require enormous amounts of data to function effectively. They learn patterns from this data but do not understand underlying concepts.
Lack of Flexibility: Inductive reasoning does not allow LLMs to adapt to new, unseen situations that fall outside their training data.
Error Susceptibility: LLMs can make errors that reveal their lack of deeper understanding. These errors, often called "hallucinations," highlight the absence of a more flexible reasoning process.
Importance of Abductive Inference
Abduction is critical in human cognition because it allows us to make educated guesses and form hypotheses based on incomplete information. This type of reasoning is essential for:
Scientific Discovery: Scientists often generate abduction hypotheses, tested through experiments and observations.
Daily Decision-Making: Humans constantly use abductive reasoning to navigate everyday life, making decisions based on the best possible explanations given the available information.
LLMs and World Models
A world model is a comprehensive understanding of how things work in the real world, encompassing physical laws, social norms, and logical constraints. For LLMs to achieve true AGI, they would need to incorporate such models to perform abductive reasoning effectively.
Current State: Present-day LLMs do not possess a world model. They operate based on patterns in the data they have been trained on without understanding the underlying principles of the world.
Future Directions: Integrating world models into AI systems could enable them to perform abductive reasoning, making them more adaptable and capable of true understanding.
LLM Errors and Hallucinations
LLM errors often manifest as "hallucinations," where the model generates plausible-sounding but incorrect or nonsensical outputs. These errors occur because:
Probabilistic Nature: LLMs generate text based on the probability of sequences of tokens, sometimes leading to logically inconsistent or factually incorrect statements.
Lack of Contextual Understanding: Without a deep understanding of the context or underlying facts, LLMs can produce outputs that seem coherent but are fundamentally flawed.
Implications for AGI Development
Understanding the limitations of LLMs and the importance of different types of inference has significant implications for the development of AGI:
Research Focus: Future research should focus on integrating abductive reasoning and world models into AI systems to move beyond the limitations of inductive reasoning.
Ethical Considerations: Recognizing LLMs' current limitations can inform ethical guidelines for their deployment, ensuring they are used appropriately and do not mislead users.
Realistic Expectations: Setting realistic expectations about LLMs' capabilities can help prevent overreliance on them and promote the development of more robust AI technologies.
Conclusion
Erik J Larson's exploration of inference in the context of LLMs highlights the fundamental limitations of current AI technologies. It points to the need for more advanced forms of reasoning to achieve true AGI. By understanding and addressing these limitations, researchers and developers can work towards creating AI systems that are not only powerful but also genuinely intelligent.
Key Quotes
On Inductive Reasoning:
“Induction is what’s reasonable to conclude given what I see in front of me and what I already believe.”
On Abductive Inference:
“Abduction presupposes a world model, a set of background facts and rules that can be applied to observations.”
On LLM Limitations:
“There’s no underlying inference that props up induction, it’s just big, big data.”
On the Future of AGI:
“We’re still stuck in fantastically complex, computationally expensive induction machines that can only infer based on the data provided to them. That’s not intelligence, even as it sometimes seems like it is.”
Why It Matters
The article underscores critical distinctions between different types of inference and highlights why LLMs, despite their advanced capabilities, fall short of true intelligence. Understanding these limitations is crucial for several reasons:
Guiding AI Development: Recognizing the boundaries of current AI technologies can steer research towards overcoming these limitations and developing more advanced forms of machine intelligence.
Setting Realistic Expectations: It helps manage expectations about what LLMs can and cannot do, which is important for both developers and users of AI technologies.
Ethical and Practical Implications: Acknowledging the gaps in LLMs' reasoning abilities can inform ethical considerations and practical applications, ensuring that AI is used appropriately and effectively.
Future Research Directions: The discussion on abduction and world models points to potential areas of innovation necessary for achieving AGI, influencing future research priorities.
In summary, Erik J Larson's article provides a nuanced understanding of LLMs' limitations and the types of reasoning required for AGI, emphasizing the need for advancements beyond inductive learning to achieve true machine intelligence.
Some things I believe (amirbolous.com)
Be patient with outcomes. Good things take time, and you need to earn the right to win (h/t @varunshenoy_). If you want to win, you need an edge. Never expect or assume you will win. The opportunity to build something people care about and use is a privilege; treat it with the respect it deserves.
We treat our parents significantly worse than they deserve to be treated. We take them for granted way more than we should.
Nutrition and physical fitness are very important. People spend more time thinking about external inputs than what they put in and how they treat their bodies.
It is very hard to accomplish anything truly great if you’re high ego. You will be wrong often and that means you need to be open and humble enough to accept that.
If you want to make progress quickly, figure out how to create as tight of a feedback loop as possible.
There are many good qualities you can have, but earnestness, resourcefulness, agency, and curiosity are rare nowadays.
Steal like an artist, but always give credit where credit is due
It’s better to actually try something out than spend a lot of time thinking about it.
Details are revealed when you start doing something, not before
Loneliness is probably the worst thing in the world.
Courage is more scarce and important than intelligence.
You are allowed to just do things.
Chasing status is almost always a bad idea.
Confidence and arrogance are fine lines, and it is better to err on the side of humility than arrogance. You may initially get less recognition, but you will attract the best type of people, and this is what matters.
Having fun is a great way to make you like working hard.
Most advice is pretty bad and almost always self-directed.
If untested, assume it’s broken.
Respect your heroes, but don’t put anyone on a pedestal; everyone is human, and they will disappoint you.
Whether you think you can or can’t, you’re probably right.
Reading for Speed, Backwards (substack.com)
In "Reading for Speed, Backwards," Admired Leadership's article explores a counterintuitive strategy for reading nonfiction materials more efficiently. Instead of traditional speed reading techniques, this method involves starting with the conclusion or the last section of a text to understand the main argument first. By knowing the endpoint, readers can return to the beginning and focus on the most critical parts of the content, enhancing speed and comprehension. The article suggests that this approach can significantly reduce reading time while improving retention of key information.
Key Points
The Challenge of Information Overload:
Active learners and leaders often need to consume large amounts of information quickly.
Traditional speed reading techniques involve focusing on keywords, using a pointer, and minimizing subvocalization.
Introduction of a New Technique:
The article proposes a non-traditional method: reading backward.
This involves reading the last chapter or final paragraphs of a book or article to grasp the central argument or conclusion.
Benefits of Reading Backwards:
Improved Comprehension: Understanding the conclusion first helps readers identify and focus on important arguments and examples throughout the text.
Increased Efficiency: Knowing the endpoint allows readers to skim less relevant sections, thus reading faster without losing crucial information.
Enhanced Retention: This method can lead to better retention of critical content, as readers are more aware of what to look for and remember.
Practical Application:
This technique is recommended for non-fiction works whose primary goal is understanding and retaining information.
It is particularly useful for professionals and students who need to efficiently process a large volume of material.
Caveats:
The article advises against using this technique for fiction, as it can spoil the enjoyment of the narrative and surprises.
Key Quotes
"They read backwards."
"If a reader knows the conclusion or ultimate endpoint of a nonfiction book, article, or essay, then returning to the beginning of the work with this knowledge makes discerning what is important to comprehend and retain more obvious."
"Reading backward in this fashion allows some learners to consume stacks of materials and retain all the important information they contain."
"Speed is replaced by efficiency in this process and will enable readers to spend the time they believe is necessary to capture the critical ingredients of a book or essay."
Why It Matters
Adaptation to Modern Information Demands: In a world where information is abundant and time is limited, finding efficient methods to process and retain knowledge is crucial.
Improved Professional and Academic Performance: For leaders, professionals, and students, this technique can enhance productivity and comprehension, leading to better decision-making and learning outcomes.
Practical Utility: The article provides a valuable tool for those who need to manage large volumes of reading material efficiently by offering an alternative to traditional speed reading methods.
Strategic Reading: This approach encourages readers to be more strategic about how they consume information, focusing on what truly matters and avoiding unnecessary details.
In summary, "Reading for Speed, Backwards" presents an innovative and practical strategy for improving reading efficiency and comprehension, particularly in the context of non-fiction materials. This method aligns well with the needs of modern learners and professionals who must navigate vast amounts of information quickly and effectively.
Is Silicon Valley Building Universe 25? - by Ted Gioia (honest-broker.com)
The article by Ted Gioia discusses the parallels between John B. Calhoun's 1968 experiment, Universe 25, and the current trajectory of Silicon Valley's technological advancements. Calhoun's experiment created a utopian environment for mice, providing for all their needs without threats. However, the mice society collapsed instead of thriving due to social dysfunctions. Gioia draws a comparison between this mouse utopia and the modern world driven by AI and technology, suggesting that removing challenges and providing an effortless life by technology could lead to similar societal breakdowns in humans.
Analysis
The Historical Context: Calhoun's Experiment
John B. Calhoun's Universe 25:
Design: Calhoun created an enclosed space in which eight mice were introduced into a utopian environment free from disease, predators, and scarcity.
Initial Success: The population grew rapidly, doubling every 55 days initially.
Decline: Once the population hit 620, social behaviors began to break down. Dominant males became narcissistic, females aggressive, and social roles eroded, leading to violence and reproductive failures.
Collapse: Eventually, no mice survived past day 600 despite the ideal conditions.
Modern Parallels
Silicon Valley's Vision of Utopia:
AI and Automation: Modern tech aims to create a world where AI handles all mundane tasks, theoretically allowing humans to pursue more fulfilling activities.
Leisure and Isolation: However, increased leisure time facilitated by technology can lead to isolation, as people become more engrossed in their devices than face-to-face interactions.
Social and Psychological Implications
Behavioral Changes in Humans:
Narcissism: Like the dominant mice, people may become more self-centered when their basic needs are effortlessly met.
Aggression and Disengagement: Social interactions might deteriorate, leading to increased aggression and social withdrawal.
Relationships and Community:
Decline in Intimate Relationships: Technology can lead to fewer intimate relationships, lower marriage rates, and declining birth rates.
Addiction to Technology: People may form stronger bonds with their devices than with other humans, leading to an addictive dependency on technology.
The Role of AI and Technocrats
Job Displacement and Purpose:
AI Replacing Jobs: AI's displacement of jobs could lead to a loss of purpose for many individuals as their roles in society become redundant.
Technocrats' Agenda: The push for rapid AI integration is driven by profit and control, often without considering the social and psychological impacts on the populace.
Anomie and Social Harm:
Durkheim's Anomie: The loss of social norms and structures can lead to anomie—a normlessness where individuals feel disconnected and purposeless.
Mental Health Crisis: The rise in mental health issues can be linked to the loss of purpose and meaningful social interactions, exacerbated by technological advancements.
Future Risks and Considerations
Immersive Technologies:
Virtual Reality: The advent of VR and other immersive technologies could further detach individuals from reality, leading to a passive existence focused on consumption rather than creation and interaction.
Social Engineering: The current trajectory of technological development risks creating a society similar to Universe 25, where human needs are met but at the cost of social and psychological health.
Why These Insights Matter
Critical Evaluation of Technology:
Balance: A balanced approach to technology that considers not just convenience but also the preservation of meaningful human interactions and challenges is needed.
Human Values: Emphasizing human values and purposes over corporate agendas is crucial to avoiding a dystopian future in which technology dictates every aspect of life.
Policy and Innovation:
Ethical AI: Policymakers and technologists must focus on ethical AI development that prioritizes human well-being.
Sustainable Development: Innovations should aim for sustainable development that enhances human life without eroding social structures and psychological health.
Conclusion
Ted Gioia's article powerfully reminds us of the potential pitfalls of unbridled technological advancement. Reflecting on Calhoun's Universe 25 highlights the importance of maintaining social structures, purpose, and meaningful interactions in rapid technological change. This perspective is vital as we navigate the future, ensuring that technological progress supports, rather than undermines, human flourishing.
Key Quotes
On Calhoun's Experiment: “Many other males became listless. Calhoun shared some details: They became very inactive and aggregated in large pools near the centre of the floor of the universe.”
On Modern Society: “Maybe you have wondered why people have become solitary, like the rodent residents of Universe 25. They have fewer friends, even as leisure time has increased.”
On Technology's Impact: “Welcome to Universe 25 for humans!”
On AI and Anomie: “Humans are so much more complicated than mice. So if mice need challenges and obstacles in order to flourish, we need them all the more.”
Why It Matters
The article highlights the potential dangers of a society dependent on technology and AI. By comparing modern technological advancements to Calhoun's Universe 25, Gioia warns that removing life's challenges and providing an effortless existence might lead to social decay, loss of purpose, and mental health crises. This perspective is crucial as it calls for critically evaluating the rapid technological changes and their long-term implications on human society. It balances technological convenience and preserving meaningful human interactions and challenges.
The Fault in Our Forecasts—Asterisk (asteriskmag.com)
"The Fault in Our Forecasts" by Susan Hough, published in Asterisk Magazine, explores the complexities and challenges of predicting earthquakes. The article delves into historical and contemporary issues seismologists face, highlighting the inherent unpredictability of earthquakes and the communication challenges scientists confront in conveying risks to the public. Hough emphasizes that while precise predictions remain elusive, significant progress has been made in understanding long-term earthquake probabilities and improving infrastructure resilience.
In-Depth Exploration
The Nature of Earthquake Prediction
Inherent Challenges:
Deep Underground Processes: Earthquake nucleation occurs kilometers beneath the Earth's surface, making direct observation impossible. Unlike meteorology, where weather systems can be observed and tracked, the underground nature of earthquakes limits our understanding.
Chaotic Systems: Some evidence suggests that chaotic processes govern the mechanics of earthquake nucleation. This means that even with perfect knowledge of fault conditions, predicting an earthquake's exact timing and location might still be impossible, similar to the unpredictability of avalanches despite visible snow conditions.
Historical Insight:
Early Studies: The 1906 San Francisco earthquake was pivotal in raising awareness about seismic hazards. Subsequent mapping of the San Andreas Fault and other faults laid the groundwork for modern seismology.
Debates in the 1920s: Scientists like Harry Oscar Wood and Bailey Willis recognized the significant risks posed by earthquakes in Southern California. Their work contrasted with industry-linked geologists who downplayed these risks to avoid scaring away investment.
Modern Approaches and Tools
Long-Term Forecasting:
Probabilistic Models: Seismologists use statistical models to forecast the likelihood of earthquakes over long periods. For example, there is an estimated 99.7% chance of an earthquake with a magnitude of 6.7 or greater occurring in California over the next 30 years.
Risk Communication: These forecasts inform the public and policymakers that while the exact timing is unknown, significant earthquakes are inevitable in certain regions.
Early Warning Systems:
Physics-Based Alerts: Early warning systems utilize the fact that seismic waves travel slower than electronic communications. By detecting the initial, less damaging seismic waves (P-waves), these systems can provide a warning before the more destructive waves (S-waves) arrive.
Limitations: These warnings are most effective for people located some distance from the earthquake epicenter. Those near the epicenter may receive little to no warning.
Communication Challenges
Balancing Act:
Avoiding Panic: Seismologists must communicate the seriousness of earthquake risks without causing undue alarm. Overstating risks can lead to public desensitization while understating them can result in unpreparedness.
Historical Lessons: The response to earthquakes in the early 20th century, such as the 1906 San Francisco earthquake, shows the importance of clear, accurate communication. Misunderstandings about earthquake risks can have dire consequences.
Infrastructure Resilience
Engineering Advances:
Building Codes: Modern engineering and construction techniques have significantly improved the resilience of buildings and infrastructure. Structures built to current codes are more likely to withstand earthquakes, reducing casualties and damage.
Predictive Models: Seismologists can predict how the ground will shake in different scenarios. These models inform building designs to ensure they can endure expected seismic forces.
Psychological and Social Impacts
Earthquake Trauma:
Lasting Effects: People who experience severe earthquakes often suffer from long-term trauma. Effective preparedness and resilient infrastructure can help mitigate these psychological impacts by providing a sense of security.
Public Involvement: Engaging the public in preparedness activities like earthquake drills can improve resilience and reduce fear.
Conclusion
"The Fault in Our Forecasts" emphasizes the dual nature of earthquake science: while precise predictions remain beyond our reach, significant strides have been made in understanding earthquake risks and improving resilience. Effective communication and continued advancements in engineering are key to mitigating the impact of these natural disasters. By recognizing the inherent unpredictability of earthquakes and preparing accordingly, communities can better withstand future seismic events.
Key Quotes
"Neither the USGS nor any other scientists have ever predicted a major earthquake. We do not know how, and we do not expect to know how any time in the foreseeable future."
"It’s not a question of if, it’s a question of when a damaging earthquake will strike."
"Major earthquakes beget earthquake predictions. In the aftermath of a damaging earthquake, with aftershocks continuing to rock the area, the public’s understandable anxieties provide fertile soil for alarmist predictions."
The Art of Asking Questions—Asterisk (asteriskmag.com)
"The Art of Asking Questions" by Adam Mastroianni explores the complexities and challenges associated with using self-report questions in social science research. Despite their widespread use, self-reports are criticized for being susceptible to biases, inaccuracies, and lies. The article delves into alternative methods like the Implicit Association Test (IAT) that aim to bypass self-reports by observing behaviors and physiological responses. However, these methods also face limitations and have not fully replaced self-reports. The piece argues that while self-reports and indirect measures have flaws, they can be complementary tools for gaining insights into human behavior and psychology.
Key Points
1. Prevalence and Criticism of Self-Reports
Self-reports are ubiquitous in social science research due to their simplicity and directness. They are used in various contexts, such as:
Mental Health: Tools like the Beck Depression Inventory and PHQ-9 rely on self-reports to diagnose and measure the severity of depression and anxiety.
Personality Assessments: The Big Five personality traits are often measured through self-report questionnaires.
Happiness Studies: Longitudinal studies on happiness frequently use self-report measures to gauge individuals' well-being over time.
However, the reliability of self-reports is often questioned because:
Respondent Biases: People may answer in socially desirable ways or misremember events.
Cultural Differences: Different cultures may interpret and respond to questions differently, affecting the consistency of self-report data.
Context Effects: The order and context in which questions are asked can significantly influence the responses.
2. Limitations of Self-Reports
The article highlights that while self-reports are easy to administer, they come with significant drawbacks:
Inconsistency: Slight changes in wording or context can lead to vastly different answers.
Confabulation: People often explain their behaviors or thoughts that may not be accurate simply because they feel compelled to answer.
Memory Issues: People’s recollections of past events can be flawed, leading to inaccurate responses.
3. Alternative Methods
Researchers have developed various methods to get around the limitations of self-reports:
Implicit Association Test (IAT): Measures implicit biases by evaluating how quickly individuals associate different concepts.
Functional Magnetic Resonance Imaging (fMRI) Tracks blood flow in the brain to infer mental states.
Physiological Measures: Observe physical responses such as heart rate, pupil dilation, and hormone levels to infer emotions and reactions.
Natural Language Processing (NLP): Analyzes language use in social media, written texts, and spoken words to draw inferences about psychological states.
These methods can provide insights that self-reports might miss, but they also have limitations, such as being expensive, time-consuming, or requiring complex interpretation.
4. Case Study: Greenwald’s IAT and the 2016 Election
Tony Greenwald’s use of the IAT in the 2016 election aimed to uncover hidden biases that traditional polls might miss. His findings suggested:
Negative Associations with Trump: Among Republicans who favored other candidates in the primary.
Positive Associations with Clinton: Among Bernie Sanders supporters.
Despite these insights, the IAT failed to predict the election outcome accurately, illustrating that indirect measures are not foolproof and may not always align with actual behaviors.
5. Complementary Relationship
The article emphasizes that neither self-reports nor alternative methods alone can provide a complete picture. Instead, they should be seen as complementary:
Self-Reports: Offer direct insights into people's thoughts and feelings, which can be valuable despite their flaws.
Indirect Measures: Provide a different perspective, helping to uncover subconscious biases and physiological responses.
The divergence between these methods can lead to deeper understanding and more nuanced research findings.
6. The Issue of Lying
People’s fear of being lied to fuels skepticism toward self-reports. However, studies indicate:
Lower Actual Lying Rates: Than what people expect.
Detection Strategies: Researchers use techniques like consistency checks and cross-referencing data to identify and mitigate lies in self-reports.
Additional Insights
Methodological Innovations
The ongoing development of new methodologies in social science is crucial. Innovations like the IAT and fMRI open new avenues for understanding complex human behaviors and mental states. However, these methods require careful validation and should be used with traditional self-report measures to ensure comprehensive insights.
Practical Implications
Understanding the strengths and weaknesses of different research methods has practical implications:
Policy Making: Accurate data is essential for crafting effective policies, especially in areas like public health, education, and social justice.
Clinical Practice: Reliable measures are crucial for diagnosing and treating mental health conditions.
Market Research: Businesses rely on accurate consumer data to make informed decisions about product development and marketing strategies.
Conclusion
"The Art of Asking Questions" underscores the importance of methodological rigor and the need for a multifaceted approach in social science research. While self-reports have flaws, they remain indispensable tools when used thoughtfully and combined with other methods. This balanced approach can lead to more accurate, reliable, and insightful findings, ultimately advancing our understanding of human behavior and psychology.
Key Quotes
“Self-reports are quick and easy to get — you simply ask someone a question. But they can also mislead you.”
“When we used both polling-type questions and the IAT, it became instantly clear that spoken and unspoken measures are no longer in sync.”
“The worst of these lies, though, are the lies people tell you just for the hell of it.”
“Neither self-report nor Stroop gives us the complete picture of self-control, or even more than a tiny sliver of that picture. But those slivers are unique.”
Why It Matters
Understanding the limitations and benefits of different research methods is crucial for advancing social science. Despite their flaws, self-reports remain valuable when used carefully and in conjunction with other methods. Exploring alternative measures can help researchers gain deeper insights into human behavior and reduce self-report biases. This balanced approach can lead to more accurate and nuanced findings, essential for fields ranging from psychology to political science. The article underscores the importance of methodological rigor and innovation in understanding the human mind and behavior.
Fundamental Questions - Lorenzo Pieri’s Blog
Fundamental Questions
Great Questions