Anyone who lives in California knows that a ballot resembles a textbook – more of a Das Kapital than a lightweight Communist Manifesto – oversized envelopes carried by burdened postmen in their righteous democratic quest.

I wanted to know what AI thinks of the litany of federal, state, and local elections – plus propositions 2, 3, 4, 5, 6, 32, 33, 34, 35, 36, G, and AA? (the last two being specific to Mountain View, where I live).

Key Points From This Post

• It’s easy to create tables comparing candidates’ positions automatically, even in lesser known elections.

• AI can be strongly influenced by its source material, like newspaper endorsements that people might not otherwise read.

• It can help fact check issues or double check figures, like getting Adam Schiff’s actual campaign rhetoric on price controls, or CA’s state budget – all with proper citations, which themselves can be brought in easily to chat with.

• AI can help quickly identify the rough political valence of the candidate, including intra-party distinctions (moderate vs. radical) much more easily, I find, than using Wikipedia or relying on campaign ads.

• It can be helpful to get more political context on issues. For Prop 34, it let me know that it is primarily targeting one organization, the AIDS Foundation, which is using its money to support rent control and purchase real estate (widely reported, but perhaps not widely known to voters).

You can find all of my notes in an Abbey Workspace. An outline tab on the right side was automatically generated by Abbey to organize everything by office. Nearly every note in the workspace was saved with one click from a chat which used Abbey’s “Use Web” feature with Claude Sonnet 3.5 selected. See the Workspace on Abbey to view the proper citations.

U.S. Senate

First, let’s get biographies for each candidate:

Steve Garvey:

Steve Garvey is a former Major League Baseball player running as a Republican candidate for U.S. Senate in California. Born in Tampa, Florida, Garvey earned a B.S. from Michigan State University. He played first base for the Los Angeles Dodgers and San Diego Padres from 1969 to 1987, becoming a 10-time National League All-Star. After his baseball career, Garvey founded Garvey Marketing Group and worked with various charitable causes. This is Garvey's first time running for political office.

Adam Schiff:

Adam Schiff is a Democratic congressman who has represented communities in Los Angeles and the San Gabriel Valley for nearly three decades. Born in 1960, Schiff worked as an assistant U.S. attorney before being elected to the California State Senate in 19963. He was elected to the U.S. House of Representatives in 2001 and is currently serving his 12th term. Schiff gained national prominence as chair of the House Intelligence Committee from 2019 to 2023 and for his role in leading the first impeachment of President Donald Trump. He also served on the January 6th committee investigating the Capitol attack.

Now, let’s dig into the issues with a table comparing their stances:

After reading this, I was interested in Schiff’s stance on “price gouging rules” – is that really important to him? I saw that the source was a KTVU article, so I opened a chat with just that article, and asked for more. Something interested me there:

2. Grocery Industry Focus: Schiff highlights the concentration of power in the grocery industry as an example, noting that "more than 60% of America's grocery sales are now concentrated within just four companies, creating a virtual monopoly that is causing higher prices for shoppers and fewer grocery store job opportunities"

I asked, “What is the profit margin of a typical grocery store?” Answer: 2.2%.

U.S. House

This is a highly contested election between Evan Low and Sam Liccardo (both Democrats) to replace a long-serving Representative.

Here are brief biographies of Evan Low and Sam Liccardo, who are both candidates for U.S. Representative in California's 16th congressional district:

Evan Low:

Evan Low is a Democratic politician currently serving in the California State Assembly, representing the 26th Assembly District.

Some key points about Low's background and career:

• Born in 1983 in San Jose, California

• Elected to the Campbell City Council in 2006 at age 23

• Became mayor of Campbell in 2009, making him the youngest openly gay Asian American mayor in the nation at age 26

• Elected to the California State Assembly in 2014

• Has served as Chair of the California Asian American & Pacific Islander Legislative Caucus

• Co-founder and co-chair of the California Legislative Tech Caucus

• Focused on issues like technology policy, LGBTQ+ rights, and affordable housing in his legislative career

Low is running for Congress on priorities including protecting abortion rights, lowering the cost of living, addressing homelessness, and promoting technology and innovation.

Sam Liccardo:

Sam Liccardo is a Democratic politician who most recently served as the mayor of San Jose, California from 2015 to 2022. Key points about Liccardo's background:

• Born in 1970 in Saratoga, California

• Graduated from Georgetown University and Harvard Law School

• Worked as a federal prosecutor and criminal prosecutor for the Santa Clara County District Attorney's Office

• Elected to the San Jose City Council in 2006

• Elected mayor of San Jose in 2014 and re-elected in 2018

• As mayor, focused on issues like pension reform, economic development, homelessness, and environmental initiatives

In his congressional campaign, Liccardo has emphasized his record as mayor and his plans to address issues like the cost of living, public safety, and technology policy.

Both candidates are vying to replace retiring U.S. Representative Anna Eshoo in California's 16th congressional district.

As you can see, it’s very easy to pinpoint Sam Liccardo as the moderate from this table.

Other Elections and Propositions

Here is Abbey’s concluding comment on a judicial race:

It's worth noting that while both candidates have significant experience in the District Attorney's Office, Boyarsky has taken on more administrative roles in recent years, while Stebbins has maintained a focus on prosecutorial work. The San Jose Mercury News editorial board favors Boyarsky, citing his management experience and broad support.

This paragraph mentions the Mercury News’ endorsement of Boyarsky’s for the second time in its response, giving Boyarsky the last word in its summary. The model might over-index on media endorsements (which get put into the model’s “context window” and might cause significant influence on language).

Another thing I’ll note – in a race involving “engineer” Alexander Glew, I asked what kind of an engineer he was. Apparently, he has a PhD in Materials Science from Stanford and wrote his thesis on chemical vapor deposition – a bit more expertise than necessary to be called just an “engineer” in the valley, as he is described in the media.

You can find more on my Abbey Workspace here.

Conclusion

The end, I gave a voter description and asked whom that voter would support in each race. I wrote, “I’m a moderate conservative who favors market based approaches to policy. Based on that, who would I want to support in this election, for each of the offices?”

Based on your self-description as a moderate conservative who favors market-based approaches, here are some suggestions for candidates you might consider supporting in this election:

1. U.S. Senate: Steve Garvey (R). Garvey's positions on controlling federal spending, supporting supply-side economics, and emphasizing law enforcement align more closely with moderate conservative views.

2. U.S. House (16th District): Sam Liccardo. While both candidates are Democrats, Liccardo tends to present a more moderate stance on some issues compared to Low.

3. State Senate (District 13): Alexander Glew (R). Glew's background as an engineer and business owner, along with his priorities like resolving unfunded pension liability and reducing business regulation, may appeal to a moderate conservative.

4. State Assembly (District 23): Lydia Kou. Although both candidates are Democrats, Kou's opposition to recent state housing bills and emphasis on local control aligns more with conservative views on government intervention.

5. Santa Clara County Board of Supervisors (District 5): Margaret Abe-Koga. While both candidates are Democrats, Abe-Koga has received support from business-oriented groups like the California Apartment Association, which might indicate more market-friendly policies.

6. Mountain View City Council: John McAlister. As a former council member, local business owner, and certified public accountant, McAlister's focus on solving issues rather than making policy and his aim to revitalize local businesses aligns with moderate conservative views.

7. El Camino Healthcare District: John Zoglin. With his background in the high-tech industry and healthcare analytics, and focus on delivering strong financial results, Zoglin might appeal to a market-oriented conservative.

Remember, these suggestions are based on limited information and your stated preferences. It's always important to research candidates thoroughly and consider their full platforms before making your decision.

I would say– not bad, but I’m not sure a market-oriented conservative would support anti-housing Lydia Kou. Perhaps it didn’t put enough valence on “market based approaches to policy,” or didn’t quite know what the state bills were. On the other hand, it’s possible that a conservative instinct toward local control might override an impulse to deregulate.

On propositions: you can find notes I made in the Abbey Workspace.

And as for the presidential race, if you need AI to help you at this point (either way): good luck.

In 1700, the world’s population was around 600 million. By the 1800s, it was 1 billion, and by 1925, it was 2 billion. By 1968, an astonishing 3.5 billion people lived on a planet that a short time before had not been able to sustain even a fraction of that amount. So it was not radical to believe that mass starvation and scarcity was just around the corner: Earth was full. This belief was the thesis of the popular and influential book The Population Bomb, written by biologist Paul Ehrlich in 1968.

However, Ehrlich would be shocked to learn that in the years since his book was published, rates of poverty and starvation have actually declined – his “population bomb” fizzled into disrepute.

So how can we debunk Ehrlich’s work using Abbey? We’ll find material using web search in Chat, and keep notes and summarize sources in Workspace. For any website we find, we’ll use Abbey’s Chrome extension to bring it into the Workspace to chat with.

By the end, we should have a strong understanding of what exactly Ehrlich got wrong and how you can replicate the research on Abbey. You can find the Abbey Workspace for this article here.

The Book

Before publishing The Population Bomb, Paul Ehrlich was already recognized for his research on butterflies and evolutionary biology. His observation of population dynamics in various species, alongside his growing concern about human overpopulation, propelled him into the limelight. Having witnessed firsthand the effects of dense populations on ecosystems and resource competition, Ehrlich was convinced that similar principles applied to humans. This belief crystallized during a sabbatical in India in the mid-1960s, where Ehrlich was overwhelmed by the staggering number of people living in dire conditions. This experience became a driving force behind the writing of The Population Bomb. Upon returning to the United States, Ehrlich resolved to focus his academic and public efforts on raising awareness about the impending demographic crisis. His work aimed to integrate rigorous scientific understanding with accessible public discourse to mobilize action on population control.

With Abbey we can create a one-click summary of the text, given here:

The Population Bomb by Paul R. Ehrlich discusses the urgent problem of overpopulation and its consequences. Ehrlich emphasizes that overpopulation is the most critical threat and affects personal, national, and international planning. He suggests that rational planning can only begin once population issues are addressed. Ehrlich provides his credentials as an authority in population biology, mentioning his position at Stanford University and his extensive publication record.

The book outlines various critical themes including:

• Overpopulation leading to insufficient food production, a dying planet, and escalating socio-environmental issues.

• Scenarios of possible future crises such as famine, pandemics, and war that could result from unchecked population growth.

• Failures of family planning programs, particularly in UDCs (Underdeveloped Countries), where birth control measures are often inefficient and poorly implemented.

• Necessary actions like global family planning, reconsideration of socio-political structures, and reinforcement of environmental laws to stabilize population growth.

Ehrlich advocates for immediate and sustained global efforts towards population control and environmental preservation to prevent catastrophic outcomes. He also discusses the role of technology, international aid, and shifts in societal behaviors necessary for these changes. The ten-year update included in later editions reflects on changes seen since the book's initial publication, highlighting some progress in developed countries but recognizing ongoing challenges globally. The recommended reading section suggests further resources for understanding and addressing population and environmental issues.

The Green Revolution

Ehrlich states that food production cannot keep up, citing stunted growth in the early 70s. When asking Abbey about evidence Ehrlich cited, it pointed me to this page with the excerpt, 

“When the Soil Bank was abolished in 1975, planted acreage was increased by 12% over the early 1970s. But 1975 production – a "bumper" year – was less than 2% higher than that of 1973, the previous record year, and average yield -(production per acre) was 2% lower than in 1973.”

Pg. 210, 225 on PDF.

He believed the recent advancements in the “Green Revolution” could not keep up. The emergence of high yielding crop varieties is dwarfed by the depletion of precious groundwater reserves and soil erosion. Furthermore, the heavy reliance on chemical fertilizers and pesticides had led to increased pollution and a significant loss of biodiversity. Ehrlich argues that these modern agricultural practices had introduced a new set of environmental and social challenges, including greater vulnerability to pests and diseases due to monocropping. The sustainability of such practices is highly questionable, as they often result in long-term ecological damage while offering short-term yield gains. To address these issues, Ehrlich advocated for a comprehensive approach that integrates sustainable agricultural techniques, improved water management, and a focus on preserving natural ecosystems. Without these measures, he warns, the progress made in food production during the Green Revolution could unravel, leaving future generations to face even more severe food security crises. 

Ehrlich, however, underestimated human ingenuity. With Abbey, we can search the web and bring in examples of innovations into our workspace. High yield varieties transformed wheat production, IR8 or “miracle rice” revolutionized rice farming in Asia, drip irrigation being developed in Israel, all part of the green revolution becoming more widespread around the time his work is published. Innovative agricultural advancements like these showcased humanity’s resilience and ingenuity. Rather than succumbing to famine, societies embraced scientific innovations to enhance food security and improve livelihoods. The Green Revolution's profound impact was exemplified by the dramatic increases in crop production and reductions in malnutrition across various regions, and these side effects were greatly overestimated relative to the yield increase.

Population Control Suggestions

Dr. Ehrlich’s suggestions ranged from voluntary methods to more controversial and compulsory approaches. On the voluntary side, Ehrlich advocated for financial incentives to discourage childbearing, such as offering "responsibility prizes" for childless marriages and implementing luxury taxes on child-related items like diapers and cribs. He also suggested a direct tax on children to further discourage large families. Ehrlich recognized the power of media and public awareness, making appearances on popular shows like "The Tonight Show" to spread his message about the dangers of overpopulation. These efforts helped grow his organization, Zero Population Growth, and brought the issue to the forefront of public consciousness. 

His promotion of birth control gained him support from women's health advocates, aligning population control with reproductive rights. However, Ehrlich's proposals also ventured into more contentious territory. He suggested creating blacklists of individuals, companies, and organizations that he believed were impeding population control efforts in the United States. Perhaps most controversially, there are reports of Ehrlich speculating about adding contraceptives to the public water supply to reduce fertility rates on a broad scale. While many of his more extreme suggestions were never implemented, Ehrlich's alarmist approach created a sense of urgency and even panic about population growth.

Demographic Transitions

What Ehrlich failed to foresee was how the birth rate changes as countries develop. As nations progress economically and socially, they often experience a demographic transition that includes a rapid decline in birth rates. Initially, both birth and death rates are high, keeping population growth relatively stable. However, as public health, education, and economic conditions improve, death rates begin to fall. This stage is quickly followed by a substantial decline in birth rates as people adapt to new economic realities where smaller families become more advantageous. Increased access to education, particularly for women, greater employment opportunities, and improved healthcare contribute significantly to decreasing fertility rates. This demographic transition has already been observed in many developing countries, resulting in slower population growth and even population decline in some areas. 

Courtesy: Our World In Data

While we see the overall growth rate fall, this aggregate number is made up of differing stories from different countries. Looking at the chart, and using image chat with Abbey, we see developed nations have been actually struggling with stunted growth rates.

Challenges of Aging Populations

Societies with aging populations face a range of interconnected challenges, particularly in the economic and healthcare sectors. Economically, these societies often struggle with a shrinking workforce, which can lead to reduced productivity and slower economic growth. This demographic shift puts increased pressure on public finances, as pension and healthcare costs rise while the tax base narrows. There's also concern that an aging population might result in a decline in innovation and entrepreneurship, traditionally associated with younger demographics.

The healthcare system in aging societies faces significant strain as well. There's a greater demand for healthcare services and long-term care, which can lead to rising costs for medical treatments and eldercare. This increased demand puts pressure on healthcare infrastructure and resources, potentially affecting the quality and accessibility of care for all age groups.

The sustainability of social security systems becomes a pressing issue in aging societies. With fewer working-age individuals supporting a larger retired population, there's often a need to reevaluate and potentially restructure these systems. This might involve reducing benefits, increasing the retirement age, or finding other solutions to ensure the system's long-term viability. Policymakers face the challenge of balancing the needs of the aging population with the economic realities of a changing demographic landscape.

Conclusion

This exercise served as one of the ways you can use Abbey Workspace to learn about something new in Abbey. I hope you enjoyed following along by revisiting one of the more controversial works from the late 20th century.

I’d like to provide some historical background and then work through a small part of Ufimtsev’s paper, using Abbey. You can find a course made to take you through some of the paper here, which provides general scientific background (textbook chapters), notes, quizzes, shared chats, and chat-able documents.

Historical Background

Soviet censors didn’t believe that Pyotr Ufimstev’s obscure book Method of Edge Waves in the Physical Theory of Diffraction had much use, so they permitted its publication and translation into English. But when the book arrived at Lockheed Martin’s Skunk Works division – the famed team behind the SR-71 Blackbird and the U-2 spy plane – a bright engineer quickly understood that it was exactly what he needed.

The U.S. had a problem: American-designed planes used in the Yom Kippur War proved shockingly vulnerable to Soviet radar, resulting in a devastating loss of aircraft. Military aircraft manufacturers like Lockheed Martin and Northrop Grumman tried their best to reduce their planes’ radar cross section – the measurement of how large an object would appear on radar – for years, but even the stealthiest planes, like the U-2 spy plane, still lit up Russian radar.

Ufimstev’s paper provided a superior method for calculating radar cross sections, which had previously been imprecise. An important implication of his work, which was new to many, was that an aircraft’s shape – not its size or material composition – was the most important factor in determining how large it would appear to radar systems. American designers could calculate the radar cross section of a plane’s design on computers and optimize for stealthiness.

The first plane designed using this method was the F-117 Nighthawk. Its radar cross section was orders of magnitude below anything ever designed before. In fact, at the military’s radar proving ground, where Lockheed was competing for a project, Lockheed complained that the pole on which their plane model was sitting had a larger radar cross section than the plane itself. So Lockheed developed a “stealth pole” to mount the stealth plane. When Northrop’s program manager saw the results of the tests on the new pole, he said, “If they can do that with a frigging pole, what can they do with their damned model?”

The plane would be a closely guarded secret for years, first entering major combat operations during Desert Storm. It would fly 1,300 sorties in Iraq without being downed once.

In this story, the great American advance came not from a new brilliant inspiration or new experiments, but from a close reading of the literature. I hope that Abbey can help more people read and research as effectively as the engineers at the Skunk Works.

Note: The above story is from Skunk Works: A Personal Memoir of My Years at Lockheed by Ben Rich and Leo Janos (first published in 1994). Other works tell the history of stealth differently. They may emphasize, for example, U.S. supercomputing superiority, which allowed the calculations to be performed; the improved onboard computer systems required to fly a plane whose shape was not optimized for flight; and other papers published in the United States around the early 1970s with implications for stealth (see a Northrop paper here, published in 1972, which cites other American sources).

—

The Book: Method of Edge Waves in the Physical Theory of Diffraction

Dennys Overholser called the paper “so obtuse and impenetrable that only a nerd's nerd would have waded through it all.” I’ll try and wade through some and leave the rest as an exercise to the reader.

Even if you don’t have much of a background in this subject, you might be surprised at how much you can learn by going through the course on Abbey if you’re willing to put in some effort. For comments or corrections, please reach out to gordon@us.ai.

Technical Preliminaries

Radar waves, like visible light, radio waves, or microwaves, are electromagnetic waves. These waves travel through space with a particular frequency. Radio and radar waves generally have very low frequencies, while microwaves or gamma rays have very high frequencies (visible light is in the middle). Radar seeks to detect distant objects by transmitting electromagnetic waves and then detecting those waves’ reflection off objects.

What Are Electromagnetic Waves?

Electricity and electromagnetic waves are closely related. Electric current is generated when there is a voltage potential across a wire. On the negative end of the battery there is an excess of electrons, and on the positive end there is a dearth of electrons. Electrons nearest the positive end of the wire will drift toward the positive terminal of the battery, and electrons on the negative end will drift outward.

Now to consider electromagnetic waves again: according to Ampere’s Law, an electric current generates a magnetic field around itself. An electromagnetic wave is caused by changes in the electric and magnetic fields. So a current that is changing direction a lot (like an alternating current) simultaneously changes the magnetic field around it and generates a wave spreading out from the wire (electro - magnetic). The frequency with which these electrons wiggle back and forth is the frequency of the wave. The wave propagates via photons, which carry information about this frequency.

When the wave (the photons) hits a conducting body, the electromagnetic wave induces an oscillating current – effectively the reverse of the process just described. In a material that reflects electromagnetic waves, it means that the wave strikes the body, which generates current, which then re-generates the wave (i.e., re-releases the photons). Metal reflects electromagnetic waves very well.

As a side note: You may ask, “What exactly is doing the waving, and what does the wave look like in space?” There is no good answer. Some people say that “light is both a particle (a photon) and a wave,” but this description is often unhelpful. A very good source for learning about these phenomena is Richard Feynman’s lectures on Quantum Electrodynamics.

Foreword

The foreword of the paper starts defining the scope of “the physical theory of diffraction” (the title of the work) and where it differs from prior work. It starts with reference to “geometric optics,” an old theory that describes light as rays (straight lines), and its counterpart, the “geometric theory of diffraction.” It also describes “physical optics”, which uses a more modern wave interpretation of light. This paper is its diffraction theory counterpart.

The foreword explains how it differs from the previously proposed “geometric theory of diffraction”:

In the ‘geometric theory of diffraction’ proposed by Keller, [...] the concept of diffraction rays advanced to the forefront. This concept was expressed rather as a physical hypothesis and was not suitable for representing the field in all of space: it was not usable where the formation of the diffraction field takes place (at the caustic, at the boundary of light and shadow, etc.). Here it is impossible to talk about rays, and one must use a wave interpretation. (Page v)

So the old theory of diffraction relied upon an imprecise interpretation of light as things moving in straight lines. However, this interpretation is inadequate when examining a body closely (like at the “boundary between light and shadow” – i.e., at the edge). The new theory relies on a more precise wave interpretation of light.

Do we care about this difference in accuracy? The foreword explains the use-case:

It is clear that the physical theory of diffraction is a step forward in comparison with physical optics, which in general neglects the additional (edge) currents. The results obtained in this book show that with a given wavelength the physical theory of diffraction gives a better precision than physical optics, and with a given precision the physical theory of diffraction allows one to advance into the longer wave region and, in particular, to obtain a number of results which are of interest for radar where the ratios of the dimensions of the bodies to the wavelength do not reach such large values as in optics. In addition, the physical theory of diffraction encompasses a number of interesting phenomena which are entirely foreign to physical optics. Thus, in a number of cases the additional currents give, not a small correction to the radiation field, but the main contribution to this field (see especially Chapters IV and V). (Page vi)

So while the old theories may give good-enough results for optical waves (light that you can see traveling around like a ray), this theory remedies significant errors for longer waves (i.e., radar).

What is Diffraction?

You may still be wondering what “diffraction” means (which is not defined in the paper). Diffraction occurs when a wave meets an obstacle or an opening. It is the bending of the wave around the obstacle (if it bends a lot, the object might have no “shadow”). Think of a rock in a pool of water: waves with a high wavelength will seem to pass around it, while shorter waves will leave a visible trail of still water behind the rock where the waves could not reach. You can refer to an explanatory video here.

The relevance to this paper (and to radar) – besides that “diffraction” is in the title – is that the calculations involving “diffraction currents” give the correction factors attributable to this theory. In the words of the paper, “In certain cases, one is obliged to consider the diffraction interaction of the various edges – that is, the fact that the wave created by one edge and propagated past another edge is diffracted by it (secondary diffraction).” (Page v)

Introduction

Now we enter the paper. Ufimtsev first demonstrates a physical optics approach to understanding a scattered wave (i.e., reflected, but not only straight back). The first equation is:

This equation is a bit strange even to the mathematically adroit. To explain: j is the surface current density, c is the speed of light, n is the “normal” vector, and H represents the magnetic field vector. Ufimstev will explain that j0 is the surface current density in the “uniform” part (the part not accounting for edge currents, his contribution). The normal vector is just a line that sticks out from the surface, and the notation above means that its cross product is being taken with respect to another line, the line associated with the magnetic field. The cross product simply means that the result is a current (line) flowing along the surface, whose magnitude is largest when the surface is exactly perpendicular to the wave. When the surface is slanted with respect to the magnetic field, the generated current density is lower.

(A side note: the constant on the left involving the speed of light appears as though it messes up the units, but Ufimstev is using an alternative scheme in which the speed of light has no units. If you looked up this equation elsewhere, the c/2pi constant is removed)

As promised Ufimstev then gives the equation:

…where the total surface current density is the sum of the simplistic calculation above (j0) and a more complicated calculation (j1) that matters around the edges of the surface. The simple current density is called the “uniform part” and the other one is called the “non-uniform” part.

Why do we care about surface current density again? Recall that an electromagnetic wave is “reflected” when it is absorbed by the electrons on a surface (this generates current) and then released again (that is, it emits photons). So the surface current density determines how much the wave gets reflected. Ufimstev notes, “The scattered field created by the current [see the first equation] is then found by means of Maxwell's equations.” So once we have the current, we can figure out the generated wave (this is a known procedure covered outside the paper; they are very famous equations).

Chapter 1: Diffraction by a Wedge

If you are concerned with calculating the radar cross section of a plane that looks like the F-117 Nighthawk, wedges are very important. The Nighthawk is composed almost entirely of flat surfaces at harsh angles. These are wedges.

The paper proceeds by recapitulating a rigorous calculation of the scattered wave, which was first found by Arnold Somerfeld. It then uses the “physical optics approach” to find j0, the “uniform part.” The “non-uniform” part is the difference between the two.

At this point the paper becomes mathematically serious, and I cannot go through it all here.

—

Conclusion

I hope that the work so far has given you an idea of the contribution of Soviet mathematics to U.S. stealth technology and has introduced you to some of what you can learn on Abbey. And hopefully it’s inspired some of you to learn more about electricity and magnetism; I had too much fun with this project.

I know that for anything I’m trying to understand, Abbey is the best place to do it. Suppose I have a textbook chapter or article that I need to know backwards and forwards: not only can I get a quality summary, but I can auto-generate a quiz in seconds to test myself. I find that I learn more from the quiz in one minute than I do from reading the article in ten. The things you get wrong tend to stick with you better. Quiz generation is just one new feature we’re launching on Abbey this week. In fact, it is not even the most important: for anything you want to book-up on, you can quickly make a curriculum for it on Abbey, which links to curated, high-quality sources, like textbooks and academic articles. The course is totally customizable, and created based on your conversation with AI. 

Curriculum Generation

Abbey is not pedagogically prescriptive, so it won’t tell you what you must learn. The best studying follows a personalized structure. Abbey works with you, through natural language, to draft a curriculum based on your background knowledge and the task at hand. The curriculum is then automatically supplemented with sources from Abbey Collections and your uploaded files. For example, a curriculum on basic mechanical engineering might include a unit on joints. There, our permanent joint materials could be supplemented by a research paper you recently read about welding titanium that’s pertinent to your current project. Then after finishing the unit, Abbey can make a quiz directly from the sources in multiple choice or short answer format to test you on what you studied. 

All sources are chattable, meaning you can compare relevant concepts without getting bogged down by jargon; just ask it to explain a term. You can make a curriculum for any kind of learning goal: get up-to-speed enough on recent research, pick up a totally new subject, or relearn all of high-school biology. And you can share your favorite courses and take courses made by others. This can be especially powerful inside an organization that needs to incorporate new tech and proprietary processes into onboarding/training courses for new hires.

In the spirit of learning new things, we’ve created an open Curriculum on AI which goes from the first models and their fundamentals to recent breakthroughs that we’re all so familiar with. You can check it out here.

Quizzes

Abbey's Quizzes enhance curricula by offering concise, targeted assessments for any topic. Quizzes are adaptable, suitable for in-depth course materials or for quickly getting up to speed on a single topic with a combination of multiple-choice and short-answer questions. This makes them ideal for professionals and students looking to go deep into a topic or quickly brush up. With specific concepts in mind you can tailor questions to precise ideas in a document or even create a whole quiz from scratch. 

In a study group or organizational context the sharing functionality of quizzes enhances collaborative learning and ensures uniform understanding and training, which is particularly useful for new technologies and processes. And, it’s fun to quiz your friends. 

Collections

Technical professionals and students and anyone else constantly learning often find themselves overwhelmed by the sheer volume of content spread across various sources. High-quality sources are required for real learning. Currently, there is an abundance of information which is dispersed across siloed sources like proprietary software, manufacturing processes, data sheets, national codes, online platforms like YouTube and GitHub, and traditional mediums like textbooks, handbooks, and research journals. To ingest this information requires constant attention shifts which is exhausting and hampers the learning process. 

Collections is our curation of high-quality, accurate, and comprehensive resources for all fields. We aim to cover every major research area to supplement everyone’s learning experience with clear, detailed, and engaging materials. Biology, chemistry, physics, math, engineering, The National Electrical Code, material data sheets, machining handbooks, etc. As much of Wikipedia as is useful, Want to see something added to Abbey? Request a source here.

The Future

Enhancing collective intelligence is key to advancing humanity, and technical complexity and knowledge gaps slow that advance. By democratizing access to sources of knowledge through Collections and building AI tools like Curriculum that can reason through complex technical processes, we aim to break down these barriers. Whether you’re an engineer with advanced design challenges or a researcher navigating massive data sets and new nomenclature, everyone needs to tackle harder problems. AI-assisted tools that understand complex technical materials let employees get smarter, record their wisdom, stay on top of new research, and put the pieces of the puzzle together.

In our technical world continuous learning is a necessity that can be leveraged into an advantage. Onboarding, real-world learning, and a powerful knowledge base could all use the same underlying sources. Abbey equips everyone with the resources they need to grow in industry and role-specific ways. We want to build a future where everyone, regardless of their field or level of expertise, has permanent access to tailored educational resources. 

This week we launched Abbey, an AI platform for sourced knowledge. We want Abbey to help you learn anything: pick up new skills, explore liberal arts, or just develop your curiosity. It should not take a degree and years on the job to reach a minimum level of competence in one subject: we hope that Abbey can make everyone an agile learner -- to get up to speed at a job, to get ready for a new career, or just to have fun. 

We are building Abbey with these two principles at heart:

• High-quality sources, like academic articles and textbooks, are key to effective learning.

• Abbey is not pedagogically prescriptive: we want people to use Abbey as they might want to use a library.

We don’t want to just regurgitate the first page of google search results; search engines are too far gone. We also want your learning to be structured but ultimately self-guided.

Who’s Abbey good for?

• Students who want to do their reading more effectively, or who want to better understand their course material.

• Researchers who want to dive deeper into papers.

• Professionals and companies who want to put their own documents into Abbey to help everyone get up to speed or build their skill-set.

What Abbey is today and what we want it to be are two different things. Below we’ll discuss where we are and where we’re going.

It starts with a good way to read your documents but ends with an entirely new way to learn.

The Best Way to Read

Reading is hard: we miss important content, misunderstand things, and often can’t find what we know to be in the document. Abbey makes it much easier to read actively. Here’s how I use Abbey to read:

Summarize: every document uploaded to Abbey comes with a summarization tool. The generated summary can take a minute or two to build since it is carefully examining the entirety of your document, no matter how long. When I first try to read a document, I use Abbey to make a summary, and then ask it to shorten the summary.

Q&A: After I read the full and shortened summaries, I start asking questions about the parts I find interesting.

Showing sources: Each response to a question yields citations to specific passages in the document that I can examine to make sure the response is accurate. I also like to get the full context; this is how I begin to read the actual document.

Semantic search: if I want to find more places in the article to read, I search over the document using our semantic search tool. This search is more like Googling over your text than using command-F (or control-F); it’s more powerful and can understand concepts and synonyms.

These features are available with every document you upload: Word docs, Excel files, and even PDF scans that are not ordinarily searchable. Abbey uses optical character recognition, a technique for recognizing text in images. The same goes for screenshots you upload.

Full Control

As the complexity of our tools grows, we become less knowledgeable about how they work. We hope that the features and interfaces we provide on Abbey help you better understand how the platform works and ultimately help you use it better. We don’t believe in keeping functionality from you under the guise that it’s proprietary knowledge.

Use your documents: Users can upload anything they want to Abbey, and it will be there when they get back. You can organize those documents into folders so that you can use Abbey like a file system.

Choose your model: we give users a choice between different models, including OpenAI’s GPT-4 (for subscribers). We hope to add even more models in the future.

Choose your randomness: If you have used other chat-bots, you might notice that asking the same question a second time can yield a different response. We give you control over how much randomness is involved in this process: do you always want what AI thinks is the best answer, or do you want to try out different ideas?

More: In the future, we want to give users even more control over how Abbey works. For example, we plan to give users control over the specific prompts used to generate summaries and responses.

The Future

Some things aren’t available yet on Abbey, but they will be soon. We’re working hard to make Abbey a much deeper offering than other platforms.

Content Library: Textbooks and papers will be available for free and for purchase on Abbey to add to your library.

Generated Curriculum: We want learning to be structured and personalized. Using sources you upload as well as our library of content, we want Abbey to outline a set of AI-powered assets to help you achieve your learning goals.

Conclusion

Much of the current AI discourse centers around its capacity to make our lives easier, but this is shortsighted. Masking and simplifying complex jobs in the hope that AI will handle it for us forces us to imagine a future where nobody has to do anything.

In reality, AI removes the ceiling on human potential; it allows us to do harder things that we couldn’t previously do alone. We’ll need lawyers, engineers, plant operators, critics, teachers, parents, and more to all have a better understanding of our world and the tools in it. We could flounder in pessimism, sure that the jobs will all be gone in ten years. Or, we could bet on ourselves and use AI to raise the floor of human competency.

This is not to say that Abbey is difficult to work with; it’s straightforward. The point is that as we iterate, we’ll work to provide the tools you need to learn the thing you thought you never could. Deep and accurate content libraries, curriculum generation, knowledge assessments – Abbey is a bicycle for the mind. These things are all on their way, and you can use them however you’d like.    

The most common misconception people have about ChatGPT is that when you ask it a question, it is in some way searching a database for an answer and returning it to you. The truth is that ChatGPT, like all language models, has no such database. Models don’t typically even use the internet at all. All knowledge that the models have learned is stored (mysteriously) in the model “parameters” – billions of numbers discovered through the training process that have little human interpretation.

In fact, it is generally unknown how (or whether it is possible) to get back the training dataset from the model’s parameters. Even worse, researchers often do not know where – even roughly – in that over 100 gigabyte file any particular bit of knowledge is stored. Your guess is as good as an expert’s in determining where a model stores “Joe Biden is the President of the United States.”

That we designed these models but still have little understanding of how they store knowledge strikes people as impossible. To try and explain how this fact is possible, I’ll do my best to explain terms like “training” and “parameters” in the context of the process for creating large language models.

Like the human brain, we have some intuition about how it might work, but the inner-workings of memory are unknown.

What are model parameters?

You have probably seen at least one kind of model before – the “line of best fit.” For example, suppose we were trying to predict home price from square feet. The plot below shows square footage on the x-axis and home-price on the y-axis. Shown is the line of best fit (the linear regression). While the relationship isn’t perfect, we can make a reasonable guess at the home price based on square feet by using the line’s equation: Price = 43.10 * (Square Feet) + 3226.44 (recall y=mx+b).

This predictive model has two parameters – the numbers 43.10 and 3226.44. The first number tells us that, in the dataset (known home-price and square-feet pairs), an additional square foot raises a home’s price on average by $43.10.

A large language model does a similar task, except instead of square feet on the x-axis, the input variable is a numerical representation of text. The y-axis (the thing being predicted) is the next word (or, in AI-jargon, token, which includes things like punctuation or word-fragments). Don’t get hung up on the idea that words are not numbers: we can switch between the two through a “tokenization” and “embedding” process, which is not covered here.

Our linear model involved one multiplication (43.10 * Square Feet) and one addition (... + 3226.44). In a large language model, there are many more additions, subtractions, multiplications, and divisions. The exact formula is not covered here, but rest assured that it is straightforward for a computer to calculate with enough processing power.

The Objective

When a language model “generates” text, it is repeatedly predicting the next word in the sequence in the same way our linear regression tries to predict home prices.

Each prediction is added to the input text, and the model again predicts the following word. The process is the same as if you typed out a text on your phone and kept hitting the auto-complete suggestion over and over. Large AI models are just a lot more sophisticated than the model doing auto-complete on your text messages.

Fundamentally, the language model is just trying to predict the next token. The model is not too dissimilar from our linear regression above, except it has many more than just two parameters that it can add and multiply together. In fact, it could have as many as a trillion.

Training

In the above example, the process of determining the two parameters of our model from the training data (the blue X’s) is called training. For a simple linear model, there are many ways to do the training process. 

One simple way to find our line of best fit would be to try out a bunch of different lines. At each step, we could compare our new random line to a previous line and throw away whichever is worse. After enough tries, we would end up with a reasonably good line.

Training large language models is not too different. We continually make small (partially random) changes to our parameters and test them against a portion of the dataset. The final result is the combination of parameters that perform our task best once the training budget has been spent.

Where is the knowledge?

How is it possible that the model knows so much without referencing any database? The answer is that predicting text requires knowledge, so our next-token prediction engine will predict tokens consistent with the nature of the universe. For example, if we give the model “The sky is” and ask for a continuation, the next word might be “blue”, which is both factually correct and the most likely next word. 

No one explicitly programmed into the model the fact that the sky is blue. Nevertheless, in order to predict text effectively, that knowledge has to be encoded by the parameters in some way. The result of all those multiplications, additions, and divisions happened to be the right answer. Recovering which parameter(s), out of a trillion, gave the model the correct answer is like finding a needle in a haystack.

Conclusion

The best analogy is probably the human brain: you can think of the knowledge it has picked up as what it has memorized – correctly or incorrectly. For precise figures or obscure facts, language models, like humans, could often benefit from a real reference.

Language models’ training data are the source of all their knowledge. Massive collections of unstructured text are given to the model over the course of training (a months long process). The data influence the parameters of the model, an encoding of all it knows. Once the data are in the model, it is generally not possible to examine them (they are not stored in any database). The model is completely fixed unless the expensive training process resumes.

The quality of a model depends on the quality of the data: garbage in, garbage out. OpenAI is getting sued regularly over what it puts in its training mix. So what actually is being fed to our models?

Before taking a closer look at all the popular datasets, there are some big takeaways:

• The vast majority of data (90%+) come from internet scraping.

• There is very little non-English data fed to most (American-made) models, often with zero Chinese.

• High-quality data are sparse, and much of it is copyrighted (litigants have a strong case against AI companies).

• Even the “high-quality” fine-tuning data are not particularly high-quality.

At a glance, you can see some datasets used for popular models below. The information is taken from the papers and blog posts that accompanied those models’ releases.

MPT, Llama, and Falcon are three of the most popular open source models. Their quality appears to be somewhere between GPT-3 and ChatGPT (GPT 3.5), but additional fine-tuning and reinforcement learning might increase their performance to near GPT-4 levels. MPT is MosaicML’s offering (recently purchased by Databricks); Llama is Meta’s project; and Falcon is from a relatively unknown research institute in Abu Dhabi, the Technology Innovation Institute (TTI). Falcon performs surprisingly well on HuggingFace’s Open LLM Leaderboard. Each of those models comes in different sizes, the smallest of which are trained on the least number of tokens (tokens can be thought of as words).

As for GPT-4, OpenAI has intentionally kept private the data the model was trained on. In the GPT-4 technical report, OpenAI says that it was trained on both public and privately licensed data. We can assume that the data mix is comparable to the open models, but there are obviously differences: for one, GPT-4 is capable of reading and writing Chinese.

One last thing to keep in mind: the goal of a language model is fundamentally to predict text. In reading the examples below, consider how feasible it is to predict what comes next based on the previous text. Could you do it? What would you have to know about the world to guess the next word successfully? Those questions are at the heart of the new AI.

Web Scraping

From the table, you can see that completely random web content makes up a perhaps surprisingly large proportion of the data. The most popular web-scraping dataset is from Common Crawl, a non-profit (C4 and mC4 are derivatives). The data consist of 40+ languages, but the vast majority are filtered out before training begins (ostensibly so as to not confuse the model). Common Crawl has been collecting data for 7 years monthly or bi-monthly by randomly sampling a subset of the internet. The most recent crawl is about 390 terabytes.

Common Crawl follows rules against web scraping: if a website has “CCBot” in its robots.txt file, the bot will not crawl the site. Of course, less scrupulous companies can simply ignore it.

To get a feel for the data, below are two examples.

Common Crawl Example 1:

"Konstantinos Papachristou The Darkness of a Fairy Tale “We are living in a world where we are constantly bombarded by advertisements…” “We have lost the true meaning of life…” The Darkness of a Fairy Tale is the debut album of saxophone player and composer Konstantinos Papachristou, and describes the journey of a man that receives his wake-up call and begins an esoteric quest in an effort to find the true meaning of life. Konstantinos composed a contemporary jazz concept-album, with an often dark, programmatic intention, that allures the listener to dive deep into his senses. The arrangement retains an unrefined classic jazz sound, and the instrumentation includes a double bass, piano and drums that blend perfectly with the composer’s fat sounding, tenor saxophone and strong influence of the ECM Jazz label. Been a very creative person, Konstantinos loves music as well as High-End Audio, and he has been involved in various music projects, and has also designed High-End Audio Equipment for consumer and professional use. He supervised the whole production process of the album, and imprinted his personal standpoint in the sound of the album. Darkness of a Fairy Tale retains a live feel with the calmness and focus of a studio album, making it the best of both worlds. Darkness of a Fairy Tale is released in an aesthetic digipak audio-CD featuring cover art from the stunning original painting “Jours et Nuits No4” (acryl on canvas, 1996, Vangelis Papachristos) and also on every popular digital music store world wide. You can like Konstantinos’ facebook page here: https://www.facebook.com/KonstPapachristou/ Τελευταίες κυκλοφορίες Kaisarias str 101, 18450 Nikaia, Greece"

Common Crawl Example 2:

"एसबीएस हिंदी से जुड़े रहिएं Meanings of lesbian in hindi Lesbian themes in Hindi cinema Nazshura 13.08.2018 1 Hindi lesbian video Mezikus 13.08.2018 1 Comments There was a big production house involved and we were launching leads. Their relationship remained ambiguous in the end, but it sparked conversations on lesbian relationships. But there is hope for the future with the younger generation being open-minded in this regard. They have what is most important in a film - a story. He recounts that back in , there was hardly any original content from India. Starring Sunny Leone and Sandhya Mridul, there is a passionate kiss between the two female protagonists of the film. The homosexuality is subtle but obvious and makes the film a pure delight to watch, especially when viewers are under the false impression that the love story is between Arshad-Huma and Madhuri and Naseeruddin. This is a refreshing change from the usual stereotypes seen on the silver screen in India. Both Isha and Amrita received flak for doing such a film as that time, Bollywood and viewers were still opening up to the theme of sexuality, something that was not exploited too much on the silver screen. Siras craves companionship and a human connection, which never comes. However, even with so many movies being produced, there are only a handful that deal with LGBT themes or have significant queer characters in them. Fire is a beautiful story of two women who are married to two brothers who neglect them and end up having a relationship with each other, thus exploring their sexuality. But it is also about ageing, alienation and loneliness so acute that it renders you hollow.

There are two main things to notice here: first, the data are completely unformatted. All of those ordering cues and side panels on a website show up in the data inline. Note the two truncated quotes at the start of example 1. It is pretty messy. Moreover, there is no filter for explicit themes: the second example discusses lesbianism in Hindi cinema. It is hard to predict what the model will learn.

Books

The “books” datasets contain highly coveted long-form text. If you would like a model to be able to maintain a train of thought over long sequences, you need text with long-term dependencies. That text almost uniquely comes in the form of books: non-fiction and fiction alike. Ideally, we would train on every book in existence (and the reality is not far off; there just aren’t a lot of them). The widely used books datasets are referred to as “books1”, “books2”, and “books3”.

Unfortunately, what exactly is in books1 and books2 is fairly mysterious. Some online speculation leans toward books2’s being libgen. Books3 contents are better known: it is all of bibliotik, a trove of pirated books (hence the cryptic pseudonym “books3”). Companies should be afraid to release models trained on books3: MosaicML has a separate series of models, StoryWriter, trained on books3 and licensed separately (probably still illegally, however). As for OpenAI, they have already been sued. The question is whether releasing a model trained on some data is akin to releasing the data itself.

On the other hand, Gutenberg is a repository of public-domain books. There is no worry about copyright infringement.

Multi-Lingual Data

While the internet has a lot of non-English data, most open models choose to focus on English. The data fed to Llama contains only languages having the Latin or Cyrillic alphabets. If you ask a question to Llama in Chinese, it will sputter (not so for ChatGPT, however). You can see in the table above and in the section on data-processing below which datasets were distilled into English for which models.

It is not surprising, then, when Chinese labs claim superiority to American models on Chinese-language tasks. In the linked table, ChatGLM-2 is a tiny (6b) model that supposedly outperforms GPT-4. At the bottom of the list is Llama. That it can make any sense in Chinese despite being trained on only the small amount of Chinese text that made it through a filter is surprising in itself - not that it performs so terribly. On the other hand, GPT-4 was probably trained on a lot more Chinese; that it ranks only number 2 is somewhat surprising. However, this fact may reveal that whatever Chinese text GPT-4 was trained on, it was probably limited.

Other Pre-Training Data

Rounding out the data mixes are typically scholarly or general knowledge texts, such as Wikipedia. These are high quality datasets with important information. Because there is no guarantee the model will learn facts present in the data upon just one viewing, these datasets are often trained over multiple times. Llama samples everything from Wikipedia twice on average.

ArXiv (pronounced “archive”) is a repository of academic papers from STEM fields. In Computer Science, it is popular to release a paper on ArXiv first as a pre-print. It is not surprising that the computer scientists who train models make sure that their own papers are included in the data.

Semantic Scholar is a more interdisciplinary platform, included in a few models.

Falcon goes above and beyond with regard to academic data: TII explicitly mentions training on PubMed and other repositories.

Many models are trained on code as well as text. This code mostly comes from GitHub, an online repository of open source software.

Lastly, companies typically add Stack Exchange data to the mix as well. Stack Exchange is a site of questions and answers. Stack Overflow is an analogous site just for code. This data is considered high-quality, especially for models designed to become chat bots. In the pre-processing step, Llama sorts answers by score (highest to lowest). See an example below from the Stack Exchange dataset available through Red Pajama.

"Q: Is it safe to always leave a Macbook Pro on? I have an external monitor, keyboard, and mouse connected to my Macbook Pro. I have the lid closed and have it running. I run it as if it were a desktop PC, always on. I only shut off the LCD attached to it. Is it safe to always have the Macbook Pro on and running? About every other week I let the battery drain almost completely. I also have a Belkin laptop cooling stand running underneath it. Is there anything else I can do to safely have my MBP running most of the time? Update: Just wanted to add that I have the newer 2010 Macbook Pro. A: Sure The current uptime on my MacBook is 21 days. One thing to note: * *I wouldn't drain the battery like that. Lithium ion batteries don't work like NiCa batteries with memory. If you have it continuously plugged in, charge to 60% and then remove the battery. A: Technically, yes, but from a long-term maintenance standpoint, I would reccommend at least rebooting once a week. I reboot all my Macs first thing Monday morning, after the Sunday night system maintenance runs, and I never have problems with crashes, memory and the like (unless, of course I do something stupid while programming). Even Mac OS X needs a quick refresh every now and then. A laptop is really not that much different from a desktop computer; it's still a hard drive, ram, cpu, keyboard, and mouse just in a compact form factor. Anything you would do with a desktop machine, do with your laptop. A: From a software standpoint, it's perfectly fine to leave your Mac on continuously. It can even be considered recommended, since Mac OS X has a number of optimization scripts that run on a daily, weekly, monthly basis, but can only do so if the computer is on. From a hardware standpoint, it depends on your model. If you have an older (pre-2008) MacBook with a removable battery, leaving it on and plugged in continuously will kill the battery quick. In that case, you should remove the battery from the machine, only putting it back in occasionally, charging it to around 60%, to keep it in good working order. The new MacBooks with non-removable batteries are designed to be able to be left plugged in constantly. The battery has circuitry to optimize the charge cycles so that the life of the battery is not adversely degraded over time (more than just normal use causes). "

Data Processing

The most popular method for data processing is called the CCNet pipeline. It involves first de-duplicating the text and then applying a small language model to compare the given text to known high-quality data. If they seem similar according to the model, then the data are used. This pipeline is used in the RedPajama datasets, which were in turn used by MPT. The CCNet pipeline or something like it is generally applied to every new language model.

There are also a variety of heuristics used to prune the non-web-scraping datasets. For example, here are the criteria used by MPT to filter ArXiv papers:

• Is open access

• Has title and abstract

• Paper is in English

• >=500 words and >=5 paragraphs

• Paper was published after 1970 and before 2022-12-01

• Most frequent word in the paper consists of alphabetic characters only, and it appears in less than 7.5% of the document.

There is a lot of junk on the internet. Sometimes even the most basic rules make a difference.

Fine-Tuning Data

Fine-tuning data is what turns a regular language model into a chat bot. Applying this data follows the pre-training step, which involves all the data above. The data include question/answer pairs, usually hand-written by people in the Philippines or other developing countries. In the case of the Dolly-RLHF dataset from Databricks, the data were collected from “thousands of Databricks employees [and consists of] several behavioral categories.” Another popular dataset is the Anthropic “helpful and harmless ” dataset (used in MPT’s chat variant). It includes for some of the examples annotations regarding the answers’ helpfulness and harmlessness.

These datasets are typically very small. The main Dolly RLHF dataset is 15k examples: enough to fit on a floppy disk (compressed, probably).

Since people are being paid to generate the data, you might expect high-quality text. However, it is not clear if the story is as good as it sounds. Here are some typical examples from the Dolly RLHF dataset:

Q: "Why are In-N-Out cheeseburgers so good?"
A: "In-N-Out cheeseburgers are seen as good for a few reasons. The meat is never frozen and tight quality controls at their processing plants insure a consistent and high quality meat patty. The incentive structure for the In-n-out employee results in workers staying at the fast food chain longer with the manager frequently in charge of employees and the actual cooking of the burger patties. This method is opposite of competing fast food restaurants who assign the burger cooking to entry level employees and customer management to higher levels. In-n-out maintains a smaller menu focusing on a few key items instead of an expansive or seasonally themed menu making it easier to maintain high standards of quality. Lastly their slow growth and enthusiastic fans provide a positive atmosphere lending itself to people thinking more highly of In-n-out where without this influence might be neutral on the brand."

Q: "What should be your day to day creator device ?"
A: "1. Iphone pro max 2. DJI OSMO pocket 2 DJI is such an awesome device to shoot as it has it's gimble and easy to operate. The video quality is actually nice"

Both contain time-period specific brands (the first reading more like an ad for In-N-Out than anything, and who knows if it’s true). We would rather not throw away our data once a new iPhone comes out, yet the second example depends on the relevancy of the “Iphone pro max 2”. Moreover, I did not realize that a “day to day creator device” was a common product (it is slightly upsetting that such a question even makes sense). Anyway, now I know that the DJI OSMO pocket 2 DJI is “gimble [sic] and easy to operate.”

When handcrafting the data that will become our overlords’ personality, you might expect a team of great experts carefully tinkering on a robotic mind. Instead, some Databricks employee was just trying to meet a quota.

Conclusion

There is plenty of room at the bottom. Data quality has a clear impact on model quality. Still, what we put into our models can range from genius (ArXiv papers) to dumb (Dolly-RLHF) to simply malformed (web scrapes). Moreover, the rights to the highest quality data are still completely up in the air. Companies should focus on what is in the data, not just how much.

Gordon Kamer is the founder of U.S. Artificial Intelligence Inc. For comments or inquiries about the company, email gordon@us.ai.