Science today stands at a precarious crossroads. While capable of solving humanity's most crucial problems, it finds itself trapped between academic prestige-seeking and corporate profit-hunting, with public funding increasingly uncertain.

Over the past two months, there has been a slew of scrutiny and cuts to scientific funding eloquently talked about elsewhere. There are three facts that have been made abundantly clear by recent events. I think you will agree with them regardless of whether you support or oppose these recent changes:

1. Science has the power to solve problems that are crucial to the future of humanity.

2. The current methods for allocating research funding are much more focused on prestige than they are on allocating funding in as efficient and problem-focused a way as possible.

3. If scientific research in the United States continues to depend almost exclusively on federal funding, its future is in serious jeopardy.

The solution may be neither more lobbying for federal grants nor appealing to corporate altruism. A third path exists—one that combines democratic prioritization of important problems with market-driven resource allocation.

I am proposing a new funding methodology that answers a simple question:
What if we harnessed the same ruthlessly efficient mechanisms that drive innovation in the private sector to efficiently fund science that solves the most important problems?

Specifically this proposal uses a modified version of a well-validated method, prediction markets—proven systems for aggregating distributed knowledge—to build a new method for bringing private sector funds into science. This isn't a theoretical proposal requiring legislative action or institutional overhaul. It's a system I believe could be implemented right now.

Funding for non-profitable innovative ideas falls through the cracks

Science advancement today finds itself caught between two flawed funding models–academia’s incrementalism and industry’s profit-seeking. Each of these generate problematic distortions in the research landscape. Valuable ideas languish because they’re too radical for conservative academic funding and too disconnected from profit for industry investment.

Academic funding excels at solving problems that require large amounts of slow incremental progress. The Human Genome Project is a great example of a time this approach worked. It was a massive undertaking where multiple laboratories systematically sequenced the human genome over years, checking and verifying each others results. These labs ultimately completed the project ahead of schedule and under budget through steady, distributed progress.

This bias towards incremental and conservative solutions becomes problematic when academic consensus clings to faltering hypotheses, as evidenced in Alzheimer’s research. Despite decades of disappointing clinical outcomes for treatments targeting amyloid beta plaques and promising alternative hypotheses, millions in public funding continue flowing into this approach. Grant decision-makers receive no penalty for giving funds to research that will likely yield no solutions for the problem at hand. Without market pressures, academic research can become entrenched in established paradigms.

The bureaucratic overhead of traditional academic funding creates additional inefficiencies. Small projects often aren’t worth the administrative burden of applying for grants, regardless of their potential impact. Academic metrics like citation counts and journal prestige determine status but often fail to translate knowledge into meaningful outcomes for the taxpayers funding the work. Incentives across stakeholders–researchers, funders, and the public–remain fundamentally misaligned, as research priorities are determined largely by institutional gatekeepers rather than what will most efficiently solve public needs.

Industry, meanwhile, demonstrates ruthless efficiency, but only when profit potential exists. Consider the market’s response to CRISPR gene editing technology in the early 2010s. Billions in capital flowed into applications within months, numerous biotech startups launched, and established companies quickly pivoted. Traditional gene therapy approaches that had consumed research funding for decades were rapidly deprioritized. While this efficiency drives rapid progress in commercially viable areas, the requirement for profitability creates massive blind spots. Fundamental research in mathematics, theoretical physics, and computational complexity enable enormous technological progress decades later. However they struggle to attract private capital because their payoffs are too distant and diffuse. Research on solutions for problems like drug addiction (substance use disorder) receives minimal private investment despite devastating societal impact, as those affected often lack resources to make such research profitable. Industry funding also tends to favor proprietary solutions over collaborative advancement of knowledge. Critical research areas can remain underfunded despite their importance, leaving the vulnerable behind as their problems lack viable business models.

What’s needed is a third path—a funding mechanism that combines academia’s ability to tackle big societal problems with industry’s ruthless efficiency. We need a system that aligns incentives for all stakeholders: researchers pursuing breakthroughs, funders seeking efficient solutions, and a public desiring tangible progress on important problems. This is precisely what Breakthrough Incentive Markets (BIMs) aim to accomplish.

The mechanics of BIMs

Background on prediction markets

Before diving into BIM’s, let’s understand the underlying mechanism that makes them powerful: prediction markets.

Prediction markets are financial markets where participants buy and sell contracts that pay out based on the outcome of future events. Examples include predicting election results, sports outcomes, or economic indicators. These markets have repeatedly demonstrated remarkable accuracy for three key reasons:

1. They aggregate diverse knowledge from many participants

2. They strongly incentivize accurate predictions (put your money where your mouth is)

3. They reward early correct predictions more than late ones

BIMs take this concept and apply it to a different problem: funding scientific breakthroughs. Instead of merely predicting outcomes, participants actively work to create them. While traditional prediction markets ask “What will happen?”, BIMs ask “How can we make this happen?”

BIMs are focused on incentivizing maximally efficient investment in research. Investors typically care about annualized return-on-investment above all else. For a BIM, final payout and initial investment are both fixed rates for any single investment. What is not fixed is the number of years before the outcome of the Research Outcome Market is reached and investors receive their payout. The shorter the time until the problem is solved, the higher the annualized return-on-investment. This means that investors can increase the value of their investment by finding and funding promising research that has a high probability of solving the problem as quickly and cheaply as possible. (More on the specific math at play here in appendix 1).

How a BIM Works

Here is how a research outcome market would work step-by-step:

1. Clearly Define the Problem: A clearly defined outcome with measurable success criteria (e.g., “A FDA-approved treatment that reduces Alzheimer’s cognitive decline by at least 40%”). You would also likely appoint a neutral party to determine when the criteria have been met.

2. Create the Outcome Pool: Establish a financial pool that pays out when the problem is solved. This pool is funded through charitable donations from individuals, communities, patient groups, and anyone who cares about solving the problem. Critically, this allows the public—not just institutions—to directly set research priorities through their donations.

3. Open the Market: Allow “Solution Investors” to buy positions in the market (with some minimum buy-in requirement). This ensures that solution investors have the means to fund potential solutions.

4. Fund Research: Solution Investors then fund research approaches they believe will increase the probability of the problem being solved—whether by their team or anyone else. Their financial incentive is to fund the projects that raise the probability of the outcome happening the most for the least amount of money.

5. Payout Upon Success: When the problem is solved (as verified by predetermined criteria), all investors who took positions are paid based on the size of the outcome pool when they entered the market, the size of other investments, and their investment. Investors receive a payout whether or not their specific research approach succeeded. (More on the specific math here in appendix 1).

I discuss the specific legal mechanisms under which a Research Outcome Market could be set up in greater detail in appendix 2. The quick version is that any solution investors would almost always be qualified purchasers under the Securities Act meaning that certain laws that normally apply to markets like this would not apply. This means that there are paths to set this up legally that are not available to more standard prediction markets.

Letting Everyone Do What They Do Best

BIMs create a division of labor between three groups where each group focuses solely on what they excel at:

• The Public (Contributors): Non-experts are good at identifying the high-level problems that need solving. A parent of a child with pancreatic cancer doesn’t need a PhD to know that finding a treatment matters. This system lets the public do what it does best: define priorities through their charitable contributions.

• Markets and Investors: Markets excel at evaluating probabilities and allocating resources efficiently. When an approach shows promise, more capital flows to it; when an approach falters, funding redirects elsewhere. With this system, markets can do what they do best: evaluate which solutions have the highest probability of success per dollar spent.

• Researchers: Scientists and innovators thrive when they can focus on deeply understanding problems and developing creative solutions. This system lets researchers do what they do best: innovate with minimal distraction from grant-writing, politicking, or conforming to institutional preferences.

BIMs provide a helpful alternative to traditional research funding mechanisms

BIMs create a third complementary funding mechanism alongside academic and industrial funding. BIMs create financial incentives tied to solving problems themselves rather than monetizing solutions. This allows investors to capture value from public goods with distant or diffuse payoffs that traditional markets cannot accommodate. It also allows for high-risk high-reward research to take place that academic funding systems might consider too unorthodox.

BIMs incentivize speed and efficiency. For most scientific challenges, a solution will eventually emerge—what investors in these markets care about is accelerating that timeline to maximize their annualized returns. This rewards innovation over methodological orthodoxy. This approach also deincentivizes the bureaucratic overhead of traditional academic research funding. It is probable that grant applications from investors in these BIMs would be similar to startup funding with brief initial applications followed by deeper vetting for promising ideas. While conventional academic grant systems might make small funding applications not worth the bureaucratic effort, these more streamlined applications make cheap, short projects that slightly accelerate timelines viable.

BIMs require researchers to make quantifiable claims about how their work will compress timelines to solutions relative to funding requested. This replaces vague significance statements with explicit accountability for results. Individual success, for a researcher, becomes defined by solving real-world problems that affect large amounts of people as quickly as possible.

While industry funding neglects valuable non-profitable areas, BIMs enable investment in solving important problems regardless of commercial potential. Since investors profit when ANY solution succeeds, the system encourages information sharing and collaboration. This incentive for collaboration also creates an incentive for more fundamental research. If one group does some fundamental research that allows another group to come up with a solution to the problem more quickly, the funders of the first group still benefit from a better return on investment. Finally, demonstrations of progress compress expected timelines as the market goes on, attracting more investment precisely when scaling becomes necessary.

The public directly influences which problems get solved through their contributions to outcome pools, adding democratic input to research priorities rather than leaving these decisions exclusively to institutional gatekeepers or profit motivations. BIMs don’t replace academic or industrial research—they provide a vital third option that channels resources toward radical innovations in neglected yet crucial areas.

The Third Model: An Aspirational Narrative

So far things have remained very conceptual. I wanted to write a very short short story to show a “potential history” of what a world with BIMs might look like. The following scenario explores how Breakthrough Incentive Markets might operate alongside traditional scientific funding mechanisms to address antimicrobial resistance. I am not an expert in antibiotics, this is just my best attempt at writing about them based on my current understanding.

In early 2026, the CDC's biennial report on antimicrobial resistance documented an acceleration of treatment failures across multiple bacterial pathogens, with untreatable gram-negative infections emerging in hospitals nationwide. Traditional responses followed familiar patterns—the NIH allocated $120 million over five years for basic research, while pharmaceutical investment remained minimal, focusing only on modifications to existing drug classes.

Against this backdrop, the Antimicrobial Breakthrough Trust launched with a clear objective: development of a novel antimicrobial effective against pan-resistant gram-negative bacteria. The initial outcome pool totaled $115 million, funded by hospital systems, public health foundations, and private citizens directly affected by treatment failures.

Dr. Maya Chen at Northwestern University had spent eight years studying bacterial quorum sensing mechanisms through NIH grants. Her approach showed promise in laboratory studies but faced translational hurdles. Her recent grant application for advancing this work toward development had been rejected, with reviewers requesting additional mechanistic studies.

In early 2027, after learning about the BIM, Dr. Chen received $3.7 million from investors who had taken positions in the market. Her funding application was unlike any one she had filled out before. She started with sending a 500 word summary of her research idea in, along with a mathematical estimate on how much this would decrease the time to solution. After several progressively smaller interview rounds, with investors, and consultants that were brought in by investors to tell sound science from bad, Dr. Chen recieved the award.

By 2029, Dr. Chen's research operated in three distinct but interconnected streams. Her NIH-funded academic laboratory continued investigating fundamental cellular biology. Her pharmaceutical consulting work helped improve existing antibiotics through targeted modifications, prioritizing predictable outcomes with clear commercial potential. The BIM-funded program operated under different constraints—with no pressure to publish or build a commercial product line, the team focused exclusively on creating a functional therapeutic as rapidly as possible.

The BIM created unique collaboration dynamics. When Dr. Chen's team encountered delivery challenges, investors introduced her to Dr. Wei Lin's group at UC San Diego, which had developed targeted bacteriophage technology under separate BIM funding. The two teams recognized complementarity between their approaches.

Where academic collaborations typically formed around publication opportunities and industry partnerships around commercial potential, this relationship formed around the mathematical calculation that combining approaches might significantly accelerate the timeline to a viable antimicrobial. The investors supported information sharing once they were persuaded that their financial returns improved if the breakthrough occurred sooner.

By 2031, the work evolved into a hybrid approach combining engineered bacteriophages as delivery vehicles for quorum-sensing disruptors, protected by specialized nanoparticle technology. The approach allowed precise targeting of pathogenic bacteria while leaving beneficial microbiome members unaffected.

As the technology progressed, boundaries between funding streams blurred. Fundamental questions generated publications from Dr. Chen's academic laboratory. Some delivery system components attracted pharmaceutical attention for broader applications. The BIM funding continued supporting the core therapeutic development that remained too risky for traditional investment.

In late 2033, the combined approach received FDA approval for treatment of previously untreatable infections. The verification committee confirmed the breakthrough criteria had been met, triggering distribution of the outcome pool—which had grown to $380 million—to all investors based on their timing and investment level.

Early investors received approximately 24% annualized returns—competitive though not extraordinary by venture capital standards, but remarkable given the public health impact and previous lack of financial incentives in this area. More importantly researchers around the world had successfully coordinated to make progress on a very large problem with potentially devastating impact.

Appendix 1: Mathematical Framework behind the incentives in a BIM

Key Incentive Mechanisms

Basic Parameters

For any Research Outcome Market, we define:

Share Pricing and Payout Calculations

When an investor contributes Iᵢ at time tᵢ​, they purchase shares that entitle them to a fixed portion of the payout pool as it existed at that time:

Share price: The price per share is set by the fund based on current pool size and investment levels. Fund managers are given significant amounts of shares at the beginning, that only vest when the outcome is reached, to align incentives.

Shares received:

Payout share:

Final payout:

Annualized return:

This structure provides investors with certainty about their final payout amount (Pᵢ) at the time of investment. The only unknown variable affecting their return is when the problem will be solved (T).

Incentive 1: Speed and Efficiency

The annualized return formula shows why investors are incentivized to minimize time-to-solution:

As T decreases, the exponent (1/(T−tᵢ)) increases, dramatically increasing returns.

Example: Accelerating Research Timeline

Consider two potential research investments:

• Investment A: $2M with expected solution in 10 years

• Investment B: $3M with expected solution in 7 years

With a payout pool of $20M:

Return for A:

Return for B:

Despite the higher investment amount, Investment B yields higher returns because it compresses the timeline more efficiently. This creates a natural selection mechanism for approaches that solve problems fastest, not just cheapest.

Incentive 2: Viability of Small Projects

Small, promising projects become viable when they can measurably compress timelines.

Example: Small but Effective Investment

Imagine a research market with:

• Current payout pool: $10M

• Total investments: $1M

• Expected solution timeline: 8 years

A researcher proposes a small $50K project that could accelerate the timeline by 6 months.

Without this project:

With this project:

The return increases despite the additional investment, making this small project mathematically worthwhile. Traditional grant systems might reject this project as too small, but Research Outcome Markets naturally accommodate any size project that efficiently compresses timelines.

When Is It Worth Investing?

For an investor with a minimum acceptable return rminrmin​, an investment is worthwhile when:

Where:

• Inew​ is the proposed new investment

• Tnew​ is the expected time to solution after the investment

Solving for the minimum timeline compression required:

This formula directly answers when an investment is worthwhile: when it compresses the expected solution timeline sufficiently relative to its cost.

Investment Decision Example

An investor considering a $500K investment with:

• Current pool: $15M

• Current total investment: $3M

• Current expected timeline: 9 years

• Minimum acceptable return: 20%

For this investment to be worthwhile, it must compress the timeline to:

So if this $500K investment can accelerate the timeline by at least 0.95 years (approximately 11.4 months), it meets the investor’s minimum return requirement.

Incentive 3: Information Sharing and Collaboration

Since investors profit when ANY solution succeeds, not just their own specific approach, they’re incentivized to share information.

Example: Collaboration Benefit

Consider two competing research teams, each with promising partial solutions:

• Team A has approach with 30% chance of success in 6 years

• Team B has approach with 25% chance of success in 7 years

If they don’t collaborate:

If they share information and combine approaches:

• New combined approach: 60% chance of success in 5 years

• Expected solution time:

This timeline compression from ~10 to ~7 years dramatically increases returns for all investors:

This creates a powerful mathematical incentive for information sharing, even among competitors.

Incentive 4: Demonstrating Progress Attracts More Funding

As research shows promise, the payout pool tends to grow, benefiting early investors.

Example: Progress Demonstration Effect

Initial market conditions:

• Payout pool at t=0: $5M

• Initial investment: $500K

• Expected timeline: 10 years

After 2 years, the research team demonstrates significant progress, causing:

• Payout pool grows to $15M

• New investors contribute additional $2M

• Timeline estimate improves to 6 more years (8 total)

For early investors:

This creates strong incentives for transparency and public demonstration of progress, as it attracts more contributions to the payout pool and compresses timelines, both benefiting early investors.

Appendix 2: Legal Implementation Through Breakthrough Incentive Trusts

Note: I am not a lawyer, I have just done a very large amount of research on my own. It is very possible I am misunderstanding several things but this is my best attempt to outline a legal structure for BIMs'. This does not constitute legal advice. Any implementation of these concepts would require consultation with qualified legal counsel.

The Trust Structure

BIMs could be funded through a Breakthrough Incentive Trust (BIT). For each BIM, a separate irrevocable trust would be established under South Dakota law.

The trust instrument would precisely define the scientific breakthrough being incentivized with clear, measurable, and objective success criteria. An independent committee of qualified experts would be chosen to verify achievement of the outcome according to these predetermined criteria.

The BIT uses a dual-pool financial arrangement. The Charitable Outcome Pool consists of donated funds committed to rewarding achievement of the breakthrough, while the Investment Pool comprises positions taken by qualified purchasers who seek to profit from the eventual breakthrough. These pools remain strictly segregated within the trust structure.

To distinguish this structure from wagering, qualified purchasers would be required to fund some amount of relevant research. This requirement would be implemented with substantial flexibility regarding both amount and specific research approaches. A small annual fee, typically 0.5-1% of the Investment Pool, would fund ongoing operations, verification processes, and administration.

Why South Dakota?

South Dakota offers several advantages that make it the optimal jurisdiction for BITs. Unlike most states, South Dakota permits truly perpetual trusts, which is critical for long-term scientific challenges that might require sustained incentives over multiple generations of researchers.

South Dakota law also allows for separation of trustee functions through its directed trust provisions. The state also provides stronger privacy protections than most jurisdictions, absence of a state income tax, and robust shields against potential creditor claims.

Most importantly, South Dakota has developed a sophisticated trust services industry capable of administering complex arrangements.

The Single-Purpose 501(c)(3)

For tax-deductible charitable contributions, a single-purpose 501(c)(3) scientific research organization would work alongside the trust. This organization would be formed with articles of incorporation explicitly defining its exclusive purpose as advancing a specific scientific breakthrough by providing outcome-based incentive funding. Its bylaws would contain detailed provisions regarding the organization's commitment to transfer funds to the trust upon verification of the breakthrough.

A scientific advisory board comprising recognized experts in the relevant field would validate the breakthrough's significance and public value for 501(c)(3) requirement purposes. Most critically, a binding legal agreement between the 501(c)(3) and the trust would specify the conditions and procedures for fund transfer upon breakthrough verification.

This structure is defensible under existing IRS guidance for scientific research organizations, particularly Revenue Ruling 76-296. This ruling establishes that scientific research organizations qualify for 501(c)(3) status when their research benefits the public rather than private interests, makes results available on a nondiscriminatory basis, and advances scientific knowledge. The X Prize Foundation, disease-specific research foundations, and scientific prize-granting institutions already recognized as 501(c)(3) entities could be used as precedence.

To ensure reliable fund flow to the trust upon breakthrough verification, several safeguards would be implemented. All charitable contributions would be accepted as legally restricted funds designated exclusively for the breakthrough incentive, creating legal obligations to use the funds only for their designated purpose. The organization would execute binding legal agreements obligating fund transfer upon verification.

Governance controls would include supermajority requirements for any purpose modification and representation from the trust on the organization's board.

Qualified Purchaser Limitation

Participation in the Investment Pool would be limited to qualified purchasers as defined under Section 2(a)(51) of the Investment Company Act. This generally includes individuals with at least $5 million in investments or entities with at least $25 million in investments.

The most significant benefit is the exemption provided under Section 3(c)(7) of the Investment Company Act, which substantially reduces regulatory requirements that would otherwise apply. Qualified purchasers are legally presumed to possess sufficient financial sophistication to evaluate complex investments, reducing concerns about investor protection that might otherwise arise with novel investment structures. Arrangements limited to qualified purchasers face fewer regulatory hurdles due to this presumed sophistication, enabling greater flexibility in structure and operations.

Operational Mechanics

The operational flow of a Breakthrough Incentive Trust would begin with the establishment of both the trust and the companion 501(c)(3) with appropriate governance structures. The 501(c)(3) would then accept tax-deductible contributions designated for the breakthrough incentive, building the Charitable Outcome Pool over time.

Qualified purchasers would take positions in the market through contributions to the Investment Pool, establishing their proportional claim on the eventual outcome payment. These investors would independently fund research they believe will advance the breakthrough, with complete discretion regarding approaches and amounts.

When achievement of the breakthrough is claimed, the independent verification committee would determine whether the predetermined criteria have been met, applying objective standards established in the trust instrument. Upon verification, the 501(c)(3) would transfer funds to the trust pursuant to its legal obligations, and the trustee would distribute to investors according to predetermined formulas based on their proportional stake and timing of investment.

Other paths

In my research I found a few other potential legal mechanisms to realize a BIM, although I didn’t end up researching them deeply when I found out about the trust mechanism. Among these would be a modified version of social impact bonds. Another option would be to use a series LLC with more conventional ownership shares as opposed to a trust.

I have developed an AI-powered cognitive co-pilot to accelerate the speed at which neuroscience researchers develop robust theories.

The human comes up with a theory. The AI system comes up with falsifiable hypotheses in natural language. The AI then codes up some experiments it can perform on a computational dataset based on the hypotheses. Finally the AI provides the results of that experiment to a human in a way that they can interpret and use to refine their theory.

Why does this matter?

According to this study drugs that target the central nervous system only succeed half as often compared to drugs that target other systems. This is a wider pattern across clinical science, clinical interventions that target the brain are much less likely to succeed. The authors of this paper (and me) make the point that one of the primary causes for the failure rates of clinical brain science is the fact that our hypotheses about the brain aren’t very robust or high-quality compared to our hypotheses about other systems in the body.

How well does this system actually work?

It works really well! In early tests computational neuroscientists are able to do theory refinement that would take 4 hours using standard methods in 25 minutes with assistance from the cognitive co-pilot.

How does the system actually work?

The cognitive co-pilot is actually many smaller and dumber AI’s in a trenchcoat. The architecture is based around “mini-experts.” A mini-expert is an AI that uses RAG and knowledge distillation to have expertise on a very specific methodology and how you implement it and interpret results (usually a mini-expert has around 15 papers as its’ background information). This architecture allows the mini-experts to collaborate to form many hypothesis and many experiments for a single theory. The architecture is also highly modular, meaninng that it will be easy for computational neuroscientists to program their own “mini-expert” for doing experiments that use a specific method.

What about safety/explainability?

I am using a framework that comes from Google Deepmind called MONA (Myopic Optimization with Non-myopic Approval). It means that a human is supervising the process instead of just the final outcome. For my system this means that at every single research step a human expert views what has been generated and either validates that the AI is doing things properly, or can select that the system might not be doing things correctly in some way. Allowing the human to act as a "checker” as opposed to the one generating the hypotheses and experimental code is what contributes to the speed up we are seeing from early results.

Key Terms

Here are some terms we used that might be confusing to a newcomer and an explanation of what they mean:

Auto-regressive Generation

Wikipedia: Autoregressive Model

This is a type of statistical process, where each new piece builds upon what came before it. The system takes its own output and feeds it back as input for the next step, creating a continuous flow of generation. Modern AI systems use this approach to create coherent text one piece at a time, similar to how humans construct sentences word by word in natural conversation.

Large Language Models (LLMs)

Wikipedia: Large Language Models

These serve as the backbone of modern artificial intelligence applications. They process and generate human language by learning patterns from vast amounts of text data. ChatGPT represents one of the most well-known examples of these systems in action.

Token

Wikipedia: Text Segmentation

Tokens function as the fundamental building blocks of text in AI systems. These discrete units might represent complete words, parts of words, or even punctuation marks. The sentence "I love AI!" breaks down into four distinct tokens: "I," "love," "AI," and "!" The AI processes these tokens sequentially to understand and generate text.

Hallucinations (in AI)

Wikipedia: Hallucination (artificial intelligence)

AI hallucinations occur when systems generate convincing but false information. These fabrications emerge when the AI fills gaps in its knowledge with plausible-sounding but incorrect data.

Transformer Models

Wikipedia: Transformer (machine learning model)

This is an AI architecture which excel at processing sequential data by understanding relationships between different elements. The system weighs the importance of various parts of the input simultaneously, rather than processing them in strict order. This architecture enables large language models to grasp complex language patterns and generate coherent responses.

Hyperdimensional Space

Wikipedia: High-dimensional space

While we live in a world with three physical dimensions (height, width, and depth), AI systems use hundreds or thousands of "dimensions" to organize information. Each dimension represents a different characteristic or feature. A children's library might organize books by just three features: reading level, subject, and age group. AI systems organize information using thousands of features simultaneously. When describing a dog, these features might include size, color, fluffiness, behavior, typical locations, related objects, and hundreds of other characteristics. This rich organizational system helps AI understand subtle differences and similarities between concepts. Two dogs might be similar in some ways (four legs, furry) but different in others (size, color), and the hyperdimensional space captures all these relationships.

Knowledge Graph

Wikipedia: Knowledge Graph

Knowledge graphs create structured networks of information by connecting related concepts, facts, and entities. These connections form a web of knowledge that mirrors human understanding of relationships between different pieces of information.

Retrieval-Augmented Generation (RAG)

Wikipedia: Prompt engineering

This hybrid approach combines the creative abilities of AI with factual information retrieval. The system accesses external databases to verify and supplement its responses, leading to more accurate and reliable output. Modern chatbots use RAG to provide answers grounded in verified sources rather than relying solely on their training data.

Connectome

Wikipedia: Connectome

A connectome provides a comprehensive map of neural connections within a brain. This intricate diagram reveals how different brain regions communicate and work together. Scientists use connectomes to understand brain function.

Working Memory vs. Long-Term Memory

Wikipedia: Working Memory
Wikipedia: Long-term Memory

Working memory and long-term memory are both theoretical ideas that are fundamental to cognition research. Working memory acts as a temporary mental workspace for immediate tasks, while long-term memory stores information for future retrieval. Cognitive scientists theorize that these two systems work together seamlessly in human cognition. Working memory holds the ingredients while cooking a new recipe, while long-term memory stores cooking techniques learned over years of experience.

Important Timestamps

2:03 Dr. Barenholtz’ main thesis: Could all human thinking and reasoning be similar to how AI language models work - taking in information and generating the next step?

13:54 Clear explanation of what "autoregressive" means

27:42 Discussion about how humans think in images and video, similar to how AI image generation works step by step

41:15 Debate about whether the brain is one unified autoregressive system or an autoregressive and retrieval system working together (Daniel’s view)

52:54 Important differences between how human brains and AI store and access knowledge

1:08:11 The potential value of theoretical brain science, even if immediate medical benefits aren't clear

1:28:16 How theoretical brain research eventually leads to medical treatments through multiple stages of testing

1:34:02 Recent exciting developments in computational brain research that make this a promising time for neuroscience

Long-context models (e.g. Gemini), designed to process extensive amounts of text within a single context window, face a critical bottleneck in the form of training data scarcity. As context lengths increase, the availability of high-quality training data diminishes rapidly. This is important because of the neural scaling laws, which have been remarkably robust for LLMs so far. Here is a great video explaining those:

One important implication is that if you run out of human-generated training data, the reasoning capabilities of your model are bottle-necked no matter how many other resources or tricks you throw at the problem. This paper provides some nice empirical support for this idea. Across all of the "long-context" models the reasoning capabilities decrease dramatically as the context length increases.

Embeddings based RAG has much better scalability but suffers from some pretty serious issues with high-level reasoning tasks. Here is a small list from this paper:

The authors not only point out many of the issues but also have a nice statement as to the core reason why towards the beginning of the paper:

1) Reasoning Failures: LLMs struggle to accurately interpret user queries and leverage contextual information, resulting in a misalignment between retrieved knowledge and query intent.

2) Structural Limitations: These failures primarily arise from insufficient attention to the structure of knowledge sources, such as knowledge graphs, and the use of inappropriate evaluation metrics.

This structural limitation is particularly problematic when dealing with documents that require deep understanding and contextual interpretation such as textbooks or legal documents. Often there will not only be an important internal structure to each document, but also an important meta-structure across documents (think of scientific papers that cite specific portions of other scientific papers). There are tricks like using knowledge graphs that try to get around some of these issues, but they can only do so much when the fundamental method shreds any structure the documents might have had before any of the secondary steps even begin.

The scalability limitations of long-context, and the reasoning limitations of embedding, lead to an important trade-off for anyone building a RAG system. Long-context models excel in creativity and complex reasoning but are limited to small document sets due to training data constraints. Conversely, embeddings-based approaches can handle vast corpuses but function more like enhanced search engines with minimal reasoning abilities. For many tasks, this trade-off is fine as the task already fits well on one side or the other of the trade-off. Many other tasks however, are simply not easily achievable with SoTA RAG methods due to the fact that they require both large amounts of documents and advanced reasoning over these documents. Solving this trade-off and creating a solution that can handle complex reasoning AND large amounts of documents, is going to be the next big step in RAG. It is a problem I have made some progress in solving personally and I would love to see more folks take a crack at.

This is an interview I did a while back and one of my favorite things that I’ve ever produced. I got to talk to Dr. Alex Ellery about some practical plans for extremely futuristic things. Unfortunately the podcast that I did the interview for has been on a long-term hiatus but I hope to pick it back up again at some point.