The Research Landscape: Most universities and companies freely share their AI discoveries. While some companies like OpenAI have become more private lately, most AI researchers still share their work openly.

The Different Parts of a Research Paper:

Abstract: This concise overview serves as your first filter, helping determine if the paper aligns with your interests.

Introduction: Here you'll find the core problems being addressed, proposed solutions, and key contributions. This section often provides the clearest picture of the paper's significance.

Methods: Focus on identifying novel concepts and approaches, even if you don't grasp all technical details initially.

Related Works: These sections illuminate how researchers have historically approached problems and highlight emerging trends. Recent papers are particularly valuable for understanding current directions.

Experiments: Concentrate on results that relate directly to your interests and business implications.

Where to Find Papers:

arXiv: Think of this as the grand library of AI research. Almost every important AI paper ends up here.

Papers with Code: This is where you find papers that come with working examples – like recipes that include video demonstrations.

Major Conferences: Major conferences like NeurIPS, ICML, and ICLR showcase cutting-edge research.

How Research Spreads: Many researchers now announce their work on Twitter/X first. Following these researchers gives you an early peek at what's cooking.

Smart Reading Approaches:

Get the Big Picture: First, read the abstracts of several papers about your topic of interest and take notes. It's like learning about Italian cuisine by reading different Italian cookbooks – you start seeing common themes and approaches.

Dive Deeper: Then pick the most relevant papers and study them carefully. Go through their introduction, methods, related works, and experiments. Don't worry if you only grasp 10-20% at first – that's normal and expected.

Be Selective: You don't need to understand every detail. Focus on the parts that matter most to your business.

Look at the Pictures: Charts and diagrams often explain ideas better than words – like how a good cookbook uses photos to show technique.

Keep a Learning List: Write down terms and concepts you don't understand to learn about later.

Join the Conversation: Find others interested in the same topics. It's like joining a book club – discussing papers helps everyone understand them better.

Making It Routine: How could you set aside regular time to read research papers, just as you might schedule time to read industry news?

Your Interests: Which parts of AI development could most help or hurt your business? Let this guide what you read.

Learning Together: Could you start a small group at work to discuss AI papers? Sometimes the best insights come from sharing perspectives.

Taking Action: How will you turn what you learn from these papers into business decisions that matter?

Core Concepts:

Environment: The space which an agent operate, whether physical (robots, drones) or digital (trading algorithms, game AI).

Policy: Internal rules and learned behaviors that guide the agent’s decisions.

Reward: Feedback (positive or negative) received by the agent, which informs its future actions.

Autonomy: An agent’s ability to operate independently using sensors and actuators.

Perception: An agent’s ability to gather data from their environments.

Historical Foundation: The journey toward agentic AI began with Norbert Wiener's cybernetics work in the 1940s, introducing self-regulating systems that could evolve through environmental interaction. This laid the groundwork for modern open-ended AI systems.

Open-endedness in AI: Agents possess the ability to continually generate new and unpredictable behaviors, solutions, or outcomes without predefined limits. Unlike LLMs that operate like vending machines, agents work more like personal chefs - planning ahead, maintaining context, and adapting to changing conditions.

Real-World Examples:

NVIDIA's Voyager demonstrating autonomous learning in Minecraft

Sakana’s AI Scientist's automated research capabilities

Google's Gemini Deep Research Agent

Windsurf IDE Agent

Agent Development Considerations:

Agent-computer interface: How should your agent communicate with computer systems? Consider whether your agent should use precise programming commands or more flexible text formats to interact with other software. This choice impacts both development complexity and system reliability.

Human-agent interface: What should users see while the agent works? Determine which information helps users understand and trust what the agent is doing, similar to how a progress bar shows you the status of a download.

LLM selection: Which AI model best follows instructions for your specific needs? Different models have varying abilities to understand and execute commands accurately, making this choice crucial for your agent's effectiveness.

Tool use: What tools does your agent need access to? Just as a mechanic needs specific tools to fix a car, your agent needs the right digital tools to complete its tasks effectively.

Environment understanding: Does your agent understand its working environment? Consider whether it needs to gather more information before taking action, similar to how a contractor surveys a site before beginning construction.

Error awareness: Can your agent recognize and fix its mistakes? This capability is crucial for reliability and trust, much like how we value employees who can identify and correct their own errors.

Planning: How does your agent approach planning? Evaluate whether it can break down complex tasks into manageable steps and create effective execution strategies.

Task breakdown: Can your agent break down complex tasks into clear, manageable steps? Just as an effective project manager knows how to divide big projects into achievable milestones, your agent needs to understand how to sequence work logically.

Exploration and search: Can your agent explore different approaches? Assess whether it can consider multiple solutions and choose the most effective one, rather than just taking the first available path.

Evaluation methods: How do you verify your agent is meeting its goals? Establish clear methods to measure whether the agent is delivering the intended results and value for your organization.

Defining ETA: ETA is when you focus on acquiring and operating an existing, cash flowing business rather than launching a new venture from 0 to 1.

Historical Context: The concept of the "search fund" formalized in 1984 by Harvard Business School's Irv Grousbeck, offering a structured pathway for MBA graduates to identify, acquire, manage, and grow a privately held company.

There Are Many Models

Traditional Search Fund: This model involves raising capital from investors across two phases: initial funding for the search process and subsequent capital for the acquisition itself.

Self-Funded Search: You’ll utilize your personal capital to finance the search for an acquisition target. This approach gives you control, freedom, and equity ownership but requires a higher degree of personal financial risk tolerance. Acquisition financing may still involve external debt or seller financing.

Sponsored Search: You’ll partner with a single investment firm, typically a family office, which provides all necessary capital for both the search and the acquisition.

Incubated Search: A relatively newer model where you’ll join an established incubator platform specializing in search fund investments.

The ETA Process

Phase 1: Search Fund Formation (2-4 months)

For traditional models, this involves investor outreach, developing a compelling investment thesis, and securing initial capital commitments.

Key Activities and Deliverables:

Create Private Placement Memorandum (PPM) and investment thesis

Network and pitch to potential investors

Secure investor commitments

Complete legal fund formation

Phase 2: Opportunity Sourcing and Evaluation (1-24 months)

A systematic process of identifying potential acquisition targets, conducting due diligence, and assessing their financial and operational viability.

Key Activities and Deliverables:

Build and execute systematic outreach strategy

Screen opportunities against investment criteria

Conduct preliminary due diligence

Build relationships with business owners

Phase 3: Transaction Financing and Closing (2-6 months)

Securing the necessary financing for the acquisition, negotiating final terms, and completing the legal and administrative processes to finalize the deal.

Key Activities and Deliverables:

Negotiate and execute Letter of Intent (LOI)

Complete comprehensive due diligence

Secure transaction financing

Execute purchase agreement and closing

Phase 4: Business Operations (4-8 years)

Assuming leadership and operational control of the acquired business, focusing on growth, efficiency improvements, and value creation.

Key Activities and Deliverables:

Execute owner transition and 100-day plan

Build/strengthen management team

Implement key strategic initiatives

Establish effective board governance

Phase 5: Exit Strategy (4-6 months)

Executing a sale of the business to realize a return on investment for both the entrepreneur and investors (if applicable).

Determine optimal exit timing and strategy

Prepare business for sale

Engage with potential buyers

Execute final transaction

How much risk are you comfortable with?

To what extent are you willing and able to deploy your own capital in pursuing an acquisition?

What level of ownership is your target?

How much day-to-day involvement do you want in the business you are seeking?

What level of external support, mentorship, and infrastructure do you perceive as necessary for your success in this endeavor?

Waymo: They are building the future of autonomous transportation. With over 20 million miles of real-world driving experience under its belt and clocking over 1 million fully autonomous miles weekly, Waymo is rapidly expanding its robotaxi service, Waymo One. They've already provided over 2 million paid rider-only trips, currently completing over 150,000 paid rides weekly. Waymo One is currently live in four major U.S. cities with plans to expand to two more. The robotaxi market is estimated to be a $100 billion+ playing field, and Waymo is poised to be a major contender.

Gemini: Think of Gemini as Google’s AI Swiss Army knife - a multimodal model that can understand, reason, and generate text, code, audio, and more. It's already in the hands of over 1.5 million developers, integrated across Google’s developer tools and boasting around 42 million active users. Their recent Gemini 2.0 release is at the top of LLM leaderboards. It’s the backbone of various Google products, like Google Cloud, Google Workspace, and Android. The generative AI market, where Gemini is positioned, is projected to explode to $1.59 trillion by 2030.

Isomorphic Labs: They are rewriting the rules of drug discovery. By leveraging AI, they're aiming to accelerate the development of new medicines across all disease areas. With collaborations with major pharmaceutical players like Eli Lilly and Novartis, potentially worth $3 billion, Isomorphic Labs is well-positioned to disrupt the $1.4 trillion pharmaceutical market and redefine how we approach medicine.

NotebookLM: Think of it as an AI co-pilot for your research. It's designed to help users make sense of their own data, acting like a virtual research assistant that can summarize, answer questions, and even generate podcast discussions. This is a $1.4 trillion opportunity by 2027, and NotebookLM is not playing small; it's eyeing the $1.96 trillion content marketing market, the $130.63 billion podcasting market, $348.41 billion education technology market, and the $2.53 trillion knowledge management market.

Financial Muscle: With a market cap of $2.371 trillion, $339.85 billion in total revenue, and a 3-year average revenue growth of 22.10%, Google has serious financial backing to support its AI ambitions. They are spending big too - $48.323 billion (TTM) on R&D - underscoring their commitment to pushing AI forward.

Beyond Search: While search may be facing new competition, Google's strategic moves in areas like autonomous driving, drug discovery, and personalized research showcase their vision extends far beyond a simple search box. Each "bet" has the potential to shake up existing industries.

Rapid Expansion: Look at the numbers: Waymo’s rapid growth, Gemini's broad adoption by developers, and Isomorphic's big partnerships are not just random occurrences. They signal a company executing on a clear, aggressive strategy. NotebookLM’s early positive user reviews also hint at a potential market disruption.

Balancing Offense and Defense: How likely is Google’s calculated approach to pay off in the long run, or will it allow more focused, bullish competitors to overtake them?

Brand Identity in the AI Era: As Google's AI ventures like Waymo, Isomorphic Labs, and others gain prominence, how will their brands impact Google's sense of identity? Will these successes reinforce the Google brand, or will they overshadow it and create a fragmented brand perception?

Market Expectations vs. Reality: To what extent are these bets already factored into the current stock price? Could the market be overvaluing these ventures, or are they still underappreciated by investors?

Latent vs. Language Space: Traditional LLMs use Chain of Thought to "think out loud" using language. Coconut reasons internally, manipulating abstract representations—like internal "thoughts"—in its latent space.

Chain of Thought (CoT): Think of CoT as training wheels for LLMs. It involves showing the model a series of explicit steps to arrive at an answer. Coconut takes off the training wheels, allowing the model to reason more independently.

Embeddings: These are like digital fingerprints for words or concepts, capturing their meaning in a dense numerical form. In Coconut, the model's internal state acts as the embedding for the next step in its reasoning process.

How Humans Think Through Language: Interestingly, Coconut mirrors how humans learn. We start by verbalizing our thoughts, but eventually, we internalize them into abstract concepts. Coconut is essentially doing the same thing.

Staged Training: Coconut learns gradually. It starts with traditional CoT examples and then progressively shifts to latent space reasoning, replacing verbal steps with internal "thoughts."

Hidden State Feedback: Instead of generating words, Coconut feeds its internal state back into itself, creating a feedback loop that drives its reasoning.

Special Tokens: Special markers, like (beginning of thought) and (end of thought), help the model distinguish between internal reasoning and external language generation.

Loss Masking: During training, the model focuses on getting the final answer right, not on verbalizing the intermediate steps.

Enhanced Reasoning: By reasoning in the latent space, the model can explore more possibilities and arrive at better solutions.

Increased Efficiency: Coconut requires less computing power and generates fewer tokens, making it faster and cheaper.

Explorative Search: Unlike CoT, which follows a linear path, Coconut can explore multiple avenues simultaneously, like a breadth-first search.

Test-Time Compute: Can we combine Coconut with approaches like o1 and DeepSeek to further enhance performance? What would the impact be?

The Nature of Thought: If LLMs can reason effectively without language, does that challenge our understanding of what "thinking" actually is? Are we overemphasizing language in our own cognitive models?

The Future of Reasoning: What will human reasoning look like when everyone has access to superhuman AI reasoning tools?

Explainability vs. Performance: As models become more efficient by reasoning in latent space, does this make them even harder to understand? Are we trading explainability for performance? What are the implications for trust and accountability?

New Frontiers: What new applications will emerge from LLMs with significantly improved reasoning abilities?

The Players: You'll primarily be dealing with the acquiring company's CEO or co-founders. Think of them as the ultimate decision-makers. They're supported by their executive team (CTO, CFO, etc.) who handle the operational details. You might also interact with a VC board member, especially if the acquirer is smaller or the deal is particularly significant. Investment bankers usually only enter the picture for very large acquisitions (think $1 billion plus).

The VC's Role: Venture capitalists are like the strategic architects behind the scenes. They help shape the overall acquisition strategy early on, ensuring it aligns with the acquirer's growth plans, and they have a strong voice in the final negotiations, making sure the deal makes sense for their investment.

The Target Profile: Your company doesn't necessarily have to be a high-growth rocket ship. VC-backed acquirers are often looking for companies with a strong product that fits their existing offerings, a talented team they want to bring on board, or a customer base that expands their market reach. They want something that helps them grow strategically.

Payment Preference: Don't expect a big cash payout. VC-backed companies often prefer stock swaps. This is because they usually have limited cash on hand and prioritize reinvesting profits back into growth initiatives. It also aligns the incentives, making you and your team invested in the combined company's future.

The "Pref Stack" and Valuation: This is where things get a bit technical. Each time a company raises funding, it creates a new layer of "preferred stock" with specific rights for those investors. This "pref stack" essentially sets a minimum price tag for any acquisition. It ensures that investors, especially those who came in later, at least get their money back. Understanding this is critical because it establishes the baseline for any offer you receive.

Strategic Fit and Long-Term Vision:

What is their primary objective for this acquisition? How does your company and product fit into their overall strategy and portfolio?

What are their long-term plans for your company and product? What is their vision for the future, and how do they see your company evolving within theirs?

What are their incentives for making this transaction successful, and what are the contingencies if it's not? This reveals their commitment and risk mitigation strategies.

Your Role and Team Integration:

What are their plans for the existing team? Are they looking to retain talent, expand the team, or rationalize roles?

What is the expected role of the founder(s) post-acquisition? Will you be expected to stay on to lead the division, transition out, or something in between?

If staying on, what is the expected duration and what are the conditions regarding non-compete clauses? This clarifies your future options and limitations.

Financial Implications and Exit Strategy:

How do they plan to structure the payment (cash, stock swap, or a mix)? This has significant financial implications for you and your investors.

If it's a stock swap, what is their exit strategy (IPO, acquisition by another company, etc.)? Your payout in a stock deal depends on the acquirer's future success and exit.

Diffusion Transformer Model: Think of it like this — it starts with pure visual noise (like TV static) and gradually refines it into a coherent video. It does this by learning how noise patterns relate to real-world images and videos. Unlike previous methods, it uses a "Transformer" architecture, known for understanding relationships between different parts of data - in this case, frames and elements within a video - to create realistic motion and interactions. This approach has proven to be highly scalable and effective in generating high-quality videos up to 60 seconds long, accurately depicting complex scenes, multiple characters, and even physics.

Multimodal Input: Sora accepts text, images, or video clips as input, offering diverse creative avenues for users to bring their visions to life.

Beyond Generation: Sora can extend existing videos, fill in missing frames, and adapt aspect ratios, making it a versatile tool for various video editing needs.

Disrupting Filmmaking: Tyler Perry halted an $800 million studio expansion after seeing Sora, highlighting its disruptive potential for traditional filmmaking.

Job Displacement: AI like Sora could displace an estimated 204,000 entertainment jobs within three years, raising significant workforce concerns.

Democratizing Video: Sora empowers smaller creators by enabling professional-quality video creation from simple text prompts, leveling the playing field.

Personalized Education: Sora can generate customized educational videos tailored to individual student needs, such as summarizing complex topics in engaging visual formats, potentially transforming learning.

Advertising Revolution: Sora could drastically reduce video advertising production costs, making video marketing accessible to businesses of all sizes.

The Rise of Video Content: Video dominates online engagement, with 3 billion viewers and 89% of consumers wanting more video from brands. The US video ad market is projected to reach $84.61 billion in 2024, reflecting this growing trend.

Competition and Data Acquisition: Google, and likely Meta, are developing similar tools like Sora, raising questions about the vast datasets used for training. Should users be compensated for their data's use in training these models?

Truth and Simulation: The ability to generate realistic video content raises profound questions about truth and reality in the digital age.

What does this mean for Truth across the internet?: As AI-generated videos become more prevalent and indistinguishable from real footage, this blurring of lines has serious implications for news, information dissemination, and societal trust.

What new realities are unlocked when Sora can be used as a simulation?: From training and education to scientific research and entertainment, the potential applications of simulated video environments are vast.