Introduction

At UC Berkeley, the top university for most venture-backed startups and #1 public university globally, artificial intelligence (AI) has become a central focus of discussion and innovation. Students, researchers, professors, founders, departments, and surrounding VCs are eager to engage with AI, which has become the hottest topic on campus. At the time of writing this thesis, I am an undergraduate senior pursuing a degree in Computer Science, with a research focus on human-centered AI and agentic AI. I teach several venture related courses under the College of Engineering’s SCET department and have witnessed firsthand how AI is transforming the workflows of frontier research and innovation at UC Berkeley. My views are shaped by my experience working as a software engineer, venture capitalist, student researcher, and UC Berkeley research and classroom experiences on human-centered AI, agentic AI, engineering ethics, responsible product innovation, venture, and more. Artificial intelligence related research began at UC Berkeley in the 1960s with the first labs focusing on machine learning and natural language processing. With over six decades of work, UC Berkeley has been at the forefront of AI with its labs driving many advancements that have shaped the field. In 2022, the release of OpenAI's ChatGPT marked a transformative milestone: an AI era where AI has become democratized and widely accessible to consumers, fundamentally changing how society interacts with technology. One of my professors shared a compelling example: a computer vision project designed to sort recycling, which took his graduate students years to develop back in the 2000s, was recently replicated by first-year undergraduates in just a matter of weeks using AI and several API calls. This rapid progress highlights AI’s immense potential to accelerate innovation while underscoring a critical challenge: how to keep educational curricula and any innovation that startups create relevant. More importantly, it emphasizes the need to equip everyone not just with today’s tools but with the adaptability and resilience required to thrive in a constantly evolving technological landscape.

This investment thesis explores the transformative potential of AI through the lens of human-centered AI and agentic AI, which represents a paradigm shift beyond traditional machine learning. By defining and contextualizing AI and related innovations within the broader AI market landscape, I will discuss the unique dynamics and opportunities in 2024, its implications for the future of jobs and society, and the ethical considerations surrounding AI’s rapid adoption. I will also challenge some conventional perspectives on AI's trajectory and offer strategic pathways for innovation and investment while navigating this unprecedented era of technological advancement and democratization.

SECTION 1. AI Revolution Explained: Timing, Trends, and Innovation

Why AI innovation is booming, market landscapes, and how investors should evaluate AI startups

1.1 Why AI, why now?

AI innovation is reaching a pivotal moment due to several converging factors:

• Advancements in Core TechnologiesLarge language models (LLMs) and generative AI systems, such as OpenAI’s GPT series, have drastically improved in capability, scaling human productivity across industries. These tools enable faster, more precise decision-making and innovation, driving businesses to adopt AI at an unprecedented rate. Moreover, agentic AI opens up many new opportunities for innovation. 
• Rapid Market AdoptionThe pace of AI adoption mirrors the swift rise of the internet. A McKinsey survey indicates that AI adoption has jumped to 72% globally, with significant interest across various regions (Mckinsey). Enterprises, small businesses, and individuals alike are integrating AI solutions to enhance efficiency and create competitive advantages.
• Economic Pressures and Demand for EfficiencyIn a challenging economic climate, AI offers solutions to cut costs, streamline processes, and unlock new revenue streams, making it a priority investment for organizations worldwide. For example, AI has enabled companies like Netflix to save approximately $1 billion by utilizing machine learning for personalized recommendations (Webster).
• Policy and Ethical ConcernsPolicymakers are actively debating AI’s societal implications, from workforce displacement to data privacy and decision-making autonomy. These discussions underscore AI’s transformative power while creating a need for companies to adopt responsible AI practices.

1.2 Market Landscape and Dynamics

AI companies are generally divided into infrastructure, middleware, and applications layer. Several venture firms have created diagrams with the major players in these layers:

Soma Capital’s AI portfolio:

The global AI market, valued at $196.63 billion in 2023, is projected to grow at a CAGR of 36.6% to reach $1,339.1 billion by 2030 (Grand View Research). With 72% of organizations adopting AI regularly as of 2024, the technology is delivering measurable ROI across industries (McKinsey).

ROI Potential:

• Rapid Realization: 83% of companies report positive ROI from AI within three months, with advanced adopters achieving 3x higher ROI than early-stage adopters (G2; Vention Teams).
• Efficiency Gains: AI-driven automation significantly boosts productivity and reduces costs (PwC).
  

1.3 How should investors evaluate AI startups?

Before we dive into the trends and technical details, I want to first address an important topic: how should investors evaluate AI startups. My take is that AI can be thought of as a fundamental utility, akin to a water source, essential yet ubiquitous in its availability. This analogy helps frame the strategic considerations around AI innovations, especially for startups that build on or leverage AI technologies.

AI technologies, particularly GPT models, have become a baseline resource for many companies, much like water is for human life. The wide availability and accessibility of these AI models mean that many early stage startups rely on foundational models as their core technology, with some training on specific datasets or using finetuning. The question for investors and businesses alike is: What happens when this 'water' becomes more expensive or less accessible? Would this startup now be in trouble? The key consideration here is whether a business has built-in resilience—can it absorb these costs or are they overly reliant.

On the other hand, advancements in AI quality would significantly enhance a startup’s value proposition. For example, more accurate models could lead to better customer insights, improved automation, or more effective marketing strategies. However, this also raises the stakes: as AI becomes more powerful, the competitive landscape intensifies. Startups must innovate continuously to maintain a competitive edge of not becoming replaced by the foundational models itself or risk being outpaced by better-funded competitors. A few questions to ask every AI startup, especially application layer:

1. Scalability: How scalable is your business model if the cost of AI increases?
2. Differentiation: What distinguishes your startup from competitors? Do you leverage proprietary data, novel algorithms, or a unique technical advantage? How defensible is your market position against competitors or large incumbents?
3. Future Risks: Are any of your competitive advantages vulnerable if foundational AI companies begin offering similar services or access to shared data? Could a larger player or internal team replicate your solution?
4. Adaptability: If AI capabilities improve dramatically (e.g., 10x performance gains), how does your solution remain valuable? Does your business model depend on a specific state of AI development, or is it adaptable to future changes?
   

Questions for AI startups in middleware or infrastructure layer:

1. Moat Against Commoditization: As infrastructure tools evolve, how do you avoid becoming commoditized? Are you adding value beyond what foundational model providers (e.g., OpenAI, AWS) might offer directly?
2. Tech Innovation & Integration: How does your product improve the efficiency, reliability, or cost of foundational AI systems? How easily can your solution integrate with existing enterprise systems or other AI tools?
3. Regulatory Compliance: How does your solution address compliance challenges in industries with strict regulations (e.g., healthcare, finance)? Does your infrastructure or middleware enable customers to meet regional data governance requirements?
   

SECTION 2. AI Technology and Ideas to Watch

My Request for Startups, Ideas and Technologies that have transformative potential

With this, here are some ideas that I have identified around October/November of 2024. Note that the AI landscape is evolving rapidly and some ideas may become outdated quickly.

2.1 Ideas To Watch:

2.2 Technologies To Watch: 

SECTION 3. Risk Assessments and Ethical Considerations

Ethical design and risk assessments as a strategic advantage and moat.

3.1 Ethical Design as a Catalyst for Market Leadership & Consumer Sentiment

Startups operate at the forefront of innovation, but with this position comes an inherent responsibility to design products that align with ethical principles and comply with regulatory standards. While regulatory compliance is often seen as a necessity to avoid legal pitfalls, ethical considerations go far beyond adherence to rules—they serve as a strategic advantage. In fact, having robust risk analysis and ethical considerations safeguards new innovations against unintended negative consequences. It might also open up new opportunities to maximize profit, as evidenced by companies like Anthropic AI, which have built their business models on safety-first principles. Consumer sentiment strongly supports this approach. A recent study by PWC indicates that over 80% of consumers are willing to pay more for products from companies committed to sustainable and ethical production [3]. Startups capitalizing on shifting consumer preferences, establishing themselves as leaders in a market increasingly driven by values-based decision-making.

3.2 Leveraging Ethical Design for Competitive and Sustainable Growth

Startups and investors can leverage ethical design and considerations to build a sustainable competitive advantage by focusing on:

1. Transparent Algorithms and Models: Invest in developing AI and systems that are auditable, ensuring users and regulators can trust the technology. There are many researchers and startups today focused on explainable AI and this will be a crucial addition to the AI pipeline.
2. Proactive Risk Assessments: Investors and startups should regularly evaluate the societal, environmental, and cultural impact of products during the design phase to anticipate and mitigate unintended consequences. It is often much more efficient and cost-effective to do so during the design phase than addressing incidents.
3. Ethics and Trust as Branding: Firms should market ethical design practices as part of the company's value proposition to attract conscious consumers and investors who prioritize sustainability and responsibility and to address the consumer sentiment shifts.
4. Sustainable Development Goals (SDGs): When appropriate, some startups can align product development with global SDGs to unlock partnerships, opportunities, and support.

3.3 Responsible Design for Resilience

Responsible design should not only address the needs of users but also consider the broader societal impacts, minimizing negative externalities that could harm vulnerable groups or perpetuate systemic issues. In this context, we draw inspiration from Winner’s assertion that “technologies are seen as neutral tools…[but] we usually do not stop to inquire whether a given device might have been designed and built in such a way that it produces a set of consequences logically and temporally prior to any of its professed uses.” Startups, therefore, have a critical role in ensuring that their innovations are designed with foresight, taking into account both the immediate and long-term implications of their technologies. A pressing example can be seen in traditional social media platforms, where the initial goal of connecting people gave rise to unintended “logically prior” consequences such as addictive algorithms, and “temporally prior” effects such as the rapid spread of misinformation. These outcomes highlight the urgent need for startups to adopt a more holistic approach to product development—one that integrates ethical considerations into every stage of the design process. For instance, startups in AI, fintech, or healthcare could implement proactive measures like algorithmic transparency, robust data privacy safeguards, or equitable access frameworks. Doing so not only mitigates risk but also builds trust, which is becoming a crucial differentiator in today’s competitive markets.

Moreover, regulatory scrutiny is intensifying as governments worldwide address the unintended consequences of unregulated technological growth. Startups that proactively adhere to, or exceed, regulatory requirements in areas like data protection, environmental impact, and social equity will not only reduce potential liabilities but also position themselves as leaders in a shifting marketplace. Ethical and regulatory alignment is not a compliance exercise; it is an investment in resilience and long-term growth.

By prioritizing responsible design and aligning with ethical and regulatory frameworks, startups can redefine industry norms, mitigate risks, and build products that generate value for society. This commitment forms a key foundation for my investment thesis, as companies that embrace these principles will achieve sustained success while driving meaningful societal change.

* This section’s insights, concepts, notes, and my writings are heavily inspired and credited towards my engineering ethics class in Fall 2024, ENGIN 125 at UC Berkeley, taught by Professor Karl A Van Bibber.

SECTION 4. Investing in the Future: Strategic Pathways in AI

Possibilities are endless, but long-term vision, diversification, and constant learning will be necessary.

4.1 Contrarian Perspectives on AI

While the rapid advancement of AI seems to promise infinite applicability, there still exists fundamental limitations that cannot be simply solved. By extension, any company proposals that guarantees a solution to these problems should be subject to additional scrutiny, but harbors tremendous potential if the promise is true. Across the Artificial General Intelligence (AGI) systems landscape, almost every implementation of AGI systems are in the form of Large Language Models (LLMs), which relies on one of the two architectures: Long Short-Term Memory (LSTMs) or Transformers. In recent years, the Transformer architecture, with the advantage of parallelizable self-attention mechanism that increases model efficiency, significantly outpaces LSTMs. Although there are advancements with the LSTM architecture recently (1), transformers still underpin most major consumer facing systems.

Modern transformer systems (ChatGPT, Gemini, Claude, etc.) require very large amounts of training data. This places constraints on their performance in highly vertical segments where training data is scarce or difficult to process in large amounts. For instance, most medical diagnosis data are stored with physical patient files, so digitizing these data at scale to train medical assistant agents will be difficult.

In addition, transformers are fixed architectures with predetermined input and output dimensions. In the case of GPT 3.5-Turbo, the input is fixed at 4096 tokens, and relies on padding to create the illusion of allowing variable length inputs. Although new models like GPT4 or DALL-E has higher limits, it is architecturally unlikely to be able to work with high resolution images. For example, DALL-E 2 can natively generate up to 1024x1024 pixels, the equivalent of 1 megapixels. While it is an improvement over the 256x256 limit for DALL-E 1, the file size is still very small by modern standards. Smart phone cameras can capture 40 megapixels and can display approximately 4 megapixels. Furthermore, this limit is also partially due to the need to minimize the sizes of training images to reduce model training time and memory requirements. That being said, DALL-E can generate higher resolution files by using additional upsampling programs or divide high resolution images to lower resolution sections, but the fundamental architecture limits the maximum amount of information that these models can generate.

Fundamentally, the limitation can be thought of as a scalability constraint. For some applications, such as speech or video generation, the problem is only cost, as repeated invocations of the neural net can amount to high costs. For other problems such as high resolution image generation, scalability may not be possible as resolution enhancement tools tend to introduce unwanted visual artifacts when pushed to the extremes.

4.2 Predictions on the future of AI, jobs, and the world

4.3 Key AI Investment Principles

First-Mover Advantage

Target sectors with low AI adoption but high disruption potential. Industries like legal tech, construction, and agriculture are ripe for transformation through automation, predictive tools, and optimized processes.

• Strategy: Focus on startups with domain expertise and proprietary datasets, as they create strong competitive moats.

Diversification

Invest across the AI value chain—foundational models, middleware, and applications—to reduce risk and capture opportunities in different market layers.

• Examples: Foundational AI (e.g., OpenAI), middleware for infrastructure tools (e.g., monitoring, APIs), and application startups solving specific problems (e.g., fraud detection, chatbots).
• Strategy: Balance high-risk foundational bets with steady middleware investments and targeted applications.

Long-Term Vision

Support startups addressing systemic challenges like data security, misinformation, and sustainability. These areas align with growing regulatory scrutiny and consumer demand for ethical solutions.

• Focus Areas:
  • Secure AI for privacy and compliance (e.g., GDPR, CCPA).
  • Tools combating deep fakes and misinformation.
  • AI for energy optimization and ESG goals.
• Strategy: Back teams exceeding regulatory standards and prioritizing ethical frameworks.
  

SECTION 5. Appendix

5.1 Bibliography

Research Papers

• A. Ng, "Data-Centric AI Initiative." Available: https://datacentricai.org/.
• ArXiv, "Synthetic Data for Machine Learning." Available: https://arxiv.org/abs/2003.08934.
• OpenAI, "CLIP, GPT Agents, and SAFE AGI Frameworks." Available: https://openai.com/research/clip.

Industry Reports

• Grand View Research, "Artificial Intelligence Market Analysis," 2023. Available: https://www.grandviewresearch.com/.
• McKinsey & Company, "The State of AI Adoption in 2024." Available: https://www.mckinsey.com/.
• PwC, "AI and Productivity: Driving ROI with AI Automation." Available: https://www.pwc.com/.
• Boston Consulting Group (BCG), "Strategic AI Value in 2024." Available: https://www.bcg.com/.

Others

• Authority Hacker, "AI Statistics Overview." Available: https://www.authorityhacker.com/.
• IBM Newsroom, "Enterprise AI Adoption Trends in 2024." Available: https://newsroom.ibm.com/.
• NVIDIA, "Generative AI for 3D Content." Available: https://developer.nvidia.com/nvidia-omniverse-platform.
• Sequoia Capital, "Generative AI Ecosystem Map." Available: https://www.sequoiacap.com/.

Antler, "Generative AI Startups by Sector." Available: https://www.antler.co/.