-AI & Us: Navigating the Human-Machine Frontier

 

AI & Us: Navigating the Human-Machine Frontier

Three years after the AI Big Bang, early galaxies are forming in the Cloud AI universe, with plenty of “dark matter” still swirling.

By Er. Kamalanathan_j

If 2023 was the AI Big Bang*, 2025 feels like First Light. The fog of the early calamity is lifting—revealing clusters of foundational companies, best practices for building, and patterns for startup success. We’re still a ways from declaring any semblance of stability, but these early AI galaxies give us more visibility than ever of the shape of things to come.

First and foremost, we ought to state categorically that we have the utmost confidence that AI is the driving force behind the greatest technological change wave ever observed. When venture capitalists speak, founders have a right to wonder how to distinguish hype from reality; however, in the case of AI, straightforward numbers provide the answer. If the most straightforward measure of startup reality is revenue growth, we’ve updated our benchmarks and focused on 20 astonishing AI startups to help define what a great AI startup looks like. While these benchmarks will undoubtedly evolve in the coming years, it’s clear that what constituted a great startup in the SaaS era doesn’t quite cut it anymore.

Naturally, the AI era does not always bring investors and startups good news. Some growth indicators can be false. Buyers are hungry, AI demos dazzle, sales can spike, but not all products deliver real value.  Particularly in times when switching costs are low, retention can be fragile. There is now less value in early hypergrowth alone than ever before. Because Big Bangs are difficult to ignore, competition is at an all-time high. Promising areas are attracting 2–3x the rivals of years past.  In addition, many of the giant SaaS companies in our portfolio, such as Intercom, which has already launched an AI product worth more than $100 million, are realizing the importance of AI. We anticipate that these leaders will provide additional competitive pressure and M&A temptation. in the years to come. Additionally, we are still in a wild period of unpredictability. Despite the fact that our heads are spinning a little slower this year, they are still spinning. We still haven't figured out what the implications of developments in AI browsers, the Model Context Protocol (MCP), and a slew of other areas we'll call "dark matter" in this report will mean for the future of AI. However, there is one certainty. There is no cloud without AI anymore. 

 Since our initial commitment in 2023, we have already invested more than $1 billion in AI-native startups at Bessemer. Additionally, AI is being used in almost every legacy SaaS company's product and operations. We are all now in this new AI universe, regardless of whether we believe we are fully prepared to navigate it.

I. Infrastructure for AI Formation of galaxies: Model layer Let's start with the obvious: a few players like OpenAI, Anthropic, Gemini, Llama, and xAI continue to dominate the foundation model landscape, improving model performance and pursuing vertical integration simultaneously. Now that they are offering agents for coding, computer use, and MCP integrations, it is clear that big labs are moving beyond just providing foundation models and model development tools. As a result of software advancements and hardware optimization from beginning to end, compute costs continue to fall predictably. The most cutting-edge open-source models, such as Kimi, DeepSeek, Qwen, Mixtral, and Llama, continue to show that the open ecosystem can still outperform proprietary models in terms of efficiency and specialized tasks. Google's recent Mixture-of-Recursions paper pushes scaling assumptions with an adaptive-depth approach that strikes a balance between inference throughput and few-shot accuracy. On the research side, we're seeing a wave of innovation. New methods for combining experts in novel ways are also being used to revive mix-of-experts architectures. Finally, inference-time methods like adaptive reasoning and test-time reinforcement learning (RL) are gaining traction, with vertical domains likely to see the most significant advancements. These model-level innovations are just one part of a larger replatforming.

A brand-new infrastructure layer that encompasses models, compute, training frameworks, orchestration, and observability has emerged as businesses develop AI-native and AI-embedded products. In our AI infrastructure roadmap for 2024, we highlighted this development. This specialized stack gives builders the speed and flexibility they need, but bundling is getting faster as players try to get more of the stack by moving into adjacent areas. Although significant progress has been made thus far, we contend that the rapid evolution of AI infrastructure is far from complete. AI Infrastructure’s Second Act

Backpropagation, convolutional networks, and transformers were major algorithmic breakthroughs that characterized the first era of AI. The field is primarily advanced by algorithmic improvements and scaling methods.As a result, this mindset has been reflected in infrastructure, which has fueled the rise of giants in foundation models, compute capacity, and data annotation. However, the following chapter may prove to be even more profound. "The second half of AI—starting now—will shift focus from solving problems to defining them," as OpenAI's Shunyu Yao recently observed. The industry will move from demonstrating that AI can solve problems to developing systems that define, measure, and solve problems with experience, clarity, and purpose in AI infrastructure's Second Act. Big labs are moving beyond chasing benchmark gains to designing AI that can interface effectively with the real world.At the same time, enterprises are graduating from proof-of-concepts to production deployments.

A new wave of infrastructure tools, not just designed for scale or efficiency but also designed to ground AI in operational context, real-world experience, and continuous learning, is emerging as a result of all of these shifts. Among the examples are:As human-generated labelled data is no longer sufficient to enable production-grade AI, reinforcement learning environments and task curation through platforms like Fleet, Matrices, Mechanize, Kaizen, Vmax, and Veris are required. New frameworks for evaluation and feedback, such as Bigspin.ai, Kiln AI, and Judgment Labs, make it possible to have feedback loops that are both specific and continuous. Systems of compound AI that combine aspects like knowledge retrieval, memory, planning, and inference optimization rather than just focusing on the raw model horsepower We are at the beginning of this transition—from AI as a proof of concept to AI as a problem-defining and adaptive system embedded in real-world experience.

The bitter lesson of AI: dark matter The "Bitter Lesson" by Rich Sutton reminds us that, rather than relying on handcrafted features or heuristics created by humans, the most effective advancements in AI have historically come from harnessing computation and general-purpose learning. Practitioners are attempting to embed context, understanding, and domain expertise in order to ensure real-world utility, but it is still unclear which techniques will prove to be the most effective or scalable as AI infrastructure enters its next chapter. II. Tooling and platforms for developers Galaxies forming: AI engineering an integral part of software development

Beyond the infrastructure stack, AI has clearly transformed software development.The new programming interface is natural language, with models carrying out the instructions. As prompts are now programs with LLMs as a new type of computer, this paradigm shift is changing the very principles of software development. AI doesn’t just mean an incremental evolution of developer tools; it has ushered in a completely new way of software development.We will cover this landscape in detail in an upcoming roadmap, Developer Tooling for Software 3.0. (Subscribe to Atlas to get it first.)

Today, the question is not whether your team uses AI, but rather how well you integrate it into a high-velocity, compounding system. This mode of software development feels like the “formed galaxy” of AI-native development.The best engineering teams aren’t just writing code with AI—they’re building systems that learn, adapt, and ship faster with every cycle.

The formation of galaxies: the Model Context Protocol Model Context Protocol (MCP), a new layer of infrastructure, will have significant effects on AI development. MCP became the standard for agents to access external APIs, tools, and real-time data at the end of 2024 and was quickly adopted by OpenAI, Google DeepMind, and Microsoft. As the MCP creators described, it can be thought of as the USB-C of AI.Persistent memory, multi-tool workflows, and granular session permissions are all supported. With it, agents can chain tasks, reason over live systems, and interact with structured tools—not just generate outputs.

For developers, MCP radically simplifies integration.It makes it possible for founders to create truly agentic products, in which AI not only assists users but also acts on their behalf across systems. It's early, and it's important to remember that MCP is not a chef's cookbook. In order to really get cooking, we need ecosystems like FastMCP from Prefect (that make it much easier to build MCP servers) and tools like Arcade and Keycard (that facilitate agentic authorization and permissioning.)We anticipate it becoming as fundamental to an agent-native web as HTTP was to the internet as the constellation around MCP connectors, governance frameworks, and agent-specific tools continues to grow. Memory, context, and beyond:Dark matter Memory is becoming a fundamental product primitive as AI-native workflows develop. Tools become indispensable when they can be remembered, modified, and customized over time. Great AI systems are expanding past recall and evolving with the user.In 2025, large context windows and retrieval-augmented generation (RAG) have enabled more coherent single-session interactions, but truly persistent, cross-session memory remains an open challenge.Startups like mem0, Zep, SuperMemory, and LangMem by Langchain are also working on memory while the foundational model companies are. Context is the data a model sees during inference.Memory refers to information retained across interactions—supporting multi-step reasoning, personalization, and agent continuity.Together, they power the next generation of AI applications.

We anticipate that the following will eventually be combined by the leading stacks:Short-term memory via expanded context windows (128k to 1M+ tokens, depending on model and architecture)

Long-term memory via vector DBs, memory OSes (e.g., MemOS), and MCP-style orchestration

Semantic memory via hybrid RAG and emerging episodic modules, designed for context-rich recall

Still, trade-offs remain: long contexts raise latency and cost.Persistent memory is brittle without smart context engineering—dynamic selection, compression, and task isolation are key.

Multi-modal memory layers and stateful workflows are being widely adopted by agentic applications—dev agents, customer copilots, and creative tools. Meanwhile, research into neural memory, continual learning, and local context buffers suggests scalable recall is within reach.

For AI application founders, context and memory may be the new moats.AI switching costs may cause builders who solve these problems to feel almost emotional. When your product is more familiar to a user's world than any other, replacing it feels like starting from scratch. Whether it's a coding assistant fluent in your team’s codebase or a sales agent embedded in your CRM and communication stack, accumulated intelligence on the user and their specific environment becomes the stickiest asset.

Many unknowns remain, but winning startups will likely need to master both the infrastructure and the interface going forward:

Creating memory-aware, adaptable, and low-latency systems Designing for implicit learning and deep integration with core workflows

Turning context into a compounding advantage—across data, distribution, and delight

Founders should treat memory not as backend plumbing but as a product.The most intelligent, personalized, and persistent AI systems of the future will be created by startups today that build with memory-awareness. III. AI for the Enterprise and the Horizon Systems of Record are under pressure as galaxies form. AI is beginning to present startups with opportunities to disrupt some of the largest horizontal systems of record (SoR) in enterprise software. Due to their deep product surfaces, implementation complexity, and centrality to business-critical functions, SoRs such as Salesforce, SAP, Oracle, and ServiceNow remained in business for decades.

The founder’s edge in the AI cosmos

We’re no longer at the dawn of AI—we’re deep in its unfolding galaxies. Today’s top startups aren’t just building faster software. They’re designing systems that see, listen, reason, and act—embedding intelligence into the fabric of work and life.

But here’s the truth: success in AI isn’t just about velocity. It’s about vectors, as in speed in the right direction. The most iconic companies won’t be those who simply ride the wave, but those who shape it—aligning exponential capability with real-world clarity.

 

AI is no longer theoretical. It’s operational. It’s generating revenue, building relationships, and rewriting industry rules. And yet, much remains unresolved: memory, context, governance, and agency. That’s the power of this moment—the map is still fuzzy, but the frontier is real.