Divya van Mahajan

The Five Vectors of AI Competition

ai-daily-brief-podcast

Overview

This episode of The AI Daily Brief (recorded May 20, 2025) presents a framework for analysing competitive dynamics among the leading AI companies—OpenAI, Microsoft, Google, and Anthropic—across five distinct vectors. The episode is structured as a preview ahead of a major week of industry conferences: Microsoft Build, Google I/O, and Anthropic’s first developer conference, Code with Claude. The host is the unnamed presenter of the AI Daily Brief podcast/video channel.

Source video: No URL provided in the transcript.

Prerequisites

  • Familiarity with the major AI labs: OpenAI, Anthropic, Google DeepMind, and Microsoft’s AI division
  • Basic understanding of large language models (LLMs) and chat-based AI assistants (ChatGPT, Claude, Gemini, Copilot)
  • Awareness of coding-assistance tools such as Cursor, GitHub Copilot, and “vibe coding” platforms (Lovable, V0)
  • General knowledge of enterprise software and cloud platforms (Azure, Microsoft 365, Google Workspace)
  • Familiarity with the concept of AI agents and agentic workflows
  • Some exposure to Model Context Protocol (MCP) and agent infrastructure concepts

Main Points

The Five Vectors of AI Competition

  • The host organises competitive analysis into five areas: consumer, enterprise, benchmarks, coding, and agents.
  • Developer competition is treated as cross-cutting and most concentrated in the coding and agents vectors.
  • This framework is explicitly described as non-scientific and non-comprehensive, but presented as a practical mental model.

OpenAI’s Codex Launch (Front-Running the Conferences)

  • OpenAI launched Codex, a coding agent powered by Codex One, a version of the O3 reasoning model optimised for software engineering.
  • Key claimed capabilities: cleaner code than O3, better instruction-following, iterative self-testing until tests pass, ability to handle multiple tasks in parallel, and availability inside ChatGPT (including the mobile app).
  • Distribution via ChatGPT is highlighted as a major differentiator over standalone tools like Cursor or Claude Code.
  • Reactions were mixed: some dismissed it as hype-cycle repetition (citing the same pattern with Devin, DeepSeek, and prior GPT drops); others (e.g., vibe coders, Ethan Mollick) noted it remains primarily accessible to technical users and could benefit from broader democratisation.
  • Josh Tobin (OpenAI agents team) argued that Codex increases the value of technical skills, since precise problem specification becomes the bottleneck for parallelised AI work.

Microsoft Build: Enterprise Deepening and Agent Vision

  • Microsoft is not positioned as a frontier model competitor; rather, it is the default enterprise platform and is deepening Copilot integration across Windows, Office, and Azure.
  • Expected announcements include semantic search in Windows settings and File Explorer, Copilot Agents enhancements, and the Windows Recall feature.
  • Model Context Protocol (MCP) was flagged as a potential keynote topic; its inclusion would signal a serious enterprise agent-building strategy.
  • Microsoft’s broader framing, drawn from their 2025 Work Trend Index, is the concept of the “frontier firm”: every employee becomes an “agent boss” managing swarms of AI agents.
  • A persistent pain point is the gap between consumer AI tools and what employees can actually access inside Microsoft-managed enterprise environments.

Google I/O: Strong Models, Uncertain Positioning

  • Google has staged a significant comeback since the prior I/O (where Gemini famously suggested glue as a pizza topping).
  • Gemini 2.5 Pro is described as a benchmark leader and is widely credited with briefly challenging Claude’s dominance in coding use cases.
  • Pre-conference product releases included an updated Gemini 2.5 Pro, a DeepMind coding agent that reportedly cut Google’s global compute by 0.7% through code optimisation, and a standalone Notebook LM app.
  • User numbers are contested in quality: 1.5 billion AI Overviews users (embedded in Search), 150 million Google One subscribers (up 50% YoY), and 350 million monthly active Gemini users—but distribution overlap with pre-installed handsets limits the interpretive value of these figures.
  • Google is described as occupying an uncomfortable middle ground between pure consumer (OpenAI’s domain) and pure enterprise (Microsoft’s domain), potentially fragmenting its prioritisation of AI initiatives.
  • Google’s Agent-to-Agent (A2A) protocol is positioned as complementary to MCP, targeting a different layer of the agent infrastructure stack (inter-agent communication vs. tool/context access).

Anthropic: Code with Claude and Model Expectations

  • Anthropic has established itself as the preferred model provider for coding tools and coding agents, with Claude consistently top-ranked by practitioners even when losing on general benchmarks.
  • Code with Claude is their first developer conference, focused on the Anthropic API, CLI tools, and MCP.
  • Reported upcoming model releases (sourced from The Information): new versions of Claude Sonnet and Claude Opus with the ability to fluidly switch between reasoning and tool use, and to self-correct mid-task.
  • Models are described as capable of automatically testing generated code, stopping to reason when errors occur, and correcting without human intervention—designed for high-level, low-supervision instructions (e.g., “make this app faster”).
  • MCP is described as having become foundational to the emerging agent-building ecosystem.

Vector 1: Coding

  • Coding is identified as the single most important breakout AI use case currently because it is an enabler of other use cases and bridges technical and non-technical users.
  • The “vibe coding” movement (non-technical users building with AI) has expanded the stakeholder base beyond professional developers.
  • Despite Gemini 2.5 Pro’s benchmark gains, habits among developers have not substantially shifted away from Claude; the host expects Anthropic to retain top-dog status among developers after its upcoming release.
  • Codex’s integration into ChatGPT’s wide distribution base is the main mechanism by which OpenAI could challenge this.

Vector 2: Agents

  • Agent competition has two distinct sub-dimensions: (1) agent infrastructure/platforms and (2) end-user agent products.
  • On infrastructure, Anthropic leads with MCP; Google is competing with A2A for the inter-agent communication layer; Microsoft is expected to incorporate these into Azure for enterprise builders.
  • On end products, OpenAI and Google are the most aggressively pursuing ownership of specific agents (Deep Research, Operator, Codex for OpenAI; various agent previews for Google).
  • OpenAI’s motivation for owning agents, not just powering them, is framed as a hedge against model commoditisation: owning the customer relationship in an agent product provides a durable moat.

Vector 3: Enterprise

  • Microsoft holds default pole position in enterprise, reinforced by its early OpenAI partnership.
  • Ramp AI Index data (based on business card/bill-pay data, skewed toward SMEs and startups): OpenAI’s business subscription adoption jumped from ~15% to 32.4% of US companies; Anthropic doubled from ~4% to ~8%; Google dropped significantly.
  • These figures represent a particular market slice and do not reflect large enterprise adoption broadly.
  • Anthropic’s enterprise growth is notable but still far behind OpenAI’s pace in this segment.

Vector 4: Consumer

  • OpenAI is the dominant consumer AI brand; for many general users, “ChatGPT” and “AI” are synonymous.
  • Weekly active users are reported at approximately 800 million, boosted significantly by the Ghibli-style image generation trend.
  • OpenAI is reinforcing its consumer lead by hiring Instacart CEO Fiji Simo as CEO of Applications.
  • Google’s 350 million Gemini monthly active users, while large, include passive distribution via pre-installed handsets, limiting interpretive value.
  • Meta’s AI user numbers are treated with similar scepticism because of forced in-app integration.

Vector 5: Benchmarks

  • Benchmarks are described as being at or near a historic low in influence on actual user or enterprise decision-making.
  • A Reddit observation is cited: Claude consistently outperforms other models on real tasks despite losing on published benchmarks to OpenAI and Google models.
  • However, the host suggests a potential partial rehabilitation of benchmarks via narrow, use-case-specific evals (e.g., LegalBench), which entrepreneurs and developers may find actionable—citing a founder who chose Gemini 2.5 Pro specifically because it ranked first on LegalBench for a legal drafting task.
  • General consumers and general enterprises are unlikely to switch models based on benchmark scores; proof-of-practice remains the dominant selection criterion.

Key Concepts

  • Codex / Codex One: OpenAI’s coding agent and its underlying model, a variant of O3 optimised for software engineering tasks with iterative self-testing.
  • Vibe coding: The practice of non-technical users building software applications using AI coding assistants, without writing code manually.
  • Model Context Protocol (MCP): An open protocol developed by Anthropic that provides AI agents with structured access to tools and contextual data; has become a foundational standard in the agent-building ecosystem.
  • Agent-to-Agent Protocol (A2A): Google’s open protocol for inter-agent communication, positioned as complementary to MCP at a different layer of the agent stack.
  • Frontier firm: Microsoft’s concept (from their 2025 Work Trend Index) of an organisation in which every employee manages swarms of AI agents, representing the end-state of workplace AI adoption.
  • Copilot Agents: Microsoft’s feature (introduced April 2025) enabling AI to handle complex multi-step tasks within the Microsoft product ecosystem.
  • Ramp AI Index: A dataset estimating business AI product adoption using transaction data from Ramp’s corporate card and bill-pay platform, skewed toward SMEs and startups.
  • LegalBench: A domain-specific AI benchmark measuring legal reasoning and drafting accuracy.
  • Code with Claude: Anthropic’s first developer conference (May 22, 2025), focused on real-world API use, CLI tools, and MCP.
  • Notebook LM: Google’s AI-powered note-taking and research tool, recently launched as a standalone app.

Summary

The host uses a week of major AI conferences—Microsoft Build, Google I/O, and Anthropic’s Code with Claude—as a lens through which to introduce a five-vector framework for tracking AI competition: consumer, enterprise, benchmarks, coding, and agents. The central argument is that these five areas are not equivalent in current importance: coding and agents are the most dynamic and consequential battlegrounds right now, with consumer leadership (dominated by OpenAI) and enterprise defaults (anchored by Microsoft) providing structural advantages that model quality alone cannot overcome. Anthropic holds a practitioner-validated lead in coding and agent infrastructure despite losing on general benchmarks, while Google has credibly recovered technically but struggles with a diffuse market position between consumer and enterprise. Benchmarks are dismissed as nearly irrelevant for most decision-makers, with a partial exception for narrow, task-specific evals. Underlying all five vectors is a deeper strategic tension: as models commoditise, the durable competitive position lies in owning the customer relationship—through consumer products, enterprise lock-in, or the agents themselves—rather than in model quality alone.