This Month in Tech: April 2025

The Second Half
The AI industry is now shifting its focus from solving problems to defining problems. Evaluation is becoming more important than training - people are now looking at how to measure real progress. A shift in mindset and skill set is required to thrive in this new era. Players in the game can now build billion- or trillion-dollar companies by making useful products out of intelligence.

OpenAI releases new simulated reasoning models with full tool access
OpenAI has released two new models, o3 and o4-mini. These models combine simulated reasoning capabilities with access to functions like web browsing and coding. It is the first time OpenAI’s reasoning-focused models can use every ChatGPT tool simultaneously. ChatGPT Plus, Pro, and Team users now have access to the models - Enterprise and Edu customers will gain access next week. OpenAI plans to release o3-pro to the Pro tier in a few weeks.

Agency Is Eating the World
AI is enabling a new wave of lean, successful companies led by individuals who leverage technology to achieve what once required large teams, challenging traditional specialization and credentialism. High-agency individuals are driving this shift, using AI to rapidly accomplish complex tasks across industries without needing deep expertise. The future favors those willing to act independently, embracing AI-driven efficiency over traditional, hierarchical business structures.

Everything we announced at our first-ever LlamaCon
At LlamaCon 2025, the Meta Llama ecosystem announced the Llama API for easier model building, new protection tools, a defender program for better security, and the recipients of the second Llama Impact Grants.

More than 1/3 of all code at Google is now generated by AI
The code is generated by AI, then reviewed and accepted by engineers.

MCPs, Gatekeepers, and the Future of AI
Model Context Protocols (MCPs) are standardized APIs connecting external data sources and applications to LLMs. The current user experience and security of MCPs are not good enough yet for widespread adoption. The power will be with client-side interfaces like ChatGPT and Claude, who control which MCPs are used and how their responses are displayed. LLM providers will act as gatekeepers, managing MCP selection and monetization similar to app stores and search engines.

OpenAI’s new reasoning AI models hallucinate more
OpenAI’s internal tests show that o3 and o4-mini hallucinate more often than the company’s previous reasoning models, as well as its traditional non-reasoning models.

Gemini 2.5 Flash
Google is releasing an early version of Gemini 2.5 Flash, its new AI model, which has improved reasoning capabilities while maintaining speed and cost-efficiency. The model uses a hybrid reasoning approach, allowing devs to control the “thinking” process and set a thinking budget to balance quality, cost, and latency. Gemini 2.5 Flash can perform complex tasks more accurately by reasoning through thoughts before responding.

A practical guide to building agents
Agents are AI systems that can perform workflows on users’ behalf with a high degree of independence. This guide, designed for product and engineering teams, explores how to build agents with frameworks for identifying promise use cases, clear patterns for designing agent logic and orchestration, and best practices to ensure agents run safely, predictably, and effectively. It provides the foundational knowledge needed to confidently start building agents. Building reliable agents means starti…

[MCP vs. A2A: Friends or Foes)
Google’s A2A (Agent2Agent) protocol complements Anthropic’s MCP (Model Context Protocol) by focusing on real-time collaboration and inter-agent communication, addressing gaps in MCP such as state management and security. While both protocols serve different purposes—A2A for agent-to-agent interactions and MCP for context integration with LLMs—there is potential overlap as the agentic ecosystem evolves.

Announcing the Agent2Agent Protocol (A2A)
Google Cloud has launched Agent2Agent (A2A), an open protocol that allows AI agents to communicate and collaborate across different platforms and vendors. It standardizes how agents exchange information and coordinate actions. A2A has support for long-running tasks and different modalities (like audio and video).

Ironwood: The first Google TPU for the age of inference
Google has announced Ironwood, its seventh-generation Tensor Processing Unit (TPU), designed for demanding AI inference workloads.

Google’s latest chip is all about reducing one huge hidden cost in AI
Google has unveiled the latest version of its Tensor Processing Unit (TPU), a custom chip built to run artificial intelligence. The Ironwood TPU is positioned for inference rather than training. Inference is considered a high-volume market in the chip world. While Google pays Broadcom for each TPU in commercial production, it still saves more money using TPUs when compared to paying Intel, AMD, and Nvidia to use their chips for inference.

AI and StackOverflow, The Changing Landscape of Developer Support
AI is changing how devs find solutions, with tools like Copilot and Claude becoming popular but still having limitations like hallucinations and outdated information. StackOverflow is declining in popularity as a result. To fill the emerging knowledge gap, devs may need to rely more on blog posts and comprehensive guides for info.

Microsoft’s AI masterplan: Let OpenAI burn cash, then build on their successes
Microsoft AI CEO Mustafa Suleyman says that it’s more cost-effective to trail frontier model builders by three to six months and build on their successes than to compete with them directly.

Anthropic Exploring Model Welfare
Anthropic launched a new research initiative to examine the potential moral relevance of AI systems, including how and when model welfare should be considered in alignment and safety efforts.

State of AI Index 2025
A great, high level, and comprehensive summary in 10 charts of the current state of AI across models, investment, and costs.