TLDR AI 2026-02-19

How can you best use AI to build better software? In 2026, this is the question on everyone's mind. At the Sonar Summit virtual conference, you'll get to hear answers from some of the most respected voices in tech.

This event will grapple with the reality of AI writing the majority of new code and what it means for teams, with featured speakers including:

• Gergely Orosz (The Pragmatic Engineer),
• Santiago Valdarrama, AWS AI Hero
• Laura Tacho (CTO at DX, co-author of the Core 4 developer productivity framework).
• And many others - see the full agenda!

It's virtual, free, and available in your timezone. Save your spot →

Google announced that the Gemini app integrated Lyria 3, its latest music generation model, enabling users to create 30-second tracks from text or images in beta. The update expanded Gemini's multimodal capabilities into consumer-facing music creation.

World Labs secured $1 billion in funding from investors like AMD, NVIDIA, and Fidelity. They aim to advance spatial intelligence with products like MARBLE, which creates cohesive 3D worlds from images, video, or text.

LangChain's coding agent went from Top 30 to Top 5 on Terminal Bench 2.0 with just a harness change. This post discusses the team's approach to harness engineering. The goal of a harness is to mold the intelligence of a model for tasks you care about. Self-verification and tracing help a lot.

OpenAI published a detailed guide explaining how prompt caching reuses repeated prompt prefixes to skip prefill compute, reducing latency and input token costs. It outlines cache mechanics, routing considerations, KV reuse, and practical strategies for increasing cache hit rates in production workloads.

Ivan Zhao frames AI as the latest “tool of life” after language, markets, and the internet. Humanity uses culture, capital, and computation to impose its values on an indifferent universe, and AI extends this by accelerating idea discovery and coordination. In this view, AI is not separate from life but a continuation of life organizing matter and energy through intelligence.

Give it a question, and it gets to work: Subagents deeply interrogate your topic, scour a wide range of credible sources, and package it all up into boardroom-ready reports, slides, docs, or websites.

Google researchers showed that randomly masking parameter updates can outperform dense adaptive optimizers, with a masked RMSProp variant exceeding recent state-of-the-art methods. Their proposed Magma optimizer uses momentum-aligned gradient masking to improve pre-training efficiency.

EVMbench is a benchmark that evaluates the ability of AI agents to detect, patch, and exploit high-severity smart contract vulnerabilities. Smart contracts routinely secure hundreds of billions of dollars in crypto assets. As AI agents improve, it becomes increasingly important to measure their capabilities in economically meaningful environments. This also means using AI systems to defensively audit and strengthen deployed contracts.

GLM-5 technical report outlines the details of the model development, combining DSA-based efficiency improvements with asynchronous reinforcement learning to decouple generation from training.

Amical has launched an open-source AI dictation app that offers privacy-focused, on-device speech recognition using models like Whisper. Unique features include context-aware formatting, custom prompts, and local data storage with a cloud fallback option for less capable devices. Available for Mac and Windows, Amical plans to expand to mobile and integrate with broader AI ecosystems.

Charles Porch, a 15-year Meta and Instagram veteran, has joined OpenAI as Vice President of Global Creative Partnerships. Porch previously led major cultural initiatives at Meta, including high-profile onboardings like Pope Francis and digital strategy for Beyoncé. In his new role, he will focus on the intersection of AI and culture, collaborating with global creative communities to develop products for learning, exploration, and entertainment.

Agents don't fail because of prompts, they fail because of systems. Take Temporal's 8-question quiz to see if yours is ready for long-running, resilient agents. Take the quiz

Compilers are deterministic, and large language models are stochastic.

Saudi Arabia's Humain has invested $3 billion into Elon Musk's xAI, acquiring a significant minority stake that converts to SpaceX shares.

Before developing or buying a security system for AI agents, consider what it monitors, what it was trained on, and how well it was tested.

David Silver plans to raise $1 billion for Ineffable Intelligence, focusing on reinforcement learning, not incremental LLM updates.