TLDR AI 2026-01-14

LTX-2 is officially out, and you need to see it to believe it. Produce up to 20 seconds of high-fidelity video with complete creative control and consistent style.

>> Native 4K, 50 FPS.

>> Stylistic consistency.

>> Coherent video and audio.

>> Precise control over motion, structure, camera.

Try it in the API Playground or locally

The Veo 3.1 update introduces new features designed for both casual creators and professionals. Users can now generate videos based on reference images. The tool now supports native vertical outputs tailored for mobile devices, and high-resolution upscales to 1080p and 4K are also available. The release is available now across multiple Google services and platforms.

Anthropic Labs has expanded its team to incubate experimental products at the forefront of AI capabilities, leveraging rapid model advancements. Co-founder of Instagram, Mike Krieger, has joined Labs for collaborative product development. Ami Vora will lead product organization alongside CTO Rahul Patil to scale Claude experiences. Recent successes include Claude Code's billion-dollar growth and the Model Context Protocol's rise to industry standard.

Google is testing an Auto Browse tool for Gemini that lets the AI manage tasks in Chrome. The tool will allow Gemini to autonomously browse the web, manage tabs, and directly interact with Chrome on users' behalf. Code references indicate that Auto Browse may be exclusive to the Gemini Ultra plan. Google is rolling out the feature incrementally to a limited audience in the US.

AI-assisted coding streamlines product creation but often sacrifices long-term design sustainability. It prioritizes immediate functionality, leading to hardcoded configurations, which are fragile for future updates. Successful product development requires experienced system design to avoid a "build-now, regret-later" scenario, transforming AI vibes into structured, sustainable solutions.

Tsinghua University and Zhipu co-hosted AGI-Next, a summit for frontier AI, in Beijing on January 10. The event included a series of keynotes by notable figures in the Chinese AI industry. This post contains a translated transcript of the event, revealing a fascinating conversation about the AI landscape in China that covers the technical side, corporate dynamics, as well as the future envisioned by China's industry Titans. The event took an honest look at whether China's open-source leadership has actually narrowed the technology gap with the US, China's emerging AI-for-business paradigm, and what it will take for Chinese researchers to take riskier bets.

The AI Chip Sales data explorer uses financial reports, company disclosures, and more to estimate compute, power usage, and spending over time for a wide variety of AI chips. The cumulative global AI compute capacity has reached the equivalent of more than 15 million Nvidia H100 GPUs. The cost to purchase these chips has rapidly escalated to tens of billions of dollars per quarter, drawing over 10 GW of power per quarter. A link to the interactive data explorer is available.

Unblocked is the AI code review that surfaces real issues and meaningful feedback instead of flooding your PRs with stylistic nitpicks and low-value comments.

“Unblocked made me reconsider my AI fatigue. Finally, a tool that surfaces context only someone with a full view of the codebase could provide.” - Senior developer, Clio

Try now for free

CaRR improves reinforcement learning for deep search agents by replacing binary rewards with citation-based rubric feedback.

GLM-Image is an open-source, industrial-grade discrete auto-regressive image generation model with a hybrid architecture that combines an auto-regressive module with a diffusion decoder. It excels in text-rendering and knowledge-intensive generation scenarios. The model performs especially well in tasks that require precise semantic and complex information expression. It supports a rich set of image-to-image tasks.

DeepSeek's Engram architecture introduces a conditional memory system that uses lookup tables for common N-gram patterns, significantly reducing computational overhead in LLMs. This approach boosts performance on knowledge benchmarks and reasoning tasks by efficiently freeing up neural resources for complex reasoning. Engram's design leverages tokenizer compression, multi-head hashing, and context-aware gating, providing a marked improvement over MoE baselines without substantial memory impact.

Vocal computing may finally become mainstream due to new AI advancements in LLMs overcoming past limitations of systems like Siri and Alexa. OpenAI's breakthroughs and its acquisition of a hardware startup hint at innovative voice-driven devices potentially moving beyond the smartphone. As tech companies refine and develop diverse hardware for AI, voice control is pivotal in advancing towards a more integrated and intuitive tech ecosystem.

AI introduces "borrowed certainty" in software, creating technical debt by replacing proof and evidence with confident-seeming predictions. Developers often mistake AI's fluency for validity, risking decisions based on unearned certainty. Best practices include treating AI output as a hypothesis, demanding traceability, ensuring human understanding, and pairing AI with execution feedback.

By 2026, advanced AI capabilities will emphasize autonomous agents and Generative UI, making UX the main business differentiator.

Zhipu's GLM-Image model is the first in China to be trained using Huawei's Ascend AI chips and Kunpeng processors.

Google expanded its Health AI Developer Foundations with MedGemma 1.5 for medical image interpretation and MedASR for speech-to-text.

This page contains the technical specification for the Claude Agent SDK, which provides programmatic access to Claude Code's agentic capabilities.