TLDR AI 2025-08-21

Most AI projects fail not because of the algorithms, but because of engineering practices. carsales discovered this when building their AI Model Finder - a vehicle recommendation tool that suggests cars based on lifestyle needs and preferences.

Working with Thoughtworks, they transformed their development processes first: implementing Test-Driven Development, refining team dynamics, and establishing clear deployment pipelines. As a result, their AI initiative delivered faster time-to-market, reduced risk, and higher ROI.

Read the full carsales case study

Anthropic has expanded its enterprise offering with Claude Code, bundled under premium seats for business plans.

Google added real-time visual guidance to Gemini Live that highlights objects on-screen when using the camera along with deeper integration across the Google Suite. The update includes more expressive speech patterns that can adapt tone based on context and take direction on pace or accent, closing the gap between Google's voice capabilities and fast-growing voice AI startups.

Character.AI has been holding discussions with possible buyers, banks, and staff over the past few weeks about raising a few hundred million dollars at a valuation of more than $1 billion. The startup's founders left the company last September to join Google, leaving employees to take over ownership. If sold, the buyer would get Character.AI's app and website, which hosts chatbots that resemble anime characters, celebrities, and historical figures. Character.AI makes most of its revenue by charging for premium features with its chatbots. It expects to reach $50 million in annualized revenue by the end of the year.

UserJot's founder argues for a strict two-tier architecture with stateless subagents that act as pure functions: no memory, no conversation history, just task-in-result-out. MapReduce patterns handle large-scale work through parallel execution, which keeps individual agent tasks under 30 seconds and prevents the debugging nightmares that plague complex hierarchies and shared state systems.

AI labs will soon favor quality over quantity when procuring RL environments and spend a surprisingly large amount of money on procurement per RL task. Procuring RL tasks cheaply results in lower quality training runs, and the compute cost per RL run soon becomes very high. AI labs will increasingly move away from training on procedurally generated RL tasks and tasks produced cheaply by massive numbers of low-paid contractors toward a more labor-intensive build process that relies on full-time domain experts who spend months developing sustained context and crafting individualized tasks. The winners will be those who deliver the highest-quality, context-rich tasks, ship them rapidly, and price them efficiently given compute tasks.

Data center construction contributed as much to GDP growth as consumer spending in the first half of 2025, a historic shift given that consumer spending typically represents 70% of economic activity. AI infrastructure spending masked what would have been a -1.5% GDP contraction in Q1. AI-sensitive sectors have driven 53% investment growth since 2019 compared to flat performance elsewhere.

From AI to automation, IBM TechXchange 2025 delivers deep technical content, live demos, and hands-on workshops.🛠️ Walk away with skills, insights, and connections that matter.

Register now →

HealthChain connects AI models to any healthcare system with just a few lines of Python. It makes integrating AI with electronic health records automatic, fast, and easy. The HealthChain API provides a secure integration layer that coordinates multiple healthcare systems in a single application. Developers can use it to build pipelines for NLP and ML tasks that easily integrate with complex healthcare systems.

Most systems rely on external documentation or pre-configured knowledge to instruct AI agents when they hit protected pages, but it would be simpler to have the instructions right there in the HTML response. llms.txt is an emerging standard for making content such as docs available for direct consumption by agents. This post proposes a convention to include such content directly in HTML responses. It is designed to fit right in with llms.txt, which has found adoption for publishing LLM-targeted content in a discoverable fashion on the web.

ByteDance's Seed Team of AI researchers today released Seed-OSS-36B on Hugging Face. Seed-OSS-36B is a new line of open source, large language models designed for advanced reasoning with both synthetic- and non-synthetic-data variants and an instruct model. The synthetic-data variant is intended as a higher-performing general-purpose option, while the non-synthetic model avoids potential bias or distortion introduced by synthetic instruction data. The instruct model is post-trained with instruction data to prioritize task execution and instruction following. The models were released under the Apache-2.0 license, which allows free use, modification, and redistribution by researchers and developers working for enterprises.

Hugging Face now enables users to generate images with Claude by integrating with Spaces and recently released models. This guide explains how to set up the workflow and highlights models tuned for natural scenes and text-rich visuals.

Google Photos will soon support natural language photo editing, starting with Pixel 10 users in the US. Users can describe edits via voice or text, and see AI-powered changes applied automatically.

You'll probably like our flagship newsletter. It's all about tech, science, and programming.

Same quick format. Still free.

Subscribe now.

xAI made Grok conversations searchable on Google without user warnings when people clicked share.

Microsoft AI CEO Mustafa Suleyman warns that "Seemingly Conscious AI" will create dangerous delusions where people believe AIs deserve rights and legal protections, disparaging academic discussions of "model welfare" as premature and dangerous.