TLDR AI 2025-06-06

Manual labeling has long been the most burdensome task in computer vision development.

Voxel51's latest ML research dives into the parameters, confidence thresholds, and configurations that unlock auto-labeling performance. These benchmarks show that zero-shot labeling with foundational models like YOLO-World and Grounding DINO achieves 95% of human performance, at 100,000× lower cost and 5,000× faster speed.

The $ difference: a human annotation service labeling 1.2 million objects would take ~2,470 hours and cost $45,725, versus 27 minutes and $0.42 with Voxel51's Verified Auto Labeling — with similar or better model mAP scores.

>> Read the research

>> Join the webinar for expert analysis

The updated preview leads all models on several major benchmarks (GPQA, Aider, and LMArena) and addresses regressions in formatting and creativity from the previous 2.5 Pro update.

Eleven Labs has launched Eleven v3, an AI text-to-speech model recognized for its expressiveness. It supports a wide range of languages, including Afrikaans, Arabic, French, and Mandarin. This advancement enhances multilingual voice applications.

Anysphere, maker of AI coding assistant Cursor, raised $900M at a $9.9B valuation as ARR surged past $500M.

LLMs aren't giving bad actors fundamentally new capabilities. OpenAI has published 10 examples of models accelerating existing hacking, fraud, and misinformation operations. This includes North Korean actors using GPT to scale fraudulent IT worker schemes, Russian groups developing sophisticated malware, and Cambodian scammers generating multilingual “task scams” promising victims $500/day for liking TikTok posts.

This 4-hour session, presented during the AI Engineer World's Fair 2025, covers recent innovations in search and recommendation systems.

To address label uncertainty in LLM-based annotation, this paper introduces a method that captures multiple possible labels and uses a teacher-student framework called CanDist to distill them into a single output.

When OpenAI's o3, Claude, and DeepSeek-R1 were tested in puzzle environments, their performance collapsed beyond certain complexity thresholds, despite generating detailed "thinking" processes. The models exhibit a counterintuitive scaling limit where their reasoning effort actually decreases as problems become more complex, and they fail to improve even when given explicit solution algorithms.

Claude Composer CLI is a tool that enhances Claude Code with automation, UX, and configuration. It reduces interruptions while giving users flexible control and tools to configure Claude. The tool provides users with system notifications to keep them informed. Users can control which permission dialogs are automatically accepted.

Modify Video allows professionals to reimagine environments, lighting, and textures in videos without altering motion or performance. It offers tools for restyling, retexturing, and editing specific elements like wardrobe and props. Outperforming competitors, Modify Video maintains motion consistency, offering multiple output variants and using advanced performance signals for high-fidelity creative control.

Google Labs launched Portraits, an AI coaching tool featuring experts like Kim Scott, to provide AI-driven guidance. The tool uses Gemini's capabilities to simulate expert advice through interactive avatars.

Anthropic trained custom models for the US government optimized for intelligence and defense use cases that have already been deployed in classified environments.

Tokasaurus is a large language model inference engine optimized for throughput-intensive workloads.

Container Use is a tool that creates development environments for coding agents, enabling multiple agents to work safely and independently with any stack.