TLDR AI 2025-05-30

Model APIs

Baseten just launched Model APIs, on-demand frontier models including DeepSeek V3, R1, Llama Maverick and Scout, purpose-built for production workloads on their optimized Inference Stack. Try them out.

📈Top of the OpenRouter leaderboard in latency and throughput.

🛠️Support tool calling and structured outputs out of the box, ideal for agents.

🤝Fully OpenAI-compatible, so Model APIs can be used as drop-in replacements.

🧪More pre-optimized models are added regularly.

Training

Because one product launch isn't enough, Baseten also released Training (Beta). Built for power users, Training provides the infrastructure to fine-tune without limits using any model or dataset. Get access.

Perplexity has introduced Perplexity Labs, allowing Pro users to go from idea to execution by generating reports, spreadsheets, dashboards, and simple apps using tools like web browsing and code execution.

Black Forest Labs released FLUX.1 Kontext, a set of flow-matching models that enable text-and-image-based in-context image editing and generation.

The tools generate "attribution graphs" to trace how large language models make decisions internally, revealing the step-by-step reasoning process behind outputs. The library works with popular open-weight models and includes an interactive Neuronpedia frontend for exploring model circuits.

Web Bench is a new dataset for evaluating web browsing agents. It consists of 5,750 tasks on 452 different websites. Anthropic Sonnet 3.7 CUA is currently state-of-the-art according to Web Bench.

Resemble AI released an open-source TTS model that outperforms ElevenLabs in benchmarks and features emotion exaggeration controls.

RenderFormer is a neural renderer that produces photorealistic images directly from triangle-based scene representations with full global illumination. It doesn't require any per-scene training or fine-tuning.

Meta researchers introduced a method called zero-shot grafting, which trains a vision encoder with a smaller surrogate model derived from a large LLM's shallow layers. This reduces VLM training costs by ~45% while preserving or even improving performance when transferred to the full LLM.

'Max-performance domains' are industries where people can be world-class by being very good at only one part of their job. It doesn't matter if you're incompetent at adjacent skills as long as you can create seminal impact. Exceptional ability at a single thing outweighs incompetence at other parts of the job. It is a luxury to work in a max-performance domain as failure is allowed and stress is usually self-imposed.

NVIDIA reports notable growth with a shift from simple AI to more complex reasoning, driving a massive increase in AI demand. Hyperscalers are deploying nearly 72,000 GPUs weekly, with Microsoft alone seeing a fivefold increase in token generation. Increased demand requires more data centers, termed "AI factories," despite efforts to reduce model sizes.

The data shows that people are making things for themselves, not the world, but there is clearly builder excitement.

MedGemma is an open-source model built on Gemma 3 that comes in 4B multimodal and 27B text-only variants.

Delaware's attorney general has hired a bank to independently assess OpenAI's restructuring plan, potentially complicating the company's move to a for-profit model.

Just a month after its acquisition of Pollen Robotics, Hugging Face has announced HopeJR, a full-size humanoid, and Reachy Mini, a desktop robot for testing AI applications.

Elon Musk was frustrated that OpenAI, led by personal rival Sam Altman, was selected as a partner in the UAE AI campus deal.

A developer tested 46 different AI coding tools, providing detailed comparisons and use cases for each platform.