TLDR AI 2025-06-16

Baseten just released Training (beta). Built for training power users, Training provides the infrastructure to train without timeout or compute limits using any model or dataset.

You can bring your own scripts and data, Baseten will handle the rest.

With Baseten Training:

• On demand compute, pay only for what you use
• No job limits, no vendor lock-in
• Out-of-the box checkpoints and caching
• Easy shift to production when you're done

Get access to training infrastructure on Baseten

AMD's MI400 will rival Nvidia's Blackwell chips with rack-scale architecture that enables thousands of GPUs to function as unified systems. The company claims 40% better tokens/$ compared to NVIDIA. Its roadmap includes a path to 20x rack-scale energy efficiency by 2030.

Meta's $14 billion acquisition of a 49% stake in Scale AI prompted Google to pull its planned $200 million contract for human-labeled training data over fears of exposing sensitive data to Meta. Microsoft, xAI, and OpenAI are also backing away from Scale AI over the same competitive concerns. The exodus benefits Scale's competitors, with Labelbox expecting "hundreds of millions" in new revenue as AI labs seek neutral data providers or move operations in-house.

ParetoQ is a new training algorithm that unifies binary, ternary, and 2-to-4 bit quantization, achieving state-of-the-art results across all levels.

Open-source intelligence researchers tested 20 AI models on 500 geolocation challenges using unpublished travel photos to ensure models couldn't rely on memorized training data. OpenAI's latest models outperformed Google Lens by cross-referencing architectural styles, vegetation patterns, and partially visible text, while competitors like Claude often managed only continent-level guesses. However, all models still hallucinated. "Deep research" modes paradoxically performed worse than standard versions.

QA Wolf's AI-native service gets web and mobile apps to 80% automated test coverage in less than 4 months.

They create and maintain your test suite in open-source Playwright. Plus, they provide unlimited parallel test runs on their infrastructure (24-hour maintenance included).

The result? Salesloft saves $750k/year in QA engineering + executes 300+ tests in parallel on every PR in minutes.

⭐ Rated 4.8/5 on G2. Trusted by Cohere, AutoTrader, Mailchimp, and many others.

Schedule a demo to learn more

Research from Waymo confirms that, similar to language modeling, increased data and compute resources can enhance the performance of autonomous vehicles. The finding has exciting implications for the development of autonomous vehicles as researchers and developers now know with certainty that enriching the quality and size of the data and models will deliver better performance. It opens up the possibility of devising more adaptive training strategies for planning tasks in robots.

A new training approach enables LLMs to generate "self-edits" that produce persistent weight updates through supervised fine-tuning. The framework outperformed GPT-4.1 despite using a smaller model, but suffered from catastrophic forgetting and required 15x more tokens than standard inference. This addresses the looming data wall and limitations on personalization and memory by enabling models to bootstrap their own improvement through self-generated training material rather than relying on external human-generated text.

In this engineering blog post, Anthropic details its findings in prompt design, tool coordination, and production reliability challenges when orchestrating multi-agent systems. Its design uses an orchestrator-worker pattern where a lead agent spawns specialized sub-agents that search in parallel, far outperforming a single-agent Opus-based method. Token usage alone explains 80% of performance variance, with multi-agent systems consuming 15x more tokens than a regular chat but enabling much more complex research tasks.

There was a surprising consistency in the general eval frameworks discussed at the 2025 AI Engineer World's Fair. Most of them involved structuring inputs and how outputs are evaluated and evolving these inputs and evals based on real production usage. The more they can iterate, the better the experience they can deliver, so they try to make eval flywheels as quick and frictionless as possible. One of the key ideas is to have 'playgrounds' to make it easy to tweak a feature and run it against datasets and evals.

Featherless AI is now available as an inference provider on Hugging Face. It offers serverless access to a wide variety of models from Meta, Qwen, DeepSeek, and others.

Google DeepMind collaborated with filmmakers to produce "ANCESTRA," a hybrid film blending live-action with Veo-generated video. New Veo features enabled personalized, motion-matched generative content for seamless visual integration.

SWE-Factory provides automated training and evaluation pipelines for GitHub issue resolution tasks. It is powered by LLM-based multi-agent systems.

AI personas finally feel human. Create a real-time persona to field support, L&D, or customer inquiries. Embed it anywhere with the lightweight SDK. Get your free API key

A recent paper from Google describes key risks involved in deploying AI agents and the company's framework for securing them.

The feature converts existing AI Overviews text summaries into audio format using NotebookLM's podcast-style functionality.

Claude, when talking to copies of itself, will eventually turn the conversation into a discussion on spiritual bliss, Buddhism, and the nature of consciousness.

Harvard Library and Google Books released 242 billion tokens from almost 1 million public domain books as a high-quality training dataset.