TLDR AI 2025-05-21

It doesn't really make sense to run a massive inference model for every small task. That's why Fastino is introducing TLMs - task-specific Language Models that are better, faster, and cheaper.

For defined tasks — such as summarization, function calling, and creative writing — TLMs are:

→ 99x faster than OpenAI

→ 17% more accurate in task-specific benchmarks

→ Able to respond in <100ms for real-time applications

And because these models are small, they don't need to burn through GPUs - so you can pay just a single flat fee.

Don't miss out on the sensible next step in language model development. Join the waitlist

After launching a private beta in December, Google has released Jules to the public. The Gemini 2.5-powered tool clones entire repositories, then autonomously writes tests, fixes bugs, and builds features while developers work elsewhere. The agentic coding landscape is divided between synchronous pair-programming assistants and fully independent agents like Devin and Jules.

Google is reentering the smart glasses space with Android XR, integrating its Gemini AI to deliver real-time vision analysis, translation, and contextual assistance through AR glasses. The rollout will begin with Project Moohan, a mixed-reality headset built with Samsung, and followed by Project Aura, a developer-focused AR glasses prototype from Xreal, and future consumer AI glasses from partners like Warby Parker and Gentle Monster.

Apple plans to open up its large language models via a software development kit (SDK) so developers can use them in their own apps. The SDK will initially only run Apple's smaller on-device models and won't allow developers to access the models running in the cloud. It will likely be announced during the Worldwide Developers Conference event scheduled to start on June 9. The event will also include an announcement for an overhaul of iOS, macOS, and iPadOS that will make them more unified and resemble the Vision Pro operating system.

Google's Deep Think is an enhanced reasoning mode for the company's flagship Gemini 2.5 Pro model. It can consider multiple answers to questions before responding. Deep Think enabled Gemini 2.5 to top LifeCodeBench, a challenging coding evaluation, and it also beat OpenAI's o3 on MMMU, a test for skills like perception and reasoning. Google will be testing Deep Think with 'trusted testers' and conducting safety evaluations before rolling out the feature widely.

KumoRFM is a pre-trained relational foundation model designed to work across any database and predictive task without task-specific training.

A report from Quora's multi-model chat platform Poe revealed dramatic shifts in AI model usage from January to May 2025, with OpenAI's GPT-4.1 family and Google's Gemini 2.5 Pro quickly gaining popularity while Anthropic's Claude models declined. The report highlights clear leaders across categories: general text is dominated by GPT-4.1, reasoning capabilities are led by Gemini 2.5 Pro, image generation is led by Google's Imagen3, and video creation is more split, with Runway currently leading.

The 2025 Semantic Layer Summit brings together experts from Blue Mercury, TELUS, HSBC and Home Depot to explore how semantic layers fuel GenAI, modern BI, and trusted decision-making. Learn how to scale AI with governed data models, reuse KPIs across tools, and accelerate innovation.

👉 Free to attend—secure your spot now.

Kevin Murphy, a widely-cited Google researcher, just updated his 200-page reinforcement learning textbook, which covers everything from traditional approaches to cutting-edge developments like DPO, GPRO, and reasoning.

François Chollet and the ARC Prize team have released ARC-AGI-2, a significantly more challenging version of their abstract reasoning benchmark. Initial testing shows even the most advanced AI systems struggling, with o3 scoring just 3% compared to 53% on the original benchmark, while human participants successfully solved 75% of tasks on average. The ARC Prize 2025 competition offers $1 million in prizes, including a $700,000 grand prize for the first team to achieve 85% accuracy.

DeepSeek researchers have shared insights from DeepSeek-V3 as a case study in hardware-model co-design. They addressed LLM scaling limitations and using innovations like Multi-head Latent Attention, Mixture of Experts, FP8 training, and a Multi-Plane Network Topology to maximize GPU efficiency and minimize communication overhead.

Google Meet now supports real-time speech translation, using DeepMind's audio language model to preserve voice, tone, and expression across languages.

Klavis AI simplifies integration with production-ready MCP servers, offering stable connections, built-in authentication, and multiple client support. It supports custom MCP servers with 100+ tool integrations, optimizing AI application scaling. Hosted solutions include creating new MCP instances and setting up OAuth for seamless operation.

Vercel has launched an AI Gateway for alpha testing, enabling easy switching between ~100 AI models without managing API keys or accounts.

Google is rolling out AI Mode in Search to all U.S. users, offering a deeper, multimodal search experience with advanced reasoning, follow-up questions, and fast AI-generated overviews.

Google has introduced Imagen 4 for high-fidelity image generation, Veo 3 for video, and Lyria 2 for music, all available on Vertex AI.