TLDR AI 2025-04-10

Leading retailers like Kiko Milano, Flaconi, and Maisons Du Monde are using AI to boost conversions, automate repetitive tasks, and gain full control of their merchandising—without relying on developers.

Want to know the secrets behind their success? Watch the Algolia webinar recording and learn how to turn your merchandising challenges into growth opportunities — from dynamic re-ranking to advanced personalization.

Watch the recording here →

Grok 3 has an API for its models, including multiple reasoning modes. This link includes pricing, which is reasonable given the assumed size of the model itself.

With improvements on podsize, bandwidth, memory, and overall flops, the 7th generation of Google's TPUs is a step change improvement over any chip on the market.

The new Max plan for Claude offers expanded usage with priority access to features and models, supporting users with demanding projects. It provides two levels: Expanded Usage for $100/month with 5x usage and Maximum Flexibility for $200/month with 20x usage. The plan is ideal for those who frequently work with substantial documents and complex data.

By treating SVGs as a foreign language, a pretrained Qwen model can generate novel SVGs based on text and images. It is easily the state-of-the-art model. An open release is coming soon.

Sakana AI had a research paper accepted to an ICLR workshop that was fully generated, executed, and written by a language model system. They improved their system by using VLMs, more general purpose search, and more.

There is a debate in the language modeling world about how much a model actually learns and how much it just memorizes. This feature, in the AI2 Playground, searches through billions of input documents, in real time, to figure out if the output of a model is just regurgitated or if it is novel. It includes the source of the utterance from multiple documents.

HybriMoE is a new framework for hybrid CPU-GPU inference on Mixture of Experts models that tackles instability and overhead with smarter scheduling and caching strategies.

This research explores how LLMs structurally internalize new knowledge by analyzing computational subgraphs. It reveals patterns in knowledge acquisition, optimization phases, and potential strategies for improving continual pre-training.

ReQFlow sets a new benchmark in protein backbone generation, delivering state-of-the-art results while being significantly faster than existing models. It is 37x faster than RFDiffusion and 62x faster than Genie2 for sequences of length 300.

Google Cloud has added enhanced generative tools to Vertex AI, including new capabilities in text-to-music (Lyria), video editing (Veo 2), voice customization (Chirp 3), and image inpainting (Imagen 3).

The Cognito team has released open-source LLMs ranging from 3B to 70B parameters, all of which outperform top open models of similar sizes, with the 70B model surpassing Llama 4 109B MoE. These models are trained using Iterated Distillation and Amplification (IDA), support both direct and reflective answering, and are available on Hugging Face, Ollama, Fireworks AI, and Together AI.

Firebase Studio is launching as a cloud-based development environment for building, testing, and deploying AI applications. It combines tools like Project IDX, Genkit, and Gemini within a unified platform, offering rapid AI prototyping and a customizable coding workspace. Developers can leverage features such as natural language prototyping, instant device previews, and easy collaboration to accelerate app development.

Adobe-backed RAD Intel has raised $37M by identifying content and creators that can drive results - and it's closing a round meant to fuel explosive growth. Last day at $0.55 per share.

Shopify CEO Tobi Lütke announced that reflexive AI usage is now a fundamental expectation for all employees.

GitHub Copilot users will face rate limits on AI models beyond the base GPT-4o model, with premium requests capped for paid tiers.

Developers can now easily build real-time voice and video apps using FastRTC, with enterprise-grade WebRTC infrastructure provided by Cloudflare and simple access via Hugging Face tokens.