TLDR AI 2025-02-21

The shift from seat-based pricing models to usage-based pricing (UBP) has been going on for a while — but AI is pushing companies to evolve their pricing models further and faster.

Kyle Poyar is an expert on SaaS pricing and on February 26th, he will be joining Metronome's new pricing webinar series, where they'll discuss:

• UBP and aligning pricing with customer value
• New trends such as AI-driven pricing and outcome-based models
• 2025 pricing predictions and how to stay competitive

Join live next Wednesday →The Future of SaaS Pricing Webinar with Kyle Poyar

Together AI announced a $305 million Series B funding round led by General Catalyst and Prosperity7, with participation from investors like NVIDIA and Salesforce Ventures. This investment will enhance Together AI's position as a leader in AI Cloud solutions, with a focus on supporting open-source AI models and NVIDIA Blackwell GPU deployment. Together AI's platform enables efficient AI application development, offering enterprise-grade inference, training, and model fine-tuning. The company has planned infrastructure expansions to support rapid ecosystem growth.

Eudia has secured $105 million in Series A funding to transform legal operations through augmented intelligence, enhancing the capabilities of legal teams rather than replacing them. By partnering with notable Chief Legal Officers, Eudia aims to streamline legal processes and shift legal departments from being cost centers to strategic value drivers. This marks a significant shift towards AI-empowered legal functions within Fortune 500 companies.

Spotify now allows audiobooks narrated with ElevenLabs' AI voice technology, enabling authors to upload AI-narrated works through Findaway Voices. The feature supports narration in 29 languages, although publishing requires passing a review process.

Report for the strongest open weights vision language model from the Qwen team.

Interesting proof of concept paper that uses Photorealistic environments rendered with 3D Gaussian splats. It achieves a 3x lower collision rate overall.

This paper introduces Model-Guidance (MG), an objective that improves diffusion model training by incorporating posterior probabilities, significantly boosting training speed, inference efficiency, and achieving state-of-the-art results on ImageNet 256 benchmarks.

Open replication of the zero reasoning paradigm. It is fully open source with training data, scripts, and weights.

Most data from web crawls gets thrown away due to poor text extraction performance. This work dramatically improves how many documents get saved and uses that to train stronger models on a fraction of the crawled data.

LLM-Oasis is a large-scale dataset designed to train and evaluate systems for assessing factuality in outputs from LLMs.

An interesting side project from the community. This was a model fine-tuned from Flux Schnell, which means it is Apache licensed. It is designed to be fine-tuned at 8B parameters and runs quite quickly while not sacrificing performance.

Development roadmap for the next year for the PyTorch compilation team.

Wenfeng Liang, a hedge fund manager, launched DeepSeek, a self-funded open-source AI platform that has rapidly gained global attention for its innovative LLMs like DeepSeek-R1, comparable to OpenAI's models. Using more cost-effective training methods and consumer-grade hardware compatibility, DeepSeek has sparked interest among both major tech companies and small institutions. Liang's focus on open-source AI development, backed by his success with Magic Square Quantitative, emphasizes collaboration and technological progress over commercial pressures.

Neuronics Medical is a medical device company developing AI for acute stroke detection.

A nice flutter app that provides a simple front end to Ollama.

Google Lens now lets iPhone users visually search directly from their screen within Chrome or the Google app using gestures like highlighting or tapping.