Parse notes
Sponsored
The world's smartest companies use Redis as the foundation of their AI infrastructure. Find out how they're building with Redis for AI and more at Redis Released: Worldwide—a virtual event [4] for anyone building the fast apps our world runs on.
> Our latest releases like Redis Copilot, Redis 8, RDI, and more can help you build simpler
OpenAI has released some work discussing improvements to consistency models (which generate in two steps as opposed to 1,000 for diffusion). They still rely on distillation from an existing diffusion model, but this work aims to improve and stabilize them as they scale.
SynthID is a way to statistically watermark generated text. It uses a pseudorandom function after top-k and top-p in the sampling process to mark text. Then you can use a probabilistic bayesian method to detect if text has been watermarked - meaning it was generated by a language model.
Transformers JS is a JavaScript library for running models. It now uses WebGPU, which is up to 1,000x faster in some cases. The new version has 1,200+ models ready to use for edge and browser-based applications.
Salesforce's new method uses a novel encoder for video that requires substantially fewer tokens for proper representation. This has been tried a number of times in the past with minimal success, the key seems to be an explicit temporal encoder along with a spatial encoder.
A team at CMU has released powerful open models with strong performance in a variety of languages.
SAM2Long addresses the "error accumulation" issue in SAM 2's memory design by introducing a training-free video object segmentation strategy.
Sponsored
Ever find yourself thinking “I wish I could turn this prompt into an automation”? Now, anyone on your team can: Bardeen is an AI agent [11] that helps companies like Deel, Miro, and Slack stay on top by fully automating sales prospecting, marketing ops, demand generation, and much more. All it takes is a single prompt. Book an exploratory
A nice little wrapper around Anthropic's computer use system that makes it much easier to use and run.
TALoS is a method that improves scene completion for autonomous vehicles by using observations made at different moments as supervision for predictions.
Anthropic's Alignment Science team developed evaluations to test AI models' potential for sabotage, including human decision sabotage, code sabotage, sandbagging, and undermining oversight. These evaluations aim to identify dangerous capabilities in future AI systems and inform necessary safety mitigations. Their findings suggest minimal risks with current models but highlight the need for more robust evaluations and safeguards as AI capabilities advance.
Microsoft's research paper on AI robustness prompted the company to invest billions in AI infrastructure, catalyzing breakthroughs with partners like OpenAI. This investment has significantly boosted Microsoft's growth in AI-driven products, exemplified by GitHub Copilot's success. Despite challenges from competition and sustainability goals, Microsoft continues to prioritize AI, with record capital spending on its AI and cloud infrastructure.
In this article, Cloudflare CEO Matthew Prince discusses AI's impact on internet infrastructure, highlighting the need for AI-capable edge computing and local inference to reduce network latency. He emphasizes the importance of regionalization in AI services due to regulatory complexities and outlines Cloudflare's approach to building a connectivity-focused network. Cloudflare aims to make internet connectivity faster, more secure, and more efficient, aligning closely with AI advancements.
OpenAI and Microsoft are negotiating the terms of their partnership, with Microsoft seeking a significant stake in OpenAI post-restructuring.
Mira Murati, former OpenAI CTO, is reportedly raising over $100 million for a new AI startup focused on proprietary AI models.
Midjourney plans to release an upgraded web tool for editing uploaded images using its generative AI.