TLDR AI 2024-06-26

With Vultr, unlock immediate access to the latest NVIDIA GPU's for AI machine learning. Receive the flexibility you need with Vultr's on-demand virtual machines or opt for the power of large-scale dedicated clusters. With 32 data center locations across six continents, Vultr delivers seamless, low-latency access to cloud infrastructure.

With Vultr, customers receive:

• Immediate access to NVIDIA GPUs, including H100 and GH200s
• Reduced latency from Vultr's global network of data centers
• Freedom from vendor lock-in with Vultr's flexible and scalable infrastructure
• Simple, affordable, and predictable pricing

Power your GenAI initiatives with Vultr.

Apple is segmenting its AI approach, differentiating between 'context' and 'world' models, with the latter, such as ChatGPT, being used for open-ended queries but without personal context. OpenAI's position is not highly defensible, given the lack of moats in large language models (LLMs) and the rise of competing models. Apple's strategy suggests a commoditization trend for LLMs that leverages on-device capabilities and private cloud to provide AI features while viewing services like ChatGPT as interchangeable components.

Etched has announced a $120M round led by Primary Venture Partners and Positive Sum Ventures, as well as angel investors including Peter Thiel, Stanley Druckenmiller, and David Siegel. It is working directly with TSMC's Emerging Business Group to produce chips on their state-of-the-art, 4 nanometer node. Its chips can achieve 500K tokens/second. One 8xSohu can replace 160 H100s.

In addition to a new CEO, Stability AI was able to secure a “significant infusion of capital” from existing and new investors.

Together AI and Morph Labs have put together a great blog post on tuning models for retrieval augmented generation. They showcase some uses of synthetic data as well.

Imbue has trained and released an extremely powerful 70B language model. It uses Imbue's custom optimizer and some great data filtering techniques. The model was trained with zero loss spikes.

EvalAlign is a new metric designed to improve the evaluation of text-to-image generative models. Unlike existing metrics, EvalAlign offers fine-grained accuracy and stability. It focuses on image faithfulness and text-image alignment.

The PyTorch team has written custom kernels to make use of sparse cores, which are usually reserved for inference only.

Microsoft's new many-to-many vision model can be tuned for specific downstream tasks. It isn't quite as powerful as PaliGemma, but is easy to run in PyTorch.

Bring inference closer to your users with Gcore. Use a single endpoint that's automatically deployed where you need it and never worry about managing GPUs and containers. Reduce latency, enable ultra-fast responses for your AI-powered apps, and double the value you deliver to users. Create a free account

FreeTraj is a tuning-free method to control the motion trajectories in video generation using diffusion models. It modifies noise sampling and attention mechanisms to guide the generated content.

Lamini has developed Memory Tuning to reduce hallucinations by 95%, embedding specific data into models more efficiently without compromising general knowledge.

GeoMFormer is a new Transformer-based model designed to improve molecular modeling by learning both invariant and equivariant features.

Researchers were able to make an LLM voice assistant more efficient and scalable by categorizing data, precomputing embeddings, and dynamically generating examples.

LLM Radar is a free tool that scans for new LLMs and delivers them straight to your inbox so you can stay ahead in the AI game without any effort.

Turn any glasses into AI-powered smart glasses.

OpenAI's advanced Voice Mode, demonstrated during its Spring Update, will launch in alpha to select ChatGPT Plus users in late July.