TLDR AI 2024-08-29

To build for the new AI era, APIs and developer platforms need to evolve. Join API Summit by Kong (free, virtual) to explore the intersection of AIs and APIs, alongside technical leaders from Kong, Goldman Sachs, Axa France, Maersk, and others.

↗️ Save your virtual spot at API Summit by Kong (FREE)

↗️ Join the pre-event hackathon to compete for cash prizes

On the agenda: Go beyond buzzwords to find out what's working today — and where the industry is headed when it comes to LLMs, dedicated clouds, microservices, and automation.

See you there!

Midjourney is expanding into hardware with a new team based in San Francisco. The move, hinted at by the hiring of ex-Neuralink and Apple Vision Pro engineer Ahmad Abbas, suggests that the hardware could be related to its ongoing development of AI models for video and 3D generation.

OpenAI is in discussions to raise several billion dollars in a new funding round led by Thrive Capital that would value the company above $100 billion, with Microsoft also expected to participate.

Google is rolling out its new features, Gems and Imagen 3, to Gemini Advanced subscribers. Gems allows users to create custom versions of Gemini tailored to specific tasks, offering premade options like learning coaches and coding partners, while Imagen 3 is Google's latest image generation model, now available to generate detailed and artistic images.

Table-Augmented Generation is a new paradigm that combines language models with databases to answer complex natural language questions.

Cohere's Aya model was able to dramatically improve its win rate vs baseline models by taking advantage of the difference between oracle model performance and routing synthetic data generation.

Reward models are generally trained as discriminative classifiers. This work from DeepMind uses a language model's logits for yes/no as the reward signal. It found that by allowing a model to use CoT and ensembling, researchers improved performance by 16%.

OctFusion is a method for generating 3D shapes with diffusion models that achieves both efficiency and high quality. It can produce 3D shapes at arbitrary resolutions in just 2.5 seconds on a single Nvidia 4090 GPU.

Researchers have demonstrated that large Transformer models can be distilled into more deployment-friendly linear RNNs, reusing weights from attention layers.

How do you know if your LLM output is any good? Human feedback is the gold standard — but it's expensive, slow, and inconsistent. This webinar by Kolena introduces cutting-edge techniques to fine-tune LLMs according to the preferences of human raters. Join the live session next Wednesday (September 4) for the best experience →

Researchers have developed a new approach to video deblurring by integrating an event camera, which captures motion with microsecond temporal resolution, to enhance the clarity of motion-blurred videos.

Cartesian has announced a number of model and system improvements. It has also released an open hybrid State space model.

Despite general headwinds for startups, AI ventures continue to secure substantial funding. U.S. AI startups have achieved nearly 30 deals over $100M in 2024, with Europe not far behind. Major investments include WAYVE ($1B), Mistral AI (~$1B), Helsing ($484M), Poolside ($400M), DeepL ($320M), H ($220M), and Flo Health ($200M).

Stephen Wolfram emphasizes the importance of incorporating philosophical rigor into AI research, addressing core ethical questions. As AI's influence grows, these questions, traditionally part of philosophy, are increasingly relevant. Wolfram advocates for deeper, classical philosophical thinking in AI development to better navigate its implications for humanity.

NVIDIA's Mistral-NeMo-Minitron 8B, a compressed version of the 12B model, achieves state-of-the-art accuracy while being small enough to run on RTX workstations.

An open source VLM designed specifically for image upcaptioning.

Boson AI has launched RPBench-Auto, an automated evaluation pipeline to benchmark LLMs' role-playing capabilities inspired by ArenaHard and Alpaca Eval.

AI's impact on HR goes far beyond CV screening. Alan Whitaker, Head of AI at BambooHR, offers some fascinating insights from the front lines of innovation. Get the learning kit