Parse notes
CEO Mark Zuckerburg recently discussed his goals with open-sourcing significant expensive AI tools in a podcast. This collection of music generation code, model weights, and evaluations is an example of just that. It allows you to generate both melody conditioned and text conditioned music. The code is fully open sourced, but model weights are not commercially open.
Bloomberg journalists found that Stable Diffusion takes racial and gender disparities to extremes that are even worse than those found in the real world, stroking fears that generative AI will perpetuate biases.
Khan Academy has introduced an AI-powered chatbot, Khanmigo, which aims to provide personalized tutoring to students. Powered by GPT-4, the bot offers guidance on a broad range of subjects and uses the Socratic method to encourage learners to solve their problems independently. While proponents believe such bots could revolutionize education, concerns about the technology's potential for error and impact on teachers' roles persist and its effectiveness is currently being studied ahead of a broader rollout in U.S. schools.
A lot of new acronyms are introduced in this work. The gist is that researchers collected a new dataset for visual instruction tuning, trained a model, and made some algorithmic tweaks along the way. The demo is really impressive, especially after the AR reveal from Apple. An exciting peak at what AI assistants might look like in the future.
Language models are well known for confidently confabulating (sometimes called hallucinating). It turns out that these models have an internal representation for truthfulness and by tweaking the model activations you can elicit a truthful response. The researchers found that by using this technique on the instruction-tuned Alpaca model, they improved TruthfulQA performance from 32.5% to 65.1%.
The paper presents Sparse-Quantized Representation (SpQR), a new format and technique that allows almost lossless compression of large language models (LLMs), overcoming the usual accuracy losses from quantization. SpQR enables powerful LLMs to run on common devices like laptops and mobile phones without performance degradation, offering over 4x memory compression gains and faster inference than traditional methods.
Advances in computer vision seemed to slow after autonomous vehicle research lost some steam, but things are picking up again. This novel technique uses a hydra of three image models to enhance the performance of natural image matting. The results are quite compelling.
In 2015, there was a breakthrough in representing continuous values with a 6D number. This work rhymes in that they propose representing videos as a 3D volume. By matching between that representation and pixel space, they can perform quite long term tracking of arbitrary pixels in a video.
Lanarky is an open-source framework to deploy LLM applications in production.
An exploration into the eerie human-like features of LLMs and how they hold promise for some fields and trouble for others.
U.S. senators introduced two separate bipartisan artificial intelligence bills on Thursday amid growing interest in addressing issues surrounding the technology. One would require the U.S. government to be transparent when using AI to interact with people and the other would establish an office to determine if the United States is remaining competitive in the latest technologies.
Two German researchers, Sophie Jentzsch and Kristian Kersting, released a paper that examines the ability of OpenAI's ChatGPT-3.5 to understand and generate humor. They discovered that ChatGPT's knowledge of jokes is fairly limited: During a test run, 90 percent of 1,008 generations were the same 25 jokes.
Tech giants, universities, and nonprofits are working to provide lessons on artificial intelligence to students in schools. There is currently no national consensus on what should be taught about the technology. Current curriculums cover a range of topics, from the basics of AI to how to program Alexa.