The week in AI: wearables becoming the next big thing

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each Thursday, we aggregate the major developments in artificial intelligence; we pass along the news, useful resources, tools or services, and exciting projects in open source. Even if you aren’t an engineer, we’ll keep you in touch with what’s going on in AI.

Amazon Web Services (AWS) announced the general availability of Amazon Bedrock, a fully managed service that provides access to a variety of foundation models from leading AI companies through a single API. Bedrock offers capabilities to build and deploy generative AI applications, while maintaining privacy and security. The service allows customers to find, test and customize high-performing foundation models to determine which works best for their use cases.

They also announced new capabilities in CodeWhisperer and QuickSight to increase developer productivity and accelerate analytics. Enterprise customers will be able to customize CodeWhisperer's code suggestions using their internal APIs, libraries, packages, etc.

Google has introduced a new user control called Google-Extended that websites can add to their robots.txt file to opt out of having their content scraped to train Google's generative AI products like Bard. By adding "User-agent: Google-Extended Disallow: /", sites can prevent Google from scraping - without impacting regular Google crawling and search indexing. It’s pretty sad that the default is being ‘opted in’, but that’s the status quo for now.

Humane touts the pin as privacy-first and stylish, but it remains unclear how the device connects to the internet and whether the (admittedly very impressive) demos reflect real-world functionality. The public reveal aimed to showcase the pin's capabilities and ‘fashionable form factor’ before the official unveiling next month.

AI wearables are starting to blow up more broadly - Rewind’s pendant and Meta’s Ray-Bans were also in the news this week.

Wharton professor and leading AI-researcher Ethan Mollick takes a look at the state of affairs in AI. He argues that while the full implications of current AI systems remain unclear, we can see the general ‘shape of their shadow’. He points to advancements like Google's upcoming Gemini model and multimodal systems with vision, voice, and knowledge connections that suggest AI’s rapid evolution. We see this too: tools and AI services are getting better - fast. Mollick contends that how this plays out depends on human decisions and regulations that empower - rather than remove - agency. He calls for responsible individual and corporate actions so that AI can uplift and simplify lives rather than cause harm.

More in AI this week:

• Slovakia’s election deepfakes show AI is a danger to democracy

• Los Angeles is using AI to predict who might become homeless and help before they do

• Microsoft CEO warns of ‘nightmare’ future for AI if Google’s search dominance continues

• Canva launches Magic Studio - all the magic of AI in one place

• Generating Change - a global survey of what news organizations are doing with artificial intelligence

• Why AI is teetering on the edge of a disillusionment cliff

• Every product unveiled at the Made by Google event

• LinkedIn introduces new AI-powered products

• Translating Latin demonology manuals with GPT-4 and Claude

While most current 3D software doesn't support Gaussian Splatting, some plugins exist for Unity and Unreal Engine. The technique is considered to have a relatively easier pathway for integration into existing renderers compared to other radiance-field-based techniques like NeRFs.

They also can’t generalize beyond their (finite) training sequence lengths. StreamingLLM addresses these challenges by only retaining the key-value states for recent tokens and "attention sinks," which are special tokens that help guide the model's reasoning. All intermediate tokens are discarded. This aids LLMs to generate coherent outputs using only the most recent context - without needing to reset their cache.

StreamingLLM achieves up to 22x faster inference (the speed at which the model can generate coherent text continuations) compared to traditional methods on streaming tasks like multi-turn dialogue. It paves the way for persistent chatbots and AI assistants that can remember details across very long conversations - pretty amazing breakthrough!

More AI open source & technical this week:

• Anthropic on the challenges in properly evaluating AI systems

• Perplexity has launched a blazing fast API for open source LLM’s

• (Open source) DocsGPT is a cutting-edge open-source solution that streamlines the process of finding information in project documentation.

• The fallacy of low-code/no-code and generative AI

• AI language models can exceed PNG and FLAC in lossless compression, says study

• Microsoft’s AutoGen framework allows multiple AI agents to talk to each other and complete your tasks

• Animate anything: new AI tool could be a game-changer for 3D animation

That’s it for this week! Thanks for reading and we’ll see you next Thursday.