TLDR AI 2024-06-27

Futuristic technology like AI requires future-friendly power sources. With governments scrambling to divest their coal and gas interests, lithium batteries are set to drive the world into a new age — where remarkable feats of artificial intelligence are made possible by electrification and renewable energy.

This means that demand for lithium is soon to drastically outstrip supply, causing this vital mineral to skyrocket in price. GM-backed EnergyX is primely placed to meet this shift. With patented technology that extracts lithium 300% faster, the company has effectively cornered the market overnight.

Their eagerly-awaited open-to-the-public investment round is now live.

Don't miss your chance. Invest in a renewable future.

Evolutionary Scale, a biology startup, just emerged from stealth with big backers. It also announced the release of its foundation model, ESM 3, which is a 98B parameter model trained on 771B biology tokens for 10^24 Flops. It used the model to discover a novel fluorescent green protein unlike any known in nature.

YouTube is negotiating with major record labels for licenses to use their music in AI song-generation tools. It is offering substantial upfront payments but facing resistance from artists worried about devaluing their work. The platform, which trialed a generative AI tool called "Dream Track," aims to expand its AI music capabilities, potentially integrating them into YouTube Shorts. Meanwhile, the music industry is pushing back against unlicensed AI usage, with leading companies suing AI startups Suno and Udio for copyright infringement.

Waymo One now allows anyone in San Francisco to hail a ride with its autonomous vehicles. The company is scaling up operations after offering tens of thousands of trips weekly. Its all-electric fleet supports the local economy and its sustainability goals. Citing enhanced safety, Waymo says its vehicles are involved in significantly fewer crashes compared to human drivers.

Actively selecting the next best batch is a challenging and open problem in pre-training. This work from DeepMind explores how to spend just 10% of the FLOPs and hard mine negative samples and still match SOTA for various tasks.

Retrieval Augmented Instruction Tuning (RA-IT) improves information extraction using large language models.

Golden Gate Claude was a powerful example of steering and interpreting models with SAEs. This work provides an intuitive explanation of how it works and has some example code for training these models.

Director3D is a framework designed to enhance real-world 3D scene generation and camera trajectory modeling. Utilizing a Trajectory Diffusion Transformer and a Multi-view Latent Diffusion Model, Director3D generates realistic 3D scenes from textual descriptions.

RES-Q is a new benchmark for assessing the ability of large language models to edit code repositories based on natural language instructions.

A great prompting toolkit almost entirely made from Sonnet 3.5 that helps test different prompts' performance.

Spam starts with your personal data being sold. The culprit: Data brokers who profit from your home address, SSN, phone number, and more. The solution: Incogni scrubs your personal info from the web. 3 minutes to set up, 180+ data brokers' and people-search sites tackled.

Use code TLDR60 for a 60% discount →

This post shares two critical approaches that help take Agents out of the playground and into the real world: long-term planning and system-level robustness. These introduce capabilities to develop higher-order plans for Agents while allowing mid-episode adaptation as well as systems approaches to intelligent orchestration of the models that provide higher accuracy and performance.

Online travel has yet to crack the end-to-end trip experience. AI offers fresh strategies for rebundling travel services, potentially shifting profits. Decision support and coordination enabled by AI will be crucial for companies to assert control over the travel rebundling process. The travel industry's future may see incumbents leveraging AI for competitive advantage, while traditional travel agents face market erosion due to AI efficiencies.

Slack engineers utilized a hybrid approach, combining Abstract Syntax Tree transformations and AI Large Language Models, to automate the conversion of over 15,000 unit tests from Enzyme to React Testing Library. The team achieved an 80% success rate in automated conversions by integrating Anthropic's Claude 2.1 AI model and capturing the DOM tree for React components. This innovative project underscores Slack's commitment to enhancing developer experience and productivity through the application of AI. It is part of ongoing efforts to stay ahead in the ever-evolving frontend landscape.

If you're looking to get in front of a highly engaged tech audience, consider advertising in TLDR. We have a subscriber base of 4+ million tech professionals to effectively reach your target audience. Learn more.

R2R is an RAG engine with a RESTful API and prod features.

This 7b model, trained on medical data, tied with GPT-4 on human preference evaluation when measured by medical doctors.

The ChatGPT desktop app for macOS is now available for all users.