TLDR AI 2024-08-20

A legal loophole allows data brokers to collect your personal data and sell it to the highest bidder. This is used for all sorts of nefarious purposes - from spam email to more convincing deepfake calls (thanks, AI!).

To get your data off these websites, you need to either:

• Contact dozens of them individually; or
• Sign up to Incogni and never think about it again.

Incogni contacts dozens of data brokers on your behalf — getting your data off the web, and making sure it stays off.

✅ Over 20,000,000 data removals completed to date

Delete your personal data today.

→ Use this link get 60% off using code TLDRAI.

Former Google CEO Eric Schmidt predicts rapid advancements in AI, with potential to create significant apps like TikTok competitors in minutes within the next few years. He highlights the unpredictability and swift progress in AI, where large language models and agent-based systems with text-to-action capabilities could converge, causing tremendous economic and technological disruption. Schmidt's view reflects the substantial investments and energy requirements anticipated for leading-edge AI development, signaling a transformative period ahead.

Waymo has unveiled its Generation 6 self-driving technology, which requires fewer cameras and sensors, integrated into Geely Zeekr EVs. The Alphabet unit plans to rapidly deploy this tech, leveraging machine learning and semiconductor advancements to withstand diverse weather conditions. This update supports Waymo's scaling efforts for its Waymo One service, which is currently providing 50,000 weekly trips.

Gemini Live, Google's AI-powered voice interaction system, mimics natural conversation but struggles with hallucinations and inaccuracies. Despite incorporating professional actors for more expressive voices, it lacks the customizability and expressiveness found in competitors like OpenAI's Advanced Voice Mode. Overall, the bot's reliability issues and limited functionality make its practicality and purpose unclear, particularly as part of Google's paid AI Premium Plan.

This paper provides an extensive review of 180 benchmarks used to evaluate Multimodal Large Language Models.

SAM2-UNet is a new framework for image segmentation that leverages the powerful Segment Anything Model 2 (SAM2) as its encoder.

BLIP was the dominant multimodal model for some time. The newest version is dramatically simpler, using a pure autoregressive loss. It achieves state-of-the-art performance on a number of captioning benchmarks.

Love waking up to TLDR? Consider advertising with us. Get in front of engaged software developers, product managers, executives and other tech professionals reading TLDR newsletters every morning. Learn more about running your first campaign.

These are the first tuned models that convincingly improve on the base Llama-3.1 models from Meta.

EasyRec is a language model designed specifically for recommendation tasks. It utilizes collaborative information from multiple datasets and innovative contrastive learning objectives to create high-quality semantic embeddings.

Mesh reconstruction from gaussian splatting allows you to take a video or image sequence and construct an editable and animatable mesh output. It does so extremely quickly and efficiently, requiring only a few steps on a single GPU.

A great article on using embeddings and clever prompting to classify every available PDF on the internet based on its semantic content.

Software professionals seeking to understand advanced ML/AI papers can follow a four-stage learning roadmap: 1) Foundational math (algebra, calculus, linear algebra, probability, statistics), 2) Classical machine learning (basic regression, classification models), 3) Deep learning (multi-layer neural networks), and 4) Cutting-edge ML (Transformers, LLMs, diffusion models). Essential resources include author-created content for stages 1-2 and recommended external materials for stages 3-4. Mastery of each stage progressively equips learners to tackle complex ML papers and stay current with evolving AI research.

This research highlights the challenges of restoring vision with AI-powered cortical implants, emphasizing that neurons in the visual cortex do not function like pixels on a screen. While high-resolution simulations show potential, achieving natural vision through cortical implants requires replicating complex neuronal patterns, a task far beyond current capabilities, leading to grainy and imperfect outcomes.

Llama file now supports Whisper models and includes a variety of speed-ups and quality-of-life improvements.

Hamming is a tool that allows you to set up Character AI-like personas and test hundreds of scenarios for your voice AI systems.

Researchers have developed an adaptive approach to deep brain stimulation that tailors electrical pulses to Parkinson's patients' varying symptoms, significantly reducing symptom duration.