TLDR AI 2024-08-09

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

💡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.

1️⃣ See the full agenda and speaker list

2️⃣ Join the pre-event hackathon to compete for cash prizes

3️⃣ Save your virtual spot at API Summit by Kong (FREE)

👋 See you there!

Hugging Face serves and stores a lot of data, most of it in LFS. XetHub has written its own, powerful, alternative for scaling Git repositories.

Humane's AI Pin faced a significant return rate post-launch with negative reviews, leading to approximately 7,000 of the sold units remaining with customers. Despite raising over $200 million from tech luminaries, Humane struggled with over $1 million in returns and internal leadership turnover. Amidst financial woes, there's speculation about a potential HP acquisition and new debt fundraising efforts.

Aligning a voice system is challenging. This post highlights the ongoing efforts to ensure the multimodal model is safe and helpful.

SAM2-PATH is a new framework designed to enhance semantic segmentation in digital pathology.

The Qwen team has released a 72B math model that outperforms all open and closed models on MATH. It also outperforms Llama-3.1-405B on a number of reasoning benchmarks. It is limited to English only, with multilingual models coming soon.

Speech-MASSIVE is a new multilingual Spoken Language Understanding (SLU) dataset. It offers a speech-based counterpart to the MASSIVE textual corpus.

PyTorch has released a new API that allows any type of attention variant to be coded and compiled to Triton. This allows for better portability, performance, and research velocity on types of attention.

Lighthouse is a user-friendly library designed to tackle the challenges in video moment retrieval and highlight detection.

The "1.5-Pints" Language Model presents a new approach to compute-efficient pre-training. By curating a high-quality dataset of 57 billion tokens, this model surpasses Apple's OpenELM and Microsoft's Phi in instruction-following tasks, as measured by MT-Bench.

Learn to harness small open source models to overpower the big closed source providers, all while controlling your own data.

Watch this recorded webinar session to see:

• Efficient LLM optimization techniques for open source models
• 1.23x more quality and 25x lower cost for Llama3.1-8B over GPT-4o
• A code-level LLM demo for redacting PII from enterprise data

Get the slides, code samples, and testing credits to try these techniques on OctoAI.

India's technology outsourcing industry is facing a pivot as AI integration threatens jobs, especially in basic tasks like call centers. Companies are challenged to ascend the value chain amidst a hiring slowdown, with some hopeful AI tools will create new opportunities in sectors like programming. The industry's future demands higher-level cognitive skills as automation continues to redefine traditional roles.

Despite impressive progress in LLMs solving complex tasks, they still fall short on consistently performing human-like tasks at scale. LLMs require detailed prompt engineering and iteration to achieve acceptable accuracy, limiting their current potential as direct human replacements in workflows. OpenAI's recent updates to JSON output control and cost reduction may address some issues, but the nuanced integration required for LLMs in organizational contexts suggests incremental productivity gains rather than immediate economic transformation.

Breakthroughs in AI pre-training have enabled models to process massive online datasets, while subsequent supervised fine-tuning with experts helps models generalize and gain specialized knowledge. Turing's approach of using "input and output pairs" generated by subject-matter experts aims to advance AI reasoning capabilities. Anticipating an "agentic" era of AI, these models could autonomously perform complex tasks by integrating specialized knowledge across domains.

Google is deploying new Gemini-powered features to Chrome Desktop, including Google Lens integration for visual searches, Tab Compare for streamlined shopping comparisons, and natural language processing for enhanced browsing history search.

Perceptive has developed an AI-controlled autonomous dental robot that performed a complete dental procedure notably faster than traditional methods.

Microsoft's new "GitHub Models" offers developers access to AI models like Llama 3.1 and GPT-4 and enables AI integration within software projects.