Episode Summary
Executive Summary: This episode surveys major 2016 machine learning and AI developments: Facebook’s Deep Text for multilingual NLP, Google’s Magenta project for machine-generated art/music, and TensorFlow’s growing technical maturity. It also covers AI business moves, bot platforms, hybrid human-AI systems, and practical perspectives on optimizing models for value and cost rather than error alone.
Main Topics: Facebook Deep Text and NLP at scale (Priority: 5/5): Facebook’s internal Deep Text engine applies deep learning to text understanding, aiming for near-human accuracy across languages and large post volumes, especially for intent detection and conversation understanding. Google Magenta and machine-generated art/music (Priority: 5/5): Google’s Magenta project explores generating original art with machine learning, emphasizing surprise, storytelling, and evaluation, while staying open source and community-driven. TensorFlow’s evolving role in ML infrastructure (Priority: 4/5): The episode highlights a new TensorFlow paper describing the system’s dataflow model and real-world performance, reinforcing TensorFlow as a major ML platform for research and deployment. AI industry investment and acquisitions (Priority: 4/5): Microsoft, Intel, Lumiata, and Findo illustrate how companies are investing strategically in machine learning, computer vision, healthcare analytics, and search assistants. Bots and hybrid human-AI workflows (Priority: 5/5): Startup bot platforms are emerging alongside a key argument that practical AI systems need humans in the loop, using active learning and hybrid interaction to manage accuracy limits. Economics of machine learning optimization (Priority: 4/5): Two articles argue for optimizing ML not just for test error but for business value and total system cost, including monetization, software/hardware, organizational, and societal trade-offs. Technical deep dives and applied ML examples (Priority: 3/5): Additional links cover reinforcement learning (Pong from pixels), deep learning for trading, and an IoT magic mirror using Cortana services and facial recognition.
Key Arguments: Deep learning is expanding beyond computer vision into natural language processing, where it can better scale across languages and data than traditional NLP techniques. Facebook’s Deep Text shows that large-scale text understanding can be performed with near-human accuracy and high throughput, making it practical for real products. Machine learning is not yet accurate enough to fully replace humans in many production settings; hybrid systems are therefore the near-term norm. Active learning improves productivity by routing uncertain cases to humans and feeding those corrections back into the model. Hybrid interaction improves accuracy by letting humans approve or rewrite machine-suggested outputs, though at the cost of throughput. AI/ML models should be evaluated on business outcomes such as revenue or profitability, not only on test error. Cost matters as much as accuracy in commercial ML; the right model is the one whose benefits justify its full implementation and operating costs. Google Magenta frames generative art as a research problem involving generation, surprise, storytelling, and human-like evaluation, not just style transfer. Open-source ecosystems like TensorFlow and GitHub-hosted projects are accelerating experimentation and community participation in AI. Corporate venture and acquisition activity show that big tech firms see ML as strategically important across vision, healthcare, security, and cloud services.
Data Points: Show episode number: 3 - The host notes this is the third show of the podcast Episode date: Friday, June 3rd, 2016 - Opening introduction to the episode Deep Text throughput: Several thousand posts per second - Facebook claims Deep Text can process text at scale Deep Text language coverage: Over 20 languages - Facebook reports multilingual support Deep Text accuracy: Near-human accuracy - Facebook blog description of the system’s performance Active learning productivity gain: 10 to 20 percent - Claire Corthell’s post claims systems can scale this much more than humans alone Human-machine task split in current AI companies: About 25 percent routed through the computer - Corthell describes current real-world operating patterns Long-term desired split: 10/90 split - Goal cited for human vs computer work, with 10% human and 90% computer Intel Capital Series B: $10 million - Intel Capital led funding for Lumiata Findo funding: $3 million - AI and NLP startup Findo announced a raise for a personal search assistant Target ML costs/metrics focus: Revenue or profitability - Patrick Hall argues models should be evaluated on business value instead of only test error
Pivotal Quotes: "Deep text is unfortunately an internal system to Facebook." — Sam Charrington: Explaining that Facebook’s new text engine is not open sourced, at least initially "we’ve gotten very good at using machines to understand and analyze ... but they’re really looking to push it further and generate wholly new art" — Sam Charrington: Describing the motivation behind Google Magenta "machine learning systems will never achieve the level of accuracy required to totally take human out of the loop" — Claire Corthell (referenced by Sam Charrington): Summarizing the core argument for hybrid intelligence in real-world AI systems
Implications: The episode suggests the near-term future of AI is practical, hybrid, and business-oriented: better infrastructure, human-in-the-loop systems, and value/cost-aware evaluation will matter more than fully autonomous “pure AI” claims.