Episode Summary
Executive Summary: This episode surveys major 2016 AI/ML developments: Apple’s WWDC added Siri integrations and a new on-device neural network API, though the host argues Apple is catching up to rivals. IBM and the Weather Company showcased Deep Thunder for hyperlocal weather and business forecasting. The episode also covers Google, Microsoft, and Amazon expanding ML teams, plus notable research on audio prediction, generative models, meta-learning, SmartReply, and deep learning architecture design.
Main Topics: Apple WWDC and new machine learning APIs (Priority: 5/5): Apple announced Siri expansion to Mac, Apple TV, and third-party iOS apps, plus Photos improvements and a Basic Neural Networks framework for running pre-trained models on-device. The host sees this as incremental and largely catch-up versus competitors. IBM Deep Thunder and weather intelligence (Priority: 5/5): IBM and The Weather Company launched Deep Thunder, a machine-learning-driven platform for hyperlocal weather forecasts and business impact prediction, leveraging massive daily weather data ingestion. Big tech talent moves in machine learning (Priority: 4/5): Google formed a dedicated ML research group in Zurich, Microsoft acquired Wand Labs for conversational platform integration, and Amazon hired CMU professor Alex Smola to lead cloud machine learning efforts. MIT research on predicting sound from video (Priority: 4/5): Researchers demonstrated a model that predicts and matches audio to silent video clips using CNNs and RNNs, aiming toward better understanding of object properties through sound. OpenAI’s first research papers on generative models (Priority: 4/5): OpenAI released four early papers centered on generative modeling, including generative adversarial networks, highlighting image generation, restoration, and related tasks. Deep learning optimization and SmartReply research (Priority: 3/5): The episode highlights Google DeepMind’s work on learning optimizers with LSTM-based meta-learning, and Google’s SmartReply paper explaining how inbox suggestions are powered by LSTMs. Architecture engineering as the new feature engineering (Priority: 3/5): A blog post argues that deep learning shifts effort from handcrafted features to designing effective neural architectures, making model architecture a key human bottleneck.
Key Arguments: Apple introduced useful ML features, but the announcements were more about integrating existing AI trends than leading them; the host believes Apple was behind Google and Microsoft in visible ML progress. IBM’s Deep Thunder is a strong example of applied machine learning because it connects hyperlocal weather prediction with downstream business impact forecasting. Major technology companies were all signaling that machine learning is strategically important by hiring researchers, acquiring startups, and building platform integrations. Modern ML progress depends not only on data and training but on choosing the right model architecture and optimization strategy. Generative adversarial networks and meta-learning were emerging as important research directions because they reduce manual effort and improve model realism or training efficiency. Audio-from-video prediction demonstrates that ML systems can learn physical-world cues beyond text and images, moving toward more general perception. On-device ML frameworks like Apple’s BNNs suggest a broader push toward running inference locally for speed and privacy, even if training remains off-device.
Data Points: Date of episode: Friday, June 17th, 2016 - The episode opening sets the timing and news window. Weather data processed daily by The Weather Company: Over 100 terabytes - Used as the data backbone for IBM’s Deep Thunder forecasting system. Forecast resolution: Two-tenths of a mile - Deep Thunder’s hyperlocal prediction granularity. Survey participants: 400 - Online survey used to validate MIT’s audio prediction model. Fake-audio identification rate: 22% - Participants identified synthesized audio only 22% of the time in the MIT study. Improvement over prior version: 2x better - The MIT audio-prediction system performed twice as well as an earlier version. Number of OpenAI papers released: 4 - OpenAI published its first set of research results that week. Google Zurich ML focus areas: 3 - Machine learning and machine intelligence; natural language processing; machine perception. IBM Weather Company acquisition estimate: $2 billion - Referenced as the rationale behind IBM’s weather intelligence strategy.
Pivotal Quotes: "Apple was caught a little bit flat-footed on machine learning and AI, and I think they know it." — Sam Charrington: Host’s assessment of Apple’s WWDC AI announcements compared with competitors. "IBM is bringing the thunder to AI." — Sam Charrington: Introduction to IBM’s Deep Thunder weather forecasting announcement. "In deep learning, architecture engineering is the new feature engineering." — Stephen Merity (referenced via blog post title): Central thesis of the discussion on neural network design as a key human task.
Implications: The episode suggests 2016 marked a turning point where AI became a platform feature and a strategic hiring race. Practical ML applications, especially forecasting and assistants, were becoming the main battleground, while research increasingly focused on better architectures, meta-learning, and generative systems.