Apr 2, 2015 · 21m · mad
Ben Medlock, SwiftKey // Building a Better Keyboard // Data Driven NYC (FirstMark Capital)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
At Data Driven NYC, Ben Medlock, Co-founder and CTO of SwiftKey, explains how probabilistic machine learning and natural language processing transform smartphone touchscreen typing into an adaptive, human-centric experience. By tracing AI history from early theoretical foundations to personalized language and touch modeling, Medlock demonstrates how SwiftKey successfully solved real-world input uncertainty at massive scale.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 6.8% of the talking time here. How this is scored →
Matt as informed peer 0.3 Guest teaching 1.8 Guest disagreement 0.2 Matt pushing back 0.3
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 18:10 Reframing user behavior adaptation
Medlock reframes an audience question by clarifying that users do not need special hacks or behavior adaptation, as the AI handles all adjustments automatically behind the scenes.
Hardest push from Matt ▶ 16:06 Host highlights iOS version glitchesMatt Turck gently pushes back on the product's quality narrative by pointing out early performance glitches in SwiftKey's iOS release.
Biggest teaching moment ▶ 6:35 Virtual keyboards as probability problemsMedlock educates the audience on why virtual keyboards represent a probabilistic inference problem under tap uncertainty rather than a simple mechanical layout issue.
Matt holds his own ▶ 16:06 Host demonstrates hands-on user experienceMatt Turck demonstrates informed user insight by calling out initial technical bugs on the iOS rollout despite praising the core product.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| SwiftKey's Vision for Human-Centric Technology | 0 | 1 | 0 | 0 | Ben Medlock delivers a solo presentation on SwiftKey's product vision and the history of AI. Because this is a monologue section, host expertise and pushback scores are set to zero. | |
| Real-World Complexity and Early AI Limitations | 0 | 2 | 0 | 0 | Medlock explains real-world complexity, probability theory, and the technical distinction between narrow and general AI. The segment is a solo presentation without host participation. | |
| Mechanical Keyboards vs. Virtual Keyboard Inference | 0 | 2 | 0 | 0 | Medlock frames touch screen typing as an inference problem rather than a mechanical layout problem. The host does not speak during this monologue segment. | |
| Mathematical Interpretations and Model Approaches | 0 | 2 | 0 | 0 | Medlock presents language modeling approaches including Ngram models and morpheme analysis. With no host interaction, host-side metrics remain zero. | |
| Data Collection and Grid Computing | 0 | 2 | 0 | 0 | Medlock details grid computing partnerships and Gaussian modeling for user touchscreen taps. The host is absent during this presentation segment. | |
| Hyperparameter Learning and Additional Language Challenges | 2 | 2 | 1 | 2 | Host Matt Turck enters at 16:06, lightly bringing up early iOS glitches before opening up an audience Q&A session. Medlock cooperatively addresses product issues such as profanity filtering and user adaptation. |
Statements from this episode (17)
Ben Medlock: Early AI symbolic logic models failed to scale
“And the systems that, that we tried to build to begin with were based on scaled down worlds where we could develop symbolic logics to manipulate sort of simplified objects within these worlds and just failed miserably to scale up to the real world.”
Medlock: 2015 AI technology cannot abstract concepts across domains
“This is what the human brain does so incredibly well, and what our technology today is currently almost entirely unable to do.”
Ben Medlock: Touchscreen typing is an inference problem, not a mechanical one
“This is a virtual keyboard, and the problem of typing was now not a mechanical problem. It's an inference problem, and that means that there is a degree of uncertainty.”
Medlock: SwiftKey is used on close to a billion devices worldwide
“We designed this in 2010, and it's now used on close to a billion devices worldwide.”
Medlock: SwiftKey chose independent probability distributions over neural networks for tractability
“And we took the first one of these because it's a lot easier and a lot more tractable.”
Medlock: N-gram models are extremely hard to beat in voice recognition
“If you've done anything in voice recognition, you will know that it's really, really hard to beat Ngram models.”
Medlock: Neural network language models are a highly promising research avenue
“Another possibility is to move away from engrams and towards neural network-based language models, which is a really promising, ah, avenue of research at the moment.”
SwiftKey used Large Hadron Collider grid infrastructure to process language models
“So we, ah, we partnered with the University of Cambridge in the UK to use the European grid, which was put together to analyze the data from the Large Hadron Collider, which is why I show you this picture.”
SwiftKey scraped the entire public internet to train its language models
“We scraped all of the publicly available texts off the internet and dumped it into language specific buckets.”
Medlock: Online profiles offer the best seed for personal predictive text models
“To get a really strong predictive model, you need to know how an individual user uses language, and online profiles give you an ideal way of getting a seed for this.”
SwiftKey's initial touch model used Gaussian distributions to map keypresses
“The system that we first built was to use Gaussian distributions to model the interaction of the person with the keyboard surface, and then we can use a linear Gaussian for the space bar, and the others are single point Gaussians.”
SwiftKey continuously retrains its touch model on-device as users type
“We build all this technology that runs on the phone. So while you're tapping on the keyboard, it's retraining via a re-parameterized version of MAP.”
SwiftKey saved users 15 trillion keystrokes out of 50 trillion typed characters
“We've had about 50 trillion characters written through the software, and we saved about 15 trillion keystrokes”
SwiftKey built Stephen Hawking's upgraded communication technology
“We just recently built Stephen Hawking's new communication technology”
Medlock: SwiftKey licenses predictive typing tech to Samsung
“We also license the technology to people like Samsung who build their own keyboards on top of it.”
SwiftKey accidentally launched its first Dutch keyboard without a profanity filter
“I do remember we, ah, we launched the, ah, Dutch version, the first Dutch version, we forgot to switch on the profanity filter. So it went out, and it was utter filth.”
Ben Medlock predicted physical keyboards would dominate programming until AGI arrives
“I think for things like programming, it's really hard to see it changing significantly. And it's sort of almost It's very difficult to imagine what would actually do a much better job of this outside of a general AI that's better at programming than we are. So…”