Jul 28, 2017 · 16m · a16z

Kathryn McKinley

0:00 / 0:00

gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions

In this presentation at the Andreessen Horowitz Academic Roundtable, Microsoft Research Principal Researcher Kathryn S. McKinley introduces Uncertain<T>, a programming model designed to gracefully handle sensor inaccuracy and probabilistic data through statistical semantics, domain-aware Bayesian inference, and probabilistic program verification.

Chapters · click to jump

2 statements ▾

Governments occasionally degrade GPS sensor accuracy intentionally, says Kathryn McKinley open →

Developers lack programming models for probabilistic ML outputs, says Kathryn McKinley open →

1 statement ▾

Phone users are least likely to be directly under the GPS dot open →

1 statement ▾

Uncertain<T> reduces GPS road-snapping code from thousands of lines to two open →

2 statements ▾

Traditional software assertions fail for probabilistic computing, requiring new approaches open →

Formal proof of differential privacy implementations remains unsolved, says Kathryn McKinley open →

How this conversation actually went

Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →

The host as informed peer 0.0 Guest teaching 5.8 Guest disagreement 0.0 The host pushing back 0.0

speaking balance: gold is the host, purple is the guest (3 minute bins)

Sharpest disagreement ▶ 5:43 Critique of traditional probabilistic programming

McKinley offers a mild technical critique of traditional probabilistic programming languages for assuming unnecessary precision, though delivered constructively as part of her lecture.

Hardest push from the host ▶ 0:10 No host pushback present

Because this transcript is a solo presentation without host participation, no host pushback occurs in the episode.

Biggest teaching moment ▶ 3:40 Explaining GPS sensor error models

McKinley educates the audience on Rayleigh distributions, explaining that GPS error forms a doughnut shape around the actual location rather than a single point estimate.

The host holds their own ▶ 0:10 No host hits back present

The host does not speak during this presentation transcript, leaving host expertise scores at zero.

the scores for every segment, with the reasoning behind each

Topic	Guest teaching	Why
Motivation: Sensor Inaccuracy and Inaccurate Speed Data	6	Kathryn McKinley presents a technical talk explaining how sensor inaccuracies lead to speed calculation errors, introducing the need for uncertain programming types. The host is absent during this monologue segment.
Statistical Semantics, Sampling, and Bayesian Networks	6	McKinley explains statistical semantics embedded in programming language runtimes and how Bayesian networks evaluate expression precision dynamically. As a monologue lecture, host-side metrics remain at zero.
Evaluating Conditionals and Reducing False Positives	6	McKinley demonstrates how sampling and hypothesis testing evaluate conditionals to systematically eliminate false positives in sensor applications. The segment is an unassisted presentation.
Exploiting Domain Context with Bayesian Inference	6	McKinley details incorporating domain context like road snapping using Bayesian constructs and sequential likelihood reweighting to compress code length. No host interaction occurs.
Probabilistic Assertions for Program Verification	6	McKinley discusses replacing deterministic assertions with probabilistic assertions for approximate computing and privacy verification. The segment is entirely monologue presentation.
Conclusion and Final Summary Remarks	5	McKinley concludes her talk by summarizing the necessity of programming language support for imperfect data across domain areas before thanking the audience.

Statements from this episode (6)

Assertion Supported

Governments occasionally degrade GPS sensor accuracy intentionally, says Kathryn McKinley

“Occasionally your government will make the sensors bad on purpose.”

Kathryn McKinley Jul 28, 2017 ▶ 0:57

Insight

Developers lack programming models for probabilistic ML outputs, says Kathryn McKinley

“Image understanding and all sorts of applications we're using today aren't giving us perfect answers, yet we don't have the tools to interpret those answers. We don't have the programming models”

Kathryn McKinley Jul 28, 2017 ▶ 1:45

Assertion Supported

Phone users are least likely to be directly under the GPS dot

“When you see that little dot on your phone, you're actually least likely to be there. The GPS error is, looks like a doughnut. It's a Rowley distribution, so you're actually likely to be in a little circle around where that dot is, alright?”

Kathryn McKinley Jul 28, 2017 ▶ 4:31

Assertion Not checkable as stated

Uncertain<T> reduces GPS road-snapping code from thousands of lines to two

“But what used to take thousands of lines of code now takes two.”

Kathryn McKinley Jul 28, 2017 ▶ 12:55

Insight

Traditional software assertions fail for probabilistic computing, requiring new approaches

“Well, now we have a, ah, traditional assertions can't do that, and so we have a new way to do this with a probabilistic assertion.”

Kathryn McKinley Jul 28, 2017 ▶ 14:21

Assertion Supported

Formal proof of differential privacy implementations remains unsolved, says Kathryn McKinley

“And so we've come, we've taken some good steps in this direction. But we haven't done the very hardest step, which is to prove in the implementation of differential privacy that you actually get it.”

Kathryn McKinley Jul 28, 2017 ▶ 15:25

Kathryn McKinley

How this conversation actually went

Statements from this episode (6)

Turn any episode into a week of clips.