Developer workflows & testing
How contracts, boundaries, state protection, testing, debugging, and controlled releases make applications dependable even when model behavior is uncertain.
240 talks

The AI Engineer knowledge library
Connected chapters on AI systems, applications, and organizations, with the conference talks behind them.
87 topics · 60 chapters · 6152 talk selections
Browse by subject
How contracts, boundaries, state protection, testing, debugging, and controlled releases make applications dependable even when model behavior is uncertain.
240 talks

How examples, objectives, fitted parameters, and independent evidence turn data into useful predictions—and why prediction alone does not tell you what action will work.
271 talks

How sequences become contextual states, how those states produce predictions, and what information access can—and cannot—guarantee.
24 talks

How learned vectors preserve task-specific distinctions, acquire geometry, support larger systems, and fail when similarity is mistaken for truth.
30 talks

Instructions state what to do. Examples show what counts as a good answer. Testing both on fresh inputs reveals which guidance the model needs.
35 talks

How objectives, data exposure, schedules, compute, and checkpoint lineage establish and extend a model’s capabilities.
37 talks

How demonstrations, preferences, and rewards change a pretrained model—and how to determine what the changed checkpoint actually does.
59 talks

How intermediate work, alternative solutions, and checking improve answers—and how to decide when that work is worth doing.
29 talks

Generated examples help only when they add verified coverage of a defined capability gap and improve performance on evidence kept independent from their creation.
48 talks

Build executable tasks whose observations, actions, rewards, resets, and recorded outcomes mean what you intend.
34 talks

Acquire new capabilities through successive updates while preserving useful behavior—and the ability to keep learning.
24 talks

Train a model from another model’s behavior—and determine what actually transferred.
24 talks

How to choose, transform, label, release, and maintain data whose meaning remains fit for an AI system’s intended use.
59 talks

A program reading a table needs to know which heading governs each value. A person checking the result needs a way back to both.
26 talks

Turn external information into useful answers without losing the conditions, sources, and limits that make those answers defensible.
515 talks

How search systems acquire an eligible collection, turn it into candidates, rank useful results, and serve them under permission, freshness, coverage, and latency constraints.
71 talks

When explicit identities, typed relationships, and connected evidence justify their construction—and how to keep every graph assertion sourced, revisable, and useful.
32 talks

A release rule tells a model how to choose a tag. A repository lookup gives it tags to choose from. Across a longer task, the application must also decide which observations still apply and how to recover the ones it leaves out.
121 talks

A saved preference, a searchable conversation, and an editable note give a future session different evidence. They also give a person different ways to correct it.
80 talks

Reusable capability exchange, from discovery and authorization to execution and uncertain outcomes.
374 talks

How to turn uncertain model generations into validated records, controlled action requests, and observed software effects.
52 talks

A model can choose what to try next. The application determines what it may change and what evidence counts as success.
457 talks

How an agent discovers a skill, opens its instructions and uses the files inside.
49 talks

How to give an agent useful feedback, preserve unfinished work, and verify the result at its destination.
90 talks

A patch, a tested build, and a working feature in a user’s hands are different outputs. Each has a receiver who needs evidence before accepting it.
62 talks

A proposed edit, a passing test, and a working change on the target branch each answer a different engineering question.
530 talks

Observe graphical state, identify the intended control, deliver the right input, and verify the authorized outcome.
54 talks

Continuing assistance that remembers what matters, carries work forward, and stays within the person's control.
37 talks

Divide work and discretion without losing task meaning, responsibility, or control of the complete outcome.
56 talks

How a serialized model request becomes a streamed completion—and how state, scheduling, memory, and workload shape determine latency, throughput, and sustainable capacity.
67 talks

Map numerical work onto parallel hardware, control data movement, and measure which changes improve the complete workload.
22 talks

Divide model work and numerical state across devices without losing control of communication, update semantics, or recovery.
25 talks

How numerical mappings, calibration and execution support turn fewer bits into useful savings—and where those savings can fail.
25 talks

A chat API can look the same whether a laptop, a nearby workstation or a hosted service generates the reply. Each placement makes a different demand on the network, the device and the user’s data.
70 talks

How to run generated or otherwise untrusted code with bounded files, networks, authority, resources, lifetime, and tenant exposure—without pretending teardown reverses effects outside the boundary.
59 talks

Choose suitable models and serving paths while preserving permissions, request meaning, resource limits, and honest failure behavior.
37 talks

A service contract tells a team what it can request, how to verify the result and who will act if the service fails.
40 talks

Reduce the resources needed for useful work without losing the quality, responsiveness, or operating limits that make the system worth using.
75 talks

Rejecting one defective output, choosing between two systems and showing that people do better work with AI each require different evidence.
402 talks

Design operational evidence that can reconstruct an AI execution, locate consequential failures, attribute time and cost, and state honestly what remains unknown.
267 talks

Trace attacker influence through data, models, tools, execution, and external effects—and place enforceable controls where that influence would otherwise gain authority.
91 talks

Turn decisions about appropriate information use into enforceable limits on collection, access, disclosure, modification, retention, and deletion.
46 talks

How to combine text, images, audio, video, and other signals without losing the relationships that make their conclusions meaningful.
40 talks

How images and video become predictions about objects, regions, space and events—and how to choose representations that preserve the information a task needs.
106 talks

How speech models, audio transport, and conversational policies cooperate to listen, respond, stop, and recover.
105 talks

How images, sound and video are synthesized, directed, revised and assessed.
72 talks

Learning how environments change—and establishing when predicted futures are useful for planning, training, and simulation.
15 talks

Connecting goals, measurements, motion, and learning in a physical system that must keep responding while the world changes.
38 talks

How to help people understand, direct, review, interrupt, and recover from AI behavior that may be uncertain, delayed, or partly wrong.
77 talks

Redesign complete processes, allocate decisions deliberately, and keep people, records and external effects under clear control.
77 talks

A customer can depend on a change that makes sense only in its own operation. The field engineer helps the product team decide which parts are worth supporting elsewhere.
58 talks

Find work worth improving, make a bounded promise, and build a business that can keep it.
36 talks

Adopt AI by changing how work is delivered, who can decide, and how useful outcomes are sustained.
263 talks

Delegate commercial choices while preserving valid terms, purchasing authority, and responsibility for what remains owed.
45 talks

Improve customer understanding, communication, and follow-through—and distinguish useful commercial work from increased activity.
38 talks

Choose worthwhile investments, fund the capabilities they require, and keep authority aligned with consequences.
212 talks

Select useful items, learn from incomplete feedback, and measure whether the resulting policy improves the product.
21 talks

Build useful financial assistance by preserving meaning, separating analysis from authority, and checking the complete workflow.
120 talks

Build useful assistance around clinical work, meaningful records, accountable people, and evidence of completed care.
49 talks

How to turn automated trials into informative comparisons, reproducible findings, and defensible research decisions.
34 talks

Explore additional talk collections. These archive topics are separate from the chapters above.