Mert

Android 17 için “Handoff” benzeri “Continue On” altyapısı duyuruldu

yazar
Android 17

Android 17Google, I/O 2026 etkinliğinde Android 17 işletim sistemi için geliştirilen, Apple’ın Handoff sistemine benzer “Continue On” altyapısını da duyurdu. Bu altyapı, kullanıcıların Android cihazlarını uygulama kullanımları özelinde birbirine bağlamasına imkan tanıyor. Örneğin telefondaki bir uygulamada açılan doküman ya da e-posta, doğrudan tablette de görünüyor ve tek dokunmayla açılabiliyor. Zaman kazandıran bu altyapı için her iki cihazda da aynı hesabın kullanılması gerekiyor.

Android 17 için “Handoff” benzeri “Continue On” altyapısı duyuruldu yazısı ilk olarak Teknoloji haberleri – LOG üzerinde yayınlandı.

Detecting Silent Model Failure: Drift Monitoring That Actually Works

yazar

TL;DR: Most drift monitoring setups alert on the wrong thing. Feature distribution drift is cheap to compute and almost always misleading. Prediction drift plus a delayed ground-truth feedback loop catches the failures that actually cost money. Here is the setup I use at Yokoy.

A model that returns HTTP 200 with a plausible-looking float is the worst kind of broken. No exception, no pager, no Slack message. The metric only moves three weeks later when finance reviews the numbers.

I have spent the last two years rebuilding the monitoring story for our expense classification models. What follows is what I kept after throwing out the rest.

The mistake I keep seeing

Teams instrument input feature drift first because it is the easiest thing to compute. Pull yesterday’s feature values, pull today’s, run a KS test on each column, alert when p < 0.05.

This generates noise. A lot of noise.

Features drift constantly for reasons that have nothing to do with model quality. A new customer onboards, the merchant category distribution shifts, you get a Slack ping at 03:00 for something that does not matter. After two weeks of this, on-call mutes the channel. After four weeks, the channel is deleted.

The problem is not the test. The problem is that input drift is a weak proxy for what you actually care about: did model performance degrade.

What to monitor instead

Three signals, ranked by cost and value.

Signal Compute cost Latency to detect False positive rate
Input feature drift Low Hours High
Prediction distribution drift Low Hours Medium
Performance vs delayed labels Medium Days to weeks Low

Prediction drift is the underrated one. If your model started returning a different distribution of outputs without you shipping new weights, something upstream broke. Could be feature pipeline. Could be a provider returning malformed embeddings. Could be a real population shift. All of these are worth investigating.

The detection logic is short: 

from scipy.stats import wasserstein_distance
import numpy as np

def prediction_drift_score(reference: np.ndarray, current: np.ndarray) -> float:
    return wasserstein_distance(reference, current)

# reference = predictions from the validation window when the model was promoted
# current = predictions from the last 24h of production traffic
# alert when score exceeds the 99th percentile of bootstrapped baseline scores

Wasserstein over KS for prediction monitoring. KS is hypersensitive to large samples and you will have large samples in production. With 500k predictions per day, KS rejects the null hypothesis for differences nobody cares about.

The feedback loop is non-negotiable

For expense classification, ground truth arrives when a human approves or corrects the prediction. Median latency is four days. P95 is three weeks.

We log every prediction with a join key and write it to a Parquet table partitioned by date. When labels arrive, a nightly Kubeflow pipeline joins them and computes per-segment performance: accuracy per merchant category, per country, per customer tier.

The per-segment view is what surfaces the failures. Aggregate accuracy stays at 94% while accuracy on a specific Swiss VAT category collapses to 71%. The aggregate view would never have caught it. 

# Simplified pipeline component spec
- name: compute-segmented-metrics
  inputs:
    predictions_table: gs://yokoy-ml/predictions/dt={{date}}
    labels_table: gs://yokoy-ml/labels/dt={{date}}
  outputs:
    metrics_table: gs://yokoy-ml/metrics/dt={{date}}
  segments:
    - merchant_category
    - country
    - customer_tier
  resource_request:
    cpu: 4
    memory: 16Gi

The cost: roughly 12 minutes of compute per day on our volume. The value: every regression we caught in the last 18 months was caught here, not by drift monitoring.

Where input drift still earns its place

I have not fully abandoned input drift. It is useful as a debugging tool after the fact. When per-segment accuracy drops, the first question is which features moved. Having the historical drift scores already computed means the investigation starts with a query instead of a backfill.

So compute it, store it, do not alert on it.

A note on LLM-based features

We added an LLM-derived feature last year for invoice text classification, routed through a gateway in front of multiple providers (Bifrost handles this for us, though others like LiteLLM or Portkey cover the same ground). The drift profile changed immediately. Provider model updates, even minor ones, shift the feature distribution in ways you cannot see from your side.

Lesson: pin the provider model version explicitly. Treat a provider model change as a feature pipeline change. Re-run the validation set. This sounds obvious until the day a default model alias updates and you find out from the metrics.

Trade-offs and Limitations

Per-segment monitoring has a cardinality problem. With three segments of 50, 30, and 5 values you get 7500 cells. Most are empty or have too few samples for meaningful metrics. We use a minimum sample threshold of 100 per cell per day and accept that long-tail segments take longer to detect issues in.

Delayed labels mean delayed detection. For models where the label takes weeks, you need a complementary fast signal. Prediction drift fills part of that gap but it is a leading indicator, not a measurement.

Wasserstein distance has no native interpretation in production units. You bootstrap a baseline and alert on deviation from it. This works but it is not as crisp as “accuracy dropped 3 points.”

Storing every prediction with features for joinability is expensive. We compress aggressively and tier old partitions to cold storage after 90 days. Plan the storage cost before you build it, not after.

Further Reading

Asya borsaları negatif seyrediyor

yazar
Asya borsaları, yükselen tahvil getirileri ve Orta Doğu’daki gerilimlerin etkisiyle negatif seyrediyor. Petrol fiyatlarının yüksek kalması enflasyon endişelerini artırırken, Japonya ve Güney Kore’de endeksler düştü. Çin’de kredi faizleri sabit kaldı, Samsung’ta grev kararı hisseleri geriletti.

Google, Gemini 3.5 model ailesini tanıttı

yazar
Google, düzenlediği Google I/O 2026 etkinliğinde Gemini model ailesinin yeni neslini tanıttı. Şirketin duyurduğu Gemini 3.5 Flash modeli; hız, düşük gecikme süresi ve agent tabanlı kullanım senaryolar…

The Double-Edged Sword: Negative Impact of AI in Software Development and How to Use AI Effectively in the SDE Process

yazar

The Double-Edged Sword: Initial Experience and Hidden Pitfalls
As a software engineer, I initially embraced AI-powered editors like Cursor for backend development with great enthusiasm. The ability to automatically generate CRUD operations and complex API logic was intoxicating. However, this convenience came with a hidden cost: a gradual decline in diligence. Over-reliance on AI led me to assume the business logic was inherently correct. This complacency resulted in significant integration issues, such as inconsistent response structures and performance-degrading N+1 queries that only surfaced during high-volume production testing.

Shifting Gears: Mastering AI for Productive Development
To reclaim control, I transitioned from passive code generation to active AI management. I learned to define strict rules within the editor to enforce consistent coding patterns and robust architecture. I also discovered that AI performance degrades with large, monolithic tasks due to token limits, often leading to missing data pointers in DTOs. The solution was to decompose tasks into smaller, manageable pieces and implement a rigorous review process for every change. Using “plan mode” for brainstorming and impact analysis allowed me to vet AI approaches before execution, ensuring smoother deployments and optimized query performance.

The Game Changer: AI-Driven Test Automation and Asynchronous Bug Fixing
Perhaps the most significant productivity boost came from using AI to tackle testing—a task often neglected due to tight deadlines. AI now helps me set up comprehensive test suites, including unit, integration, and E2E tests with real databases. For a single API, I can generate 8 to 10 test cases covering edge cases, invalid inputs, and duplicates in minutes. Furthermore, tools like Jules AI have revolutionized my workflow by allowing asynchronous development. I once even managed to brainstorm a bug fix and merge a PR via my phone during a doctor’s appointment, with the AI handling the heavy lifting and triggering the deployment pipeline while I was away from my laptop.

Conclusion: The Path to AI-Assisted Software Engineering Excellence
AI in software development is a powerful tool that requires a careful, observant hand. When used with skepticism and structured oversight—through task decomposition, strict rules, and automated testing—it provides a massive positive impact. The key to excellence is not just letting AI write the code, but carefully observing and guiding what it produces.

💥 Tech Talks Weekly #105

yazar
AURI by Endor Labs

Free security tools for AI-powered coding (Sponsored)

AURI by Endor Labs

Security teams talk risk. Developers talk function. AURI by Endor Labs is the common language between them, grounded in shared context and evidence.

AURI by Endor Labs brings security into the architecture of agentic coding, with free tools for developers and AI agents to detect vulnerabilities, block malware, and fix security bugs.

Learn more and try it yourself →

Happy Thursday 👋 and a warm welcome to Tech Talks Weekly #105!

Before we start, if you found something useful, leave a comment. Thank you!

That said, there’s tons of great content this week, so let’s jump right in!

Join 9,500+ Software Engineers & Engineering Leads who receive a free weekly email with all the recently published conference talks and podcasts. Stop scrolling through messy YouTube/RSS feeds. Stop FOMO. Easy to unsubscribe. No spam, ever.

🏆 Featured this week

Here are our top recommendations this week. Watch now or bookmark!

1. “Java 26: Better Language, Better APIs, Better Runtime” from JavaOne 2026

Conference ⸱ +8k views ⸱ May 17, 2026 ⸱ 00h 50m 25s

tldw: After JavaOne 2025, JDK 25 became the latest LTS release. But Java keeps moving forward, and now JDK 26 is here. Nicolai Parlog, a Java Developer Advocate at Oracle, demoes the key changes from Java 21 to 25, including unnamed patterns, module imports, the foreign-function and memory API, stream gatherers, and the class-file API. He also takes a closer look at JDK 26 and its preview features, such as primitive patterns, improved structured concurrency, and HTTP/3 support.

https://youtube.com/watch?v=86eK0Or2eIE

2. “Beyond Code Coverage: Functionality Testing with Playwright MCP — Marlene Mhangami, Microsoft” from AI Engineer Singapore 2026

Conference ⸱ +7k views ⸱ May 16, 2026 ⸱ 00h 19m 45s

tldw: When an LLM writes tests after the code, it often produces tests that simply confirm the current implementation rather than checking whether the user experience really works. As a result, your test suite can pass even while the app is broken in ways the tests never cover. Marlene Mhangami, Senior Developer Advocate for Python and AI at Microsoft, argues for reversing that workflow. Get the agent first to write failing Playwright tests, then generate the code needed to make those tests pass. The demo walks through this approach using GitHub Copilot and the Playwright MCP server on a toy store search feature. With the browser open, you can watch the agent interact with filters and validate search results in real time.

https://youtube.com/watch?v=FWEInOtngmM

3. “Tech Truth: Agile Evolution & the Future of SW Engineering • Martin Fowler & Kent Beck • GOTO 2025” from GOTO Copenhagen 2025

Conference ⸱ +1k views ⸱ May 18, 2026 ⸱ 00h 53m 56s

tldw: Martin Fowler and Kent Beck, longtime friends and leaders in software development, recently sat down to talk about their experiences over the past thirty years. They shared how AI now plays a real role in their work: Kent sees it as a patient tutor for trying out ideas, while Martin points out that AI helps us communicate with computers, but people skills are still essential for building software that meets real needs. They also looked back at the Agile Manifesto, almost 25 years later, discussing what Extreme Programming got right and how some of its ideas changed as more people adopted them. The conversation ends with helpful advice for junior developers.

https://youtube.com/watch?v=ii_rLjQfjp0

4. “Why Most Event-Driven Systems Fail in Banking (And How to Fix It)” from InfoQ Dev Summit Munich 2025

Conference ⸱ +1k views ⸱ May 18, 2026 ⸱ 00h 51m 25s

tldw: Event-driven, cloud-native architectures promise ultimate scalability, but highly regulated industries like fintech and healthcare introduce strict boundaries where you cannot afford to lose or duplicate a single event. Chris Tacey-Green, Head of Engineering at Investec, walks us through the foundational principles of lean events versus fat events, the critical difference between commands and events and why you do not need event sourcing to succeed. You will also learn how to implement patterns like Inbox and Outbox patterns to stop data duplication, maintain event contracts across domain boundaries and tackle event ordering.

https://youtube.com/watch?v=8VUXniuTTBA

5. “Structured Concurrency: The paradigm shift that changes modern development – Android Makers” from Android Makers 2026

Conference ⸱ +100 views ⸱ May 14, 2026 ⸱ 00h 38m 38s

tldw: For decades concurrent programming has meant wrestling with complexity, resource leaks and stray processes. Structured concurrency is the paradigm shift that changes how you write, read and reason about concurrent code, and this session cuts through the hype to reveal the core principle: concurrent tasks should have a clear beginning, end and scope, just like any other code block. You will see how that simple idea organizes tasks into clear hierarchies that enable automatic, reliable cancellation, so when one task fails its entire family is gracefully cleaned up. Watch to see how this paradigm, now being adopted in languages from Java to Kotlin and beyond, makes concurrent code robust, readable and easier to manage than ever.

https://youtube.com/watch?v=SQFQyyA1VOI

Found something useful? Leave a comment. Thank you 🙏

6. “Formal Methods as Agent Guardrails” from Software Engineering Daily

Podcast ⸱ May 19, 2026 ⸱ 00h 48m 32s

tldl: Formal methods are a branch of mathematics and computer science focused on proving the correctness of systems and they are promising a more rigorous foundation for software. Their complexity has kept them confined to a small community of specialists, but that is now changing as agentic AI systems take on increasingly autonomous roles. This episode of Software Engineering Daily explores how formal methods can act as guardrails for AI agents, giving them the verifiable safety that autonomous software increasingly demands.

https://softwareengineeringdaily.com/2026/05/19/formal-methods-as-agent-guardrails/?utm_source=rss&utm_medium=rss&utm_campaign=formal-methods-as-agent-guardrails

7. “What is harness engineering? — Thoughtworks Technology Podcast” from Thoughtworks

Podcast ⸱ May 14, 2026 ⸱ 0 days 00:40:51

tldl: Harness engineering is one of the most significant terms to emerge in software engineering in 2026. It broadly refers to the work done to control unpredictable AI agents and coding assistants, and its rise signals growing attention on what it takes to make agents reliable and consistent enough for real-world production software. Listen to this episode to see what harness engineering actually is and how it should be done.

https://youtube.com/watch?v=KQJLPsbnKiQ

📆 New talks

Here’s the complete list of all the talks published since the last issue, grouped by conference, and ordered by number of views for your convenience.

AI Engineer Singapore 2026 (18), Android Makers 2026 (17), Conference Talks 2026 (3), Devnexus 2026 (11), Devoxx Poland 2024 (2), GOTO Copenhagen 2025 (2), GOTO Serverless Bangalore 2025 (1), InfoQ Dev Summit Munich 2025 (2), JavaOne 2026 (1), Node Congress 2026 (1), QCon AI New York 2025 (1), QCon San Francisco 2025 (2), Spring I/O 2026 (3)

📆 New podcasts

Here’s the complete list of all the podcasts published since the last issue.

Go to 📆 New podcasts

📈 Most-watched talks

Find the most watched talks from the past 7d, 30d, 90d, 6m, and 12m.

Go to 📈 Most-watched

🗄️ Talks and podcasts by category

This issue’s talks, sorted by view count and organized by topic.

AI Engineering

  1. “Harnesses in AI: A Deep Dive — Tejas Kumar, IBM” from AI Engineer Singapore 2026 — Conference ⸱ +38k views ⸱ May 17, 2026 ⸱ 00h 20m 27s

  2. “Build Agents That Run for Hours (Without Losing the Plot) — Ash Prabaker & Andrew Wilson, Anthropic” from AI Engineer Singapore 2026 — Conference ⸱ +28k views ⸱ May 18, 2026 ⸱ 01h 15m 40s

  3. “Agents Don’t Do Standups: Building the Post-Engineer Engineering Org — Mike Spitz, PFF” from AI Engineer Singapore 2026 — Conference ⸱ +22k views ⸱ May 15, 2026 ⸱ 00h 17m 50s

→ See 31 more AI Engineering talks on Tech Talks Weekly Plus

AWS

  1. “How to Build Your Full-Stack Applications With CDK & AWS Amplify • Erik Hanchett • GOTO 2025” from GOTO Serverless Bangalore 2025 — Conference ⸱ +200 views ⸱ May 19, 2026 ⸱ 00h 25m 40s

DevOps

  1. “Mind the Gap (In your Agent Observability) — Amy Boyd & Nitya Narasimhan, Microsoft” from AI Engineer Singapore 2026 — Conference ⸱ +2k views ⸱ May 14, 2026 ⸱ 01h 20m 07s

Java

  1. “Java 26: Better Language, Better APIs, Better Runtime” from JavaOne 2026 — Conference ⸱ +8k views ⸱ May 17, 2026 ⸱ 00h 50m 25s

  2. “Core Resilience Features in Spring Framework 7 by Juergen Hoeller @ Spring I/O 2026” from Spring I/O 2026 — Conference ⸱ +1k views ⸱ May 18, 2026 ⸱ 00h 46m 46s

  3. “Beyond local tools: Deep dive into MCP with Spring AI by James Ward / Maximilian Schellhorn” from Spring I/O 2026 — Conference ⸱ +1k views ⸱ May 18, 2026 ⸱ 00h 51m 55s

→ See 12 more Java talks on Tech Talks Weekly Plus

Kotlin

  1. “Offline Should be the norm building local first apps with CRDTs & Kotlin Multiplatform” from Android Makers 2026 — Conference ⸱ +100 views ⸱ May 14, 2026 ⸱ 00h 44m 51s

  2. “Koin + Kotlin Compiler = ♥️” from Android Makers 2026 — Conference ⸱ +100 views ⸱ May 19, 2026 ⸱ 00h 43m 40s

  3. “The Lord of Collection Functions – The Fellowship of Kotlin” from Android Makers 2026 — Conference ⸱ +100 views ⸱ May 18, 2026 ⸱ 00h 38m 16s

→ See 1 more Kotlin talk on Tech Talks Weekly Plus

Kubernetes

  1. “Product Thinking for Cloud Native Engineers” from InfoQ Dev Summit Munich 2025 — Conference ⸱ <100 views ⸱ May 18, 2026

Misc

  1. “How I Doubled eBay’s Productivity – And Still Got Fired” from QCon San Francisco 2025 — Conference ⸱ +3k views ⸱ May 14, 2026 ⸱ 00h 49m 18s

  2. “Tech Truth: Agile Evolution & the Future of SW Engineering • Martin Fowler & Kent Beck • GOTO 2025” from GOTO Copenhagen 2025 — Conference ⸱ +1k views ⸱ May 18, 2026 ⸱ 00h 53m 56s

  3. “Ye Cannae Change the Laws of Physics • Kevlin Henney • GOTO 2025” from GOTO Copenhagen 2025 — Conference ⸱ +900 views ⸱ May 15, 2026 ⸱ 00h 45m 37s

→ See 4 more Misc talks on Tech Talks Weekly Plus

Mobile

  1. “Shipping “Smart UI” with RemoteCompose – Android Makers by droidCon 2026” from Android Makers 2026 — Conference ⸱ +200 views ⸱ May 16, 2026 ⸱ 00h 45m 48s

  2. “The ViewModel Performance Trap: When State Management Bites Back – Android Makers by droidCon 2026” from Android Makers 2026 — Conference ⸱ +100 views ⸱ May 17, 2026 ⸱ 00h 48m 57s

  3. “Ways to reduce your app’s memory and storage footprint – Android Makers by droidCon 2026” from Android Makers 2026 — Conference ⸱ +100 views ⸱ May 15, 2026 ⸱ 00h 40m 44s

→ See 14 more Mobile talks on Tech Talks Weekly Plus

Python

  1. “Beyond Code Coverage: Functionality Testing with Playwright MCP — Marlene Mhangami, Microsoft” from AI Engineer Singapore 2026 — Conference ⸱ +7k views ⸱ May 16, 2026 ⸱ 00h 19m 45s

Security

  1. “Hack Your Own App Getting Started with Mobile Penetration Testing – Android Makers by droidCon 2026” from Android Makers 2026 — Conference ⸱ +100 views ⸱ May 16, 2026 ⸱ 00h 32m 00s

Software Architecture

  1. “Why Most Event-Driven Systems Fail in Banking (And How to Fix It)” from InfoQ Dev Summit Munich 2025 — Conference ⸱ +1k views ⸱ May 18, 2026 ⸱ 00h 51m 25s

  2. “Relevance of Unix Philosophy in Modern Architectural Solutions • Daniel Pokusa • Devoxx Poland 2024” from Devoxx Poland 2024 — Conference ⸱ <100 views ⸱ May 15, 2026 ⸱ 00h 47m 27s

  3. “From SQL to Flow: clean offline-first architecture with Kotlin 2.x” from Android Makers 2026 — Conference ⸱ <100 views ⸱ May 19, 2026 ⸱ 00h 17m 03s

→ See 1 more Software Architecture talk on Tech Talks Weekly Plus

🔓 See everything in Tech Talks Weekly Plus

Never miss great conference talks and podcasts again for $7 $5/month.

  • 📁 Search, filter, sort the full archive of talks & podcasts since 2020.

  • 🗄️ Complete category view of new talks & podcasts.

  • 📈 Most-watched talks from the past 7d, 30d, 90d, 6m, and 12m.

Subscribe to Tech Talks Weekly Plus →

This issue is free for everyone, so feel free to share or forward it!

Enjoy the weekend ☀️ and see you next week!