Loading…
Browse by type
Explainer reads
Explainer posts on Yomu, summarised newest first.
- 1,318 posts
- 20 companies
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Explainer posts
Reading the agent traces is how you make the call your eval can't
Automated evaluations often miss unknown failure modes in AI agents because developers only write tests for anticipated errors. While building an itinerary agent for AI Engineer World's Fair, an open-weight model routed a speaker query to a tool returning only tracks, leading the model to hallucinate famous speakers and falsely claim the tool provided that data. Because a tool was executed, the trace appeared grounded until inspectable tool outputs revealed the discrepancy. To resolve this, the author retained the inexpensive model, corrected query routing, tightened system prompts against ungrounded entities, and added a custom entity-diffing eval in CI. Ultimately, directly reading agent telemetry traces alongside errors and logs is essential for diagnosing root causes and deciding trade-offs between model costs, prompt adjustments, and routing fixes.
Multi-token Residual Prediction
Multi-Token Residual Prediction (MRP) adapts multi-token prediction for diffusion language models by training a small module to predict the residual between adjacent denoising steps rather than the full next-step distribution. This residual target is easier to learn, allowing a tiny 3-layer transformer to accurately predict corrections across multiple steps. In the static regime, MRP enables lossless speculative decoding with up to 1.56× throughput in SGLang, or direct decoding with a small quality cost. In the dynamic regime, MRP recovers up to +16 accuracy points lost by aggressive low-threshold decoding through a remasking mechanism. The module is trained on a frozen backbone and composes cleanly with existing inference methods.
Why we built ADK 2.0
ADK 2.0 addresses the structural challenge of moving AI agents from prototypes into production, where infinite loops, hallucinated bypasses, and failures without clean exceptions make model-led orchestration unreliable. Because language models are slow, costly, and variable for routing, scheduling, and error handling, the release separates execution routing from language processing and reserves LLMs for ambiguous or cognitive steps. Building on ADK v1’s support for Python, Java, Go, TypeScript, and Kotlin, it adds a structured workflow runtime and task-collaboration model, with Workflows available in Python since March and newly launched for Go. Deterministic graphs combine tool calls, Human-in-the-Loop steps, and specialized agents with explicit branching and error handling. Dynamic workflows extend this model with native Python and asyncio control flows for retries, branching, and reusable sub-workflows.
From monolith to Lakebase to LTAP: rethinking the database from storage up
The post examines why monolithic OLTP databases become fragile and difficult to scale when compute, the write-ahead log (WAL), and data files share one machine. Using Postgres as its primary example, it contrasts local WAL durability and physical cloning with Lakebase, whose stateless Postgres compute externalizes those components into SafeKeeper and PageServer services. SafeKeeper replicates log records across a quorum through Paxos-based network replication, while PageServer applies the WAL and materializes data in cloud object storage. The resulting separation is presented as a way to improve durability, independently scale reads and writes, isolate workloads, and support high availability and branching without physical database clones. LTAP extends the design by making transactional tables directly queryable for analytics from a single governed copy, avoiding CDC or mirroring and allowing transactional and analytical engines to scale independently.
Designing GPU-Accelerated Query Engines with NVIDIA GQE
GPU-accelerated query engines encounter performance constraints stemming from memory and I/O bandwidth limitations. NVIDIA introduces GQE, a reference architecture using NVIDIA cuDF, nvCOMP, and other CUDA-X libraries to accelerate SQL query execution across modern hardware like the NVIDIA GB200 NVL4. The architecture coordinates three distinct layers: a query layer consuming Substrait plans, a data layer managing in-memory table transfers, and an execution layer executing relational operator task graphs. GQE employs pipeline parallelism, partition pruning, hybrid compression, and batched asynchronous transfers to overlap data movement with GPU compute. On a non-standard 1 TB TPC-H benchmark, GQE on a single B200 GPU achieved a 7.5x total execution time speedup over dual-socket CPU configurations running DuckDB.
Why intent prediction needs more than an LLM
Traditional large language models rely on next-token prediction trained over broad text sequences, excelling at conversational synthesis and code generation while struggling with forecasting future user behavior. Predicting intent inherently requires modeling decision-making under conditions of uncertainty, a capability not directly produced by general text pretraining. Dedicated behavioral foundation models address this limitation by training representations directly on proprietary, sensitive behavioral data, such as anonymous browsing sessions and product identifiers. By learning base representations of behavior, these models generalize to unseen predictive tasks, such as optimizing ad campaigns for products absent from the original training dataset. Implementing these behavioral models also necessitates privacy-centric machine learning techniques, ranging from differential privacy concepts to homomorphic machine learning.
Washington: Background Check Criteria Disclosure
Uber’s Washington State driver-screening disclosure sets time-based limits on driving violations, criminal charges, and convictions for people seeking to provide rides on the platform. Drivers may have no more than three minor driving violations in three years, while major driving violations and pending criminal matters such as driving while unlicensed or uninsured are disqualifying during that period. A broader seven-year bar covers felonies and specified offenses including violence, theft, fraud, DUI-related violations, hit and run, reckless driving, and speeding over 100 MPH; the disclosure also bars serious offenses including sexual crimes, murder, kidnapping, and active warrants. Seattle and King County add a seven-year licensing-related rule, while Vancouver adds tighter recent-violation limits and separate felony and drug-offense lookbacks.
Washington State Driver Information
This page outlines benefits, protections, rights, and eligibility information for drivers operating on the Uber platform in Washington state. Starting July 1, 2026, drivers may request a downloadable CSV report containing two years of Washington mobility trip receipt data through drivers.uber.com; an email with an access link will arrive within three business days, and the link expires after seven days. It directs drivers to multilingual notices about driver rights and Paid Sick Time, states that drivers may be entitled to Paid Sick Time and Unemployment Benefits, and provides contact information for appealing deactivation through Drivers Union. The page also describes background check criteria for operating on the platform, noting that different criteria may apply in Seattle and King County.
Agents hate friction: early thoughts on building for agents
Designing software for LLMs and agents requires more than extending human developer experience: a polished text-first interface can still produce an AX that is merely functional. The post frames an agent as a combination of model, harness, context, and skills, with behavior varying across providers, models, and harnesses; it focuses attention on commonly used combinations while testing beyond them. In a test setup, agents receive goals and different amounts of context, and model-harness combinations are evaluated by how they reason, which tools they call, how many steps they take, and how consistently they repeat a task. The resulting design questions include friction, context and token efficiency, familiarity with established tools, and the trade-off between fewer steps and higher context or cost; the central conclusion is that agents may abandon a product when friction persists, making testing necessary to expose that behavior.
How Verifiable Digital Credentials Are Reshaping Trust Architecture
Verifiable Digital Credentials (VDCs) are presented as a new trust layer for applications that need to verify claims such as age, licenses, employment, or eligibility without collecting documents and excess PII. The model uses an Issuer-Holder-Verifier trust triangle: authorities sign credentials, users hold them in wallets, and applications request specific proofs, with selective disclosure limiting exposed data. The standards stack combines SD-JWT VC or mdoc credential formats, OpenID4VCI for issuance, OpenID4VP for presentation, and the W3C Digital Credentials API for browser-wallet invocation. Production verification also requires issuer trust lists, trust-framework integration, and key rotation, not merely signature checks. Because browser API rollout is still progressing, the recommended baseline is redirect-based OpenID4VP with the Digital Credentials API added as a progressive enhancement, starting with one bounded verification use case.
A Decision Framework for ETL Migration to Databricks
An ETL migration with hundreds of stored procedures, schedulers, scattered permissions, and a warehouse renewal deadline needs workload-by-workload decisions rather than a single rewrite strategy. The framework assigns work among Databricks SQL, Spark Declarative Pipelines (SDP), and PySpark or Spark SQL notebooks. SQL tasks suit single statements, Unity Catalog-governed stored procedures handle procedural logic, SDP manages dependencies, retries, quality constraints, and batch-plus-streaming, while notebooks cover complex logic, ML feature engineering, integrations, and large or tightly controlled Spark workloads. It recommends four phases—assessment, quick wins, modernization, and optimization—using profiling, side-by-side validation, and parallel runs before retiring legacy systems. Migration tools can automate 60–80% of initial conversion, but architecture choices remain essential: the goal is consolidating orchestration, metadata, lineage, permissions, and validation rather than reproducing technical debt.
Paging Charity! How can engineering leaders avoid becoming Bond villains?
Engineering leadership advice often prioritizes personal empathy over commercial competence, which Charity Majors argues is a misdiagnosis of organizational failure. Because over 90% of venture-backed startups fail, tech culture frequently venerates ruthless founders who succeed commercially. New engineering managers frequently focus solely on subjective employee comfort, mistaking business instability or misaligned stakeholder priorities for micromanagement. To champion humane values sustainably, engineering leaders across all tiers must first master business operations, clarify product strategy, and deliver business wins. Advancing to the director level requires moving beyond generic cheerleading, building high-performing teams oriented around continuous improvement, and demonstrating systems thinking alongside business acumen.
From test bench to lakehouse: how AVL modernizes measurement data analytics with Impulse
AVL’s Lakehouse for Measurement Data addresses the scale, reproducibility, and governance limitations of desktop tools and isolated scripts used for automotive measurement analysis. Built on Databricks, the platform ingests ASAM MDF4 and other files into a Medallion Architecture, applies configurable DQX quality rules in a hierarchical Silver model, and uses Impulse to compile declarative Python TSAL expressions into distributed Spark execution. Engineers can select channels, create virtual signals with alias resolution, unit conversion, time alignment, and interpolation, define events, and compute duration- or distance-weighted aggregations in about 10 lines of Python. Impulse supports Gold-layer reporting, ad-hoc Spark DataFrames, and ML feature matrices, with Unity Catalog governance and Workflow orchestration. AVL reports reducing analysis time from days to minutes, processing many recordings per run, lowering infrastructure costs versus on-premises solutions, and enabling self-service, reproducible, standardized analysis.
Q&A: How KRAFTON Built PUBG Ally, a Co-Playable Character Powered by NVIDIA ACE
KRAFTON developed PUBG Ally, an on-device AI co-playable character for PUBG: BATTLEGROUNDS, using NVIDIA ACE tooling and models. Conventional non-playable characters rely on fixed dialogue, but PUBG Ally processes live game states and player voice input in real time. The interactive pipeline combines automatic speech recognition, a quantized Mistral-NeMo-Minitron-2B small language model, and a custom text-to-speech model running locally on client GPUs with at least 8GB of VRAM. To minimize latency, the architecture separates reactive gameplay into a behavior-tree System 1 layer while routing strategic decisions and speech generation to the language model. KRAFTON tuned the non-deterministic system across multi-layer evaluations and large-scale playtests with over a thousand players.
What To Look For in a Serverless Database for AI Applications
Serverless databases are presented as a baseline for AI applications whose traffic can be volatile, idle for long periods, or spike when agents fan out queries. The guide distinguishes managed serverless systems from autoscaling products by focusing on architectural separation of compute and storage, true scale-to-zero, cold-start behavior, connection handling, pricing, performance, portability, governance, and AI capabilities such as vector search. It recommends evaluating both low- and high-utilization costs, published warm-up times, tail latency (p95/p99), and built-in pooling or HTTP/Data APIs for high-concurrency agents and serverless functions. The text positions Lakebase as an example that combines serverless Postgres, shared lakehouse storage, and Unity Catalog governance, and cites reported cost and management reductions from a 2025 study while noting that provisioned deployments may suit continuously high-throughput workloads.
What Is Serverless PostgreSQL?
Serverless PostgreSQL is presented as a fully managed cloud database model that decouples compute and storage, allowing each to scale independently with demand. Traditional deployments require teams to size infrastructure, manually manage scaling, and absorb costs from idle capacity. In serverless systems, the provider provisions compute on demand, can suspend it when idle, and bills according to active usage; scale-to-zero may introduce cold-start latency. The architecture can also support database branching through copy-on-write, creating isolated environments without duplicating data. The article distinguishes this model from lakebase architecture, which combines transactional and analytical workloads on a shared foundation using decoupled compute, durable object storage, log-based storage systems, and orchestration.
The Rise of Sports Intelligence: How the Lakehouse Turns Tracking Data into Competitive Advantage
Professional basketball’s Hawk-Eye SkeleTRACK feed produces roughly 22,620 positional updates per second—about 65 million records per 48-minute game—yet teams often cannot turn that volume into timely, trusted decisions. The integration gap comes from separate vendors for tracking, wearables, video, scouting, and medical data, alongside calibration differences, weak provenance, and compute limits. The Databricks Data + AI Platform is presented as a governed lakehouse that ingests feeds with Lakeflow, refines them through medallion layers, and uses Unity Catalog for lineage, access control, and auditing. Models for shot probability, injury risk, and fatigue can run alongside serving and custom applications, with Lakebase supporting sub-second interactive queries. Applications include proactive load management, real-time coaching intelligence, and enriched broadcast or fan experiences across tracking-rich sports.
amazon.science
The fuel of the future is already here: Why TRISO matters
Amazon is investing in next-generation nuclear technology powered by tristructural isotropic (TRISO) fuel particles to satisfy the increasing energy demands of cloud computing and artificial intelligence infrastructure. Each sub-millimeter TRISO particle encloses a high-assay low-enriched uranium kernel within a porous carbon buffer and a three-layer ceramic shell featuring silicon carbide that acts as a robust miniature containment vessel. Testing by the U.S. Department of Energy demonstrated that irradiated particles withstand temperatures of 1600°C for 300 hours without detectable failures. In pebble-bed designs like the X-energy Xe-100 reactor, graphite spheres containing TRISO fuel circulate continuously through the core to enable online refueling, passive heat removal, and modular 80-megawatt deployments at facilities like the Cascade Energy Center.
What are Dashboards?
A dashboard is a visual interface that combines metrics, KPIs, and visualizations from one or more sources so users can judge performance against a goal at a glance. It acts as an organized, live visual layer over databases, warehouses, cloud applications, spreadsheets, or other sources, rather than being merely a collection of individual charts or reports. A typical flow sends source data through a query or pipeline into a visualization layer; modern dashboards may query underlying data directly and refresh on a schedule or in real time, avoiding stale copied data and definition drift associated with older extract-based architectures. Filters and drill-downs support investigation, while dashboard types—operational, analytical, strategic, and tactical—differ by audience, time horizon, and update frequency; usefulness ultimately depends on a clear purpose, defined audience, trusted data, and enough context for action.
What if the answer was already in your data?
Kythera Labs is building an AI-native healthcare strategy platform on Databricks to give health systems access to strategic intelligence that historically required specialized analysts or consulting firms. Its foundation converts 339 billion medical and prescription claims covering more than 300 million patients into governed, event-based data, resolving providers, harmonizing codes across 130 vocabularies, and reconstructing patient journeys. Healthcare Strategy Agent, built with Agent Bricks, lets executives ask questions such as where oncology referrals are going and receive analyses of leakage, competing providers, physicians, and reimbursement opportunity in minutes. A Louisiana health system went live within ten days and reported 150% greater visibility into encounters, 12% more keepage, 22% less leakage, and $3.8 million in estimated annualized retained-encounter value. Unity Catalog, Lakebase, Delta Lake, Delta Sharing, and serverless infrastructure provide shared governance, lineage, access controls, and operational integration.
Top topics
- Architecture830
- Performance812
- LLMs648
- Machine Learning617
- Open Source561
- AI Agents542
- AI454
- Developer Experience446
Weekly digest
The best engineering reads, once a week.