Loading…
Browse by type
Explainer reads
Explainer posts on Yomu, summarised newest first.
- 1,315 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
Sidekick's continual learning loop
Frontier models can accelerate an AI product launch, but their serving cost and latency can become impractical as usage grows, while their weights do not learn from production failures. Shopify’s proposed flywheel captures low-quality conversations, defines quality with a rubric and calibrated judges, improves the surrounding harness through autoresearch, and then converts repaired trajectories into supervised fine-tuning and GRPO training signal. A panel of frontier reasoning models critiques failures, an arbiter produces repair instructions, and unresolved cases go to expert annotators; gist compression then shortens the agent’s static prompt without measured judge-quality loss. In production, the GraphQL agent serves up to 2,000 requests per minute and reportedly achieves a 96% serving-cost reduction versus an estimated frontier-model cost, with lower latency and higher throughput. The approach is presented as a continual-learning system that turns production experience into model-weight updates.
Beyond VLAs: How World Action Models Reshape Robot Manipulation
Robot policies built on vision-language-action models often struggle with physical generalization because vision-language backbones describe scenes instead of predicting physical dynamics. World action models address this limitation by building policies on video world models that jointly predict actions and future world states. NVIDIA Cosmos 3 provides an open foundation using a Mixture-of-Transformers architecture where an autoregressive transformer guides a diffusion transformer to synthesize continuous modalities including video and robot actions. Specializing Cosmos 3 on the DROID platform produced policies capable of simultaneously outputting action trajectories and predicted video observations while boosting RoboLab success from 28.1% to 36.8%. These models support flexible deployment tiers, from workstation serving on RTX PRO 6000 GPUs to real-time 15 Hz on-device execution on NVIDIA Jetson Thor embedded hardware.
The Software Supply Chain Is Under Siege. Devs Are Still the First Line of Defense
A research report by Omdia examines escalating software supply chain security risks driven by growing dependencies on third-party code, open-source software, and artificial intelligence. Over three-fourths of surveyed organizations experienced a software supply chain incident in the preceding year, with known vulnerabilities in third-party software representing the most common exploit vector. Surveyed respondents ranked AI technology as their top supply chain concern, alongside broader issues involving vulnerability remediation and code composition identification. While shifting security left is a high priority for 98% of organizations, 45% still lack confidence in the robustness of their supply chain defenses. Although secure container libraries and Software Bills of Materials provide measurable mitigation benefits, many organizations only generate inventories on a case-by-case basis.
The 9 best tools for your early-stage startup tech stack in 2026
Early-stage startups face a crowded software market and limited budgets, making technology selection a question of leverage rather than completeness. The guide proposes three selection principles: choose tools that scale with usage, integrate with the existing stack, and solve a clearly defined problem. It then recommends Stripe for payments, Slack for collaboration, Granola for meeting notes, Fin and Intercom for customer service, Linear for planning, Datadog for observability, MongoDB for data, Rippling for workforce administration, and Apollo for outbound sales and CRM. The examples emphasize operational coverage, from Stripe’s subscriptions and MongoDB’s flexible schema to Fin’s stated ability to resolve up to 76% of support conversations. The Fin Startup Pack is presented as a way to reduce adoption costs through partner credits and discounts, including more than $500,000 in credits and specified offers for each recommended tool.
GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model
Meta’s Generative Ads Recommendation Model (GEM), a hybrid foundation model for ads recommendations on Instagram and Facebook, trains across several thousand latest-generation GPUs and combines trillions of sparse embedding parameters with billions of dense parameters. The post frames its engineering problem around recommendation-specific jagged and asymmetric sequences, memory-bound operations, numerical sensitivity, communication overhead, activation pressure, and load skew that make conventional LLM training techniques insufficient. Its co-design approach combines custom kernels such as Jagged Flash Attention (JFA), BlockAttention, and Generalized Dot-Product Attention (GDPA), mixed ultra-low-precision training including MXFP8 attention and MLP, topology-aware 5D parallelism, SM-free collectives, and memory and load-balancing techniques. The reported result is a doubling of end-to-end training efficiency to 20–25% Model FLOPs Utilization (MFU) while total training FLOPs increased fourfold over 12 months; Base Batch Shuffling (BBS) alone delivered a reported 4% efficiency gain, with 4% QPS improvement and 4% peak memory reduction.
The New Monday Morning Report: How Generative AI can deliver the insights your executives need.
Retail and CPG joint planning meetings often spend time reconciling fragmented, stale, and inconsistent data instead of deciding how to adjust spend, inventory, or forecasts. The proposed Monday Morning Brief replaces the manually stitched weekly deck with an overnight agent-generated brief built from continuously refreshed point-of-sale, shipment, and inventory data alongside trade, promotion, forecast, and external signals. It uses Delta Sharing for a shared view, Genie Ontology for business context, Unity AI Gateway for permissions, guardrails, logging, rate limits, and human approval, and supports models across AWS, Azure, and Google Cloud. The agent scans item-store combinations, ranks material watchouts, answers cited natural-language follow-ups, and can draft actions such as purchase-order changes, while humans retain approval for consequential decisions. The proposed progression from report to ritual to intelligent decision system includes a 90-minute workshop, a day-30 pilot scope, and a first live Monday targeted for day 90.
Scaling real-time AI agents with session-aware load balancing
Real-time AI agents require infrastructure that manages continuous, bidirectional conversations rather than isolated requests, including interruptions that halt speech generation and update context without dropping the connection. The post argues that QPS and instantaneous CPU utilization are insufficient because they miss long-lived committed work and silent sessions that can become active simultaneously. It proposes application-level active-session tracking around the stream lifecycle, with cleanup in finally blocks and consistent snapshots, then combines session pressure with CPU and memory utilization in a hybrid routing model. The model normalizes session counts into a rate and uses target utilization, smoothed utilization, cost per session, and a safety scaler to estimate effective capacity and regulate new assignments. Validation should use long-lived, varied workloads and metrics such as session distribution, overloaded assignments, startup latency, dropped sessions, and counter behavior; high-concurrency trackers may require sharded or LongAdder-style aggregation.
How AI is transforming analytics at Grab
Grab describes analytics as central to business decisions and presents AI agents as a way to democratise access to trustworthy answers. Over two years, increased model capability has enabled agents to prepare data, write and run queries, analyse opportunities, design experiments, interpret results, and draft commentary in minutes rather than a week. The proposed ladder ranges from AI-assisted work through human-planned and agent-planned workflows to end-to-end autonomy, while people retain accountability for framing, definitions, causal interpretation, decisions, and stakeholder relationships. Grab describes Spartan, which routes natural-language Slack requests across more than 50 skills and 120 analysis frameworks, plus Scarlet for pipeline failure triage and recovery. The programme also maintains agent context through ContextIQ and reports lower analyst intervention, a roughly 33% reduction in cycle times, and greater self-serve resolution in its support channel.
Backstage with Lakebase, part 3
This third part describes joining Backstage’s live operational Postgres catalog with analytical billing data so teams can query resource ownership and cloud spend together. Lakebase isolates compute per workload, allowing portal traffic and FinOps queries to use the same underlying storage; benchmarked catalog queries ran at 55–65 ms end-to-end and searches at two to four ms. Because Lakehouse Federation’s Postgres connector accepts static credentials while Lakebase applications use OAuth JWTs, the proof-of-concept creates a separate SCRAM-SHA-256 Postgres role for federation. This enables a single query to join Backstage resource names with Unity Catalog’s system.billing.usage without data movement, while branch billing remains visible per branch and endpoint. The operational considerations include password rotation, read-only grants, branch TTLs, and scale-to-zero endpoints that stop billing when idle.
Co-Designing AI Model Attention for Fast, Interactive Long-Context Inference
As context lengths increase in long-context and agentic workloads, attention increasingly dominates GPU inference runtime. Dense attention performance depends on architectural parameters like group size, head dimension, and sequence length, alongside GPU parallelization strategies. Prefill remains compute-bound and scales quadratically with input sequence length, showing minimal runtime sensitivity to group size variations. Conversely, decode generation is memory-bound and scales linearly with KV cache length, where doubling group size roughly halves decode execution time. To maximize throughput and interactivity on NVIDIA hardware, model architects should push group sizes higher, select head dimensions of 128 or 256, reduce effective KV state, and keep tensor parallelism less than or equal to the number of KV heads.
Using Triggers On Schemaless, Uber Engineering's Datastore Using MySQL
Schemaless triggers provides a scalable, fault-tolerant, and lossless way to listen for changes in Schemaless, Uber Engineering’s datastore using MySQL, and drives workflows including trip billing, analytics-warehouse ingestion, and cross-datacenter replication. Schemaless stores immutable cells and all their versions, allowing each shard to function as a partitioned change log with monotonically increasing added IDs while retaining random-access APIs. The framework tails shards through get_cells_for_shard, assigns them to worker processes under a leader, persists offsets in shared storage, and invokes column-specific callbacks such as bill_rider; handlers must be idempotent because retries can occur. Failed workers’ shards are reassigned, repeatedly failing cells can move to a separate queue, processing continues during leader or shared-storage outages although redistribution pauses, and replicas insulate triggers from database-node failures.
The Architecture of Schemaless, Uber Engineering's Trip Datastore Using MySQL
Schemaless is Uber's scalable, fault-tolerant datastore whose immutable cells are identified by row key, column name, and ref key, while JSON content remains schemaless and supports eventually consistent secondary indexes. Its worker nodes accept HTTP requests, route them to storage nodes, and aggregate results, while storage data is partitioned into typically 4096 shards, each replicated across one master and two minions in multiple data centers. Writes go to the master and replicate asynchronously, with idempotent retries, circuit breakers, failover, and configurable reads from masters or minions. To reduce loss during master failure, buffered writes persist each request to a primary and secondary cluster, though buffered data may not be immediately readable. Each shard maps to a MySQL database, where an auto-increment added_id primary key supports linear insertion and trigger processing, and MessagePack with ZLib compresses cell JSON blobs.
Best Delivery App to Work with in 2026 | Uber Uber
Choosing a delivery platform in 2026 depends on hourly earning potential, payout speed, vehicle requirements, scheduling flexibility, and extras beyond base fares. The guide compares six US apps—Uber Eats, DoorDash, Instacart, Amazon Flex, Grubhub, and Walmart Spark—across restaurant, convenience, grocery, retail, and package work. It describes Uber Eats features including per-delivery fares, 100% tip retention, Tip Guarantee, Uber Pro Card access to earnings after each delivery, promotions, vehicle flexibility, and scheduling options. Competitors use different models: Instacart shoppers pick groceries, Amazon Flex pays guaranteed block fares, and some services impose market-specific conditions. The guide cautions that earnings are market-dependent, figures may not be directly comparable, and net income must account for fuel, maintenance, phone data, and self-employment taxes.
Dispatches from O'Reilly: The best risk mitigation strategy in data? A single source of truth
Data organizations face severe operational risks from inconsistent metrics, sprawling access controls, and incomplete change management across disparate BI tools and warehouses. Traditional remediation relies on centralized BI gatekeepers and fragmented governance frameworks, which introduces costly bottlenecks and fails to scale. Implementing a semantic layer mitigates these risks by consolidating metric definitions, business logic, and calculations into a single governed location. Centralizing access shrinks the governance surface area and allows metric updates to propagate automatically to downstream consumers, including Tableau, Power BI, Excel, and AI tools. Structured metadata captures context directly alongside models, enabling reliable self-service and federated hub-and-spoke delivery while keeping definitions auditable under version control.
Autoscaling endpoints for LLM inference
Dedicated LLM inference deployments face a trade-off between paying for idle GPUs and suffering sharp p95 latency when replicas hit batching limits. The autoscaling loop converts observed load into desired replicas with ceil(N × observed/target), then applies scale-up and scale-down windows, replica bounds, and GPU placement. It presents inflight_requests as a leading, engine-agnostic default, while TTFT and end-to-end latency target responsiveness but trail demand, and GPU or token utilization prioritize efficiency with less latency headroom. Cold starts can take several minutes because replicas require GPU placement, weight downloads, VRAM loading, and warm-up, so policies should scale early; streaming is required for TTFT, decoding speed, and throughput-per-replica metrics, whereas end-to-end latency supports both streaming and non-streaming workloads.
Foundations for an AI-forward healthcare organization
Healthcare organizations adopting AI face fragmented data, mismatched governance, and no repeatable operating model, rather than a shortage of ideas or vendors. The piece defines an AI-forward organization as one where AI can be built, trusted, and scaled through a foundation of unified data, visible guardrails, and an operating model that helps teams prioritize and move pilots into production. It describes patient identifiers differing across source systems, requiring manual reconciliation and creating recurring integration costs. Governance must avoid both untrusted outputs and approval processes so rigid that nothing leaves the sandbox, while self-service users need controlled access to clinical, operational, and financial data. The source says modern tooling can centralize permissions and enable a governed first-use case in days rather than quarters when scope and data are ready.
Agentic media buying cannot scale without the right foundation. See how buyers and sellers get there on Databricks.
Media buying remains slowed by fragmented coordination across emails, spreadsheets, PDFs, and phone calls, leaving teams to research inventory, compare pricing, negotiate, and issue orders manually. The post presents an agentic workflow in which buyer and seller agents discover inventory, negotiate prices, and book deals through IAB Tech Lab’s AAMP standards, including AdCOM, OpenDirect, OpenRTB deals, and registry-based discovery and trust. Its Databricks implementation runs self-contained applications on Databricks Apps, with CrewAI agent crews using Databricks Foundation Model APIs and Claude models, Lakebase providing transactional Postgres state, and MLflow tracing capturing decisions and tool calls. The example processes a $200,000 Q3 Brand Launch across CTV and Linear TV with a $38 CPM ceiling, and is packaged as a Databricks Automation Bundle accelerator deployable with a single command.
GPU Management: Why Idle GPUs Are the New Grounded Aircraft
Enterprise AI is shifting from a primary focus on model capability to the utilization of scarce, expensive GPU capacity. Like aircraft, GPUs accrue financing, depreciation, power, and cooling costs by calendar hour, while producing value only during useful compute hours; unlike aircraft, they cannot freely absorb workloads with different memory, latency, and duration profiles. The article explains how peak-sized clusters, mixed workloads such as training, fine-tuning, quantization, real-time and batch inference, embeddings, and evaluation can leave capacity stranded even when dashboards show high occupancy, and argues that API pricing can motivate enterprises to own infrastructure. It presents specialized models and GPU orchestration as complementary levers: specialization reduces each workload’s footprint, while management reallocates freed capacity across workloads and queues. The conclusion is that installed capacity alone does not determine returns; enterprises need both strategies to close the gap between GPUs owned and useful output.
How we set up our cloud agent environment
Setting up a cloud-agent environment for the Cursor monorepo required treating development infrastructure as a product for agent users. The team aligned cloud VMs with local development by adapting utilities and setup scripts for Ubuntu, adding dependencies to a Cursor-defined Dockerfile, and introducing security controls for secrets and network access. To make the repository easier to operate, it built anydev to start services, route utility scripts, explain subcommands, and supervise long-running builds. Cursor Cloud MCP and Cloud Doctor then inspect environments, distinguish transient from salient failures, perform root-cause analysis, open high-confidence fixes, and improve workflows using traces. The resulting end-to-end testing and recorded demos helped cloud agents grow from about one in ten merged PRs in December to more than half, with the post identifying tool parity, documented skills, and core-workflow verification as readiness checks.
How to use Google microbenchmarks for evaluating TPU performance
Evaluating TPU performance requires measurements beyond product specifications, because real-world behavior depends on architecture and workload constraints. The suite measures network collectives, compute, HBM bandwidth, host transfers, and Ragged-Paged Attention primitives, using metrics such as GB/s, latency, TFLOPs, MFU, and TTFT-related throughput. Its src and configs directories provide implementations and tunable parameters, while sample Kubernetes recipes connect measurements to TPU architecture and runtime behavior. Roofline analysis uses peak MXU throughput and HBM bandwidth to distinguish compute-, memory-, and network-bound workloads and guide kernel, sharding, or rematerialization choices. In a 110B MoE workload on a 4x4x4 TPU 7x configuration, microbenchmark-guided changes addressed ICI stalls and memory-bound attention, reducing training step time by 21.2%.
Top topics
- Architecture830
- Performance812
- LLMs648
- Machine Learning617
- Open Source561
- AI Agents540
- AI454
- Developer Experience446
Weekly digest
The best engineering reads, once a week.