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
Education Report: How educators use Claude
An analysis of educator AI adoption in higher education examines how faculty use Claude across teaching, research, and administration. The researchers analyzed approximately 74,000 anonymized conversations from Claude.ai Free and Pro accounts associated with higher education email addresses during May and June, automatically filtering educator-specific tasks and matching them to O*NET categories; they also drew on survey and qualitative research with 22 Northeastern faculty. Curriculum development was the leading use, appearing in 57% of analyzed conversations, followed by academic research at 13% and student-performance assessment at 7%. Faculty also used Claude Artifacts to build interactive games, quizzes, simulations, data visualizations, scheduling tools, and budget planners, shifting from conversational assistance toward deployable educational resources. The analysis found more augmentation for lesson design, advising, and grant writing, while routine administrative work was more automation-heavy; grading was automated in 48.9% of relevant cases despite faculty rating it AI’s least effective application.
Protecting the wellbeing of our users
Claude’s safeguards for conversations involving emotional support focus on suicide and self-harm, sycophancy, and an 18+ access requirement. For suicide and self-harm, the approach combines system-prompt guidance, reinforcement learning informed by human preference data and expert input, and a classifier that can display country-specific crisis resources supplied by ThroughLine. The post reports that Claude Opus 4.5, Sonnet 4.5, and Haiku 4.5 respond appropriately to clear-risk single-turn prompts 98.6%, 98.7%, and 99.3% of the time, while multi-turn results are 86% and 78% for Opus 4.5 and Sonnet 4.5. Prefilled real-conversation stress tests produced 91% and 73% for those models, versus 36% for Opus 4.1. The work also describes efforts to reduce sycophancy, evaluate model behavior, identify underage users, and continue publishing methods and results.
Three principles for building a vector platform at Thumbtack
Thumbtack describes a vector platform that lets ML engineers stand up production vector search without separately negotiating database access, building ETL, or writing a query service. The design uses an existing Aurora Postgres cluster with pgvector, while Airflow jobs generate embeddings into BigQuery and managed reverse-ETL syncs them into vector tables; a query API validates inputs and returns ranked results. Three principles guide the layers: reuse trusted infrastructure, treat embeddings as ordinary data flowing through established pipelines, and lower adoption costs for subsequent teams. Production examples include semantic customer-support escalation using cosine similarity and an internal knowledge assistant combining vector retrieval, often with keyword search, before a language model responds; planned directions include self-serve provisioning, vector-backed feature retrieval, and framework-neutral agent memory.
Unifying governance across engines and catalogs in the Open Lakehouse
Apache Iceberg’s REST Catalog additions—read restrictions and catalog labels—address different cross-platform governance problems: delegating enforcement to trusted engines and carrying governance context across federated catalogs. Read restrictions let a catalog evaluate policy for a principal and return bounded column-projection actions and row-filter expressions that a trusted engine must enforce, while leaving trust establishment outside the protocol. Catalog labels exchange opaque key-value metadata at table and column level, allowing a consuming catalog to map classifications locally, preserve native policy expressiveness, and maintain its own audit trail. Together with centralized enforcement through Cross-Engine ABAC, they provide distinct options for direct access and catalog federation, although read restrictions cannot express policies outside their standardized vocabulary.
Why don’t machine learning research agents overfit?
Repeatedly optimizing against long-lived machine-learning benchmarks appears to invite overfitting, yet the post notes that gains on heavily reused benchmarks have largely transferred to newly created test sets. It describes a paper that resets LLM-based research agents, gives an explorer repeated access to validation feedback, and passes the resulting strategy through a token bottleneck to a fresh reproducer. Across eight datasets, strategies from hundreds of experiments were reproducible from prompts as short as 16 to 32 tokens; at shorter budgets, performance diverged, showing that the compressed message carried data-dependent choices rather than only prior knowledge. Deliberate overfitting occurred in 38 of 102 runs, with validation accuracy more than 10% above held-out accuracy, but those gains disappeared after compression; possible pretraining side channels remain unresolved.
From token consumer to token provider: Building your org's AI API
Enterprise AI usage can become costly and fragmented when teams rely on separate provider accounts, personal keys, and independently operated inference stacks, leaving spending, workloads, and governance difficult to see. The post presents Model-as-a-Service (MaaS) in Red Hat OpenShift AI 3.4 as a shared service in which a platform team deploys approved models and exposes them through a standard, OpenAI-compatible API. Administrators can use the model catalog, distributed inference through llm-d on NVIDIA L40s, scoped API keys, rate limits, and dashboards for token use, latency, errors, GPU utilization, and queue length. Developers obtain self-service endpoints and can connect tools such as Claude Code, OpenCode, or Continue with existing clients. The architecture combines OpenShift, vLLM, an AI gateway, Prometheus, Grafana, and Jaeger to centralize inference, governance, and cost attribution across local and external model providers.
To Infinity and Beyond: ThunderKittens Now on NVIDIA Vera Rubin NVL72!
Together’s kernels team ported ThunderKittens to NVIDIA Vera Rubin NVL72 and investigated why existing Blackwell GEMMs underuse the platform. The work targets NVFP4 and FP8 GEMMs: although Rubin preserves Blackwell’s programming model, the original kernels reach only about 42.1% and 44.4% of roofline because tensor cores consume operands twice as fast as the old pipeline supplies them. The team integrates Rubin’s 64-byte K step, exclusive tensor memory extending to 576 columns, oversized shared memory up to 328 KiB, B-side collector buffers, and earlier A-release signaling into its NVFP4 kernel. The resulting kernel exceeds 22 PFLOPS and is described as competitive with cuBLAS and CuTe DSL, while the post also compares Rubin’s specifications with NVIDIA HGX B200.
Async GRPO with LoRA across HF Jobs: a bucket, a proxy, and no NCCL
An AsyncGRPO setup uses LoRA adapter-only synchronization to separate vLLM Jobs, rather than moving full model weights over NCCL. A Storage Bucket mounted at the same path lets the trainer publish versioned adapters that replicas load, while a proxy adds authentication, routes rollouts toward KV-cached prefixes, and broadcasts adapter updates. Versioned adapter names prevent cached prefixes from mixing policy versions, and max_staleness=4 requires six LoRA slots during replacement. Five runs identify client-side concurrency, checkpointing, and packing as important throughput factors: raising in-flight requests to 384, disabling checkpointing, and packing improved median step time from 22.9 to 4.8 seconds. At 500 steps, the final run took 53 minutes instead of 3 hours 27 minutes, trained 31% more samples, and had basically the same reward curve.
What Is SMS Pumping Fraud and How to Stop It
SMS pumping fraud, also called SMS toll fraud, uses bots and fake phone numbers to trigger automated text messages from public-facing endpoints, leaving businesses to pay for traffic that produces no genuine customers. The attack can target OTP requests, sign-up and lead forms, transactional notifications, marketing campaigns, and chatbot endpoints; it may appear as a sudden spike or as distributed low-volume traffic that accumulates over days or weeks. Suggested detection measures include monitoring traffic by use case and geography, reconciling spend with conversions, examining clustered phone-number patterns, and analyzing incomplete forms. The prevention guidance covers CAPTCHA, behavioral bot detection, rate limits, geographic controls, delayed OTP requests, alternative verification methods, and Twilio products including Verify Fraud Guard, SMS Pumping Protection, and Lookup SMS Pumping Risk Score.
Enterprise AI Accuracy: Building a More Trustworthy RAG Application
An enterprise RAG pipeline scored above 90% on standard text-retrieval benchmarks but approximately 46% on complex, customer-representative documents, with failures spanning parsing, chunking, enrichment, embeddings, retrieval, and answer generation. The proposed diagnostic traces wrong answers backward from generated response to retrieved context, chunks, and parsed source, using human-validated end-to-end evaluation and changing one stage at a time. Intelligent Parsing routes pages according to complexity and preserves relationships in structures such as multi-page tables; in a staged benchmark, it raised accuracy from 65.1% to 84.4%, while Enriched Indexing raised it to 86.8%. Retrieval improvements include dynamic metadata pre-filters for explicit constraints and GraphRAG for multi-hop evidence, with GraphRAG slated for general availability in November 2026. Across broader improvements, complex-document accuracy exceeded 90%, and service-based Just-in-Time indexing reduced sub-100 MB payload indexing from 30+ minutes to approximately seven to ten minutes.
When to Use Encode-Prefill-Decode Disaggregation to Accelerate Multimodal Model Serving
Encode-prefill-decode (EPD) disaggregation separates a multimodal model’s vision encoder from LLM prefill and decode, addressing contention when image or video processing delays language-model work. Using NVIDIA Dynamo, the post compares aggregated serving with colocated encoder workers and a heterogeneous, disaggregated encoder tier, with vision embeddings transferred to PD workers through NIXL over UCX RC/TCP Ethernet. EPD is most useful for media-heavy requests, short-to-medium outputs, smaller or quantized models, and mixed text/multimodal traffic; long decode sequences or light media processing reduce its value. In the reported image-heavy test, colocated and heterogeneous EPD reduced TTFT by 58% and 50%, while the heterogeneous tier served 70% more traffic at the same under-100-ms ITL SLO. Mixed-traffic measurements also reduced mean TTFT for text requests from 92.3 to 53.3 ms.
Dario Amodei's remarks at the AI Safety Summit
At the AI Safety Summit, Dario Amodei explains why rapid AI progress makes dangerous capabilities difficult to predict and motivates Anthropic’s Responsible Scaling Policy (RSP). The policy combines AI safety levels (ASLs), modeled loosely on biological safety levels, with recurring tests for dangerous capabilities along the compute scaling curve. ASL-2 covers current models and calls for model cards, external red-teaming, and strong security, while ASL-3 would require unusually strong protection against weight theft and safeguards preventing operationally useful CBRN information; ASL-4 adds escalated misuse and autonomous-system risks. Amodei also describes executive involvement, integrating RSP deadlines into product and research planning, board and external accountability, and the goal of refining company policies into regulatory testing and auditing regimes.
From One Horizontal Scaling Controller to Many: Evolving Uber's Compute Platform
Uber’s failover redesign exposed a limitation in its long-standing single-controller model for scaling stateless workloads. To reuse idle capacity during regional failover, the platform needed multiple orchestrators to influence replica counts without overloading UDC, which remained on the normal deployment lifecycle path. The solution introduced the ServiceScale CRD and Service Scale Controller, allowing separate actors to record steady-state and temporary failover intent in Kubernetes and letting SSC reconcile the combined decision. Production work then exposed risks from informer caches that lagged reality and near-simultaneous multi-writer updates, including ReplicaSet metadata-spec drift that disrupted proportional scaling and could leave workloads stuck. Uber added fleet-wide detection, an automated UDC healer, a long-term scaling-path fix, and extensive kind integration tests, staging, and canaries; the year-long rollout avoided customer-impacting outages while establishing a more extensible control plane.
Why Rider and ReSharper Were Slow to Start, and How Microsoft Helped Fix the Problem
After ReSharper’s out-of-process (OOP) architecture launched, users reported slower Windows startup, and profiling traced the delay to Microsoft Defender scanning the separate process. Defender spends only a couple of seconds on processes in write-protected paths but can add tens of seconds when ReSharper loads DLLs from a user installation directory. The investigation used Microsoft Defender ETW logs and direct CPU-time measurements, with cold-start tests repeated 10 times inside an 8-vCPU, 8-GB Hyper-V VM; JetBrains IDEs took roughly 10–40 seconds to scan, while Microsoft IDEs and other editors stayed under one second. Collaboration with Microsoft led to Defender optimizations in version 1.449.454.0, improving scanning for ReSharper OOP and Rider in write-protected directories. JetBrains also released the Defender Performance Tool for live monitoring, offline snapshots, and CSV export.
A practical approach to end-to-end Solvency II reporting in Databricks
Solvency II reporting connects data, actuarial reserving, capital calculation, quantitative reporting templates (QRTs), the ORSA, approvals, and disclosure, but fragmented systems make readiness and cross-process questions difficult to manage. The described Databricks implementation provides a governed workflow and control tower around existing actuarial and capital suites, automating ingestion checks, data-quality review, reconciliation, orchestration, reporting, and audit trails. It uses MLflow-managed reserving models alongside engines such as Prophet, RAFM, and Igloo, while narrow-scope AI agents can investigate QRT mismatches, suggest remediation, route questions, and support human review. An LLM can draft an ORSA from current numbers, while scenario analysis assesses doubling a cyber book over 12 months; reviewers retain decision authority. The workflow is presented as easier to monitor and investigate.
Beyond embedding: How to secure AI/BI Dashboards for every viewer
Embedding a Databricks AI/BI Dashboard in a customer-facing application is straightforward; authorization determines which rows each viewer may see. The pattern supports external partners and internal Databricks teams with one dashboard and shared dataset, while applying different regional access and masking rules. An entitlements table stores partner IDs and IdP-synchronized group names alongside regions and PII-mask flags; a secured view filters on the signed __aibi_external_value from the embed token. Direct SQL access is separately protected by Unity Catalog row filters and column masks, which evaluate the querying user's identity and groups. The design provides a default-deny, pooled logical multi-tenancy model that centralizes rules while covering both application and workspace query paths.
How we rebuilt Cloudflare Workers’ module registry for Node.js compatibility
Cloudflare has rebuilt the module registry in workerd, the open-source core of the Workers runtime, to improve speed, standards compliance, and alignment with Node.js. The change addresses compatibility beyond runtime APIs by making module resolution, loading, and caching behave more like Node.js across ESM, CommonJS, and WebAssembly. The new registry uses URLs as specifiers, supports import.meta.url, import.meta.main, and import.meta.resolve(), distinguishes modules by query strings and fragments, validates import attributes, follows require(esm) rules, and compiles modules lazily while sharing caches across isolates. It also supports WebAssembly source phase imports and consistent loading errors; the existing registry remains in place, and the new implementation must be enabled with the new_module_registry compatibility flag.
Rust AI in Practice: Building LLM Applications With Rig
Rig is an open-source Rust library that provides a common interface for LLM providers including OpenAI, Anthropic, and Gemini, avoiding application rewrites when switching services. It structures applications around a provider client, completion model, agent, preamble, and registered tools, with async APIs handling provider network requests. In the Rat Code demo, built with Ratatui, separate tools read and write files and run shell commands, while Rig’s streaming API updates the terminal as responses arrive. The library also supports RAG through vector store abstractions, local inference via Ollama, llama.cpp, or Candle, and cassette-based integration tests that replay recorded HTTP interactions in CI; live model quality requires separate scheduled testing.
How we built OCA, our AI on-call assistant
Ramp built OCA, an AI on-call assistant that investigates incidents, shares findings in Slack, answers follow-up questions, and can ask Inspect to prepare code-fix pull requests for human review. In the past four weeks, it assisted with 575 of 1,220 incident-related merged PRs; among incidents with merged fixes, OCA-assisted fixes averaged 37% fewer engineers and about 50% less estimated engineer time, based on Slack activity. Investigations run in a monorepo through Claude Code or Codex, with Temporal managing workflow, while prompts, custom tools, skills, and six one-shot hooks provide successive evidence and reasoning checks. Session-driven testing exposed errors such as misleading one-hour log windows, leading to just-in-time instructions and interim updates that balance accuracy with response speed. OCA remains human-in-the-loop: it has broad read-only access, narrowly scoped writes, and cannot merge changes, although the team is building toward end-to-end routine incident handling.
The Evolution of WSL Support in JetBrains IDEs
JetBrains IDEs have supported WSL through several architectures, but the 2026.2 release makes Native mode the recommended entry point for projects stored there. In Native mode, the Windows IDE delegates filesystem and process operations to the Rust-based IJent agent inside WSL, with EelApi abstracting local and remote environments for IDE and plugin code. Earlier approaches included 9P filesystem access, WSLg, and Remote Development, with problems involving symlink handling, latency, GUI limitations, installation overhead, and client–backend communication costs. Performance tests on a cold first open of spring-framework measured Native mode at 11.5 seconds to become ready versus 18.5 seconds for 9P, while project-tree scanning took 3.7 versus 8.1 seconds. Remote Development remains available, but direct project opening is now preferred in IntelliJ IDEA, WebStorm, and PhpStorm.
Top topics
- Architecture830
- Performance812
- LLMs648
- Machine Learning617
- Open Source561
- AI Agents542
- AI454
- Developer Experience446
Weekly digest
The best engineering reads, once a week.