Loading…
Latest reads
Every post, summarised
Grounded summaries of engineering blogs, newest first. Each links to the original.
- 4,671 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.
Latest posts
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.
Loading Parquet data into MySQL with ClickHouse
The post demonstrates how to move Parquet data into MySQL and query the resulting table using ClickHouse table functions. It sets up MySQL 8.4 in Docker, runs clickhouse-local with a configuration file and named collection, and inspects StackOverflow votes stored in a 2024 Parquet file on S3. An INSERT INTO FUNCTION mysql statement loads the file into a MySQL InnoDB table with 12,296,977 rows, while subsequent examples query MySQL from ClickHouse using either translated ClickHouse syntax or an opaque MySQL SQL string. The post also shows named-collection credentials and a query that identifies posts with both upvotes and downvotes, returning ten ranked results. It concludes that this workflow stays within ClickHouse and notes that table functions also support Postgres, SQLite, and MongoDB.
ClickHouse release 26.8
ClickHouse 26.8 is a long-term support release with 98 new features, 128 performance optimizations, and 556 bug fixes. It adds background queries, PostgreSQL-style regular-expression operators, array subscripting with multiple indices, expanded Parquet and data-lake integrations, and improvements for GROUP BY and joins. Background queries return immediately while continuing after client disconnection, with progress exposed through system.processes and completion details recorded in system.query_log; the example exports 243,619,704 rows to Parquet. Performance work includes dictionary-filter pushdown for Parquet, adaptive parallel GROUP BY processing, smaller hash-table cells for single-string keys, parallel merge joins, IEJoin for two-inequality conditions, and a cost-based optimizer for distributed plans. The PostgreSQL wire protocol supports psql commands such as \d, \dt, and \dv while the release adds text tokenizers.
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.
How to Build a Voice AI Agent with OpenAI's GPT-Live, Twilio Agent Connect, and Python
This tutorial shows how to build a phone-based voice AI agent with OpenAI's GPT-Live, Twilio Agent Connect, and Python. It configures TAC's GPTLiveProvider to bridge Twilio Programmable Voice Media Streams to GPT-Live-1 in the OpenAI API, using 8kHz mu-law audio and avoiding custom WebSocket plumbing. The tutorial sets a proactive greeting, exposes a Python weather function through Responses delegation, and runs the function outside the real-time voice loop so a simulated three-second backend delay does not block the conversation. The resulting setup handles inbound calls through a Twilio webhook and outbound calls with an E.164 number, while the tutorial also notes per-call customization and alternative voice providers.
Build Voice AI Experiences with Twilio and GPT-Live-1 in the OpenAI API
The post introduces a Twilio Agent Connect integration for building production-ready voice AI experiences with OpenAI’s GPT-Live-1 speech-to-speech model. It addresses the gap between real-time speech intelligence and customer-facing communications by connecting GPT-Live-1 to Twilio Voice and Conversations while preserving context across channels and supporting reliable scale. The SDK supplies a native GPTLiveProvider that links Twilio Voice Media Streams directly to the OpenAI API without custom WebSocket plumbing, and handles turn-taking, audio streaming, interruption handling, and tool delegation. Tutorials, documentation, and example repositories cover inbound and outbound agents implemented with Python and Node.js, giving developers materials to build, test, and deploy voice applications.
Native is now the future of mobile at Shopify
Shopify is moving its mobile apps from React Native to Swift and Kotlin after reevaluating the assumptions behind its 2020 decision as coding models improved. React Native had reduced duplicated implementation and feature-parity work, but agents now handle much of the cross-platform implementation, translation, testing, and review work. Shopify is choosing greenfield rebuilds, with Helix breaking migrations into reviewed checkpoints that require tests, visual comparison, adversarial code review, and human approval. The Shop app reached the app stores as a fully native rebuild in 12 weeks, while the Shopify app migration is underway. The company will continue selected React Native library support while migrating all its apps, and will judge success by product velocity, app quality, and autonomous agent work.
121+ Cyber Monday email subject line examples in 2026
Cyber Monday email marketing depends on subject lines to help messages stand out in a crowded inbox, even when designs and discounts are strong. The guide recommends pairing subject lines with preview text, emphasizing exclusivity and urgency, stating discount amounts, using emojis selectively, and keeping copy concise. It also covers personalization, segmentation, pre-event notices, abandoned-cart messages, and reminders. More than 121 examples and templates are organized into Cyber Monday, pre-event, FOMO, abandoned-cart, and emoji-heavy categories, with additional seasonal resources. The post positions Twilio SendGrid for campaign execution and reports 75.1 billion emails processed during Cyber Week in 2025, a 14.6% increase over the previous record-breaking sending, but provides no comparative performance results for the example lines.
Migrating Shop app from React Native to native
Shop migrated its React Native app to native iOS and Android implementations in SwiftUI and Jetpack Compose after coding-agent advances changed the economics of a shared mobile codebase. A one-week proof of concept showed that one engineer using agents could recreate screens, interactions, and flows closely enough to make a full migration feasible although the result was not production-ready. Six engineers then built the native foundations and main journeys with feature teams validating edge cases; they preserved sign-in state, push notifications, analytics events, and behavior while simplifying some screens. Compared with React Native, native startup was 23% faster on iOS and 50% on Android, stability rose from 99.5%+ to 99.95%+, Android size dropped 109 MB, release builds fell about 75%, and scrolling reached 120 FPS. The migration’s workflow combined specialized agents, plan approval, Tardis live-app feedback, screenshot and event comparisons, and human review while feature parity remained a release requirement.
Build an AI Voice Assistant with Twilio Voice and Media Streams, OpenAI's GPT-Live API, and Node.js
This tutorial shows how to build an AI voice assistant that answers phone calls using Twilio Programmable Voice, Twilio Media Streams, OpenAI's GPT-Live-1, and Node.js. GPT-Live-1 acts as the speech-to-speech model for listening, talking, and handling interruptions, while the delegated gpt-5.6-terra model handles deeper reasoning and tool use. The implementation uses a Fastify server with WebSocket support, dotenv configuration, a marin voice, and separate prompts for the voice and backend roles. Callers can request a city's top headline through the custom get_top_headline function or ask factual questions through OpenAI's hosted web_search tool. The guide exposes port 5050 with ngrok, points a Twilio number at an /incoming-call webhook, and concludes that the working example fits in under 200 lines.
Make Outbound Calls with Twilio Voice and Media Streams, GPT-Live-1 in the OpenAI API, and Node.js
This tutorial shows how to build an outbound AI voice agent that uses Twilio Programmable Voice and Media Streams to call a permitted phone number, rather than waiting for inbound calls. It connects the call to GPT-Live-1 in the OpenAI API through a Node.js server, with Fastify, ws, dotenv, and the Twilio helper library handling the application and relay. The implementation reads the destination from a --call flag, validates E.164 formatting, exposes the server through a public tunnel, and uses environment variables for credentials and the calling number. Owlie identifies itself as an AI assistant, retrieves a callback note with get_callback_reason, delegates factual questions to hosted web_search, and can be interrupted during speech. The completed demo rings the chosen number, explains why it is calling, and can return random callback notes or web-searched answers, while the tutorial stresses verified numbers, model access, and outbound-calling compliance.
Why You're Getting 429s in Production Using Auth0 and How to Stop Them
Auth0 429 responses in production are tied to rate limits, commonly when applications call the Management API for every request or obtain a new machine-to-machine (M2M) token for every request. Because limits are per tenant and vary by plan, logging x-ratelimit-limit, x-ratelimit-remaining, and x-ratelimit-reset headers exposes approaching exhaustion. User roles, permissions, and metadata that can be treated as access-token snapshots should be added as custom claims by a Post-Login Action, while APIs read them from the JWT. M2M tokens, valid for 24 hours by default, should be cached and refreshed with a 60-second expiry buffer; data requiring freshness should sit behind a server-side cache with a 30-to-60-second TTL. When a 429 occurs, retry at the Unix timestamp in x-ratelimit-reset rather than applying exponential backoff, because the reset time is deterministic.
Rebuilding AUTOMATIC1111 with Gradio Workflow
Workflow1111 rebuilds most of AUTOMATIC1111’s stable-diffusion-webui feature set as a single Gradio Workflow canvas. It combines eleven media pipelines and 73 nodes for text-to-image, refinement, image-to-image, prompting, interrogation, detection masks, annotators, upscaling, background removal, PNG metadata, and image-to-video. The implementation uses four operator kinds—Python fn functions, model calls through InferenceClient, Space calls, and Hub dataset rows—with parallel branches for independent work; 32 of 36 operator nodes are fn nodes, and 22 run in-process without network calls. Models can run through Inference Providers, Spaces, or a local GPU, while output nodes automatically become REST endpoints and MCP tools. The result is a browser-accessible, OAuth-enabled workflow that can also be duplicated, rewired, deployed, and called from code.
ClickHouse Cloud vs. Snowflake: What drives the real-time performance-per-dollar gap
The benchmark compares ClickHouse Cloud with Snowflake under a continuous real-time workload, asking whether ingested data becomes query-ready quickly enough for interactive analytics. Both systems received the same 113.2-billion-row NBBO dataset at a target of 1 million rows per second, with matched sorting or clustering keys, pre-aggregations, query schedules, and approximately 16 CPUs of query-serving compute. ClickHouse used native asynchronous inserts and maintained ordering and pre-aggregation during ingestion, while Snowflake used Snowpipe Streaming and Interactive MV refresh as separately metered services. In the end-to-end CostBench comparison, ClickHouse Cloud delivered 412× better performance per dollar: Snowflake’s path cost 3.32× more and its matched queries took 124× longer in total. The comparison attributes the gap to query-readiness work, refresh lag, reconciliation, and growing drill-down latency.
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.
Our compliance framework for California's SB 53
California’s Transparency in Frontier AI Act (SB 53), taking effect January 1, requires frontier AI developers to publish frameworks explaining how they assess and manage catastrophic risks. Anthropic’s public Frontier Compliance Framework (FCF) covers cyber offense; chemical, biological, radiological, and nuclear threats; AI sabotage; loss of control; model-weight protection; and safety-incident response. It describes a tiered system for evaluating model capabilities across those risk categories and the mitigations used, while building on the company’s Responsible Scaling Policy and model system cards. Anthropic says the FCF will serve as its compliance framework for SB 53 and other regulatory requirements, while the RSP remains a voluntary policy that may go beyond current rules. The company also argues for a federal standard requiring public secure-development frameworks, deployment system cards, whistleblower protections, flexible minimum standards, and coverage limited to established frontier developers.
Introducing Labs
Anthropic is expanding Labs, a team dedicated to incubating experimental products at the frontier of Claude’s capabilities. The announcement describes a product-building model that combines edge experimentation, testing unpolished versions with early users, and scaling successful ideas into dependable customer products; it cites Claude Code, MCP, Skills, Claude in Chrome, and Cowork as examples. Mike Krieger is joining Labs to build alongside Ben Mann, while Ami Vora will lead Product in partnership with CTO Rahul Patil to scale Claude experiences. Anthropic says the structure is intended to support both experimental discovery and responsible scaling for enterprise customers and a growing user base, and it is hiring builders focused on emerging technology.
Irina Ghose named Managing Director of India
Anthropic has appointed Irina Ghose as Managing Director of India as it prepares to open its first office in the country. Ghose brings more than three decades of experience scaling technology businesses and most recently led Microsoft India, where she drove enterprise AI adoption across banking and financial services, healthcare, manufacturing, and government. In her new role, she will help grow teams, engage policymakers and academic institutions, strengthen developer engagement, and build partnerships with enterprises and organizations addressing local challenges, with an emphasis on local insight, trust, safety, and long-term impact. India is Claude.ai’s second-largest market globally, and Anthropic says nearly half of usage there is concentrated in computer and mathematical tasks, according to its fourth Economic Index.
Mariano-Florentino Cuéllar appointed to Long-Term Benefit Trust
Anthropic’s Long-Term Benefit Trust has appointed Mariano-Florentino (Tino) Cuéllar as a new member, adding experience in law, governance, international affairs, and technology’s effects on public institutions. Cuéllar is a former Justice of the Supreme Court of California, former leader of Stanford’s Freeman Spogli Institute, current President of the Carnegie Endowment for International Peace, and chair of the William & Flora Hewlett Foundation; he plans to leave Carnegie in July 2026 for Stanford roles. The independent Trust selects members of Anthropic’s Board of Directors and advises Anthropic’s Board and leadership on maximizing advanced AI’s benefits and mitigating its risks, with no Trustee holding a financial stake in Anthropic. Kanika Bahl and Zachary Robinson, who served since the Trust’s founding, concluded their terms.
Anthropic partners with the UK Government on GOV.UK
Anthropic has been selected by the UK's Department for Science, Innovation and Technology (DSIT) to help build and pilot a dedicated AI-powered assistant for GOV.UK. The assistant will initially support people entering or re-entering the workforce by helping them find work, access training, understand available support and resources, and route them to relevant services. Powered by Claude, the agentic system is designed to provide individually tailored guidance, maintain context across interactions, and let users control what data is remembered or opt out; personal information will be handled under UK data protection law. Using DSIT’s “Scan, Pilot, Scale” framework, the partners will test and iterate the assistant before wider rollout, building government expertise to maintain it independently.
Top topics
- Architecture834
- Performance816
- LLMs660
- Machine Learning622
- Open Source563
- AI Agents553
- AI470
- Developer Experience448
Weekly digest
The best engineering reads, once a week.