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
Introducing the Palmyra-mini family: Powerful, lightweight, and ready to reason!
WRITER released three open models in the Palmyra-mini family ranging between 1.5B and 1.7B parameters, designed for efficient inference and specialized reasoning tasks. Built on the Qwen architecture, the release includes the standard palmyra-mini base model alongside two Chain of Thought variants, palmyra-mini-thinking-a and palmyra-mini-thinking-b. In benchmark evaluations, palmyra-mini achieved 52.6% on Big Bench Hard, while palmyra-mini-thinking-a reached 82.87% on GSM8K and palmyra-mini-thinking-b reached 92.5% on AMC23. For palmyra-mini-thinking-b, applying reinforcement learning fine-tuning to an OpenReasoning-Nemotron-1.5B base improved single-shot pass@1 accuracy while reducing sampling diversity and majority@64 performance. The models are available in GGUF and MLX-BF16 quantizations and support inference engines including vLLM, SGLang, TRTLLM, and TGI.
Seamlessly connect Stripe events to your frontend
Stripe events can connect backend state changes to responsive frontend interfaces for payments, subscriptions, connected accounts, and Stripe Terminal workflows. The post contrasts polling, which can increase traffic, latency, and delta-computation complexity, with push subscriptions, while noting that a hybrid of initial API loading and incremental updates is common. It describes webhook endpoints and Amazon EventBridge as ingestion choices, then emphasizes idempotency, ordering, duplicate detection, and failure handling. Frontend delivery options include WebSockets, MQTT, gRPC, Server-Sent Events, polling, AWS AppSync, and Momento, selected according to messaging pattern, direction, latency, scale, and operational trade-offs. An abbreviated lounge example uses a PaymentIntent and Terminal reader, routes payment_intent.requires_payment_method through EventBridge to AWS AppSync Events, and lets a React client mark the reader ready to collect payment.
Workflows: Automate compliance safeguards by flagging transactions
Stripe Workflows can automate additional compliance safeguards by flagging transactions that need review before order processing. The described pattern starts when a Payment intent succeeded event occurs, retrieves the customer, and checks whether the customer’s custom_status metadata is not “Approved” while the payment amount exceeds $5,000. When these conditions are met, the workflow emails one or more designated team members with instructions for assessing the order. The examples motivating review include unusually high order values, unfamiliar sales regions, and behavioral patterns commonly associated with fraud, such as a new customer order above a monetary threshold. Workflows uses a visual building interface in the Stripe dashboard to sequence actions across multiple Stripe products, helping enforce business rules, reduce manual effort, and support risk management.
Tricks from OpenAI gpt-oss YOU 🫵 can use with transformers
To support OpenAI's GPT-OSS series of models, the transformers library introduced several performance upgrades that apply across supported architectures. The release integrates zero-build custom kernels downloadable directly from the Hub, reducing external dependency bloat and compilation friction for operations like Liger RMSNorm and MegaBlocks MoE. Native support for MXFP4 quantization groups vector elements into 32-value blocks with shared scales, allowing GPT-OSS 20B to fit in roughly 16 GB of VRAM and GPT-OSS 120B in roughly 80 GB. In addition, transformers incorporates Flash Attention 3 with attention sinks, continuous batching via the generate_batch API for experimentation, and automatic memory pre-allocation to speed up model loading on GPUs.
Fine-tune Any LLM from the Hugging Face Hub with Together AI
Together AI and Hugging Face have introduced an integration enabling developers to fine-tune compatible Hugging Face Hub models directly on Together AI's managed infrastructure. To launch a fine-tuning job via the Python SDK, users provide a base model from Together's catalog as a configuration template alongside the target Hugging Face repository identifier. This base model template dictates GPU allocation, memory configuration, training pipelines, and inference setup for custom models with matching architectures and sizes. The workflow operates bidirectionally, pulling from public or token-authenticated private repositories and optionally pushing completed checkpoints back to the Hub upon completion. Geared toward CausalLM models under 100 billion parameters, the capability enables faster iteration cycles and domain adaptation without custom DevOps infrastructure.
Securing Salesforce Integrations with Heroku AppLink
Salesforce integrations commonly use Connected Apps or External Services, but long-lived OAuth tokens and public vendor endpoints can expose org data to replay and misuse. Heroku AppLink provides a managed bridge in which a service mesh admits requests only from explicitly connected Salesforce orgs, validates context, and exposes services as native Salesforce actions. For calls back to Salesforce, its SDK obtains short-lived, scoped credentials at runtime through user, user-plus, or authorized-user modes, avoiding refresh-token storage in vendor systems. The mesh uses the x-client-context header to convey org and user information, while the SDK keeps bearer credentials out of application code. The post describes setup commands for the AppLink add-on, service-mesh buildpack, External Service publication, and authorizations, concluding that AppLink improves governance, least privilege, and replay resistance.
Jupyter Agents: training LLMs to reason with notebooks
Small language models often struggle to compete with large frontier models on complex, agentic data science tasks. To improve notebook-based reasoning, researchers simplified agent scaffolding down to roughly two hundred lines of code with dedicated execution and final answer tools, boosting baseline easy accuracy on the DABStep benchmark from 44.4 percent to 59.7 percent. They constructed a curated training dataset by deduplicating two terabytes of Kaggle notebooks, automatically fetching five terabytes of linked datasets, and scoring educational value and relevance using Qwen3-32B. Question-answer pairs grounded in verified execution traces were generated to fine-tune compact Qwen3-4B thinking and instruct models. The team released the trained models alongside the dataset and execution sandboxes, establishing a foundation for reinforcement learning and distillation on notebook workflows.
The Region Consolidation Project
Fly.io is consolidating its region map and deprecating 17 regions across North America, Europe, Asia-Pacific, South America, and Africa. The announcement lists 17 deprecated locations, including Madrid, Bucharest, Warsaw, Atlanta, Boston, Denver, Guadalajara, Miami, Phoenix, Querétaro, Seattle, Montreal, Bogotá, Ezeiza, Rio de Janeiro, Santiago, and Hong Kong, and names 18 consolidated regions: seven in North America, five in Europe, four in Asia-Pacific, and one each in South America and Africa. For apps migrated by Fly.io, most moves should be seamless, although applications with region-specific logic must be updated to use the new region. Manual moves generally require changing primary_region in fly.toml, scaling up in the destination with fly scale count, and scaling down in the original region. If volumes are involved, the guidance recommends stopping existing machines, forking volumes into the new region, and creating machines from those forks; support is offered for issues that arise.
mmBERT: ModernBERT goes Multilingual
mmBERT is a massively multilingual encoder model trained on more than 3T tokens across over 1,800 languages to improve upon existing multilingual architectures like XLM-R. Built upon ModernBERT with a Gemma 2 tokenizer, mmBERT employs a three-phase training curriculum consisting of pre-training on 60 languages, mid-training on 110 languages, and a final decay phase covering 1,833 languages. The training pipeline integrates an inverse mask ratio schedule, dynamic language temperature annealing, and TIES merging across three decay variants. Benchmark evaluations demonstrate strong natural language understanding on English GLUE and multilingual XTREME, as well as competitive retrieval performance on MTEB v2 and CoIR. The release includes standard base and small models alongside open training data and checkpoints.
Triage and Fix with Confidence: heroku run and OTel on Heroku Fir
Heroku Fir addresses production troubleshooting when a release causes failures, such as a background worker breaking after a database schema change, without requiring access to a live formation dyno. Its interactive heroku run command launches a dedicated one-off dyno with the application code and environment variables, relaying terminal input and output for tasks such as rake db:migrate; using it requires a public SSH key and a deployed application. The platform’s native OpenTelemetry (OTel) signals add router request-duration summaries and percentiles, represent heroku.router.connect and heroku.router.service as floats, and attach heroku.app.name to application signals. Together with service.name filtering for first-party services, these capabilities support identifying affected applications, applying isolated fixes, and verifying latency and errors through an observability platform.
Migrating to React Native's New Architecture
Shopify successfully migrated two largest apps, Shopify Mobile and Shopify Point of Sale, to React Native’s New Architecture while continuing weekly releases for millions of merchants. The strategy prioritized minimal code changes, dual-architecture compatibility, and performance and stability parity, with TopHat generating both architecture builds for every pull request. The team upgraded React Native first, updated or removed incompatible dependencies, and changed only native modules that failed under the new architecture rather than migrating all 40-plus modules to TurboModules. Migration issues included state batching, UIManager-related blank screens, shadow-tree manipulation, view flattening, and main-thread legacy module initialization; the reported outcome was maintained development velocity with no feature-development disruption, followed by planned optimization work.
Updating sales restrictions for unsupported regions
Anthropic is updating its Terms of Service to address access by companies from regions where its products are not permitted, including entities operating through foreign subsidiaries. The company says ownership or control from jurisdictions such as China can create legal obligations to share data, cooperate with intelligence services, or support military and intelligence objectives. Under the change, companies or organizations more than 50% owned directly or indirectly by companies headquartered in unsupported regions are prohibited from using the services, regardless of where they operate. Anthropic frames the update as aligning its restrictions with real-world national-security risks and reiterates support for export controls, US energy projects, and rigorous evaluation of AI models for capabilities relevant to national security.
Marketplace monetization with Stripe
Online marketplaces connect buyers and sellers while monetizing commissions, shipping, marketing, payment operations, insurance, or other value-added services. Greens & Dairy Mart illustrates a model in which farmers receive payment after delivery, sellers pay a 15% rate on sales, and shipping, marketing, payout, and insurance options create additional revenue streams. Stripe Connect supports transaction-linked fees through the Platform pricing tool or application_fee_amount, while Stripe Billing provides usage-based meters, subscriptions, and invoices for recurring or one-off charges. The example records pay-per-click, conversion, and penalty events against farm customers, charges monthly subscriptions automatically, and bills damaged-goods insurance annually, with the stated conclusion that programmatic disbursement and monetization give marketplaces a central point to monitor and report these activities.
Welcome EmbeddingGemma, Google's new efficient embedding model
Google DeepMind released EmbeddingGemma, a multilingual embedding model with 308 million parameters and a 2048-token context window designed for on-device applications. Based on the Gemma3 transformer backbone, the architecture replaces causal attention with bidirectional attention to function as an encoder, followed by mean pooling and two dense layers producing 768-dimensional vectors. The model incorporates Matryoshka Representation Learning, allowing outputs to be truncated down to 512, 256, or 128 dimensions for reduced memory and storage footprints. Trained on approximately 320 billion multilingual tokens across more than 100 languages, the quantized model operates under 200 MB of RAM. In domain-specific evaluations on the MIRIAD dataset, fine-tuning increased NDCG@10 from 0.8340 to 0.8862, outperforming larger baselines.
Building Etsy Buyer Profiles with LLMs
Etsy is experimenting with large language models (LLMs) to create anonymous buyer profiles from browsing and purchasing behavior, capturing interests such as styles, categories, and shopping missions. The pipeline combines recent searches, item views, purchases, favorites, session data, and listing information from internal feature-store and BigQuery sources, then prompts an LLM to generate structured profiles with explanations and confidence information. To scale processing, Etsy moved listing data to clustered and partitioned BigQuery tables, shortened input history from two years to nine months, increased batch sizes and compute, added managed concurrency, and used Airflow to stagger tasks. For 10 million users, these changes reduced generation time from 21 days to three and cut estimated cost per million users by 94%; experiments apply profiles to query rewriting and refinement pills.
Meet Mick Hodgins, Notion’s new EMEA GM
Notion announces Mick Hodgins as its new general manager for EMEA, linking the appointment to a priority of expanding its international footprint; 80% of users are outside the US. Hodgins brings 13 years at Google, where he scaled the EMEA ads business and led teams serving customers in more than 20 EMEA countries across changes including mobile, programmatic advertising, the pandemic, and AI. His vision combines local execution with a unified regional team, including new office hubs in London, Paris, and Munich, continued growth of Dublin HQ, and EU data residency. He aims to support students, startups, and enterprises across EMEA while building high-performing teams and positioning Notion as a go-to solution for knowledge workers.
Build Secure Agent-to-App Connections with Cross App Access (XAA) Using OIDC
The guide presents Cross App Access (XAA) as a protocol for secure agent-to-app and app-to-app access, giving enterprise IT centralized control over connections, policies, and accessed resources. It walks through an OIDC demonstration using Todo0 as the Resource App, Agent0 as the Requesting App, and Okta as the enterprise identity provider, covering app registration, trust establishment, token exchange, and API calls. The flow uses the OAuth 2.0 Identity Assertion Authorization Grant: Bob logs in once, Agent0 obtains an identity assertion, Okta exchanges it for a Todo0 access token, and Agent0 calls the protected API. The demonstrated result is an agent that can read pending tasks, add tasks, and mark tasks complete, with changes reflected in Todo0. The guide is marked out of date because its Admin Console instructions have been replaced by AI agent registration under Directory > AI Agents.
SAIR: Accelerating Pharma R&D with AI-Powered Structural Intelligence
SandboxAQ released the Structurally Augmented IC50 Repository (SAIR), an open-source dataset containing 5.24 million computationally co-folded 3D protein-ligand structures paired with empirical IC50 binding potency data. The repository addresses training data scarcity in structure-based drug discovery, where over 40 percent of the included target proteins lack experimental structures in the Protein Data Bank. To construct the dataset, engineers executed over 130,000 GPU hours of the Boltz1 co-folding model on 760 NVIDIA H100 processors hosted via NVIDIA DGX Cloud on Google Cloud Platform. Infrastructure optimizations maintained over 95 percent GPU compute utilization, compressing the generation timeline from three months to three weeks. Quality validation using PoseBusters confirmed that 97 percent of the predicted complexes met physical plausibility and chemical sanity standards.
1.5x faster MoE training with custom MXFP8 kernels
An upgrade from Hopper H100s to Blackwell B200s exposed the Mixture-of-Experts (MoE) layer, implemented with MegaBlocks, as a major training bottleneck: it consumed nearly 53% of forward-pass time and 27% of backward-pass time. The team rewrote the layer at the GPU-kernel level using CUDA, PTX, and some ThunderKittens, while moving from BF16 to microscaled FP8 (MXFP8) with nearly zero loss in training quality. The implementation addresses Blackwell-specific tensor memory constraints and quantization overhead, including custom MXFP8 quantization that sustains more than 6.2 TB/s and emits scale matrices directly compatible with tcgen05.mma. On an internal model, MoE execution became 3.5x faster in both passes, producing a 1.5x end-to-end speedup on Blackwell and 2x versus the original Hopper setup, measured in tokens per second per GPU.
Engineering stories behind the Medium Daily Digest Algorithm: Part 2
Medium’s Daily Digest filtering had become costly and difficult to maintain after a single Bloom Filter instance accumulated billions of items for read, muted, and presentation rules. The system could face up to 15,000 filter questions per feed and more than 15 million per second across thousands of feeds, while Bloom Filters could not delete or enumerate entries and introduced derived-data drift. The team replaced the muted-writer filter with a user-based database query that fetched all muted writers for the current user and cross-referenced them with the story shortlist, then split read filtering between early user-based queries and later pairwise lookups. The resulting implementation removed Bloom Filters entirely, was reported as ten times cheaper in their situation, made costs independent of shortlist size, supported reading-history clearing, and was easier to control and debug.
Top topics
- Architecture834
- Performance816
- LLMs660
- Machine Learning622
- Open Source563
- AI Agents553
- AI470
- Developer Experience448
Weekly digest
The best engineering reads, once a week.