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 Auth0 SDK for TanStack Start
Auth0 released a dedicated SDK for TanStack Start, a full-stack React framework, to replace hand-rolled OIDC flows or adapted SDKs. The SDK provides hooks, components, route guards, server middleware, and server helpers that integrate with TanStack Start’s server functions, middleware, and file-based router. Setup involves installing the package, configuring environment variables for the Auth0 application, creating a server-side Auth0 instance, registering the auth0Middleware in the start file, and wiring the router context and Auth0Provider in the root route. The SDK automatically mounts auth routes for login, callback, logout, profile, and backchannel logout, and supports features like login, logout, protected routes via requireAuth, session reading on the server with getSession, and additional capabilities such as passkeys, organizations, MFA, and API scope checks. The fastest way to try it is the official quickstart, which scaffolds a fresh app and adds the SDK step by step.
Mix, Don't Tune: Bilingual Pre-Training Outperforms Hyperparameter Search in Data-Constrained Settings
In data-constrained language model pre-training, two remedies exist: aggressive hyperparameter tuning such as high weight decay, and mixing in data from a high-resource auxiliary language to expand the training distribution. This study systematically compares these levers across four model scales from 150M to 1.43B parameters, using Arabic as the low-resource target and English as the auxiliary, over approximately 1000 pre-training runs. Mixing yields larger improvements than hyperparameter tuning on both validation loss and downstream task accuracy, and the gap grows with model size. Mixing boosts performance by an amount equivalent to 2–3× the unique target data on validation loss and 2–13× on downstream task accuracy, with the gain scaling steeply with model size. The authors recommend mixing in a high-resource language, prioritizing the mixing ratio over hyperparameter tuning, and transferring hyperparameters from a small proxy model via μP, noting that target-language validation loss systematically underestimates mixing’s value.
OCP Summit 2026: Powering AI with Open Innovations in Scale-Up, Scale-Out, and Scale-Across Networks
At OCP Summit 2026, Meta unveiled next-generation open networking infrastructure for AI, including IcePack, its first liquid-cooled network rack, and IceCube, a 102.4T switch blade delivering 128×800G non-blocking capacity. The Flattened Planar Fabric (FPF) introduces a multiplane scale-out design where each GPU connects via eight independent 100G links to different switches, reducing latency and enabling larger clusters. Meta also shared its BAG super-spine for multi-site aggregation and Tahan, an Ethernet-based 51.2T scale-up switch for MTIA 300 systems. The company launched the Open Switch Platform Architecture (OSPA) initiative with Microsoft and Alibaba to standardize switch platform hardware and software interfaces. Additionally, Meta announced open-source implementations of MetaRoCE, its RDMA transport protocol optimized for AI cluster networks over Ethernet.
The data lakehouse is a better data warehouse: 2026 benchmarks and proof
Data warehouses have been the standard for structured SQL analytics for two decades but cannot serve modern AI, ML, streaming, and unstructured data workloads. The data lakehouse now matches warehouse SQL performance while serving all these workloads from a single governed foundation. Databricks Lakehouse improved production SQL workload performance by 77% from a 2022 baseline, with BI queries improving 14% over a recent five-month period. Customer migrations show 90% faster queries and 30–40% cost reduction at Lumen Technologies, and 80–90% faster analytics at Trek Bicycle. The lakehouse eliminates redundant data copies, supports open formats like Delta Lake and Iceberg, and extends governance to AI agents through Unity Catalog and Unity Gateway.
How Flo Health built safer, more useful open models with Databricks AI Runtime
Flo Health, a women’s health app with 81 million monthly active users, aimed to improve generative AI responses for medical safety and usefulness while operating at scale. Relying on general-purpose models created cost, dependency, and domain-expertise limitations, so Flo built a post-training pipeline with Databricks AI Runtime. The team defined 26 AI judges across three categories: 12 medical-safety judges acting as hard constraints, 6 micro judges for nuanced behaviors, and 8 usefulness criteria. Using these judges as a reward signal, they generated over 15,000 synthetic questions from privacy-first principles and applied reinforcement fine-tuning with supervised fine-tuning (SFT) and direct preference optimization (DPO). After fine-tuning on 32 NVIDIA H100 GPUs using PyTorch and Fully Sharded Data Parallel (FSDP), medical safety scores rose from 34% to 61%, usefulness from 42% to 76%, and the combined metric improved nearly threefold to 47%.
Brickster Voices: Meet Clara Kurian
Clara Kurian, Databricks’ first university recruiter for APJ, leads the company’s early career hiring in India. Databricks has invested over $250 million in India, treating it as a strategic hub rather than a satellite office, with workforce growth of 270% since 2022 and leadership expansion of 350% across all business units. Clara designed the first structured India Intern Program with 51 interns, aiming to give them a genuine experience of the company’s pace, people, and culture through Speed Mentoring with engineering leadership and activities like Sneaker Art and Pickleball socials. She advises early career candidates to build strong fundamentals, curiosity, and a builder mindset, noting that the field moves fast and thrives on continuous learning. Candidates stand out by showing genuine curiosity, authenticity, and a willingness to figure out unknown problems, as the technical bar is almost a given.
Can we change an AI model’s architecture without retraining it from scratch?
The Transformer architecture’s memory cost scales with token count, making long-context AI agents expensive. In controlled comparisons of three state space models - Gated KalmaNet, Gated DeltaNet, and Mamba-2 - Gated KalmaNet performed best, offering variable-length inference without retraining. Priming identifies which attention layers can be replaced by state space models using a sliding-window attention proxy, then reuses pre-trained weights to convert them into hybrid architectures without retraining, reducing pretraining cost by over 99%. A 32-billion-parameter hybrid model matches the original Transformer’s reasoning quality while decoding up to twice as fast or reaching correct answers in two-thirds of wall-clock time. The team extended hybrid context to 256K tokens without additional training and released primed models at 8B and 32B scales along with training code, kernels, and a vLLM plugin under Apache 2.0.
Shipping Media Over QUIC at Meta Scale for Instagram Live Ingest
Meta deployed Media over QUIC (MoQ), an emerging IETF standard built on QUIC, as the transport for live video ingest on Instagram Live across iOS and Android. MoQ provides independent streams that avoid head-of-line blocking, native relay-based fan-out and caching, tunable per-subscription latency from about 100 ms to several seconds, and track-aware delivery where relays can drop or prioritize objects. To replace the in-house RUSH protocol without disrupting production, Meta integrated the open-source moxygen library, placed a translation layer at edge endpoints to convert MoQ frames to RUSH, and used server-side routing for rapid experiment ramp-up and rollback. After months of experimentation, the MoQ path matched RUSH across all streaming-quality indicators and reliability measures, achieving metric parity as the rollout target. This production-tested foundation enables Meta to extend MoQ to end-to-end live delivery and real-time interactive AI services, simplifying multiple protocol stacks into one.
ReplaySSM: Rethinking State Storage for Faster Hybrid-Model Decoding
Hybrid models interleave state-space model (SSM) layers with attention layers, but SSM layers can become a decoding bottleneck at serving batch sizes due to memory-bound state updates. ReplaySSM represents the SSM state as a checkpoint plus a buffer of recent inputs, separating output computation from state updates. This reduces state memory traffic by roughly half, simplifies rollback for speculative decoding by avoiding state snapshots, and supports low-precision state reads without propagating rounding error through the recurrence, enabling separate precision for output computation and state updates. Evaluated on Nemotron 3 and Qwen3.5 models from 4B to 120B parameters on H100 and B300 GPUs, ReplaySSM achieves 1.20× to 1.47× faster decode steps at batch size 256, with end-to-end gains growing with batch size. The technique is integrated into TensorRT-LLM, vLLM, SGLang, and FlashInfer, and runs in production in GLM-5.3-Flash and Kimi-K3.
Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations
Meta built a system that aggregates fragmented idle inference capacity from a globally constrained fleet into a shared opportunistic pool to enhance ads recommendations. Idle capacity arises from four mismatches: peak provisioning with demand variation, model development timing with capacity delivery, regional resource ratios, and accelerator types. Two elasticity levers exploit this preemptible, volatile capacity: Rank More Ads increases candidates evaluated by ranking stages, using both CPU and accelerator surplus, and Serve Better Models routes more requests to a more powerful model, using accelerator-only surplus. The system protects baseline serving through isolated routing, minute-scale feedback loops that adjust candidate volume and model coverage, and millisecond-scale value-aware load shedding that removes lower-value work during sudden preemption. This allows optional work to expand and contract with opportunistic capacity, driving better ad relevance and advertiser outcomes without requiring additional capacity orders.
CoSupport's Sarah Hatter on customer support - The Intercom Blog
Sarah Hatter, founder of CoSupport and author of the Customer Support Handbook, discussed the evolution of customer support with Intercom. She noted that customers now prefer self-service over contacting support, so simple FAQ pages and email forms are no longer sufficient. Hatter stressed that support must prove company empathy through personalized language, and tracking feature requests, bug reports, and customer tone positions support as the front line for product improvement. When scaling, instead of simply adding hires to handle rising queries, teams should refine processes to reduce the need for support. In hiring, process can be taught but light-heartedness, self-management, and getting to the root of empathy cannot, and Hatter’s Elevate Summit achieved 60% women and 30% non-white by featuring front-line workers instead of only founders and CEOs.
Harnessing user research with Intercom - The Intercom Blog
TrackDuck, a visual feedback tool for developers, used Intercom to collect user feedback and track behavior, bridging the gap between developer knowledge and user understanding. They simplified a ten-step registration form to just four steps with three entry fields, resulting in a 120% increase in customers who completed registration and started using the system. After discovering through Intercom that some new users never returned after adding their first project, they implemented a seven-step onboarding tour immediately after registration and used auto-messages to track where users dropped off. To satisfy commercial client requests for mobile access without the cost of native apps, they made the control panel adaptive using Media Queries, ngTouch, and matchmedia-ng with a vertical block design. They also added recognizable icons to inline editing fields after users failed to understand that static-looking text was editable, and removed seldom-used features to reduce clutter and focus on valuable functionality.
Inside Intercom X Black Tech Unplugged - The Intercom Blog
Intercom partnered with Black Tech Unplugged, a monthly podcast highlighting Black innovators in tech, to amplify underrepresented voices after George Floyd’s killing. The joint episode features an interview with startup lawyer Lynnise Pantin, discussing her piece “From Fyre Fest to Theranos, the Invisible Racial Subtext of Raising Capital” and biases in the tech ecosystem. Host Deena McKay, an IT Consultant in Chicago, founded the podcast to provide representation and career advice for Black professionals. The episode also recommends other Black Tech Unplugged episodes, including a raw conversation with friends (Episode 22) and job tips from a recruiter (Episode 30). This collaboration represents a small step by Intercom toward racial equity by connecting its audience with voices that need to be heard.
Communicating the value of content design - The Intercom Blog
In her talk at Write the Docs Prague, Kelly O’Brien demystifies content design, a discipline coined by Sarah Richards that focuses on structuring and presenting information. She compares the hidden depth of content design work to an iceberg, where only the surface copy is visible. To communicate the value, she introduces a house-building metaphor: the foundation is the product system, walls and doors are interactions, and the roof is surface copy. Without early involvement at the foundation level, teams may discover conflicts and ambiguity late in the process, leading to frustration and ineffective copy. O’Brien advises content designers to build relationships with designers and product managers, join early meetings, and sell the win-win: early involvement saves time, improves quality, and makes everyone happier.
What an Emerging Design Pattern Tells Us About Disruption
The post identifies an emerging design pattern where smaller companies build ‘best of breed’ products targeting sophisticated users from higher up the market, described as disruption from above. These products, Superhuman and Linear, focus on a specific niche and deliver deep, opinionated design that larger incumbents cannot match. A key example is the Cmd-K keyboard shortcut pattern used by several apps, enabling fast, learnable power-user navigation for tasks like inbox triage. The post argues this approach is economically viable because even a small percentage of a massive market can sustain a significant SaaS business, as shown by 20,000 users at $30 per month yielding over $7 million in annual recurring revenue. Drawing a historical parallel to the Canon Cat, which sold only 20,000 units, it concludes that low barriers to entry allow innovative design to succeed where earlier hardware failed.
Stars In Their Eyes: Why We Changed Our Emoji Ratings
Intercom’s conversation rating feature, built over two years ago, used a five-point emoji scale with 😍 as the top option, but after launch 30% of conversations requested its removal. Research conducted a couple of months later showed users associated 😍 with romantic partners, and male users were uncomfortable sending it to male agents, which artificially lowered agent ratings. This issue was also noticed by Intercom’s internal customer support team, who observed the same pattern. Intercom replaced 😍 with 🤩, the starry eyes emoji, which was perceived as excitement rather than romance, and both customers and the team approved. The change increased “Amazing” ratings by 11% overall and 17% for male agents, eliminating the gender gap and demonstrating how emoji interpretation varies across cultures and genders.
5 books every marketer should read in 2019 | Inside Intercom
The Intercom marketing team, a group of lifelong learners, compiled a reading guide of five books that have influenced their careers and thinking. Kevin Kelly’s The Inevitable categorizes 12 technological forces like AI and streaming, offering an optimistic view of future change. Byron Sharp’s How Brands Grow argues that brand growth depends on reaching new customers through broad reach, not on increasing loyalty from superfans, using Harley-Davidson as an example. James Twitchell’s 20 Ads That Shook the World explores historic campaigns such as De Beers’ creation of diamond engagement rings, providing context for modern advertising. William Zinsser’s On Writing Well and Jake Knapp’s Make Time round out the list with advice on clear writing and focusing on a single daily highlight for better productivity.
Bridget Harris on bootstrapping your startup - The Intercom Blog
Bridget Harris bootstrapped YouCanBook.me, a SaaS scheduling tool, to their $1 million in annual recurring revenue and profitability without taking any venture capital funding. The company, now serving more than 10,000 customers including Atlassian and Shopify, emerged from a freemium scheduling tool called When is Good after businesses showed willingness to pay. Harris argues that VC interests often prioritize rapid growth and timely exits, which can conflict with a founder’s long-term vision, but bootstrapping still requires exposure to external ideas. She draws on her past as a busker to explain the need for adaptability and risk tolerance in bootstrapping, and she has banned internal email at her remote company to encourage productivity and collaboration. The key to bootstrapping, according to Harris, is finding problems that can be solved incrementally through agile and lean principles, enabling sustainable growth without external funding.
Empire Selling's Dan Swift on leveraging social media for sales
Dan Swift, CEO of Empire Selling, explains that cold calling is nearly obsolete and that successful sales prospecting now relies on social media and live chat to build relationships. He launched LinkedIn’s social selling business and later led sales at Sprinklr before founding Empire Selling, which trains B2B companies in digital and social selling techniques. Swift advises being transparent with employers about side businesses and using that time to test content and market viability before going full-time. He recommends using The Challenger Sale methodology to challenge traditional industries like financial services, and stresses the need for a common sales language and strong culture when scaling. Swift emphasizes optimizing LinkedIn profiles to appear as trusted advisors, connecting with everyone in the organization, using Twitter as a research tool, and exploring video as an emerging channel with rich analytics.
Skyliner's Marc Hedlund On Engineering Management
Marc Hedlund, a veteran engineering leader and former VP at Stripe and Etsy, discusses his experience shutting down his startup Skyliner and the lessons learned. He explains that Skyliner failed because its value proposition was not compelling to a wide enough market, and because Amazon released a competing product, Code Star, which he advises entrepreneurs not to build solely on AWS. Hedlund contrasts the rational decision to close Skyliner with the emotional difficulty of shutting down his earlier startup Wesabe, where customer impact was deeply felt. He emphasizes that startups should focus on making users happy quickly rather than pursuing long-term hoped-for outcomes, and that founders must be honest about market signals. Hedlund addresses tech diversity through his board work with Code2040, sharing that simple actions like smiling and saying hello to Black and Latinx visitors can significantly impact recruitment.
Top topics
- Architecture834
- Performance816
- LLMs660
- Machine Learning622
- Open Source563
- AI Agents553
- AI470
- Developer Experience448
Weekly digest
The best engineering reads, once a week.