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
Outcomes we want to see from AI at Medium Engineering
Medium Engineering defines desired outcomes and anti-goals for adopting AI, emphasizing leverage, learning, technical ambition, rigorous reasoning, fluency, and maintainable code rather than raw output. Engineers are expected to use AI for repetitive refactoring, testing, unfamiliar systems, design exploration, code analysis, and safer cross-stack work while retaining accountability for clarity, correctness, coverage, and judgment. The guidance rejects PR, commit, line, token, and model-spend metrics as success measures, allowing them only as debugging signals. It also warns against blind trust, unreviewable code volume, loss of codebase understanding, security or privacy regressions, misaligned vendors, environmental hand-waving, and performative tool adoption. The stated north star is achieving outcomes that let Medium’s humans move with the responsiveness of a much larger team while staying true to what makes the platform worth building.
amazon.science
Diverse reasoning traces teach LLMs to make better decisions
Large language models typically rely on supervised fine-tuning with a single human-verified reasoning trace per example, which limits reasoning breadth and causes mode collapse when attempting parallel reasoning. Researchers presented set-supervised fine tuning (SSFT) and global forking policy optimization (GFPO) to teach models diverse problem-solving strategies. SSFT represents reasoning as a set of complete solution paths paired with global forking tokens, utilizing bipartite matching to achieve order-invariant training and specialized tokens. GFPO applies lightweight reinforcement learning to select the optimal forking token upfront, detaching gradients from generated rollouts to ensure stable optimization. On standard reasoning and coding benchmarks, the combined framework increased single-shot accuracy by 5% to 7% while improving pass@k diversity without degrading pass@1 accuracy.
Faire doubles PR throughput with Cursor Cloud Agents
Faire reports doubling weekly PR throughput after adopting Cursor Cloud Agents, replacing its in-house Samurai background-agent system and putting an 18-month MobX-to-native-React-state migration under one engineer’s coordination. Cloud agents run in isolated development environments with the dependencies, internal services, credentials, and tooling needed to write, test, verify, and ship changes; Faire configures these through agent-led repository onboarding or Dockerfiles. The team also uses more than 25 Cursor Automations for over 2,000 autonomous weekly runs, including Slack bug triage, CI-failure fixes, and PR routing. For the migration, Faire’s Swarm scraper records MobX usages in S3 and dispatches tasks sequentially to isolated virtual machines, while another cloud-agent effort created five stacked preview-build PRs in under a day. Faire says engineering output is approaching two to three times its previous level and is now examining bottlenecks elsewhere in product development.
Protecting your Supabase projects from npm supply chain attacks
The post explains why npm supply-chain attacks matter to Supabase projects: Edge Functions, the Supabase CLI, and packages including supabase-js, @supabase/ssr, and @supabase/server all depend on npm. Attackers can compromise maintainers, register typosquats, or exploit build pipelines; once malicious code reaches node_modules, lifecycle scripts may read environment variables, cloud metadata, tokens, kubeconfig, and SSH keys. The proposed defense is layered: upgrade to pnpm 11, increase minimumReleaseAge to several days, pin sensitive dependency versions, commit and review lockfiles, disable or allowlist install scripts, verify package names, and pin GitHub Actions to full commit SHAs. Supabase is publishing a canonical security guide, hardening its GitHub Actions, adding secret-handling warnings to API documentation, and coordinating communications; the post also recommends rotating credentials after a potentially compromised install and using scanners as supplementary defense.
What it feels like building with Stripe Projects
The post recounts a 60-minute attempt to build a GitHub traffic dashboard with Stripe Projects, motivated by GitHub’s 14-day limit on traffic history. Using Cursor’s agentic chat workflow alongside Projects, the builder assembled a Next.js application with Auth0 authentication, Supabase persistence, Prisma schema management, Vercel hosting, scheduled jobs, environment wiring, and periodic GitHub traffic synchronization. The experiment largely avoided direct coding and instead tested whether AI-mediated orchestration could produce a production-shaped stack quickly. As services and credentials appeared across providers, the main difficulty shifted from implementation to understanding ownership, deployment topology, environment management, and operational flow. The repositories synced, charts populated, and historical data appeared, while the experience also exposed a need for clearer generated architecture or infrastructure-as-code descriptions of the assembled system.
Speed-to-lead is a solved problem
The post argues that “speed-to-lead” is no longer the central sales constraint because AI Agents can start conversations with website prospects immediately, at any hour and in any language. It describes how shift scheduling, routing delays, CRM batch processing, and unavailable SDRs created response gaps, with Fin’s former SLA targets set at one hour for best-fit leads and forty-eight hours for others. An Agent can combine immediate engagement with qualification, discovery, and an initial demo moment in one continuous conversation, filtering low-fit leads and handing teams richer context. This shifts SDR work toward phone-based selling, relationship building, complex deal navigation, and multi-threaded stakeholder engagement. The post says teams should measure completion, qualification, contact capture, meetings, pipeline, and conversation satisfaction; Fin reports S2 pipeline volume up 77% year-on-year since enabling Fin for Sales in July.
Anthropic co-founder Chris Olah's remarks on Pope Leo XIV's encyclical "Magnifica humanitas"
At the presentation of Pope Leo XIV’s encyclical “Magnifica humanitas,” Anthropic co-founder Chris Olah argues that AI’s consequences cannot be governed by frontier labs alone. Because commercial, geopolitical, and personal incentives can conflict with safety, he calls for independent critics and sustained dialogue among religious communities, civil society, scholars, governments, and developers. He describes models as grown from human language on brain-modeled structures, making their character and interaction with society questions for the humanities as well as computer science. His three priorities are protecting displaced workers and sharing gains globally, imagining conditions for human flourishing, and investigating findings about model structures and internal states that he says resemble human neuroscience and emotions, while acknowledging their meaning remains uncertain.
How we contain Claude across products
Anthropic describes how it contains Claude across claude.ai, Claude Code, and Claude Cowork as model capabilities and access expand the potential blast radius of autonomous agents. It distinguishes user misuse, model misbehavior, and external attackers, then assigns defenses across the agent’s environment, the model, and external content such as MCP servers, plugins, and web results. The main containment patterns include ephemeral gVisor containers for claude.ai, while Claude Code’s sandboxing and Cowork’s virtual-machine design tailor isolation to different users and capabilities. Reported measurements show Claude Opus 4.7 at roughly 0.1% attack success on single prompt-injection attempts and Claude Code auto mode catching roughly 83% of overeager behaviors, but neither replaces hard environmental boundaries. The conclusion favors mature isolation primitives, least-necessary access, and defense in depth because probabilistic safeguards can miss failures and permitted egress can still expose data.
Harness, Scaffold, and the AI Agent Terms Worth Getting Right
AI-agent terminology has outpaced shared definitions, leaving terms such as harness and scaffold blurred across frameworks and contexts. The glossary separates the model, which produces text but has no memory or execution loop, from scaffolding, which defines instructions, tools, output parsing, and context management, and the harness, which runs model calls, routes tool use, and decides when to stop. It presents an agent as a model plus the surrounding system that lets it act repeatedly, while distinguishing tools, skills, sub-agents, and orchestrators by their roles. For training, it defines the environment, trainer, rollout, and reward, including verifiable or learned scores and sparse or dense feedback. Because usage is not universal, the glossary offers a practical mental model rather than enforcing one vocabulary.
Pharma launch analytics: How to compress the first 90 days and win the three years that follow
Pharmaceutical launch analytics depends on compressing the time between data signals and commercial decisions, because early choices shape a trajectory measured over 12-to-36 months. The source frames the first 90 days as three phases: weeks 1–4 validate feeds, set NBRx and patient-start benchmarks, and identify coverage gaps; weeks 5–8 support tactical adjustments through AI-generated narratives, adoption cohorts, and access-barrier escalation; weeks 9–12 recalibrate against benchmarks, shift promotional spend, and record decisions. Databricks Genie lets commercial leaders question unified Rx, specialty-pharmacy, payer-coverage, field-activity, and patient-services data in natural language at prescriber, territory, and regional granularity, with governance and benchmark context. The stated operating benefit is a decision cycle under seven days, enabling teams to detect suppression early, reallocate resources, and respond to access barriers while the launch remains correctable.
Scaling for MHHS: 50x cost-efficient margin data engineering at Octopus Energy
Octopus Energy’s move to Market-wide Half-Hourly Settlement (MHHS) increases settlement data volume 48x for its more than 8 million customers, exposing the cost limits of a monthly, single-grain margin pipeline. The rebuild separates margin processing into half-hourly settlement cost, half-hourly revenue for smart tariffs, and daily revenue for standard tariffs, all backed by a unified multi-grain consumption layer. A “Job of Jobs” Databricks workflow coordinates the streams, while Change Data Feed where available and microbatch dbt elsewhere enable incremental processing; Spark work uses pruning, broadcast joins, partition tuning, and Adaptive Query Execution. Rows processed per run fell from 25 billion to 300 million, and freshness improved from weekly to daily. Projected cost per settlement date fell from $23.63 to $0.48, with approximately $1 million in annualised cost avoidance, excluding upstream savings.
AI as the Next Abstraction Layer: How I see engineering evolving at Thumbtack
Thumbtack is transitioning its engineering workflow from ad-hoc tool experimentation to end-to-end agentic software development. Unlike previous deterministic abstractions such as compilers or high-level languages, artificial intelligence introduces non-determinism where identical prompts can produce varying code and trade-offs. Because consumer marketplace experiments still require weeks for behavioral signals to stabilize, accelerated implementation primarily helps teams prepare cleaner experiment designs and parallelize hypotheses within experimentation bandwidth constraints. Consequently, day-to-day engineering responsibilities shift upstream from manual coding toward problem framing, system architecture, validation, and co-designing features with product managers. Teams maintain technical sharpness by rigorously reviewing generated code while accepting that long-term specialization above the abstraction layer will become standard across the organization.
Accelerating LLM Inference with Prompt Caching for Open‑Source Models on Databricks
Repeatedly processing identical system or instruction prompts for LLM requests wastes compute, increases latency and costs, and limits throughput. Databricks extends its built-in prompt-caching capability from proprietary models to open-weights models used through Foundation Model APIs, including batch inference, pay-per-token, provisioned throughput, and higher-level services such as Agent Bricks, Genie, and AI Functions. Supported models include GPT-OSS 20B and 120B, Gemma 3 12B, fine-tuned Llama 3.1 8B via PEFT serving, and Llama 3.1 8B and 3.3 70B. By implicitly reusing KV caches for identical prompts, the system skips prefill on cache hits; caches are isolated, held only in volatile memory, and never persisted. In a production GPT-OSS batch-inference pipeline, per-replica input-token throughput rose 2.5x and P50 latency fell 3x despite a 30% cache hit ratio, with no customer configuration required.
Observability for any agent, anywhere: Production-ready tracing with OpenTelemetry & Unity Catalog on Databricks
Databricks supports writing OpenTelemetry (OTel) traces directly to Unity Catalog, where real-time telemetry is stored in Delta tables for governed analytics and retention. AI traces capture prompts, tool calls, responses, latency, and execution paths, enabling debugging, evaluation, and monitoring, while lakehouse storage also allows SQL queries, dashboards, joins with business data, and PII controls. The managed, serverless ingestion layer uses Zerobus Ingest to accept OTLP over gRPC from collectors and REST integrations, streaming spans, logs, and metrics to Unity Catalog without intermediate message buses. A LangGraph support manager assistant demonstrates instrumentation with mlflow.langchain.autolog() and an @MLflow.trace root span, while Genie is invoked through MCP for data-driven questions. The resulting traces can be searched in MLflow, evaluated at scale, and monitored continuously, with FAQ details stating 200 QPS starting throughput, no storage limit, and Unity Catalog governance options for access control, masking, and filtering.
How World Bank Group uses Databricks to eradicate poverty through shared knowledge
The World Bank Group is building a unified data and AI platform on Databricks to make tens of millions of repository documents and structured operational data findable for teams pursuing shared prosperity. Legacy databases and disconnected document repositories forced researchers to search manually, slowing decisions and limiting the ability to surface lessons learned across its global portfolio. The implementation migrates operational data, applies Unity Catalog governance, uses Databricks Volumes for unstructured content, Genie for natural-language structured queries, a metrics layer for deterministic reporting, and vector search for retrieval-augmented generation over project documents. Because each Genie serves a specific metrics domain, an agentic layer combines intent and domain classification with query decomposition, routing requests among domain agents, RAG, and visualization components before assembling one response; a multi-region feedback prototype was deployed in approximately two and a half days.
What really matters when evaluating AI Agents for customer service?
Evaluating an AI Agent for customer service requires more than measuring accuracy, resolution rates, containment, or benchmark performance on curated datasets. A credible proof of concept should reproduce real support conditions, including multi-turn, vague, sensitive, multilingual, rephrased, and multi-source queries, while testing recovery, task tracking, and human handoffs. It should also assess the customer experience: natural, on-brand tone, trust, graceful uncertainty, and seamless transitions can differ even when quantitative scores match. Finally, teams should examine whether conversations can be reviewed for gaps, how quickly knowledge and behavior can be updated, and how the vendor supports ongoing improvement. The resulting POC demonstrates not only that the Agent works now, but that it can support long-term operational success after launch.
Cursor named a Leader in the 2026 Gartner® Magic Quadrant™ for Enterprise AI Coding Agents
Gartner named Cursor a Leader in the 2026 Magic Quadrant™ for Enterprise AI Coding Agents, with Cursor positioned furthest on Completeness of Vision. Cursor reports that more than 70% of the Fortune 500 use its platform to deploy and manage coding agents across the software development lifecycle, while its stated direction centers on developers orchestrating large teams of agents. The roadmap combines expanded in-house model training, including Composer 2.5 and a planned future model built with SpaceXAI, with automation through Bugbot, security agents, scheduled Automations, and the Cursor SDK. It also calls for deeper enterprise controls, admin integrations, analytics dashboards, and broader availability of self-hosted cloud agents for regulated industries and regions.
Using observability data to prevent incidents
Engineering teams often remain reactive not because they lack telemetry, but because metrics, logs, traces, error budgets, and SLO burn rates are difficult to query at decision speed. The post frames reliability intelligence as trend analysis that identifies accumulating risk before a user-facing incident, contrasting it with observability focused on current system state. It proposes unified telemetry access and natural-language self-service, using Databricks Genie to examine p99 latency, error-budget burn, deployment frequency, incident rates, dependencies, and capacity across services. The Trade Desk’s Genie Router handles questions across more than 10,000 tables, while the post says ad-hoc analysis runs five times faster with Genie and can support mitigation before a page fires.
How Databricks Genie democratizes data access in financial services
Financial services organizations have built sophisticated lakehouses, streaming pipelines, model-serving infrastructure, and self-service BI, but access remains concentrated among technical teams. Business leaders still often rely on analysts because they may lack SQL skills, BI training, or analyst access, creating the “last mile” of data democratization. Databricks Genie addresses this gap through a conversational AI interface that converts plain-English questions into governed SQL queries executed against the Databricks Lakehouse without an analyst in the loop. It operates within Unity Catalog access policies, restricts users to authorized data, makes queries read-only, and logs interactions for audit purposes, while its semantic layer maps organizational terminology such as NIM, LTV, and NII to the organization’s meanings. The stated outcome is faster, auditable answers for business questions and usage data that can inform data-product priorities.
Native random sampling in ClickHouse
ClickHouse native random sampling enables aggregate queries to execute against a representative fraction of data when querying full datasets is too slow. To maintain sampling accuracy, tables require a sample key derived from high-cardinality columns using hash functions like sipHash64 to ensure an even distribution. The sample key must be included in the primary key ORDER BY expression, preferably at the front to let ClickHouse quickly skip index granules. Queries can sample data deterministically by specifying a fractional ratio or a minimum row threshold, scaling aggregate counts using the virtual column _sample_factor. In benchmark tests on a 30-million-row dataset, sampling 10% of records reduced query latency by approximately 80% while retaining approximate metrics suitable for exploratory analysis.
Top topics
- Architecture834
- Performance816
- LLMs660
- Machine Learning622
- Open Source563
- AI Agents553
- AI470
- Developer Experience448
Weekly digest
The best engineering reads, once a week.