Claude Duet 🤝, ChatGPT’s ad stack 📣, the browser’s untapped power 🌐
TLDR
Together With 
TLDR Dev 2026-10-22
Products that matter. Dedicated learning time. Real stability. (Sponsor)
Looking to build your technology career in an environment wheretech investment is a constant? Discover what it's like to build at Fidelity:
Looking to build your technology career in an environment wheretech investment is a constant? Discover what it's like to build at Fidelity:
- Systems with a broad reach: Architect, design, and deploy platforms powering financial security for millions.
- Continuous learning: Dedicated time during work hours forupskilling, certifications, and experimentation.
- Solid foundation: Startup-level tech ambition backed by a long history of organizational stability.
Build a career that keeps moving forward.Find out more about working in tech at Fidelity.
Fidelity is an equal opportunity employer.
🧑💻
Articles & Tutorials
How ChatGPT's Ad Stack Works (12 minute read)
ChatGPT's ad system uses a post-trained LLM to orchestrate retrieval, ranking, relevance checks, and auctions as user intent evolves across conversations. Stateful ad sessions preserve context, while separate models, services, and code keep ad selection outside ChatGPT's answer-generation path.
ChatGPT's ad system uses a post-trained LLM to orchestrate retrieval, ranking, relevance checks, and auctions as user intent evolves across conversations. Stateful ad sessions preserve context, while separate models, services, and code keep ad selection outside ChatGPT's answer-generation path.
Front-end Observability at Scale with ClickHouse (11 minute read)
OpenReplay's largest customers ingest billions of events and roughly 6 TB daily on a single ClickHouse instance, with sub-second queries. Moving behavioral analytics out of Postgres also replaced disruptive retention cleanups with TTL rules and simplified data pipelines through materialized views.
OpenReplay's largest customers ingest billions of events and roughly 6 TB daily on a single ClickHouse instance, with sub-second queries. Moving behavioral analytics out of Postgres also replaced disruptive retention cleanups with TTL rules and simplified data pipelines through materialized views.
Computing Bounding Boxes (6 minute read)
tldraw transforms shape vertices into page coordinates, then takes their minimum and maximum values to compute fast axis-aligned bounding boxes. The walkthrough covers nested transforms and rotated selection handles, with approximate bounds filtering candidates before more expensive geometry checks.
tldraw transforms shape vertices into page coordinates, then takes their minimum and maximum values to compute fast axis-aligned bounding boxes. The walkthrough covers nested transforms and rotated selection handles, with approximate bounds filtering candidates before more expensive geometry checks.
🧠
Opinions & Advice
Who Holds the Off Switch? (24 minute read)
Five imagined coffee purchases in 2035 illustrate different models of data control, from state surveillance and privacy regulation to commercial data markets, public infrastructure, and biometric identity. Privacy increasingly depends on who can stop collection. AI assistants and paid privacy tiers will make that question more pressing.
Five imagined coffee purchases in 2035 illustrate different models of data control, from state surveillance and privacy regulation to commercial data markets, public infrastructure, and biometric identity. Privacy increasingly depends on who can stop collection. AI assistants and paid privacy tiers will make that question more pressing.
The Web Is More Than Enough (6 minute read)
Modern browsers already offer cryptography, native-speed code, graphics, and audio, while AI makes those capabilities easier for individuals to use. Falling development and hosting costs could expand personal software creation and challenge app-store economics.
Modern browsers already offer cryptography, native-speed code, graphics, and audio, while AI makes those capabilities easier for individuals to use. Falling development and hosting costs could expand personal software creation and challenge app-store economics.
Asking LLM for Confidence (8 minute read)
An LLM's stated confidence is generated text, so a precise-looking number should not be treated as a calibrated measure of correctness. The proposed alternative derives scores from observable evidence and historical outcomes, then checks calibration regularly on the same kinds of tasks.
An LLM's stated confidence is generated text, so a precise-looking number should not be treated as a calibrated measure of correctness. The proposed alternative derives scores from observable evidence and historical outcomes, then checks calibration regularly on the same kinds of tasks.
🚀
Launches & Tools
payroc.ai: the fastest, most developer-friendly way to embed payments in your app (Sponsor)
Your next payments integration shouldn't take a quarter. Payroc for Developers gives you AI-ready docs, integration skills and a self-service sandbox. Minutes from sandbox to first transaction, rev share from day one, AI- and MCP-native. Start building right away - it's free.
Your next payments integration shouldn't take a quarter. Payroc for Developers gives you AI-ready docs, integration skills and a self-service sandbox. Minutes from sandbox to first transaction, rev share from day one, AI- and MCP-native. Start building right away - it's free.
Claude Duet (Website)
Claude Duet is a proactive agent that delivers morning briefs, handles tasks while you are away, and asks for approval when needed. The research preview uses Opus 5.5, carries over Claude memory, skills, and apps, and is offered through a waitlist for US Pro and Max users.
Claude Duet is a proactive agent that delivers morning briefs, handles tasks while you are away, and asks for approval when needed. The research preview uses Opus 5.5, carries over Claude memory, skills, and apps, and is offered through a waitlist for US Pro and Max users.
Omnara (GitHub Repo)
Omnara is an open-source managed agent platform that can run with different model providers and execute tools on sandboxes or your own machines. A REST API, TypeScript SDK, CLI, and dashboard expose agent runs, while a Postgres event log lets workers resume them after crashes.
Omnara is an open-source managed agent platform that can run with different model providers and execute tools on sandboxes or your own machines. A REST API, TypeScript SDK, CLI, and dashboard expose agent runs, while a Postgres event log lets workers resume them after crashes.
🎁
Miscellaneous
Evaluating Agents on Highly Subjective Tasks (20 minute read)
Reliable evaluation of open-ended agent tasks combines calibrated human judgments, automated judges or rubrics, and production monitoring. Power analysis, confidence intervals, and repeated runs help distinguish improvements from noise, while realistic, diverse tasks, and production feedback keep benchmarks useful.
Reliable evaluation of open-ended agent tasks combines calibrated human judgments, automated judges or rubrics, and production monitoring. Power analysis, confidence intervals, and repeated runs help distinguish improvements from noise, while realistic, diverse tasks, and production feedback keep benchmarks useful.
Inside OpenAI's Jalapeño AI Chip (18 minute read)
OpenAI's Jalapeño inference chip links 64 dedicated cores to individual HBM memory banks to reduce data movement and improve utilization. This interview explains the tradeoff in memory-management complexity, the choice to keep prefill and decode on fungible hardware, and how AI-assisted optimization accelerated chip design and kernel development.
OpenAI's Jalapeño inference chip links 64 dedicated cores to individual HBM memory banks to reduce data movement and improve utilization. This interview explains the tradeoff in memory-management complexity, the choice to keep prefill and decode on fungible hardware, and how AI-assisted optimization accelerated chip design and kernel development.
Why Hardware Specialization Is Back (11 minute read)
Slower transistor scaling, memory bottlenecks, and power limits are renewing the case for specialized AI inference hardware. Matching different inference phases to different accelerators can improve performance per watt while adding deployment and software complexity and leaving room for GPUs to absorb successful ideas.
Slower transistor scaling, memory bottlenecks, and power limits are renewing the case for specialized AI inference hardware. Matching different inference phases to different accelerators can improve performance per watt while adding deployment and software complexity and leaving room for GPUs to absorb successful ideas.
⚡
Quick Links
Switch Harnesses, Not Models: 2x OpenCode's Fixes, 4x Claude Code's (Sponsor)
Same open model, very different results: on DeepSWE, CodeAF solved 2× OpenCode's issues and nearly 4× Claude Code's, at up to 4.8× less per fix. A new open-source coding harness.
Try CodeAF on GitHub →
Same open model, very different results: on DeepSWE, CodeAF solved 2× OpenCode's issues and nearly 4× Claude Code's, at up to 4.8× less per fix. A new open-source coding harness.
Try CodeAF on GitHub →
Railway Security Disclosure (2 minute read)
Railway reports possible unauthorized access to customer environment variables, including some decrypted contents, and recommends rotating stored secrets and private registry credentials while reviewing service logs.
Railway reports possible unauthorized access to customer environment variables, including some decrypted contents, and recommends rotating stored secrets and private registry credentials while reviewing service logs.
Google Rust Style Guide (Website)
Google's Rust style guide covers code organization, formatting, ownership, API design, traits, errors, and concurrency.
Google's Rust style guide covers code organization, formatting, ownership, API design, traits, errors, and concurrency.
OpenAI Models Are Coming to Google Cloud (2 minute read)
OpenAI frontier models and Codex workflows are coming to Google Cloud's Model Garden.
OpenAI frontier models and Codex workflows are coming to Google Cloud's Model Garden.
Chain of Thought for Humans (3 minute read)
Narrating your reasoning aloud can aid problem-solving by exposing assumptions, organizing dependencies, making rules explicit, and clarifying choices for backtracking.
Narrating your reasoning aloud can aid problem-solving by exposing assumptions, organizing dependencies, making rules explicit, and clarifying choices for backtracking.
Love TLDR? Tell your friends and get rewards!
Share your referral link below with friends to get free TLDR swag!
Want to advertise in TLDR? 📰
If your company is interested in reaching an audience of web developers and engineering decision makers, you may want to advertise with us.
Want to work at TLDR? 💼
Apply here, create your own role or send a friend's resume to [email protected] and get $1k if we hire them! TLDR is one of Inc.'s Best Bootstrapped businesses of 2025.
If you have any comments or feedback, just respond to this email!
Thanks for reading,
Priyam Mohanty, Jenny Xu
Thanks for reading,
Priyam Mohanty, Jenny Xu