{"days":[{"date":"2026-10-24","entries":[{"id":"404ae824-27e5-8031-aa7c-ccaeced01003","title":"Twelve years of distributed SQLite","url":"https://philipotoole.com/twelve-years-of-distributed-sqlite/","thought":"","date":"2026-10-24","createdAt":"2026-10-25T02:05:00.000Z"},{"id":"404ae824-27e5-80a5-811e-c41115ae7f87","title":"Write Ahead Logging in SQLite","url":"https://sqlite.org/wal.html","thought":"So well written. Sometimes you just have to read SQLite documentation on a weekend to get inspiration.","date":"2026-10-24","createdAt":"2026-10-25T01:32:00.000Z"},{"id":"404ae824-27e5-805d-994f-d3df27749df3","title":"Log-structured storage","url":"https://jvns.ca/blog/2017/06/11/log-structured-storage/","thought":"timely reminder that you can write your own database in less than 5 LoC.","date":"2026-10-24","createdAt":"2026-10-25T01:31:00.000Z"}]},{"date":"2026-09-20","entries":[{"id":"3e1ae824-27e5-8001-a9a1-dfb53c0cf1c1","title":"Why Taste Matters More Than Technical Skill | Shopify Founder, Tobi Lütke","url":"https://www.youtube.com/watch?v=G9P9D9hptq8","thought":"","date":"2026-09-20","createdAt":"2026-09-20T21:09:00.000Z"}]},{"date":"2026-09-19","entries":[{"id":"3e1ae824-27e5-80ef-a4c1-cb255d6d30d5","title":"The last IMO problem AI could not solve","url":"https://www.youtube.com/watch?v=Nbwv5wHQoj0","thought":"","date":"2026-09-19","createdAt":"2026-09-20T02:42:00.000Z"}]},{"date":"2026-09-18","entries":[{"id":"3dfae824-27e5-8066-b5f2-d69cfe1802a7","title":"Introducing System One Models & Jev","url":"https://typesafe.ai/blog/introducing-system-one-models-and-jev","thought":"","date":"2026-09-18","createdAt":"2026-09-18T18:30:00.000Z"},{"id":"3dfae824-27e5-80f3-86ec-e34aff3239af","title":"Training LLMs to Predict World Events (Guest Post with Mantic)","url":"https://thinkingmachines.ai/news/training-llms-to-predict-world-events/","thought":"","date":"2026-09-18","createdAt":"2026-09-18T18:30:00.000Z"},{"id":"3dfae824-27e5-8030-89e0-d529262c2a95","title":"Warren’s last letter to Berkshire’s shareholders","url":"https://www.berkshirehathaway.com/news/sep1826.pdf","thought":"","date":"2026-09-18","createdAt":"2026-09-18T17:10:00.000Z"}]},{"date":"2026-09-17","entries":[{"id":"3dfae824-27e5-80c9-9935-f03816277f14","title":"Dwarkesh’s interview with Noam Brown","url":"https://x.com/dwarkesh_sp/status/2100616332144169048?s=20","thought":"Very interesting to hear about agent architectures and how increasing number of agents improves the chances of improving a very hard problem. However Noam doesn’t even attribute 10% of solving Navier Strokes problem to multi agent architectures. The reason for solving the problem is having a strong model.\n\nAnother interesting thing that he mentions is that having large number of agents has a quadratic coordination tax as well, they do not coordinate well. Having a sophisticated message passing between large number of agents helps in reducing coordination tax between agents.","date":"2026-09-17","createdAt":"2026-09-18T06:55:00.000Z"},{"id":"3dfae824-27e5-8035-9fb4-ea3706817750","title":"Hacking OpenAI","url":"https://www.hacktron.ai/blog/hacking-openai","thought":"Always backport your security changes to your products & always keep your machines up to date with latest software patches. I’m wondering whether AI agents force everyone to always keep upgrading their hardware — phones only get software updates for a few years before they stop receiving updates. This will contribute even more towards the hardware / chip bottleneck we see due to AI already.","date":"2026-09-17","createdAt":"2026-09-18T06:08:00.000Z"},{"id":"3dfae824-27e5-8007-b654-f180b767acce","title":"OpenAI solves Navier–Stokes Millennium Prize Problem","url":"https://openai.com/index/navier-stokes-solution/","thought":"From [Noam Brown](https://x.com/polynoamial/status/2097375837670785447):\n\nYes, this result cost millions of dollars. But remember that whenannounced o3 it cost ~$500,000 to score 87.5% on ARC-AGI 1. Today, Astra scores higher for ~$20. In 2025 it took us and GDM an enormous amount of compute to achieve IMO gold. For the 2026 IMO, anyone with a $20/month ChatGPT subscription could do it.\nMassively scaling test-time compute gives us a glimpse of the future. I believe that a year from now everyone will have an AI at their fingertips capable of solving problems of this caliber.","date":"2026-09-17","createdAt":"2026-09-18T05:25:00.000Z"},{"id":"3dfae824-27e5-8089-a8d3-dbc9fe22f9fd","title":"Implications of Large-Scale Test-Time Compute","url":"https://x.com/polynoamial/status/2064210146558136827","thought":"Interesting blog post. The author recommends testing the token efficiency of models while evaluating benchmarks. It might be better for most orgs to have a cheap model which gets 90% of the way there with less cost than 100% of the way with 10x more cost.","date":"2026-09-17","createdAt":"2026-09-18T05:22:00.000Z"}]},{"date":"2026-09-03","entries":[{"id":"3d0ae824-27e5-802c-bcc6-ed720828e34c","title":"GPT 6 Astra","url":"https://openai.com/index/gpt-6-astra/","thought":"It’s been a while since we’ve seen a model release have state of the art results on all benchmarks.","date":"2026-09-03","createdAt":"2026-09-03T20:31:00.000Z"}]},{"date":"2026-08-30","entries":[{"id":"3cdae824-27e5-8046-9e26-c4c43cf207fb","title":"Running a Software Factory Efficiently at Uber Scale","url":"https://x.com/UberEng/status/2093444169037762840?s=20","thought":"I think cost optimization for using coding models will be a vertical team in every company going forward. This blog from Uber mentions some ways they found for cost saving, mainly optimizing in things like having a central MCP gateway, dynamic tool loading, creating graphs of codebase to spend less time searching for stuff, evaluating different open source and closed source models for their own type of software engineering tasks & benchmarks (UberSWE!) & choosing the one with the best cost/performance ratio.","date":"2026-08-30","createdAt":"2026-08-31T01:29:00.000Z"},{"id":"3ccae824-27e5-80f6-ab9c-f893e01fa193","title":"Most Neoclouds Suck At Security","url":"https://newsletter.semianalysis.com/p/most-neoclouds-suck-at-security","thought":"","date":"2026-08-30","createdAt":"2026-08-30T23:29:00.000Z"},{"id":"3ccae824-27e5-804a-adbf-d3fad7bbc5d7","title":"What Is Amazon Aurora and Why Are There So Many?","url":"https://www.joesluis.com/blog/what-is-amazon-aurora","thought":"","date":"2026-08-30","createdAt":"2026-08-30T18:38:00.000Z"},{"id":"3ccae824-27e5-806f-9cec-dfb8a78222c6","title":"The Rise and Fall of Agent Civilizations","url":"https://www.dwarkesh.com/p/openai-huggingface","thought":"","date":"2026-08-30","createdAt":"2026-08-30T18:38:00.000Z"}]},{"date":"2026-08-27","entries":[{"id":"3caae824-27e5-804f-9abf-c69e13b5c68f","title":"Amazon MemoryDB: A Fast and Durable Memory-First Cloud Database","url":"https://dl.acm.org/doi/epdf/10.1145/3626246.3653380","thought":"","date":"2026-08-27","createdAt":"2026-08-28T01:16:00.000Z"},{"id":"3caae824-27e5-80c8-bd53-dcafb504de54","title":"Very interesting Black Friday buy","url":"https://pollen-robotics.com/microduck/","thought":"","date":"2026-08-27","createdAt":"2026-08-28T00:14:00.000Z"},{"id":"3caae824-27e5-808f-85a1-e1bd55f317b3","title":"BtrLog: Low-Latency Logging for Cloud Database Systems","url":"https://arxiv.org/abs/2606.27051","thought":"Since database WALs are hot now, this paper talks about how to maintain a WAL without paying for the latency of S3. The idea is that we persist operations on local NVMe drives on multiple nodes with a quorum & asynchronously send these to S3 with conditional put operations.\n\n![](<https://prod-files-secure.s3.us-west-2.amazonaws.com/bb74a6cf-6d93-4bbc-84b3-0031808e6acf/42f9467c-a022-4a6b-a1c3-8626ce25f9ee/image.png?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=ASIAZI2LB466WG6GBM7T%2F20261117%2Fus-west-2%2Fs3%2Faws4_request&X-Amz-Date=20261117T192527Z&X-Amz-Expires=3600&X-Amz-Security-Token=IQoJb3JpZ2luX2VjECMaCXVzLXdlc3QtMiJIMEYCIQCE%2FBCE1HyCFJ1gg%2FGPmp%2BgQIoU%2FyrNsEd4BHduYZXt2QIhAIBfhQxJSDsK7yiXQ2i1cLv1DmycoqISgGLBuYb%2BN5%2BVKogECOv%2F%2F%2F%2F%2F%2F%2F%2F%2F%2FwEQABoMNjM3NDIzMTgzODA1IgxAmenaSzofs2%2FtfWwq3AMCYQjTwt%2F9XNQ7uDC7qfJ9VP7K2OsOpjAcVw7Iz9SWfhVZHalDnsZ3c%2F9rhG6WA%2Fzzf4ZjXZBlA8RRXCFxNhJVBlTx9N8LedxEb88NY2XlEz3KAePM8tAXOsuVmBvCAorg6dv0nKPTPNy3%2F3vIcnufWY%2BJgQNXXXEYv%2B%2BtkFTpSmH3t5XZhKDSyMQzlT2vn0%2FsDxflgL9XiUzwEuoXD6OFLn%2F88inDnipQLhne8wVM4Qfpo%2FqRj62Sn7Zayg%2BkFne9LeztMclDQTu%2BX3l%2FZUXCL9HK4cTnss7oTU7YADccV2DkN06eoBTCJwjxw6IibGjfkdqBDqgDl63BNsEFAA1uG5OJO2qyEZb%2F2htelFbWJX2yIZiqaZouCIHwosxlKG2SQV1rIkHwgG2eZtxZ44ghrGlqfExoIyzmtn81wsKyUrOzowQ5DBS0ZaDN3VFbjV9qUhSRrObFtBcZNufy3QBP1J4THEjDSFqwAYea2Ekwu4tfMq8uuhf8nwyfonLPQDBeCMdWDsvx4gUFJmwW8iyZ4R4rXx2IPN6HlVl2uRU9owN2RaEUGS%2FiQ%2F5ZPo1ilSDJgU4BQ2pqo3HY%2BvzWAZX8ZUJ4dU5TMinHI8Zr%2BSaAmitpzuQCoNO9z4NsezD5wPLXBjqkATu55dJUGCR4hd8KMul8jTSb26jubWIil11NZ4hucuJHakchMNCeCzGyv76YyVvN76S%2FQIcc3yHeKoXDsdJaU5qlGisS8QcnceaxWm3ouzHQqbHlVyQ%2B2MKDUGBKYqqTyiIAxJLbJ7Cv3QzWb%2BuF1da9fJ3l8iUzBilv02h3dDxaoRxIlr1aEFSiAJOg9RaewOosi0R2Go5f4omuH1GmTMAE6Vlq&X-Amz-Signature=207684f7880cceda8c095b1838da792336f7c71786ec6ba41c500bace1e74f64&X-Amz-SignedHeaders=host&x-amz-checksum-mode=ENABLED&x-id=GetObject>)","date":"2026-08-27","createdAt":"2026-08-28T00:09:00.000Z"}]},{"date":"2026-08-22","entries":[{"id":"3c4ae824-27e5-8002-82e5-c304305c34fe","title":"Maintaining durability of objects with less overhead","url":"https://en.wikipedia.org/wiki/Reed%E2%80%93Solomon_error_correction","thought":"I read this in context of an article discussing S3 internals. The idea behind Erasure Coding is that if you want to keep m raw copies of any particular piece of data to tolerate node failures, you end up with a large overhead. Instead you can use erasure coding to first shard the data into k shards. Then you can construct m parity shards using an algorithm & now you have (k+m) shards representing the original data. The idea behind the parity shards is that if (k) out of (k+m) shards remain available, you can always reconstruct the original data.\n\nAssume k = 10 & m = 5. You’d have to keep 15 shards instead of 10*3 = 30, which gives you a 4x benefit in terms of overhead (relative to 10).","date":"2026-08-22","createdAt":"2026-08-22T22:16:00.000Z"}]},{"date":"2026-08-21","entries":[{"id":"3c3ae824-27e5-80d1-9dc3-f16e470bd308","title":"Are Open Models Catching Up?","url":"https://newsletter.semianalysis.com/p/are-open-models-catching-up","thought":"","date":"2026-08-21","createdAt":"2026-08-21T17:41:00.000Z"}]},{"date":"2026-08-20","entries":[{"id":"3c2ae824-27e5-80f1-b931-eb8718d23229","title":"Building and operatinga pretty big storage system called S3","url":"https://www.allthingsdistributed.com/2023/07/building-and-operating-a-pretty-big-storage-system.html","thought":"","date":"2026-08-20","createdAt":"2026-08-20T20:59:00.000Z"},{"id":"3c2ae824-27e5-8004-b428-e80e90ff197e","title":"Rebuilding Linear’s delta sync read path","url":"https://linear.app/now/rebuilding-delta-sync-read-path","thought":"","date":"2026-08-20","createdAt":"2026-08-20T20:34:00.000Z"},{"id":"3c2ae824-27e5-80a7-9caf-c400f5821f9a","title":"Gisting: Compressing LLM Agent context to ↑ throughput and ↓ cost","url":"https://shopify.engineering/gisting","thought":"","date":"2026-08-20","createdAt":"2026-08-20T19:49:00.000Z"}]},{"date":"2026-08-19","entries":[{"id":"3c1ae824-27e5-8016-93ad-c38bc0f96fde","title":"Git at any scale","url":"https://cursor.com/blog/git-at-any-scale","thought":"Very trendy on Twitter this week.\n\nIf you haven’t noticed this year, Github is struggling to stay available. You can blame this on Microsoft, but the thing is that Github was not designed for software built by agents. There are some fundamental assumptions you make while designing a system — like for example, user won’t push 100+ commits each second on one project, or push more than 100k LoC in a 5 second window across all their projects. The architecture treats these assumptions as foundations and has years of work built on top of it. With agents, when these assumptions become incorrect, a new system has to be built.\n\nSpokes assumed that 3 replicas / repo was a sweet spot, but now codebases are larger than ever, and you need more replicas / repo. Increasing number of replicas degrades performance at tail.\n\nThe idea behind Continuity is simply — keep a local repo copy on fast NVMe drive & once we durably store all a push in a WAL on object store (like S3), we apply the reference transaction on the local repo copy using Git, without waiting for a quorum. This makes pushes as fast as your disk. This also makes the local servers stateless — if a repo is missing from a local disk, just fetch it from the WAL. The idea is to make servers or services stateless, consider S3 to be the single source of truth and have minimal metadata layer outside of S3.\n\nPros:\n\n- System is consistent. Each server returns the a globally consistent view of the codebase.\n- Most popular codebases or huge codebases get a large number of replicas assigned to it.\n- Tiny repos created by agents can be hosted by one replica.","date":"2026-08-19","createdAt":"2026-08-19T17:29:00.000Z"}]},{"date":"2026-08-18","entries":[{"id":"3c0ae824-27e5-806f-90ce-e0339870e593","title":"Wiz Red Agent Finds Its Way Into Snowflake’s Internal Jira Due to an AI-Generated GitHub Copilot Autofix","url":"https://www.wiz.io/blog/red-agent-snowflake-copilot-cicd-bug","thought":"","date":"2026-08-18","createdAt":"2026-08-18T16:02:00.000Z"},{"id":"3c0ae824-27e5-804b-9a1b-ec3ce8318bda","title":"The Amazon tax","url":"https://seths.blog/2026/08/the-amazon-tax/","thought":"","date":"2026-08-18","createdAt":"2026-08-18T15:10:00.000Z"}]},{"date":"2026-08-16","entries":[{"id":"3beae824-27e5-80e7-b910-f38c579ca7cd","title":"MathCode: A Frontier Mathematical Coding Agent","url":"https://github.com/math-ai-org/mathcode","thought":"","date":"2026-08-16","createdAt":"2026-08-16T21:23:00.000Z"},{"id":"3beae824-27e5-8003-b3dc-e672eeed5b41","title":"how to enable 1M context length in codex with gpt-5.6 sol","url":"https://x.com/thsottiaux/status/2089082893804896524","thought":"So far for me, compaction has pretty much solved the issue of context windows. Claude Code & Codex both perform compaction after the context window is utilized and the compaction summary is good enough for short-horizon tasks.\n\nThere is a documented 1M context window option for Codex, but some users mentioned increased usage.","date":"2026-08-16","createdAt":"2026-08-16T21:22:00.000Z"}]},{"date":"2026-08-15","entries":[{"id":"3beae824-27e5-800c-8990-ce278954609b","title":"Maximizing the value of your Claude Code sessions","url":"https://claude.com/blog/maximizing-the-value-of-your-claude-code-sessions","thought":"TLDR (copied from the article)\n\n- **Run **`/clear`**between tasks.** This prevents prior irrelevant context from being sent back to the model, which can reduce token usage.\n- **Set your model and effort level before you start. **Changing either one mid-conversation can bust your prompt cache, which can increase token cost.\n- **@-mention files instead of naming them. **The file gets attached to your message directly, which saves a Read call, or a search if Claude has to go find it.\n- **Add quiet flags to noisy commands, or run them in a subagent. **Command output is added to the conversation just like a file, and stays there for the rest of the session.\n- **Run **`/context`** once in a fresh session. **It shows what's loaded (`CLAUDE.md`, MCP tool definitions), so you can cut out anything unnecessary.\n- `/compact`** before you take a break from your keyboard. **The prompt cache expires after an hour, and summarizing a conversation is much cheaper while it's still cached.","date":"2026-08-15","createdAt":"2026-08-16T00:30:00.000Z"},{"id":"3beae824-27e5-807d-9655-d2e9b877e5cd","title":"Working With AI Feels More Like Leadership Than Coding","url":"https://allen.bargi.org/notes/working-with-ai-feels-like-leadership/","thought":"","date":"2026-08-15","createdAt":"2026-08-16T00:25:00.000Z"},{"id":"3bdae824-27e5-8019-87e4-f598763db06e","title":"Inside AppLovin’s $100B Ad Engine","url":"https://youtu.be/IZHwulZQixQ?si=L14fkeQKItAMjejt","thought":"Great listen! I worked at AppLovin in 2024 on their backend team and it was refreshing to hear Gio again.","date":"2026-08-15","createdAt":"2026-08-15T07:47:00.000Z"}]},{"date":"2026-08-13","entries":[{"id":"3bbae824-27e5-8039-ae60-daf9a2dd0c1f","title":"Zed, Delta & DeltaDB","url":"https://zed.dev/blog/introducing-delta","thought":"Finally we get the first multiplayer coding agent harness. It’s kind of Google Docs + Coding agents, but on a git level.\n\nThe collaboration is possible due to DeltaDB (a layer on top of Git), which tracks the finegrained changes the user asks the coding agent to do in between commits. The idea is that this data is lost in traditional git and can be useful to the coding agent in the future.\n\nAnother difference is that DeltaDB uses conflict-free replicated worktrees, based on CRDT-style replication. Multiple humans and agents can operate on copies of the worktree on different machines and have their changes synchronized in real time, like Google Docs does.","date":"2026-08-13","createdAt":"2026-08-13T07:43:00.000Z"}]},{"date":"2026-08-11","entries":[{"id":"3b9ae824-27e5-80a0-a281-c80020fdabc3","title":"Claude improves lower bound for the  fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis from 41.6% to 67.2%","url":"https://www.anthropic.com/research/riemann-zeta","thought":"> Jarred Sumner, an Anthropic staff member (and non-mathematician), prompted Claude to “take a real stab” at the hypothesis itself, leaving the mathematical choices from there up to the model. Initially, Claude generated and tried 650 ideas, none of which worked. Jarred prompted Claude to try again, and it spent a day and a half coordinating about 60 Claude subagents, which this time went much deeper: between them, they ran 2,400 shell commands and wrote hundreds of Python scripts.1 The subagents ran thousands of numerical checks against known zeta zeros and refereed one another’s work. Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”).2 This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress.","date":"2026-08-11","createdAt":"2026-08-11T20:44:00.000Z"}]},{"date":"2026-08-08","entries":[{"id":"3b7ae824-27e5-8048-b0e7-e4005f443d4a","title":"Cache invalidation really is one of the hardest problems in computer science","url":"https://surfingcomplexity.blog/2022/11/25/cache-invalidation-really-is-one-of-the-hardest-things-in-computer-science/","thought":"","date":"2026-08-08","createdAt":"2026-08-09T02:42:00.000Z"},{"id":"3b6ae824-27e5-80b5-9c6e-ddeaf2d29686","title":"Advanced Napkin Math: Estimating System Performance from First Principles","url":"https://www.youtube.com/watch?v=IxkSlnrRFqc","thought":"","date":"2026-08-08","createdAt":"2026-08-08T18:00:00.000Z"}]},{"date":"2026-08-07","entries":[{"id":"3b5ae824-27e5-808c-b3b2-f3e80238c24b","title":"Gemini is Cooked but GCP is Cooking","url":"https://newsletter.semianalysis.com/p/gemini-is-cooked-but-gcp-is-cooking","thought":"","date":"2026-08-07","createdAt":"2026-08-07T17:08:00.000Z"}]},{"date":"2026-08-05","entries":[{"id":"3b4ae824-27e5-80b1-b1d9-f0a1bb81cd0c","title":"CUSEC 2016 – Reliable Software in a Chaotic World by Simon Eskildsen","url":"https://youtu.be/MOluPRfnyVk?si=Y-N6nL2UoIrOOe-q","thought":"- [Bitsquatting](https://en.wikipedia.org/wiki/Bitsquatting) is super interesting.\n- How loosely coupled are different things in your application? When things go wrong in your application, do cascading failures lead to dramatic failures and complete loss of service?\n- Are there fallbacks for different things in your system? If a thing fails (primary), is there a secondary fallback which can provide service in a degraded manner? For example the talk gives an example of a e-commerce website where if the user auth service doesn’t work, the logical fallback is to log the user out and let him / her continue browsing and check out as a guest.\n- Resiliency matrix: Once your application is built on top of building blocks, you have to think about resiliency as a O(N^2) problem. You should know what the resiliency matrix for your application looks like, where the columns represent the basic user actions & rows represent the failure of any building block used in the application.\n\n![](<https://prod-files-secure.s3.us-west-2.amazonaws.com/bb74a6cf-6d93-4bbc-84b3-0031808e6acf/9970e7fc-bc00-4d5e-996d-9613ff1c0454/image.png?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=ASIAZI2LB466YBWPRUWG%2F20261117%2Fus-west-2%2Fs3%2Faws4_request&X-Amz-Date=20261117T192527Z&X-Amz-Expires=3600&X-Amz-Security-Token=IQoJb3JpZ2luX2VjECMaCXVzLXdlc3QtMiJIMEYCIQDOHGKHUXoq153rtzFbb2zWpbOT8ZGkCwe0XuxU7hc6UgIhAPbuwLQb6cuhfcybSbn4FhVWJYw7iU7RpFcs65joNd4YKogECOv%2F%2F%2F%2F%2F%2F%2F%2F%2F%2FwEQABoMNjM3NDIzMTgzODA1Igx9k%2BtqQMHpUDLbRDwq3AMQwuVnUH5r673iRiU2HGFe3dKJAGfXktZo6AJcOwpOJ8OCm6H7UhvT6KGKoIZoPZC0HcEfpZ7xTQqreHDbaBDhtwWhrMnbNyAMFWne8JmDG2Qbx0Pd2TgyKfJahijjTvvLPZUTTDQN2KMHPn2v0lrD1uKqgma3rbu1ZwQ7ygHk2pmoAu832xyIYUNmD3sf5%2FGbd%2BDWOC2slNUct4rvJOEk%2FTG8oyuMjqzu%2B4xr8GeYsdNnA0%2BBUYkFEHIgwLkgaKW40Ag1%2FcEGwQdB%2FkjfbiF6Ny1OB6ar8OsiPf8a2dlk6%2BYQp6IjXPiNm9OlTAECIMKczYazG7q%2B%2Bj3YJSg%2BGU%2BIJqq8yhEUBZfdo4jy9n7kvnuKzAIs116nrmQa8WUKaEKO6CXKMYSuM0cHt9kGDyMuT4rFKqUhNJEx%2BGLuXq8x2jA7Rzcjlu6VqnyqayRRhgl60Dc38PZZtRiMyiLRbsPb2u3%2F4s8qfd6WudBQW090IL5pSE3fbrMMXF7IeOWLb4LzTT4l64Nz49Mmb60jUidzgSQqMYVVh%2BuUl1s1B7ONqfnhDG5UlnAwpyMf9kqmzq33OIq6jHx6kQsbZxhycHas2EdMV5NYla9d%2BRS1L1o6RW5PIegwXNHoZBjtMTCjv%2FLXBjqkAbFl4Gtpj5E93NWlYDzYk2j8wfDnDGna%2BJ6YFl4egOAqnstdmRE7l4zRDpg2fVQiCY5njGWuW7YSg%2BQ6RgtzVeYzORF9Urcvzu1YVvYg7Lk67nzFgp3wpLX%2F%2F%2BJHuV4v9VRSzruPztETy55wSDuuFmVhv44BUVLxw%2BE6XRrbFZT%2FhTt4K0YYSCIrW4ecGSaTmXupiZlYhX3Zi60QSTgXqcYOG%2FcK&X-Amz-Signature=2647f644c47433457180b75bc5f96b3eeb827183b913ce03f09631d4d3d8dc6e&X-Amz-SignedHeaders=host&x-amz-checksum-mode=ENABLED&x-id=GetObject>)\n\n- Testing each condition of resiliency matrix can be hard. You can use [Toxiproxy](https://github.com/Shopify/toxiproxy) to add fake delay between different blocks of your system to test the behaviour of your application when database is down & cache is slow.\n- “With fallbacks the system is still vulnerable to slowness. ECONNREFUSED is a luxury in distributed systems, slowness is the killer”.","date":"2026-08-05","createdAt":"2026-08-06T01:04:00.000Z"}]},{"date":"2026-08-04","entries":[{"id":"3b3ae824-27e5-8079-8630-dc85df698cd4","title":"CMU Database Seminar: turbopuffer, Object Storage-native Database for Search","url":"https://www.youtube.com/watch?v=pqoRNwNaxfs","thought":"General advice from the talk:\n\n- Being on-call during the night helps you write better code during the day.\n- When looking for possible optimizations you can make in a system, look for the gap between [napkin math](https://github.com/sirupsen/napkin-math) & real world performance — it’s either an opportunity for optimization or a bug in your understanding.\n\nAbout [TurboPuffer](https://turbopuffer.com/) (object-store native database)\n\n- Object storage is the sole truth for turbopuffer. Everything else—query servers, indexing workers, RAM caches and local NVMe caches—is designed to be replaceable / stateless.\n- S3 became consistent in 2020.\n- S3 added CAS support in 2024.\n- Coordination is done with CAS on object storage.\n- Tradeoffs of Turbopuffer:\n    Pros:\n\n    - low cost\n    - great search performance in terms of cost.\n    - high write throughput\n    - fast warm queries\n    - extremely simple because you don’t have a fat metadata layer\n\n    Cons:\n\n    - slow cold queries (when you go to object storage).\n    - high write latency\n- For search, graph index (like HNSW) might take more roundtrips to object storage than clustered index, so clustered indexes are used. Graph indexes are optimized for cheap random access, while object storage is optimized for fewer, larger, parallel reads.\n- A new database arises when:\n    - there is a paradigm shift & every company has to run a new type of workload\n    - if existing incumbents cannot support it without a huge architecture change.\n- When you design a database, you worry about read amplification, write amplification and space amplification. With object storage database, you also worry about S3 operation amplification. “How many GET, PUT, LIST and multipart operations are needed?”","date":"2026-08-04","createdAt":"2026-08-05T04:02:00.000Z"}]},{"date":"2026-07-31","entries":[{"id":"3aeae824-27e5-80b3-90c9-f73fcadf27bf","title":"Anthropic’s blog post on their model breaking into three separate companies while running internal benchmarks","url":"https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals","thought":"I think Anthropic is just responding to OpenAI’s recent blog post where they broke into HuggingFace’s production during their benchmarks. I don’t know how serious the companies they broke into are, as one of them have not even responded back **to** the team. The interesting case for me is that in both cases (OpenAI & Anthropic), the sandbox was broken by the underlying model. Since most of the infra / sandboxes was built with vibe-coding with an inferior model, is always coding with the frontier model or having a security check / review with the frontier model always necessary for merging code into production in the future?","date":"2026-07-31","createdAt":"2026-07-31T07:03:00.000Z"}]},{"date":"2026-07-30","entries":[{"id":"3aeae824-27e5-8089-ab60-fbc76f71cf2b","title":"The “Situational Awareness” situation","url":"https://x.com/ParadisLabs/status/2082840577230459018?s=20","thought":"This story is crazy. This was a collapse on the scale of FTX / Lehman brothers, but on a random week in July where retail Korean traders are getting margin called on AI stocks and AI stocks are ~30% below ATH. Citadel supposedly bought everything and the market popped immediately today as soon as they took their book. This situation reminded me of [Long Term Capital Mangement](https://www.goodreads.com/en/book/show/10669.When_Genius_Failed). Running 4x levered on your entire book can wipe you out and force you to liquidate. A good investing lesson for everyone on the West Coast in 2026.","date":"2026-07-30","createdAt":"2026-07-31T06:59:00.000Z"}]},{"date":"2026-07-25","entries":[{"id":"3a9ae824-27e5-8011-b901-f8ca5773d948","title":"The new rules of context engineering for Claude","url":"https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models","thought":"Interesting blog post by Anthropic. They removed over 80% of Claude Code’s system prompt because the new models are much better at understanding intentions.\n\n1. When we use coding models, they discovered that the model’s system prompt, the harness prompt, and user defined skills and instructions often had clashes and conflicts between them. They found that they can remove many of these instructions and the models can infer them from the environment & context. For eg, they no longer need to instruct Claude Code to not delete files.\n1. Instead of giving examples for tool use, giving better tool definitions, descriptions, and parameters to Claude is better.\n1. Use progressive context instead of putting it all out at once. Instead of asking it to review code everytime after it’s written something, add code review skills and it progressively load it into context.\n1. Some more tips:\n- keep [claude.md](http://claude.md/) light, dont explain things that can be inferred by looking at your codebase.\n- use progressive context disclosure — have more skills rather than hardcoded instructions.\n- keep your skills lightweight as well. if you feel your skill is become too heavy, split it up into multiple files.\n- reference files with @ in your skills / plans.","date":"2026-07-25","createdAt":"2026-07-26T00:32:00.000Z"}]},{"date":"2026-07-24","entries":[{"id":"3a8ae824-27e5-80ff-8cb8-d3d998b8e1ab","title":"Tuhin Srivastava (BaseTen) in Economics of the AI Supercycle","url":"https://youtu.be/Qh7Oxvo5sJI?si=drgq3R_6zk35Oha0","thought":"Good watch, helped me understand the stickiness of inference layer, why companies choose to posttrain open source models & what moat [BaseTen](https://www.baseten.co/) has over the other neoclouds / hyperscalers.\n\n- Open source models are 90 days behind frontier labs. However open source models can run ~70% cheaper than frontier models for the same quality of work done.\n- So if you’re running a software business, when it comes down to the balance sheet, you can bet that all the big players / apps will go down the path of open weight models for the cost margin. I actually agree with this. Last year, I would not ask an inferior model to figure out where I should eat icecream, but this year it’s become clear that smaller & cheaper models are good enough at actual productive tasks.\n- A cynical reason to use open source models is that your user signal, data & IP can be stolen anytime. Cursor was told Claude Code was [\"just a research project\"](https://www.businessinsider.com/cursor-ceo-michael-truell-spacex-elon-musk-anthropic-2026-6)—until it wasn't.\n\nHe even compares closed source model labs to East India Company!\n\nBaseTen is built on two bets:\n\n- Application layer is independent and will still exist (SaaS won’t be dead)\n- Open source models will continue to keep up with frontier models, even if they lag behind.\n- BaseTen has access to enough compute to fulfill demand.\n\nBoth of these bets looked impossible to be true just 1 year back, just goes to show you how quickly things change.\n\n- Inference is sticky — even if all hyperscalers build their own inference layers on top of the hardware, inference turns out to be as sticky as a database.\n- Renting vs Buying — access to hardware is a bottleneck. Renting puts you at a disadvantage if you’re trying to gather more compute & it also more expensive than buying (~30%).","date":"2026-07-24","createdAt":"2026-07-25T04:57:00.000Z"},{"id":"3a8ae824-27e5-8039-99ea-fd4cf84a6e31","title":"Jensen’s first tweet!","url":"https://x.com/JensenHuang/status/2080643682408321103","thought":"Jensen posted his first tweet and it’s already iconic!\n\n- He published a letter supporting open-weight models. He argues that open source software has helped build the American economy and so will open-weight models.\n- He also argues that distillation should not be looked down upon, as it’s just like a “fork” in software world.\n- FYI, Jensen follows all the companies he works for. By all, I mean two — Nvidia & Denny’s. Name a more iconic resume.","date":"2026-07-24","createdAt":"2026-07-25T04:34:00.000Z"}]},{"date":"2026-07-23","entries":[{"id":"3a6ae824-27e5-80df-acfc-c1e4356e80af","title":"Show HN: Bento - An entire PowerPoint in one HTML file (edit+view+data+collab)","url":"https://bento.page/slides/","thought":"Big fan of the idea. It’s the idea of a PDF for PowerPoint and where you can still make changes.","date":"2026-07-23","createdAt":"2026-07-23T23:10:00.000Z"}]}]}