Thomas Pelletier @pelletier.dev · 18/04/2026I agree bsky.app/profile/pell.... We can of course always criticize the quality of the benchmarks, but it’s what we have today and what I would have used if I were to do that manually. 000
Thomas Pelletier @pelletier.dev · 25/03/2026IIRC its approach was to rip out unsafe uses, replace with memory copies, to get the tests to pass, then optimize the parser’s data structures to be more efficient and reduce the amount of memory allocations required. I’ll try to dig up the transcript of that session. 100
Thomas Pelletier @pelletier.dev · 25/03/2026Couple things likely helped a lot here imo. Extensive test coverage that are fast to run, and the fact that I had already purposefully isolated the unsafe bits to a specific area of the code base. 002
Thomas Pelletier @pelletier.dev · 24/03/2026Next experiment: ask an agent to rewrite go-toml from scratch given just the public API and test suite, and set up agentic automation. If one of the labs wants to support this, reach out! 000
Thomas Pelletier @pelletier.dev · 24/03/2026I've tried a few agents in this context (Claude Code, Cursor, Codex). Claude Code seems to match my flow and work best for me out of the box. 100
Thomas Pelletier @pelletier.dev · 24/03/2026Being able to kick off work in the background and review results from my phone changed the game for me. Instead of needing a block of focused time at a desk, I could get drafts for most issues and outstanding todos. 100
Thomas Pelletier @pelletier.dev · 24/03/2026geomean overhead of only ~1.4% vs v2.2.4, zero additional allocations, all tests passing. The key enabler was go-toml's extensive test and benchmark suite. The agent could make changes, run the tests, check the benchmarks, and iterate without me hovering over it. 100
Thomas Pelletier @pelletier.dev · 24/03/2026"Remove unsafe, keep performance within 10% on each benchmark, don't touch tests or public APIs." None of them could pull it off until ~last month. Opus 4.6 via was the first to actually land it: 100
Thomas Pelletier @pelletier.dev · 24/03/2026I don't have much time to work on open source these days. Removing unsafe has been on my list for years but I never had the time to sit down and work through it. Every couple months I'd ask whatever the latest model or agent was to try it like this: 100
Thomas Pelletier @pelletier.dev · 24/03/2026Just released go-toml v2.3.0. github.com/pelletier/g... It no longer uses the unsafe package. Nearly 300k repositories depend on it transitively (k8s, Docker, Hugo,...). Removing that hidden risk feels right. This is also the first release built largely with AI coding agents.github.comRelease v2.3.0 · pelletier/go-tomlThis is the first release built largely with the help of AI coding agents. Highlights include the complete removal of the unsafe package. go-toml is now fully safe Go code, with a geomean overhead ... 420
Thomas Pelletier @pelletier.dev · 04/02/2026Interested? bedrockrobotics.com/careers or dm me.bedrockrobotics.comCareers | Bedrock Robotics 000
Thomas Pelletier @pelletier.dev · 04/02/2026Thrilled to be on this truly incredible journey! Building the platform to tap into the staggering, rapidly increasing amount of data has been the learning experience of my lifetime. It takes everything one can do with a computer to tackle that scale. www.nytimes.com/2026/02/04/b...nytimes.comBedrock, an A.I. Start-Up for Construction, Raises $270 Million 100
Thomas Pelletier @pelletier.dev · 16/07/2025Out of stealth! With an absolutely incredible team, we've been building out the platform that supports state of the art ML and robotics at bedrockrobotics.com. To me it’s an adventure like no other! If you like scale, we are actively hiring! buff.ly/ZynwFKtbedrockrobotics.comBedrock RoboticsAdvanced Autonomy for the Built World 020
Thomas Pelletier @pelletier.dev · 15/05/2025Dude you know cursor AND windsurf? Principal level material 110
Reposted by Thomas PelletierZed @zed.dev · 14/05/2025This week we're launching the beta program for our long awaited debugger. If you want to join, you can sign up here: zed.dev/debugger. We'll email you once you're in. 1688
Thomas Pelletier @pelletier.dev · 02/04/2025Anyone knows of a public database of past @aws spot instances pricing? 010
Thomas Pelletier @pelletier.dev · 02/03/2025Would have loved to but you chose the week I’m in SF! 110
Thomas Pelletier @pelletier.dev · 28/02/2025Terraform but with resources knowing about AWS limits. 010
Thomas Pelletier @pelletier.dev · 04/02/2025That! and the ability to suggest a code change over multiple lines. 110
Thomas Pelletier @pelletier.dev · 19/01/2025Really enjoying this thread. Looking forward to a follow up if you end up experimenting with it! 010
Thomas Pelletier @pelletier.dev · 27/11/2024New favorite Claude prompt: "generate a ragebait article about poor kubernetes decisions with mostly bullet points. make it look like a review/lessons learned. no paragraphs" 1102
Reposted by Thomas PelletierAlex Miller @alexmillerdb.bsky.social · 06/11/2024My shoddy ASCII art about writing data to disk was surprisingly popular, so I finished off an old set of notes talking more comprehensively about durably writing data to disk and added a better version of the diagram. transactional.blog/h... 26311
Thomas Pelletier @pelletier.dev · 01/11/2024That’s great thank you! Familiar with arrow but only know data fusion by name. 010
Thomas Pelletier @pelletier.dev · 01/11/2024#rust community: any public codebase you consider high quality to learn what good rust code looks like? bonus points if it's concurrency heavy 210
Thomas Pelletier @pelletier.dev · 23/10/2024One day I might just create a new build system. In anger. 010
Thomas Pelletier @pelletier.dev · 19/10/2024A self-contained editable file that runs everywhere on Cosmopolitan. That was a fun evening hack last month 💻 github.com/pelletier/myfile 000
Reposted by Thomas PelletierRuss Cox @swtch.com · 02/04/2024A walkthrough of the xz attack shell script. An RC4 variant in Awk, what more could you want? research.swtch.com/xz-scriptresearch.swtch.comresearch!rsc: The xz attack shell script 07527
