William Callahan

New year, new personal website! With an added in-site terminal - williamcallahan.com

@RenjiePi @RenjiePi I’ve been a big user of other Nemotron models. Its biggest weaknesses seem to be around its (lack of) reasoning. One could send the task planning/investigation from the execution to different LLMs, but that’s not really practical.

Introducing Nemotron-Terminal: a systematic data engineering pipeline for scaling LLM Terminal Agents. We bridge the gap between open models and proprietary models with a fully open synthetic-to-real trajectory pipeline. 🤯The payoff: SFT on our Nemotron-Terminal-Corpus boosts Qwen3-32B from 3.4% → 27.4% on Terminal-Bench 2.0 (+24.0), rivaling models multiple its size. What makes it work? 🌟Terminal-Task-Gen: A lightweight data curation pipeline that seamlessly combines the adaptation of existing datasets with robust synthetic task construction. 🌟Nemotron-Terminal-Corpus: A massive, open-source dataset covering diverse terminal interactions, which contains explicit planning and execution traces for complex long-horizon tasks. And we’re releasing everything: 📦 Nemotron-Terminal-Corpus (Large-scale dataset) 🤖 Nemotron-Terminal models (8B, 14B, 32B) Paper: arxiv.org/abs/2602.21193 HF Daily: huggingface.co/papers/2602.21… Models & Data: huggingface.co/collections/nv… Our tech report just hit the #1 spot on Hugging Face Daily Papers! We're also incredibly excited to see the open-source community putting our work to the test, with the Nemotron-Terminal-Corpus dataset currently trending at over 1,800 downloads and counting. We can't wait to see what the community build with it!

A beautiful rust-based IDE TUI! Great work @TysonGeorge.

@nikunj Yeah the blockage points seem to be the driving loops caused by surprise-closed routes? Part of the problem seems to be the peculiar way they’re parking the meter enforcement vehicles in intersections, making it unclear what is open/closed, especially to Waymos!

Pro-tip: If you’re trying to get around downtown SF, don’t drive or Uber this week. It’s absolutely pointless. Lyft e-bikes are the way to go. Saved 20+ mins last evening.

@ajambrosino A nice kickoff to start at OpenAI and this team @inferencetoken. Great work @ajambrosino and team!

Today, we’re introducing the Codex app, our flagship Codex experience. Work on multiple things in parallel, extend Codex with skills, and automate repetitive tasks. The most exciting part for us has been using the app to build itself. This is the first of many new things coming to Codex– more to come.

How I feel explaining SF and AI right now every time I go back to my home town 😂

@ArturSkowronski @myfear @Jfokus @gAmUssA Hah - yeah, that’s definitely why I shared and thought you’d appreciate it. Love it, let’s show ‘em what we can do with JVM languages!

@WilliamCallahan @myfear @Jfokus Just in time - that was my comment to @gAmUssA talk today 🙈

JVM in Age of AI - 2026 edition slides are here! First time I ever updated an existing talk - at least to that extent 😉. I ended up with 200+ slides, 11 demos, and a six-JDK-version magnum opus. Even with the 90 minutes the amazing @Jfokus crew gave me, I still barely squeezed it in 😁 slideshare.net/slideshow/jvm-…

@mattyp Hey that’s my house! And yes, it’s magical here 🌉

you can just move to san francisco and live 15 minutes away from this

Wow, new record! github.com/WilliamAGH

@MaximeRivest @willccbb Definitely Gemini 2.5 Pro was the big step jump… but coding with Gemini models is definitely a love-hate relationship 😂

I genuinely don't remember exactly what LLM/AI tool was released in May but something clearly affected the behaviors of many technical people. @willccbb do you remember?

@ctatedev I took a stab at this last year… this looks great! A worthwhile project done well.

Introducing json-render AI-generated UI. Deterministic output. 1. Define your component catalog 2. AI steams JSON 3. Render interactive UI Let users prompt dashboards, widgets and apps - safely constrained to components and actions you define

@nicolaygerold Not to mention memory. I have a MBP M4 w/ 128gb of memory that regularly runs out with many simultaneous sessions, largely due to the combination of each session spawning its own batch of npx instances for each MCP and language servers. The ghostty killers 😂

Just shipped lazy-loading for MCPs. Now your agent can use skills to load the MCPs they actually need, only when they need them. Don't forget to take the MCPs out of the global configuration.

@belicosus @ybhrdwj *lose/lose, as loose means another thing 😉

@ybhrdwj The sad thing is, that if Vibe coding kills Tailwind, all the created applications will be broken going forward. Currently it's a loose/loose situation

Tailwind lays of 75% of their team. the reason is so ironic: > their css framework became extremely popular w AI coding agents, 75m downloads/mo > that meant nobody would visit their docs where they promoted paid offerings > resulting in 40% drop in traffic & 80% revenue loss

@zaidmukaddam Likewise! williamcallahan.com/bookmarks

@skeptrune @hahnbeelee @DanielWolkowitz Or bring back a Jimmy John’s

@hahnbeelee @DanielWolkowitz i'll stop complaining once they open a Jersey Mike's in FiDi

software developers merging code written by Opus 4.5

@benhylak @herakleitos137 lol at the cars having the right of way here 😂

@herakleitos137 not needed, they're from the east coast. cars have the right away.

my siblings are moving to san francisco, so my girlfriend and i wrote them 10+ page survival guides (complete with a kit of gift cards)

Opencode is one of the most elegant repos I’ve seen I recommend checking it out if you’re learning AI

@Pidud_ Some are interesting for the sheer size of their trade and diaspora travel overwhelming tourism. Including the burgeoning cross border semiconductor activity for Arizona <> Taiwan and presumably the large Latin America diaspora for Arkansas <> San Salvador.

@mrkaran_ @zerodha Hey @mrkaran_, the “vector integration” link at the bottom of the onboarding docs first page is broken. Great work on the app!

🎉After 8 months of daily use at @zerodha, ironing out bugs along the way, Logchef is finally v1.0. Logchef started as a log viewer I actually wanted to use. Now it has alerting, an actual query language, AI assistant, and a field sidebar that makes log exploration actually fast.

@Lethain @simonw Likewise! Well written/thought out. And to those saying it was written in part by AI and edited by Simon, that would seem completely congruent with his statements… so I don’t see the issue.

Love this post from @simonw simonwillison.net/2025/Dec/18/co…