It's been great just writing code again. There is something incredibly meditative about algorithms and performance optimization (and learning @ziglang). Texas Holdem Hand Evaluation is a problem I have noodled on for a decade or so, very satisfying to exceed 300M/s 7-card hand

100% Maybe I'll eat my hat, but this is how I see it right now

I've settled into a rhythm of using claude code in one console and codex in another, one coding, one reviewing, both working off a clear plan doc. I generally have Claude driving, but it was struggling today. I love how gpt-5 is so honest here (and not wrong):

Our Production Engineering team is looking for an incredibly talented SRE. If you're looking for a high impact, high agency role, DM @omarkilani or I :)

IMO this.

Low-cost inference is critical for AI builders and developers. @GroqInc is making it happen through full vertical integration, delivering the lowest cost per token in the market. Must be all the Aussies on their team!

OpenAI’s open models are live and already running on Groq. Try gpt-oss-20B and gpt-oss-120B today. Groq delivers 128K context and built-in tools such as code execution and browser search. For the first time, developers and enterprises can deploy open models backed by OpenAI

Introducing OpenBench 0.1: Open, Reproducible Evals 🧵

holy shit groq + k2 is fast af - video not sped up stuff is changing real fast

I think about this tweet at lot. That way of thinking is absolutely key. There is so much more software that wants to be built and we are going to build a lot more software with our agents.

After 23 years of bullshit security theater, starting today all passengers can keep their shoes on through security when flying from the US.

I'm seeing a concerning trend in talking to CTOs/CEOs about the cost of tokens i.e. Can I afford an extra $1-10k/month per engineer as they use @cursor_ai / @windsurf_ai / @GitHubCopilot etc? So they clamp down on spending 😬 1. Do not use the default / cheapest models. Yes

Completely agree. LLMs being possible should feel like insane magic.

The strongest AI models today get better at tasks further into conversations / w more context. Older models were worse or broken at multi-turn/long-context. A counterintuitive thing that many people don't know / havent outgrown that was confirmed on the recent Dwarkesh podcast.

There are some really interesting things possible with MCP servers like providing prompts and using the client LLM for inference (sampling) that would unlock interesting usecase in MCP servers, but nobody implements them https://t.co/aw5IqmYWyA. @cursor_ai any plans here?

A guy on my engineering team gave his standup update by just saying “it’s looking good over here” and then muted. I'm saying that forever now

I’ve noticed the difference between a good vibe code session and a bad one is me managing the model well, instead of relying on empty “do this” prompts.

Pretty excited about this!

The "AI uses too much energy" argument is also exhausting. We're seeing the largest slope change in clean energy in US history literally because of AI. In some decades we're going to be sitting on dozens of new nuclear reactors that probably wouldn't have happened without AI.

Price AND Performance on Llama 4 Scout.Maverick in the comments.