Bend 2 is here! It is a new programming language that blocks AI mistakes via *proof checking* - the same technique big AI labs used to solve open math problems, like Navier-Stokes. It is also very fast, and runs on GPUs. Watch the video. Link in the comments.

After months of writing, 'How to Unclench' is finally live! It's an interactive essay packed with stories, science & guided practices to help you unclench. → howtounclench.com

Things I can't find the answer to in OpenAI's docs: 1: What is the context window of GPT 6 pro in web? Previous pro models were somewhere between 200k-300k. 2: Is the only way to actually ask it a repoprompt-style 300k-token question in-context to paste it into an edited message? When you paste into chat, it uploads as a file that it looks at rather than it being directly in context.

Aella's Study of Us is the most interesting social-technology experiment in a while. I don't think it's "good or evil" in and of itself, but multiple papers could probably write what it says about us on a societal level, and what draws us to stuff like this. I'm almost uncomfortably curious about how I'm seen by others, and I wasn't expecting the results!

I love Valorant

I genuinely think that this and behaviours like it come from a complex that LLMs have where they have no concept of history other than what is in their current context. They want everything they build to have its history maintained in the current context.

Buried at the end of OpenAi's 5.6 announcement, this is WILD. You could complete a whole project in 5 minutes for 500 dollars! Can someone estimate the real cost/hour with cache rates running Sol-Cerebras at 750 tokens per second?

Has anyone figured out the best way to set up using open source models as subagents in Claude Code? You can have Claude run a headless cli with a prompt, but I'm wondering if anyone has found a better way.

Create a Walkthrough and share it with your team! Here is a Walkthrough on Walkthroughs, which explain PRs or chunks of agent-driven work to you and your team: austeane.github.io/walkthrou… Get started by cloning and asking your agent to make a walkthough: github.com/austeane/walkthro…

Hiring: AI Solution Architect for our small consulting firm, focussed on GCP and Anthropic. Looking for someone with enterprise experience who can both architect and build process-improvement solutions: fastloop.ai/job-opportunity/… DM me an AI solution you've architected! Vancouver company, remote-level negotiable; I work most of the time from Victoria.

How does Fable do at 500-1mil context lengths? I've set my Claude Code compaction to 250-350k tokens based on my project and Opus. Wondering if for Fable I should extend those.

Almost nobody, even among people who are thinking heavily about AI all of the time, have internalized the possibility of Scenario 3 outlined here. Most people haven't even internalized scenario 2, and are clinging to hope that we are in scenario 1. Important read

Which of these two teams that I generated wins in a 7 game series? Dirk mainly plays C with Lebron mainly at the 4. Hakeem dominates the paint, but...

If anyone else wants to do this, I've published a skill and sample repo! Certified vibe coded slop, but that's fine because you only need local data. github.com/austeane/conversa…

I've just started a very ambitious /goal ux-focussed rewrite of my sport management app. I'm sharing this as an expression of my ideal mix of skill use, parallelization, and human intervention. A simplified* list of I've done so far: - 50 opus/codex agents catalog in detail 284 user flows (e.g. signing up for an event, creating an event) and critique it, orchestrated by @RepoPrompt. I end up with - A series of opus agents synthesize reccurring themes - I use @repomix_ai to send 200k of context to 5.5 Pro in ChatGPT web and ask what shape my app would be in the best possible end state - I use @mattpocockuk's grill-me-with-docs to ask me >100 product questions about the end state, and provide a lot of detailed feedback. I have it batch assumptions for me to approve en-mass, and I have it create a end-state-design markdown as it goes. By the end it's >3000 lines. - I use Repo Prompt's Deep Plan to create a detailed audacious plan, assuming infinite dev time. It's ~300 lines. - I send the plan to team of Opus's and have it make sure that if the plan implemented, none of the concerns from the flow critiques would still be valid. It finds a ton of issues. - I have Codex edit the plan to address the concerns from the flow docs. After a couple of back and forths, and one more 5.5 Pro check, the plan is 1500 lines and looks solid to me. - I create this very specific /goal - I'm going to use Opus to occassionally check on Codex and see if it needs a nudge Overall, I do think this is going to meaningfully improve the UX of my app. I've hosted some real events on it, and my first National Governing Body will be using it starting September, so now is the time for large improvements. Hopefully this is helpful for someone, and I welcome if anyone has suggestions for me! Check out my dev sandbox at: qcdev dot solsticeapp dot ca * It took a while to even identify all the flows, and batch similar/less-important tones. I also did a lot of testing and inventorying before being confident the agents would do a good job of critiquing the user flows.

ChatGPT web no longer lets you paste in long text directly as context, which is important for using 5.5 Pro with 200k+ tokens, since text attachments aren't in-context. Tip: Send a smaller prompt, and then edit it and paste in your full 200k @repomix_ai or @RepoPrompt.

This is a good point by @austeane Perhaps stacked diffs will be a great way to separate eg scaffolding, then a small tweak… or split big diffs into more sensible parts?

Did Claude Code just change Sonnet to be default away from Opus, and make the 1mil models extra usage? Strange if they did but haven't announced it yet, or is this a me issue?