Mr. N @synthbeat

I always think that Grok models should be left in the mix, for that specific reason all NSFW stuff, but close to 85% SOTA technical stuff. I haven't applied that within coding context yet deeply, but was experimenting with adding Spanish profanities via Grok API and it complied. I hope they haven't neutered it as it was Grok 4.5 or something.

@petergyang @simonw Well, that would an equivalent of updating it PelicanBench 2.0 as current results are near saturation already

Astra is BEHIND MUSE SPARK 1.3 MAX ON ARTIFICIAL ANALYSIS??? I was starting to get skeptical of their intelligence index, but there's no way I can trust this anymore

Thank you for writing this post. Some of these nuances, are really the missing docs from OpenAI. But it gets even more confusing, if you have enabled "Developer Mode" so your ChatGPT Chat can access your server via private MCP. And then you have Remote, to run against Codex limit, on your devices, and then you have Grok/Anthropics/Gemini subscriptions on top of that, so even though the product use case is clearer, the choice for us (human operators) is still non trivial🙈

Refine your approach until it’s so seamless you don’t need to be in the mood to do it.

Your visual style for your projects, is pretty recognizable, and if you get the background behind them it's pretty nice .The aesthetics shouldn't distract from the projects' value. Getting into realistic, can get too pretentious, like some crypto-utopian futuristic kinda of thing.

One of the best things you can do to read well is to use memory first, then bring in comprehension. Take a research paper. You are not going to truly comprehend this in one shot. It is deep, and using a fair amount of jargon outside your immediate familiarity. If you attempt to comprehend the paper in one pass, it will take you longer than what is justified. Worse, the time it takes you will destroy your ability to see the connective tissue that binds the work into a single cohesive story. Only a swift read can give you that necessary holistic view. So instead, your first pass should use memory techniques to anchor the scaffold *you* think is important. Yes, reading someone else's work is still a *you* thing. Now you have anchors that resonate with you. There's a reason you felt these were important, you just cannot quite articulate why yet. Your wealth of unique experience gave you these anchors. Take them seriously. They are yours. Now, for comprehension, you must take these anchors into dialogue. Dialogue with yourself (go for a walk), with someone else (go for coffee), or with AI. Use each anchor to start your conversation. What does this really mean? Why is it important? What background is needed to truly comprehend this point? Now you are blending the swift memorized scaffold with deep comprehension. You have both narrative and understanding. Story and appreciation. This applies to any book. Do not sit with each sentence as though it were gold. Too much filler in books. Instead, move swiftly to see the whole. What your mind wants to keep it will keep. Use spatial locations in the mind. Then being those locations into dialogue. Enrich the anchors with highly variable information that hits them from many angles. Make them robust in the face of relentless questioning. Make it you. Make it memorable. Make it meaningful. That is how one reads well.

AI exposes what computer science was really valuable for all along: reasoning about computation and systems. The new scarcity is “what should the system do, how should it be structured, and why should I expect it to behave correctly?” State, recursion, search, optimization, compression, feedback, representation, protocols, computational complexity and so on transfer very well to organizations, biology, economics, AI agents, scientific workflows, and other systems. Time for industry to ditch the stupid leet code and whiteboard exams, and focus on what actually matters now.

The reason people are selling you ways to do what they did, is because what they did is no longer making them money.

For anyone with endless ideas, this agent age is nirvana as those ideas are met with endless execution, endless exploration. I've never had has much fun working with computers as I do right now. What a time to be alive.

@doodlestein Is it already 30x, I thought it was 10–15x?

Yes, that's what I thought about. Especially if you dig further into it, it doesn't come up as a free option (a stylistic toggle), like a deterministic option. I don't know what went into training, in context of such artificial English. I can guess it will dumb down the models, beyond the restricted subset of vocabulary.

@doodlestein you think it's worth it? Haven't seen your opinion on this. I'm very skeptical, looks way too superficial. Unlike some small prompts, like your fresh eyes trick.

Daniel Lockyer @DanielLockyer

a month ago

I've updated my rules to tell AI to use Simplified Technical English Will report back

72 43 1K 460K 2K

@_alphashark_ @yoheinakajima @ActiveGraphAI @doodlestein Well, define swarm, if working on one repo then. You either deal with merges from worktrees later or with agent coordination, if working on same branch.

@_alphashark_ @yoheinakajima @ActiveGraphAI You should really follow @doodlestein work. He is having exactly swarms of agents working on every repo throughout his 50+ active projects or so.

@_alphashark_ @yoheinakajima @ActiveGraphAI I'll just leave it here

Jeffrey Emanuel @doodlestein

7 months ago

@jarredsumner Worktrees suck ass, I’ve been saying this for months. The best solution I’ve found is my agent mail system with advisory file reservations. Plus beads. I’ve been using it for months now to produce ridiculous amounts of high quality software by myself.

11 1 60 16K 49

@doodlestein @korenmiklos Thanks for promoting it heavily in replies. It's great when rm -rfing their home dir, happens on someone's else machine, and you can still overcome procrastination, and take it seriously, like I did in one of those last public cases.

I just wanted to tag @doodlestein in this. If you go the the GPT 5.6 xhigh way it doesn't take an effort to spend 20-30% per day on 1-2 projects. Use beads, then you can expect any feature to be done via 7–8 phases, with proper prompts. All testing, E2E, will make the agents work for even quite trivial stuff for 12+ hours.

The only way to make a thing truly consistent is to strip the life out of it.

No, you do not understand. What Bryan does is not trial-and-error. In trial-and-error, you try something, observe the outcome, and let the system itself reveal what works. That's how evolution works. It's how machine learning works. The key is that the system generates the evidence. What Bryan Johnson is doing is very different. He starts with a theory that certain biomarkers represent health. He intervenes in his body to optimize those biomarkers. He measures whether the biomarkers moved in the desired direction. He concludes the intervention was beneficial. The crucial difference is that the measurements are embedded in the intervention itself. He is changing a complex adaptive system according to a reductionist model of health. Biomarkers are not causal variables. In a highly interconnected system, changing one measurable quantity can be expected to produce compensatory unknown changes elsewhere that aren't measured/known. Bryan Johnson's approach says "I have a model of the solution, so I'll optimize the variables my model says matter." That is NOT trial-and-error. Completely different epistemologies. Trial-and-error discovers the objective function through repeated interaction with reality. Bryan Johnson assumes he already knows the objective function and optimizes a handful of measurable variables within it. The risk is that in a complex adaptive system, the true objective function is vastly richer than the biomarkers being optimized. He is doing naive intervention based on simplistic notions of measured health, not trial-and-error. If you optimize a bad map, you don't reach your destination faster, you just get lost more efficiently. And here, "lost" is the degradation of one's true health.