Applied cosmology toward machine precise solutions to replace humans with autonomous systems in all domains.
Frisco, TX
Joined June 2009
- Tweets37.1K
- Following665
- Followers1.5K
- Likes17.5K
Pinned Tweet
As AI agents become more capable of accomplishing real-world work at human-equivalent competencies, we will need to reorganize our division of labor. Some changes will disrupt how a job role is performed. Other changes will change what job roles are needed or how many people are needed to do the work.
AI agents that automate tasks can be seen as tools for accelerating a person's work. A person can do more and they can do it faster with AI tools. However, the job role remains firmly the human's responsibility, because their know-how, judgement, and accountability are irreplaceable.
AI agents have no moral or ethical grounding. An AI agent cannot be held liable. The buck cannot stop at an AI agent. A human job role exists for those purposes.
As work is accelerated, we will see responsibilities roll up and concentrate into human job roles that have expanded scope of control. Work bottlenecks are primarily due to the slow speed of knowledge transfer, collaboration, and consensus-building between humans. The more that multi-person collaboration can be replaced by solo work, the more efficient thr work becomes.
Job roles will become more multi-disciplinary, as responsibilities concentrate into solo work for improved efficiency. If coders no longer write the code and they no longer read the code, the job of engineering software shifts from coding to overseeing requirements, design, quality assurance, and building machines that do those activities along with doing the coding.
This disruption and displacement is emotionally taxing. Egos get bruised. It's very difficult for highly-paid professionals to accept that their knowledge, skills, and competencies can largely be automated. Their experience remains valuable, but it needs to be applied differently. Instead of toiling as they used to, the knowledge, skills, and expertise need to be directed into training, building, and running machines that do the job role they used to do by hand.
It feels like OpenAI made an adjustment to how usage is counted, because I'm seeing a dramatic difference on the $20 Plus plan. Three days of usage 5-7 days ago knocked my weekly allowance down to 10% remaining. Since that time, I've only consumed down to 7% remaining, and it feels like there is an infinite allowance remaining when using gpt-6-luna high. I'm wondering if they started charging zero for cached reads.
AI coding depends on answering questions very carefully. Otherwise, the agent goes in the wrong direction, and you derail very quickly. Unfortunately, the agent uses language that is incredibly dense and obscure. Each word in a sentence implies a library of knowledge and expertise. Half the time, I only understand half the meaning. For example, I know enough to decode AS as Authorization Server, the OAuth 2.0 token issuer. I can guess confidently that grant-provider means the authorization logic within Keycloak for adding a permitted claim on the token.
I need another AI on the side to explain what the hell I'm even looking at.
Presumably, my past decisions on Storage is summed up as "private grant-provider/context/mapper". Somehow that is connected with RFC 7662 (whatever that standard is about). Reading machine-generated questions is like translating from a foreign language even though the words are all in English.
Crawling down a rabbit hole so you don't have to. In case you wonder why the tiered ranking has S+ above traditional school grades (A-F), the S is for 秀 (shū) which means excellent or exemplary.
This huge jump in strength will greatly benefit all those users who are solving open math problems.
Is she the only dissident alive in Canada? @RealAndyLeeShow
‼️Meet the new "DOGE" Minister in Quebec!
Jacinthe-Ève Arel was just appointed by Plamondon to CLEAN UP government.
She routinely mocks Carney as an authoritarian who kneels before Buckingham Palace.
Here are some of her best quotes about Carney and Trump!
On floor crossing: “They called Trump a threat to democracy. What do they call it when politicians switch sides overnight and no one blinks?”
On throne speech: "Mark Carney, the man who claims to stand up to Trump, invites the King of England to affirm… Canada’s sovereignty. When you think you’re strong enough to defy Washington, but you still kneel before Buckingham to have our speeches read. Weakness dressed up as majesty.”
On Carney saying 50%+1 is absolutely not sufficient for independence: “And it’s the President of the United States Donald Trump who threatens democracy, who has authoritarian tendencies and who remakes the rules his own way? The emperor has no clothes in Canada.”
She has also compared certain COVID measures to totalitarianism.
Canada needs more Jacinthe-Ève Arel's!
In stark contrast to professional work, where I keep the code under close supervision, apply human understanding, taste, and judgement carefully, I let my personal projects progress largely unsupervised without human understanding and occupying little to no attention to detail. Yolo-slop it until I have time to look at it later.
What a difference. When I came back to the personal project days later, I see upward of 50 worktrees across a half dozen open issues, all partially implemented with a web of blockers and no comprehensible path to finishing. More concerning is that the worktrees hold divergent code with no clear path to merging. What on earth is the agent doing? I don't think even the agent knows.
Obviously, I want to experiment and learn from applying riskier techniques and practices, so that I can refine those into a mature methodology. I find it helpful to journal what I do like this out in the open, because solo dev is lonely. That being said, team dev is even lonelier, because working with humans is exhausting.
After spending a ton of tokens on github.com/jetpen/decent-key… without any understanding of where it is taking me in pure hopium that it will land a PR after acceptance criteria are met, I have a free moment to look at the code. Oh boy, did I have feedback. I asked it accumulate my list of grievances and guidance. The result in the amendment under section "Required lifecycle and documentation remediation before closing Issue #5".
After this remediation plan is implemented, hopefully this will look better. Right now, this repo is an ugly mess of slop. That's to be expected, given that I started doing this with no prior knowledge of the base technology, only a vision for the outcome, and not a clue about the implementation approach.
I had set the coding agent off without even the barest of guidance or direction, except the desired outcome. The coding agent has churned and churned making incremental progress, and at times backsliding, when the final destination looked to be within reach. Finally, the weekend has arrived, and I have a few minutes to give this some attention.
The quality of the README is a reflection of the overall repo right now. It's laughably bad. Let's see if this can be cleaned up without starting again from nothing, maybe mining the progress so far toward a fresh redo.
github.com/jetpen/decent-key…
I should have been more precise in directing it to develop a lifecycle control cli targeting local build, test, and acceptance prior to pushing to create a PR and executing CI pipelines. Now steering it to do that, instead of targeting CI.
Every time I've used a banked or purchased reset, tibo has granted a free reset an hour or two later, so I've had the worst luck. Debating whether to purchase one today...
We need a yaml descriptor for declaratively describing a shape for multi-agent orchestration. Describe each agent in terms of its name, purpose, model, and permissions for sandboxing. From this portable descriptor, we should build tooling that allows this shape to be materialized in any coding agent.
19 AI coding rules, all from 14 of my posts that 𝗲𝗮𝗰𝗵 𝗽𝗮𝘀𝘀𝗲𝗱 𝟮𝟬𝟬𝗞 𝘃𝗶𝗲𝘄𝘀. Together those posts have 8.7M+ views and 94K bookmarks.
Last night I went back through six months of posts, broke them down rule by rule, rewrote each one, and took several drafts to get to the list below.
Kickoff
1. When you hand it a task, give it a goal: how far to take it, how to verify each step, and what counts as done.
2. When unsure, ask first. Don't silently pick one interpretation and run with it. Touch only what's required, and don't fix up neighboring code on the way.
3. For any task over 5 steps or 30 minutes, have it build a progress dashboard before it starts and update it after every step: progress, what's stuck, questions waiting on you, and what it'll do by default if you haven't answered. In Claude Code, make it an artifact: the page refreshes in place, and Ctrl+] opens it anytime. In other tools, a local HTML page works.
In AGENTS.md / CLAUDE.md
4. Delete lines like "think carefully" and "take a deep breath" that push it to think more. Opus 5.5 already thinks before every reply and decides how much. Delete them and 𝗿𝗲𝗽𝗹𝗶𝗲𝘀 𝗰𝗼𝗺𝗲 𝗯𝗮𝗰𝗸 𝗳𝗮𝘀𝘁𝗲𝗿, with no drop in quality.
5. If code is only used in one place, don't pull it out into an abstraction, adapter or config layer. Add one when a second place actually needs it.
6. Don't design for rare or future cases nobody asked about. Delete the old code you replace, unless compatibility is an explicit requirement.
7. Before adding a dependency, changing a public API or data format, or deleting user data, stop and ask me.
8. When you switch models, have the new one review your AGENTS.md and skills against the official guide: overly broad triggers, docs it has to read for every task, redundant checks, conflicting instructions, and rules that make it stop and check with you before the work is done. Recommendations first, no edits yet.
UI
9. Give it a DESIGN.md first that spells out colors, type sizes, spacing, corner radius, shadows and component rules, and have it read the file before every UI build. Don't want to write one? Take a ready-made DESIGN.md from a product you like and drop it in your project root.
10. Codex has image generation built in. If you think the UI it made is ugly, tell it: "Use imagegen to reimagine this page, then implement it."
11. Pick one page and have it redo it in the style of a product you like. Pin down the style first: a restrained palette, a clear type hierarchy, plenty of whitespace. Visuals only, no new features, and have it walk you through every change.
12. Cut taglines, helper text and secondary labels. If a screen needs them to make sense, fix the layout.
When you're stuck
13. When fixing a bug, have it find the root cause and explain the fix before it touches the code.
14. If Opus 5.5 is your main model, run /advisor fable to put Fable on the sidelines as an advisor. Before committing to a plan, when the same error shows up a second time, and before saying it's done, Opus checks with it first. Opus still writes the code.
Before you ship
15. Have Opus 5.5 send out 5 subagents to check the 20 things worth checking (design, mobile, plus the loading, empty and error states), then fix them all in one pass.
16. Then send 4 subagents after the 20 things that slow a site down. Run Lighthouse in mobile mode three times before and after the fixes and compare the medians. Don't cut features to gain speed.
17. Last, have it play the bad guy: with two regular user accounts and one admin account, actually run all 14 attacks, keep the evidence, then fix what it finds.
Saving usage
18. Start effort at medium: low for mechanical edits, medium for everyday work, high when medium gets stuck, and save xhigh and max for work where you've measured a gain.
19. Subagents that only search, read logs or run tests need their own model, or they all run on your main model. Unused MCP servers and a CLAUDE.md well over 200 lines 𝗲𝗮𝘁 𝘂𝘀𝗮𝗴𝗲 𝗲𝘃𝗲𝗿𝘆 𝘁𝘂𝗿𝗻. Turn those MCPs off and cut CLAUDE.md down.
Each rule's original post is linked in the first reply. The full checklists and prompts are in those posts.
I continue to disparage folks for becoming overly dependent on the strongest frontier models for ordinary work, because I see it as costly, wasteful, and lazy. Reliance on strength of inference is mostly a substitute for human skill and effort, especially in regard to applying that human skill and effort toward encoding that expertise into automating SDLC with deterministic tooling.
One area where I am highly reliant on a strong model is for diagnosing and fixing failures. Nobody wants to add debugging code and examining outputs from test execution to isolate bugs. It's toil meant for interns, although the work benefits greatly from human experience and intuition.
I still don't understand agent long-term memory systems like openviking or MemPalace.
For coding, I want everything to be reproducible. Therefore, the git repo is the source of truth always, with the ticketing system being for planning and coordinating the work. I don't want any memory anywhere else, especially not in a human brain or in some other compute location.
It's inevitable that Bitcoin will rise to over $1M per coin, just as it is inevitable that a Big Mac will cost more than $126,080 given enough time with monetary inflation.
Bitcoin will never break its all-time high again." — Jason Calacanis
Jason, I'll bet you $1,000,000 it will break its ATH. Here are the terms:
Bitcoin hits or exceeds $126,210.50 on or before October 28, 2028, 12PM PST
I’m happy to trust you to pay or if you prefer we can each escrow the dollars with independent party.
The bet is either settled when Bitcoin hits the ATH or on October 28, 2028.
@jason Do we have a bet?
In contrast to all the reporting from folks saying how trash gpt-6.1-terra is, we have this. The reality is, each model seems to be superior at some tasks in accuracy, speed, or cost and inferior at other tasks. People making broad conclusions about "better" or "worse" are liars.
GPT-6.1 Terra just ranked ABOVE Claude Fable 5.1, while costing almost 14x less per benchmark task!
Artificial Analysis has completely overhauled its Intelligence Index, and the results are wild!
Opus 5.5 still leads with 58, but Astra is just ONE point behind at 57! Sol follows with 55, Sonnet 5.5 with 54, and somehow Terra takes fifth place with 51, beating Gemini 4 Argon and Fable 5.1
My first reaction? These rankings must be broken!!!
But the more I looked into the new benchmarks, the more sense everything made!
Opus is still incredibly strong at coding, but Astra is much more competitive in other areas than its previous ranking suggested. And Muse Wonder Full dropping to 50 after all the hype? Honestly, not that surprising!
The real shock is Terra. It managed 42% on the new FrontierCode 2.0, matching Astra and Sonnet 5.5! The new index puts much more emphasis on practical coding, research and agentic tasks, which could explain why Terra beats Fable!
And here’s the funniest part: Terra costs $0.96 per Intelligence Index task, while Fable costs $13.04
Maybe the rankings aren’t broken after all. Maybe some of our assumptions about these models are!
