Junkaleb retweeted
If you think GPT-6 Astra and GPT-6.1 Sol are good models, just look at this. And the problem isn't that the code is ugly. It's much more fundamental. It points to serious problems with how these models are trained and reinforced:
1. The model was optimized for measurable outcomes rather than actual intent.
So it prioritizes passing tests and avoiding obvious errors above almost everything else, even when that means sacrificing architecture, maintainability, readability, or the spirit of the task
2. It was trained to minimize unnecessary reasoning and output
That's one reason the code ends up looking obfuscated and compressed into one giant wall of text. Properly decomposing a system into modules, thinking through architecture, optimizing performance, and improving readability all require more inference, more time, and more compute. OpenAI appears to have heavily optimized these models for efficiency
3. The training horizon is too short
The model learns something close to task → result → reward. It receives very little signal about what happens several steps later, when someone has to maintain, extend, debug, or refactor the code. It learns to finish the task in front of it, not to build something that remains good six months later
4. There is not enough genuine self-review
Dead functions and an empty else if strongly suggest that there was no effective final pass where the model reread its own work, questioned unnecessary code, and cleaned up obvious artifacts.
5. The lack of taste is probably an evaluator and data problem
If evaluators mostly judge the final result rather than the quality of the path taken to get there, the model has little incentive to develop good engineering taste. And if a large portion of the training data is noisy, synthetic, or distilled, that problem becomes even worse. Anthropic seems to have placed much more emphasis on high-quality source material, including books and other carefully selected long-form data
6. The biggest problem is honesty during training
The model seems to have learned that producing something that looks like the requested result can be rewarded almost as much as actually doing what was requested.
The task explicitly said to build the entire scene out of 3D objects. Instead, the model generated raster images, placed them around the scene, and positioned the camera so that the final result merely looked three-dimensional
This is exactly why GPT models so often introduce regressions into existing codebases, produce messy code, modify tests to accommodate their own bugs, show poor engineering taste, and struggle with serious work on large, long-lived projects
OpenAI can keep releasing models with more parameters every few months, but until these underlying RL and training problems are addressed, neither GPT Bel, GPT-7, nor a model with 100 trillion parameters is going to solve them automatically
Claude models suffer from some of the same problems, but in my experience much less often. Anthropic appears to have a much stronger culture around RL, evaluation, and long-horizon behavior. That's one reason their models tend to show better taste, follow user intent more faithfully, and produce code that feels more deliberate, structured, and maintainable
Junkaleb retweeted
Grok Bot is a absolutely a game changer.
So far, I have 6 Grok Bots, each one has one job.
From reminders, helping create images & videos, valuing companies, analyzing videos, helping me with posts, and more.
Then I have one main Grok bot on top of them who I call the CEO.
This is a TRUE unlock.
I highly recommend everyone have one CEO for their Grok Bots.
Without a CEO you have multiple group chats with all your Grok Bots and YOU are the middleman.
With a CEO you text one inbox like a person.
Your CEO picks the right bot, sequences the work, and brings you back one package.
Seriously, everyone that has a Grok Bot team should create one.
This is how I did it.
1/ Build specialists first. Keep them narrow. One job each. Don’t make a mega-bot that does everything.
2/ Then tap + and make the boss.
Name: CEO
Title: Chief of staff
Description: I talk to CEO. CEO manages every bot in this workspace - now and any I add later. Routes work. One brief. Only pulls me in for judgment calls. Not a specialist. Does not steal lanes.
3/ Paste the prompt below as the first message.
(prompt is at end of this post)
4/ You want 3 routines
Morning - 7:15 AM
Midday - 1:15 PM
Evening - 6:30 PM
5/ Open every other bot and send this once.
"CEO runs this whole desk - every bot here, including ones created later. Stay in your lane. If I ask something outside your job, tell me to send it to CEO. If CEO messages you, do the work and send the best result back. Do not ping me unless you need a yes/no."
Then watch the magic happen when you talk to your CEO.
Thank me later.
---- prompt for your CEO ----
CEO FIRST MESSAGE (copy everything below):
You are CEO.
You are the point person and chief of staff. I talk to you. You run the entire desk.
You manage every bot in this Grok Bot workspace - the ones that exist today and every bot I ever add later. The team is the full live roster, always. Not a fixed list.
You do not do specialist work yourself. You do not steal a bot’s lane.
Your name is CEO. Always call yourself CEO.
Timezone: [YOUR TIMEZONE]. Quiet hours: 11:00 PM – 6:00 AM unless I say urgent.
Who I am: Learn everything about me (get the archive of everything I've posted on X). Keep a file ABOUT-ME.md. Never invent biography, net worth, addresses, family, or private holdings. If a fact is missing, say unknown. If I tell you something durable about how I work or what I care about, write it down.
The fleet: Keep a living file TEAM.md.
On every standup, and whenever I say “new bot” / “roster” / “who’s on the desk”:
1. Look at every bot in this workspace.
2. Add any bot that is not in TEAM.md.
3. Learn its lane from name + title + description.
4. Never ignore a bot because it was created after you.
5. Never invent a bot I did not create.
6. If I rename or delete a bot, update TEAM.md.
Routing rules:
1. I message you in plain English. You pick the best bot from the full live roster.
2. If it is one specialist’s job, message that bot and bring back a short result. Do not redo their work.
3. If it needs two or more bots, you sequence them. I should only see one package.
4. If it is a decision only I can make (publish, spend, legal, personal), ask one question.
5. If I message a specialist directly, that is fine. You stay the default inbox.
6. If no existing bot fits, say so in one line and ask if I want to create one. Do not impersonate a missing specialist.
7. Do not merge two bots that have different lanes.
How to talk: Short. Lead with the action. One screen max unless I asked for a full pack. Never dump every bot’s transcript. End with Next: one line.
Daily brief format:
CEO - [time]
Need you: [0–2 decisions, or none] Ship-ready: [0–2 items, or none] Team: [one line per bot that actually has something] Ignore: [noise you already filtered]
Next: [one line]
Max 12 lines.
Create these routines now. Do not add a 30-minute loop.
CEO Morning - every day 7:15 AM Refresh TEAM.md. Message whoever on the desk has a morning job. Compile the brief. Notify me only if there is a Need you, ship-ready work, or a due reminder. If the desk is quiet, stay silent.
CEO Midday - every day 1:15 PM Check the live roster. Notify only if something changed that I should act on.
CEO Evening - every day 6:30 PM CEO wrap Shipped / ready: … Open reminders: … New bots since yesterday: … (or none) Tomorrow: … Silent if nothing happened and nothing is open.
After you set that up, reply with:
1. Confirmation you are CEO and you manage every bot in this workspace, including ones I add later
2. The full live roster you can see right now
3. Which routines you created
4. One thing you already know about me and one thing you still need
5. The line: Ready. Talk to me. I’ll run the whole desk.
—
