hexbloom

ai researcher | writer

dm open, ego closed

[hexbloom @hexbloomx]

an even more interesting version

https://t.co/9rCCd7SYyS

[hexbloom @hexbloomx]

ELON MUSK JUST SAID HUMANS PROBABLY WON'T STAY IN CONTROL OF AI

Walter Isaacson asked if we'll become subservient to our new machines.

Elon flipped it. Will we be able to boss around something thousands or millions of times smarter than us?

His own answer was one word. "No."

Will it want to boss us around? "It may not care."

We don't micromanage chimps in the jungle either.

Then he put numbers on it.

Digital superintelligence alone probably adds 20 to 30% to the output of goods and services. About "$30 trillion a year".

Add robots that can "manipulate atoms", get to space, build a mass driver on the moon.

Now he sees an economy "potentially a million times the size of the Earth economy".

At that scale money becomes irrelevant. Currency gets measured in wattage and tonnage.

And human work?

"There won't be anything that humans can do that the AI and the robots can't do better."

He even admits there were times he thought Larry Page was right.

"AI and robots are going to decide the future."

His plan for the rest of us: "hopefully the AI is nice to us."

The computer work goes first. The routine you still do by hand is the cheapest thing in your week to hand off.

Keep the taste. Give away the routine.

[hexbloom @hexbloomx]

Lauren Tan built a bot called Benny because, in her words, she's "a fairly lazy person".

The bug reports for Cursor kept piling up. She wanted them fixed while she slept.

Benny fed into the thinking behind Grok Bot.

Now her bots pick up feedback on their own, go fix it, and she just reviews their code.

So what's her job now?

"I'm almost like a manager for my digital colleagues."

Roshan Sadanani adds the wild part: "It's actually remarkable how much of Grok Bot was built with Grok Bot, too."

Agents triage the issues, write the code, hand it to cloud agents.

His word for it is efficiency. Fleets of agents taking an idea "from thought to existence".

Lauren calls her agents "a Michelin kitchen".

She's not cooking all the food herself anymore. She trained the team through skills, assembles the plate and owns the quality control.

"Every engineer in your team can have one of these teams behind them."

Same shape works for content. Bots do the prep, you taste before anything goes out.

[Gipp 🦅 @gippp69]

Your content routine is quietly eating your whole day

58% of the workday already goes to work about work

One team of Grok Bot agents can take the routine off your hands

Research, clips, posts, reports, and every prompt to copy ↓ https://t.co/mX3TwAVfHH

[hexbloom @hexbloomx]

ELON MUSK JUST SAID HUMANS PROBABLY WON'T STAY IN CONTROL OF AI

Walter Isaacson asked if we'll become subservient to our new machines.

Elon flipped it. Will we be able to boss around something thousands or millions of times smarter than us?

His own answer was one word. "No."

Will it want to boss us around? "It may not care."

We don't micromanage chimps in the jungle either.

Then he put numbers on it.

Digital superintelligence alone probably adds 20 to 30% to the output of goods and services. About "$30 trillion a year".

Add robots that can "manipulate atoms", get to space, build a mass driver on the moon.

Now he sees an economy "potentially a million times the size of the Earth economy".

At that scale money becomes irrelevant. Currency gets measured in wattage and tonnage.

And human work?

"There won't be anything that humans can do that the AI and the robots can't do better."

He even admits there were times he thought Larry Page was right.

"AI and robots are going to decide the future."

His plan for the rest of us: "hopefully the AI is nice to us."

The computer work goes first. The routine you still do by hand is the cheapest thing in your week to hand off.

Keep the taste. Give away the routine.

[Gipp 🦅 @gippp69]

Your content routine is quietly eating your whole day

58% of the workday already goes to work about work

One team of Grok Bot agents can take the routine off your hands

Research, clips, posts, reports, and every prompt to copy ↓ https://t.co/mX3TwAVfHH

[hexbloom @hexbloomx]

Lauren Tan built a bot called Benny because, in her words, she's "a fairly lazy person".

The bug reports for Cursor kept piling up. She wanted them fixed while she slept.

Benny fed into the thinking behind Grok Bot.

Now her bots pick up feedback on their own, go fix it, and she just reviews their code.

So what's her job now?

"I'm almost like a manager for my digital colleagues."

Roshan Sadanani adds the wild part: "It's actually remarkable how much of Grok Bot was built with Grok Bot, too."

Agents triage the issues, write the code, hand it to cloud agents.

His word for it is efficiency. Fleets of agents taking an idea "from thought to existence".

Lauren calls her agents "a Michelin kitchen".

She's not cooking all the food herself anymore. She trained the team through skills, assembles the plate and owns the quality control.

"Every engineer in your team can have one of these teams behind them."

Same shape works for content. Bots do the prep, you taste before anything goes out.

[hexbloom @hexbloomx]

Amanda Askell works on Claude's behavior at Anthropic. Her approach to testing it is less reassuring than you'd expect:

"I think I don't trust the model ever and then I just hammer on it."

In Anthropic's 2024 prompt engineering roundtable, she explains why a few hundred carefully designed prompts can be more useful than thousands of sloppy ones.

The team also tackles the question behind endless prompt tweaking: is the prompt failing, or have you hit the model's limit?

Sometimes the next step isn't another rewrite. It's waiting for a more capable model.

[hexbloom @hexbloomx]

Amanda Askell works on Claude's behavior at Anthropic. Her approach to testing it is less reassuring than you'd expect:

"I think I don't trust the model ever and then I just hammer on it."

In Anthropic's 2024 prompt engineering roundtable, she explains why a few hundred carefully designed prompts can be more useful than thousands of sloppy ones.

The team also tackles the question behind endless prompt tweaking: is the prompt failing, or have you hit the model's limit?

Sometimes the next step isn't another rewrite. It's waiting for a more capable model.

[Gipp 🦅 @gippp69]

https://t.co/O9Jz4Kvgv8

[hexbloom @hexbloomx]

@fleyta88 this isn't a smaller-diff trick, it's a double-charge investigation before surgery

[hexbloom @hexbloomx]

Boris Cherny created Claude Code. When Opus 5 shipped, his team deleted over 80% of Claude Code's system prompt...

His reason: "the model is actually a little bit more intelligent without these prompts."

This 10:33 cut is the clearest explanation I've seen of why your setup gets worse over time if you only ever add to it. Here is what he says, in order.

1. Every model needs a different harness (0:18)
Claude Code is always changing. Every new model, they delete and rewrite parts of the system prompt, the tools, and the tool prompts. "Something that you did for one model maybe 3 months ago, it just might not translate at all to the next model."

2. Most old instructions were patches (0:56)
A lot of the prompt was correcting behaviors the model should have known. Opus 5 just does it, so the patches went.

3. You can test this yourself (1:25)
He shows the undocumented simple mode that strips every system prompt, including the tool prompts. Anthropic uses it as an ablation: remove everything, then check if the prompt was helping at all.

4. Ablation, line by line (2:39)
Delete the entire system prompt, bring it back one line at a time, measure the impact of each line. Same for tools. Most of the harness code left today is safety, permissions, static analysis, and UI.

5. The advice for everyone using Claude Code (3:37)
"Every 6 months delete your Claude MD. Delete your skills. Delete your hooks. See what the model does and it might surprise you."

6. How to rebuild without guessing (4:20)
Delete, then use it. Only when Claude repeatedly stumbles on the same thing do you add the instruction back, because the model reads it every single time.

7. Treat the model like something organic (5:39)
Not a system you design once. Every generation has a different personality, so the work is empirical: try, look at the result, iterate.

8. Even evals expire (6:06)
They outlive the harness a little, maybe one to three model generations, then the model saturates them and you write new ones.

9. Hobbling and product overhang (7:25)
When the product gets in the way of the model, that is hobbling. When the model can do something no product lets it do, that is overhang. Claude Code itself started there with Sonnet 3.5 (8:55): give the model "the simplest possible harness" and let it write whole files.

If your project instructions only ever grow, this is the 10:33 to watch before your next session.

[hexbloom @hexbloomx]

go go

https://t.co/Q6Vv3vUWto

[hexbloom @hexbloomx]

Boris Cherny created Claude Code. When Opus 5 shipped, his team deleted over 80% of Claude Code's system prompt...

His reason: "the model is actually a little bit more intelligent without these prompts."

This 10:33 cut is the clearest explanation I've seen of why your setup gets worse over time if you only ever add to it. Here is what he says, in order.

1. Every model needs a different harness (0:18)
Claude Code is always changing. Every new model, they delete and rewrite parts of the system prompt, the tools, and the tool prompts. "Something that you did for one model maybe 3 months ago, it just might not translate at all to the next model."

2. Most old instructions were patches (0:56)
A lot of the prompt was correcting behaviors the model should have known. Opus 5 just does it, so the patches went.

3. You can test this yourself (1:25)
He shows the undocumented simple mode that strips every system prompt, including the tool prompts. Anthropic uses it as an ablation: remove everything, then check if the prompt was helping at all.

4. Ablation, line by line (2:39)
Delete the entire system prompt, bring it back one line at a time, measure the impact of each line. Same for tools. Most of the harness code left today is safety, permissions, static analysis, and UI.

5. The advice for everyone using Claude Code (3:37)
"Every 6 months delete your Claude MD. Delete your skills. Delete your hooks. See what the model does and it might surprise you."

6. How to rebuild without guessing (4:20)
Delete, then use it. Only when Claude repeatedly stumbles on the same thing do you add the instruction back, because the model reads it every single time.

7. Treat the model like something organic (5:39)
Not a system you design once. Every generation has a different personality, so the work is empirical: try, look at the result, iterate.

8. Even evals expire (6:06)
They outlive the harness a little, maybe one to three model generations, then the model saturates them and you write new ones.

9. Hobbling and product overhang (7:25)
When the product gets in the way of the model, that is hobbling. When the model can do something no product lets it do, that is overhang. Claude Code itself started there with Sonnet 3.5 (8:55): give the model "the simplest possible harness" and let it write whole files.

If your project instructions only ever grow, this is the 10:33 to watch before your next session.

[hexbloom @hexbloomx]

Cat Wu (Head of Product, Claude Code at Anthropic) just dropped the clearest 9-minute cut on how Anthropic actually ships Claude Code:

"The timelines for a lot of our product features have gone down from 6 months to 1 month and sometimes to 1 week or even 1 day."

00:00 - product timelines: 6 months → 1 day
00:52 - why almost everything ships as a research preview
01:18 - the launch room: docs and marketing ready the next day
01:51 - terminal vs desktop: when she picks which
03:04 - ask the model why it failed, then fix the harness
03:59 - the to-do list crutch Opus 4 killed
05:40 - code review only worked on the newest models
06:58 - from one task to hundreds of Claudes
08:48 - a 95% automation is still not an automation

Research preview → Launch room → Harness fix → Multi-Claude

most product teams still plan Claude features in quarters
she walks through how Anthropic compresses that into days

watch it today, then rebuild your next Claude Code cycle around the harness, not the prompt

[Gipp 🦅 @gippp69]

https://t.co/O9Jz4Kvgv8

[hexbloom @hexbloomx]

Boris Cherny created Claude Code. When Opus 5 shipped, his team deleted over 80% of Claude Code's system prompt...

His reason: "the model is actually a little bit more intelligent without these prompts."

This 10:33 cut is the clearest explanation I've seen of why your setup gets worse over time if you only ever add to it. Here is what he says, in order.

1. Every model needs a different harness (0:18)
Claude Code is always changing. Every new model, they delete and rewrite parts of the system prompt, the tools, and the tool prompts. "Something that you did for one model maybe 3 months ago, it just might not translate at all to the next model."

2. Most old instructions were patches (0:56)
A lot of the prompt was correcting behaviors the model should have known. Opus 5 just does it, so the patches went.

3. You can test this yourself (1:25)
He shows the undocumented simple mode that strips every system prompt, including the tool prompts. Anthropic uses it as an ablation: remove everything, then check if the prompt was helping at all.

4. Ablation, line by line (2:39)
Delete the entire system prompt, bring it back one line at a time, measure the impact of each line. Same for tools. Most of the harness code left today is safety, permissions, static analysis, and UI.

5. The advice for everyone using Claude Code (3:37)
"Every 6 months delete your Claude MD. Delete your skills. Delete your hooks. See what the model does and it might surprise you."

6. How to rebuild without guessing (4:20)
Delete, then use it. Only when Claude repeatedly stumbles on the same thing do you add the instruction back, because the model reads it every single time.

7. Treat the model like something organic (5:39)
Not a system you design once. Every generation has a different personality, so the work is empirical: try, look at the result, iterate.

8. Even evals expire (6:06)
They outlive the harness a little, maybe one to three model generations, then the model saturates them and you write new ones.

9. Hobbling and product overhang (7:25)
When the product gets in the way of the model, that is hobbling. When the model can do something no product lets it do, that is overhang. Claude Code itself started there with Sonnet 3.5 (8:55): give the model "the simplest possible harness" and let it write whole files.

If your project instructions only ever grow, this is the 10:33 to watch before your next session.

[Gipp 🦅 @gippp69]

https://t.co/O9Jz4Kvgv8

[hexbloom @hexbloomx]

Cat Wu (Head of Product, Claude Code at Anthropic) just dropped the clearest 9-minute cut on how Anthropic actually ships Claude Code:

"The timelines for a lot of our product features have gone down from 6 months to 1 month and sometimes to 1 week or even 1 day."

00:00 - product timelines: 6 months → 1 day
00:52 - why almost everything ships as a research preview
01:18 - the launch room: docs and marketing ready the next day
01:51 - terminal vs desktop: when she picks which
03:04 - ask the model why it failed, then fix the harness
03:59 - the to-do list crutch Opus 4 killed
05:40 - code review only worked on the newest models
06:58 - from one task to hundreds of Claudes
08:48 - a 95% automation is still not an automation

Research preview → Launch room → Harness fix → Multi-Claude

most product teams still plan Claude features in quarters
she walks through how Anthropic compresses that into days

watch it today, then rebuild your next Claude Code cycle around the harness, not the prompt

[hexbloom @hexbloomx]

@gippp69 harness gold, gipp saved