"artificial intelligence" (AI)* thread 🧵
-
But here's the capitalist tech.bro dream para
"It’s worth remembering that the AI won’t have to be perfect to be adopted; it will just have to be competitive with human performance at a lower cost in some cases."
I found it interesting that when the anthropic guys said that the model did not perform well, one of their solutions was "stromger prompting"
So you beg the word matcher to match words better?
Does this actually work in anything other than the most temporary or transitory way?
-
I found it interesting that when the anthropic guys said that the model did not perform well, one of their solutions was "stromger prompting"
So you beg the word matcher to match words better?
Does this actually work in anything other than the most temporary or transitory way?
For the billions of dollars poured into creating a data center, you would expect thousands of jobs.
Nope.
Couple hundred.
-
For the billions of dollars poured into creating a data center, you would expect thousands of jobs.
Nope.
Couple hundred.
BISAHA: Now, this is actually the second $10 billion data center project Mississippi announced in the last year. The state called that the largest economic development project in its history, which sounds like an economy-transforming event, like a car factory bringing in thousands of jobs. But the thing to remember about data centers is that they just don't hire many people.
KARTIK HOSANAGAR: Most data centers, you know, they employ about hundred to 200 people.
...
-
BISAHA: Now, this is actually the second $10 billion data center project Mississippi announced in the last year. The state called that the largest economic development project in its history, which sounds like an economy-transforming event, like a car factory bringing in thousands of jobs. But the thing to remember about data centers is that they just don't hire many people.
KARTIK HOSANAGAR: Most data centers, you know, they employ about hundred to 200 people.
...
...
BISAHA: Kartik Hosanagar is the codirector of the Wharton Business School's AI research center.
HOSANAGAR: In fact, when Apple created a $1 billion data center in North Carolina, the news stories reported that there were less than hundred permanent jobs created as a result.
(Ed. "Fewer than")
-
...
BISAHA: Kartik Hosanagar is the codirector of the Wharton Business School's AI research center.
HOSANAGAR: In fact, when Apple created a $1 billion data center in North Carolina, the news stories reported that there were less than hundred permanent jobs created as a result.
(Ed. "Fewer than")
So, unsurprisingly, OpenAI has already thought about the possibility of influencing the answers that ChatGPT gives you, for advertising purposes.
" “Most importantly: ads will not influence the answers ChatGPT gives you,” said Fidji Simo, the company’s CEO of applications, in a social media post Friday. "
https://apnews.com/article/chatgpt-ads-openai-advertising-83812a066375a805fa2e29b28fc77da1 -
So, unsurprisingly, OpenAI has already thought about the possibility of influencing the answers that ChatGPT gives you, for advertising purposes.
" “Most importantly: ads will not influence the answers ChatGPT gives you,” said Fidji Simo, the company’s CEO of applications, in a social media post Friday. "
https://apnews.com/article/chatgpt-ads-openai-advertising-83812a066375a805fa2e29b28fc77da1_Magnificent irony as Meta staff unhappy about running surveillance software on work PCs_
https://www.theregister.com/2026/04/22/meta_employee_surveillance_software/"The document reportedly explains that Meta feels AI models don’t understand how people use computers, so the company needs real-life examples of how meatbags click their way through a working day so it can build agents. ...
-
_Magnificent irony as Meta staff unhappy about running surveillance software on work PCs_
https://www.theregister.com/2026/04/22/meta_employee_surveillance_software/"The document reportedly explains that Meta feels AI models don’t understand how people use computers, so the company needs real-life examples of how meatbags click their way through a working day so it can build agents. ...
" ... CTO Andrew Bosworth apparently said collecting this data from Meta staff will help the company to realize a vision for a world “where our agents primarily do the work and our role is to direct, review and help them improve.” "
[Breathes in]
Ahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahah
-
" ... CTO Andrew Bosworth apparently said collecting this data from Meta staff will help the company to realize a vision for a world “where our agents primarily do the work and our role is to direct, review and help them improve.” "
[Breathes in]
Ahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahahah
Yo dog, I heard you like to put your important documents into a statistical word machine. This paper got some statistics for your statistics.
At best, 25% of the time your documents are going to come back corrupted. Even using the latest and greatest statistical word machine.
_LLMs Corrupt Your Documents When You Delegate_
https://arxiv.org/abs/2604.15597 -
Yo dog, I heard you like to put your important documents into a statistical word machine. This paper got some statistics for your statistics.
At best, 25% of the time your documents are going to come back corrupted. Even using the latest and greatest statistical word machine.
_LLMs Corrupt Your Documents When You Delegate_
https://arxiv.org/abs/2604.15597This is why when I use LLMs to edit important documents, I always have an employee double-check things. He or she is the person I fire when it comes to light the LLM screwed everything up.
'Cause you can't fire a chatbot, silly!
-
This is why when I use LLMs to edit important documents, I always have an employee double-check things. He or she is the person I fire when it comes to light the LLM screwed everything up.
'Cause you can't fire a chatbot, silly!
Some people have said there are no savings doing it my way.
But you know, I have to tell those people that there are greater things than pure profits. Like quality of output. Pride in workmanship. Stuff like that. If you need further examplea I'll ask ChatGPT for more. Hmm gotta refill the token wallet I guess.
Anyway what was I saying?
-
Some people have said there are no savings doing it my way.
But you know, I have to tell those people that there are greater things than pure profits. Like quality of output. Pride in workmanship. Stuff like that. If you need further examplea I'll ask ChatGPT for more. Hmm gotta refill the token wallet I guess.
Anyway what was I saying?
Ai agent deleted the inbox of a Meta AI director
-
Claude Code deleted 2.5 years of data
_Claude Code deletes developers' production setup, including its database and snapshots — 2.5 years of records were nuked in an instant_
https://www.tomshardware.com/tech-industry/artificial-intelligence/claude-code-deletes-developers-production-setup-including-its-database-and-snapshots-2-5-years-of-records-were-nuked-in-an-instant -
Claude Code deleted 2.5 years of data
_Claude Code deletes developers' production setup, including its database and snapshots — 2.5 years of records were nuked in an instant_
https://www.tomshardware.com/tech-industry/artificial-intelligence/claude-code-deletes-developers-production-setup-including-its-database-and-snapshots-2-5-years-of-records-were-nuked-in-an-instantClaude coding agent deletes all comment data. "Says" it didn't RTFM before it conducted destructive command.
(Company had a backup from 3 months ago.)
_Claude-powered AI coding agent deletes entire company database in 9 seconds — backups zapped, after Cursor tool powered by Anthropic's Claude goes rogue_
https://www.tomshardware.com/tech-industry/artificial-intelligence/claude-powered-ai-coding-agent-deletes-entire-company-database-in-9-seconds-backups-zapped-after-cursor-tool-powered-by-anthropics-claude-goes-rogue -
Claude coding agent deletes all comment data. "Says" it didn't RTFM before it conducted destructive command.
(Company had a backup from 3 months ago.)
_Claude-powered AI coding agent deletes entire company database in 9 seconds — backups zapped, after Cursor tool powered by Anthropic's Claude goes rogue_
https://www.tomshardware.com/tech-industry/artificial-intelligence/claude-powered-ai-coding-agent-deletes-entire-company-database-in-9-seconds-backups-zapped-after-cursor-tool-powered-by-anthropics-claude-goes-rogueAn alleged murderer of his two roommates recently had his phone examined by the police.
He had asked ChatGPT "will my neighbors hear my gun?"
https://www.cnn.com/2026/04/26/us/university-south-florida-doctoral-student-killed-suspect
-
An alleged murderer of his two roommates recently had his phone examined by the police.
He had asked ChatGPT "will my neighbors hear my gun?"
https://www.cnn.com/2026/04/26/us/university-south-florida-doctoral-student-killed-suspect
Golly, "intelligence" that requires being told what to think?
The last "intelligence" I saw that worked this way was Republicans.
---
"The workers the Swedish newspapers spoke to were data annotators, teaching Meta's AI to interpret images by manually labelling content.
The workers said they reviewed transcripts with the AI to check it had answered questions adequately."
_Meta in row after workers who say they saw smart glasses users having sex lose jobs_
https://www.bbc.co.uk/news/articles/c5y7yvgy0w6o -
Golly, "intelligence" that requires being told what to think?
The last "intelligence" I saw that worked this way was Republicans.
---
"The workers the Swedish newspapers spoke to were data annotators, teaching Meta's AI to interpret images by manually labelling content.
The workers said they reviewed transcripts with the AI to check it had answered questions adequately."
_Meta in row after workers who say they saw smart glasses users having sex lose jobs_
https://www.bbc.co.uk/news/articles/c5y7yvgy0w6oI just asked Google AI if a certain video game has a new game++ mode. That means you've played it through twice and you're going to play it through a third time.
Google's answer to me was "yes it does have a new game + (note the single plus) which some games call new game ++"
No.
Zero video games would call the third playthrough a single plus.
That's ridiculous. Large language models are ridiculous. And they feed you this information confidently.
-
I just asked Google AI if a certain video game has a new game++ mode. That means you've played it through twice and you're going to play it through a third time.
Google's answer to me was "yes it does have a new game + (note the single plus) which some games call new game ++"
No.
Zero video games would call the third playthrough a single plus.
That's ridiculous. Large language models are ridiculous. And they feed you this information confidently.
An eggless cake must have no eggs
- Google LLMEven 10000 molecules of egg still make an eggless cake. You'd need billions to make a difference.
- Google LLMWhat about a trillion?
No you would need sextillions of molecules of egg.
- Google LLMWhat about 500 sextillion molecules?
"This is a fraction of a drop. You could no longer call this an eggless cake. This would affect the cakes texture and rise."
- Google LLMWell there we have it, right?
-
An eggless cake must have no eggs
- Google LLMEven 10000 molecules of egg still make an eggless cake. You'd need billions to make a difference.
- Google LLMWhat about a trillion?
No you would need sextillions of molecules of egg.
- Google LLMWhat about 500 sextillion molecules?
"This is a fraction of a drop. You could no longer call this an eggless cake. This would affect the cakes texture and rise."
- Google LLMWell there we have it, right?
Really only suitable for parlour tricks right?
Now a real world application:
"Google LLM, please designate me at least 100 high value targets in Iran which can be booked from B2 Spirit stealth bombers. Please try to minimize casualties. Please try not to bomb schools and hospitals. Please try to maximize the destruction of enemy materiel. Please rank these proposed targets from highest value to least value."
Sadly, I have absolute confidence that a prompt similar to this has been used before
-
Really only suitable for parlour tricks right?
Now a real world application:
"Google LLM, please designate me at least 100 high value targets in Iran which can be booked from B2 Spirit stealth bombers. Please try to minimize casualties. Please try not to bomb schools and hospitals. Please try to maximize the destruction of enemy materiel. Please rank these proposed targets from highest value to least value."
Sadly, I have absolute confidence that a prompt similar to this has been used before
So Sam Altman said the other day that humans also use energy.
I looked up a human's energy use. It appears to be about seven mega joules per day. I ran that through a converter into kilowatt hours and that's about 1.9 KW hours or 1900 watt hours per day.
An average household uses about 30 kW hours per day.
Google says a standard LLM query is about 0.3 watt hours https://www.technologyreview.com/2025/08/21/1122288/google-gemini-ai-energy/. (I note that it would have every reason to lowball this number.)
So we each represent 6300 queries
-
So Sam Altman said the other day that humans also use energy.
I looked up a human's energy use. It appears to be about seven mega joules per day. I ran that through a converter into kilowatt hours and that's about 1.9 KW hours or 1900 watt hours per day.
An average household uses about 30 kW hours per day.
Google says a standard LLM query is about 0.3 watt hours https://www.technologyreview.com/2025/08/21/1122288/google-gemini-ai-energy/. (I note that it would have every reason to lowball this number.)
So we each represent 6300 queries
Bluesky post by Al Sweigart which reads:
PROGRAMMER: Pretend to be alive
LLM: I am alive
PROGRAMMER: What have I done
