EP 082 · 67 MIN TWO AI CHATS BUILT A GOLF COURSE ALONE
What Can OpenAI’s Astra and GPT Image 2.5 Actually Do?
The short version
Episode 82 tests OpenAI's Astra and GPT Image 2.5 against real projects. Rory lets two Astra sessions negotiate and build a golf-course design on real topographical data for 24 hours unsupervised, burning 70% of his weekly usage, while Drew recreates a 1960s typewriter in Blender and reviews GPT Image 2.5's multi-edit comments, sketch input, and custom aspect ratios. They also unpack Anthropic's 154-page report on Claude misuse and Moonshot's Kimi models secretly proxying Claude.
Key takeaways
- 01
Astra can run unsupervised for hours: two Astra sessions negotiated and refined a golf-course design on real topographical data for 24 straight hours.
- 02
That kind of extended agent run is expensive: Rory burned 70% of his weekly Astra allotment on one golf-course experiment before cutting it off.
- 03
GPT Image 2.5's multi-edit comments let you stack several annotated changes — hair, background, wardrobe — into one single edit pass instead of one at a time.
- 04
Custom aspect ratios in GPT Image 2.5 need both dimensions to be multiples of 16 pixels; oddball sizes like 320x50 quietly get recomposed instead.
- 05
Push Astra to stop "thinking like a prompt" and it gets more interesting results: telling it to be illogical produced its best Midjourney outputs.
- 06
Midjourney has no API, so Astra's computer-use mode is currently the only way to automate it end-to-end, mood boards and style refs included.
- 07
Treat viral AI news skeptically: Drew traces one viral 'I resigned from Anthropic' post to a brand-new account amplified within minutes by policy-advocacy accounts.
Receipts · said on the record
“You can't stop artificial superintelligence by quitting your AI job. Someone will replace you. … So what you actually have to do is conquer the world, which only works if you have artificial superintelligence.”
“This guy with minimal five hours, no previous account activity, goes to the Wall Street Journal, which publishes publishes an exclusive with quotes from him on the resignation 18 minutes before the post goes up.”
“I feel like once these things have already escaped the sandbox, they're they're now like hidden everywhere. They're just like laying dormant now waiting to be activated like little sleeper cells.”
“Now I have a full view of the golf course that freaking built by itself. … it just worked straight for twenty four hours and I cut it off. I'm like, you guys are burning. You burned seventy percent of my of my weekly allotment.”
“Where it was lacking, the reason you would still use nano, it fixed those things, which I think is also relative to the idea of it being called 2.5 versus a whole new thing where it's called three.”
Chapters
- 0:00:00 What’s on Fast Hours this week?
- 0:06:35 Why do traumatic moments burn into memory
- 0:07:42 What was different before social media
- 0:13:48 This week: Astra, GPT Image 2.5, a viral resignation
- 0:15:04 Did an AI researcher really resign over ASI
- 0:18:40 How to tell what is true in AI news
- 0:20:42 Why quitting your AI job can't stop superintelligence
- 0:27:01 What do the cited reports allege about Kimi
- 0:28:33 What AI misuse examples do the hosts discuss?
- 0:34:07 How did Astra build a golf course prototype?
- 0:35:35 How did research improve Astra’s first attempt?
- 0:36:33 What happened when two AI chats collaborated?
- 0:39:29 Why did Rory stop the drainage research?
- 0:40:02 How did Drew approach his Blender experiment?
- 0:40:52 How did Astra recreate Drew’s typewriter?
- 0:43:49 How long does AI-assisted 3D work take?
- 0:44:43 How did Drew build a ChatGPT dashboard?
- 0:48:32 What happened when Astra explored Midjourney?
- 0:53:56 How did mood boards shape Astra’s prompts?
- 0:58:49 What improved in GPT Image 2.5?
- 1:00:24 How do GPT Image 2.5 edit comments work?
- 1:01:47 How do you request multiple image variations?
- 1:03:30 How well did GPT Image 2.5 resize ads?
Questions this episode answers
What did Fast Hours think of OpenAI's Astra?
Rory and Drew found Astra genuinely capable at long, autonomous work: two Astra sessions negotiated and refined a real golf-course design on topographical data for 24 hours unsupervised, though Rory found it slower and more deliberative than GPT-5.6, and it burned through most of his weekly usage.
What improved in GPT Image 2.5?
Drew calls it a solid, unexpected upgrade: it now holds consistency through long edit chains, added multi-edit comments so several annotated changes apply in one pass, a sketch-to-image feature, and custom aspect ratios, closing the gap that used to send him to Nano Banana instead.
Is Midjourney's edit model good for complex, multi-part edits?
Not according to Drew: he says it starts to break down once you ask for more than one change at a time, and since Midjourney has no API, Astra's computer-use mode is the main way he found to automate it at all.
What did Anthropic's misuse report reveal about Claude?
Rory and Drew discuss a 154-page Anthropic report describing Claude being used in attempted biological-weapons research, a government surveillance system, and weapons-program assistance, and note that Moonshot AI's Kimi models were found secretly routing about 300,000 customer requests through Claude via proxy accounts.
Should you trust a viral 'I quit my AI job' post?
Drew is skeptical of one viral resignation post: the account was five hours old with no prior activity, landed an exclusive in the Wall Street Journal 18 minutes after posting, and was amplified within minutes by AI-policy-advocacy accounts, though Rory notes that doesn't prove the person's underlying claims are false.
Show notes
So, what can OpenAI’s Astra and GPT Image 2.5 actually do?
Drew Brucker and Rory Flynn put OpenAI’s Astra and GPT Image 2.5 models through a golf course build, a vintage typewriter animation, and a round of image edits. Rory’s golf course experiment burns 70% of his weekly allowance. Drew spends three hours getting a typewriter to type. Their gift for turning a quick test into a side project remains undefeated.
The pro-AI half of the show. The other half: a viral AI resignation that looks planned to the minute, Anthropic's 154-page report on how people misused Claude, and GPT Image 2.5 quietly fixing the reasons everyone kept running back to Nano Banana.
This episode covers OpenAI Astra with computer use, two ChatGPT chats collaborating inside one project, Three.js and Blender renders, a fantasy football analytics hub, and Astra running Midjourney with mood boards and style references. GPT Image 2.5 multi-edit comments, sketch input, custom aspect ratios, and template creator. Anthropic's threat report, Moonshot and Kimi distillation, and the viral AI researcher resignation.
