Max Woolf
@minimaxir.bsky.social
Senior Data Scientist at BuzzFeed in San Francisco // AI content generation ethics and R&D // plotter of pretty charts
https://minimaxir.com
https://minimaxir.com
Pinned
Nano Banana can be prompt engineered for extremely nuanced AI image generation
Nano Banana allows 32,768 input tokens and I’m going to try to use them all dammit.
minimaxir.com
New blog post up: I spent a lot of time researching Nano Banana, Google's new generative AI model, and not only is it substantially better than ChatGPT, it is capable of taking extremely nuanced prompts even thousands of tokens long to generate exactly what you want. minimaxir.com/2025/11/nano...
Reposted by Max Woolf
Are we not even making our own disinformation? What happened to American manufacturing?
Twitter pays people based on engagement (views, retweets, comments, etc). It appears that many MAGA accounts are based abroad and they use AI technology to generate low-effort rage bait.
My guess is that this will get worse as AI tech improves. For instance, fake videos of minorities doing crime.
My guess is that this will get worse as AI tech improves. For instance, fake videos of minorities doing crime.
November 24, 2025 at 12:24 AM
Are we not even making our own disinformation? What happened to American manufacturing?
October+November 2025 Update on my Patreon www.patreon.com/posts/144253...
November 24, 2025 at 12:25 AM
October+November 2025 Update on my Patreon www.patreon.com/posts/144253...
shitposted on techcrunch blog posts and then i got an email from a hiring manager at Apple
November 22, 2025 at 10:06 PM
shitposted on techcrunch blog posts and then i got an email from a hiring manager at Apple
This entire image is a single Nano Banana Pro generation, with a relatively simple and unoptimized prompt.
November 22, 2025 at 5:09 PM
This entire image is a single Nano Banana Pro generation, with a relatively simple and unoptimized prompt.
There's more than enough new happening with Nano Banana Pro to justify a followup to my Nano Banana blog post. I'll try to get that out after the Thanksgiving holiday since a) I have a feeling it will be long and b) there are issues on Google's end with NBP I'll see if they fix.
November 21, 2025 at 6:03 PM
There's more than enough new happening with Nano Banana Pro to justify a followup to my Nano Banana blog post. I'll try to get that out after the Thanksgiving holiday since a) I have a feeling it will be long and b) there are issues on Google's end with NBP I'll see if they fix.
I...worked on the detailed Nano Banana prompt engineering analysis for months...and...Google just...Google released a new version.
November 20, 2025 at 4:18 PM
I...worked on the detailed Nano Banana prompt engineering analysis for months...and...Google just...Google released a new version.
I'm the SQL sicko who does ORDER BY created_at ASC instead of just ORDER BY created_at.
November 20, 2025 at 1:25 AM
I'm the SQL sicko who does ORDER BY created_at ASC instead of just ORDER BY created_at.
It's funny that Gemini 3 can do low and high thinking, but can't do medium thinking.
November 18, 2025 at 11:53 PM
It's funny that Gemini 3 can do low and high thinking, but can't do medium thinking.
I attempted to use Gemini 3 in Copilot and it had fun so I had to add "NO FUN ALLOWED" to the AGENTS[.]md
November 18, 2025 at 11:27 PM
I attempted to use Gemini 3 in Copilot and it had fun so I had to add "NO FUN ALLOWED" to the AGENTS[.]md
do not question prompt engineering
protip: Yes, it actually does help to talk to your AI like this. Not a joke
November 18, 2025 at 10:16 PM
do not question prompt engineering
Reposted by Max Woolf
Microsoft advertises Copilot Vision as a potential AI gaming aid, so we tested it out with some Balatro.
www.theverge.com/report/82244...
www.theverge.com/report/82244...
November 18, 2025 at 5:21 PM
Microsoft advertises Copilot Vision as a potential AI gaming aid, so we tested it out with some Balatro.
www.theverge.com/report/82244...
www.theverge.com/report/82244...
Reposted by Max Woolf
There’s a very surreal conversation I keep having on here where people seem unable to hold two ideas at the same time, and I’m not sure why. It’s simply true that the Big AI platforms like ChatGPT are:
1. Extremely bad for society, in many ways.
2. Very genuinely popular with lots of people.
1. Extremely bad for society, in many ways.
2. Very genuinely popular with lots of people.
November 18, 2025 at 4:59 PM
There’s a very surreal conversation I keep having on here where people seem unable to hold two ideas at the same time, and I’m not sure why. It’s simply true that the Big AI platforms like ChatGPT are:
1. Extremely bad for society, in many ways.
2. Very genuinely popular with lots of people.
1. Extremely bad for society, in many ways.
2. Very genuinely popular with lots of people.
Reposted by Max Woolf
Interesting changes from Grok 4 to Grok 4.1. Decreases in harmful responses but also increases in sycophancy and deception.
It isn’t clear how to interpret the sycophancy score, but the MASK score for deception is quite high compared to big models.
Sycophancy leads to higher LMArena scores…
It isn’t clear how to interpret the sycophancy score, but the MASK score for deception is quite high compared to big models.
Sycophancy leads to higher LMArena scores…
November 18, 2025 at 2:55 AM
Interesting changes from Grok 4 to Grok 4.1. Decreases in harmful responses but also increases in sycophancy and deception.
It isn’t clear how to interpret the sycophancy score, but the MASK score for deception is quite high compared to big models.
Sycophancy leads to higher LMArena scores…
It isn’t clear how to interpret the sycophancy score, but the MASK score for deception is quite high compared to big models.
Sycophancy leads to higher LMArena scores…
I can confirm Grok 4.1 has effectively no content filters: even on the web UI which should have its own safety prompts, it's *extremely* permissive and I suspect that the other safety filters in its model card can be defeated.
Also, wtf at those next prompt suggestions.
Also, wtf at those next prompt suggestions.
November 17, 2025 at 9:51 PM
I can confirm Grok 4.1 has effectively no content filters: even on the web UI which should have its own safety prompts, it's *extremely* permissive and I suspect that the other safety filters in its model card can be defeated.
Also, wtf at those next prompt suggestions.
Also, wtf at those next prompt suggestions.
Reposted by Max Woolf
speaking entirely for myself, but someone who is preeminently qualified to speak about LLMs (I have used them for over a decade) and how games are sold (13 years of doing that):
it costs more resources to convince people the benefits of your game using AI than the money you'd "save" using AI
it costs more resources to convince people the benefits of your game using AI than the money you'd "save" using AI
"Surely there's a middle ground here?" The CEO of Helldivers 2 studio Arrowhead says people "jump to extreme takes" when debating generative AI, and that there needs to be more nuance in the discussion.
www.videogameschronicle.com/news/surely-...
www.videogameschronicle.com/news/surely-...
‘Surely there’s a middle ground here?’ Helldivers studio CEO says people ‘jump to extreme takes’ when debating GenAI | VGC
“This is such a trigger point for big parts of the games industry today”…
www.videogameschronicle.com
November 17, 2025 at 3:49 PM
speaking entirely for myself, but someone who is preeminently qualified to speak about LLMs (I have used them for over a decade) and how games are sold (13 years of doing that):
it costs more resources to convince people the benefits of your game using AI than the money you'd "save" using AI
it costs more resources to convince people the benefits of your game using AI than the money you'd "save" using AI
testing LLM rhyming and the goddamn LLM hallucinated a proper noun just to maintain the rhyme scheme
November 16, 2025 at 3:29 AM
testing LLM rhyming and the goddamn LLM hallucinated a proper noun just to maintain the rhyme scheme
Reposted by Max Woolf
New blog post up: I spent a lot of time researching Nano Banana, Google's new generative AI model, and not only is it substantially better than ChatGPT, it is capable of taking extremely nuanced prompts even thousands of tokens long to generate exactly what you want. minimaxir.com/2025/11/nano...
Nano Banana can be prompt engineered for extremely nuanced AI image generation
Nano Banana allows 32,768 input tokens and I’m going to try to use them all dammit.
minimaxir.com
November 13, 2025 at 5:40 PM
New blog post up: I spent a lot of time researching Nano Banana, Google's new generative AI model, and not only is it substantially better than ChatGPT, it is capable of taking extremely nuanced prompts even thousands of tokens long to generate exactly what you want. minimaxir.com/2025/11/nano...
New blog post up: I spent a lot of time researching Nano Banana, Google's new generative AI model, and not only is it substantially better than ChatGPT, it is capable of taking extremely nuanced prompts even thousands of tokens long to generate exactly what you want. minimaxir.com/2025/11/nano...
Nano Banana can be prompt engineered for extremely nuanced AI image generation
Nano Banana allows 32,768 input tokens and I’m going to try to use them all dammit.
minimaxir.com
November 13, 2025 at 5:40 PM
New blog post up: I spent a lot of time researching Nano Banana, Google's new generative AI model, and not only is it substantially better than ChatGPT, it is capable of taking extremely nuanced prompts even thousands of tokens long to generate exactly what you want. minimaxir.com/2025/11/nano...
LLMs when you set temperature = 2.0
"skin as part of brain?"
This is that great intellect people talk about?
November 13, 2025 at 5:34 AM
LLMs when you set temperature = 2.0
November 13, 2025 at 12:41 AM
My 25 minute Nano Banana blog post will be up tomorrow, in which I highlight a funny failure state.
November 12, 2025 at 10:02 PM
My 25 minute Nano Banana blog post will be up tomorrow, in which I highlight a funny failure state.
OpenAI's announcement of GPT-5.1 immediately triggered the flame war detector on Hacker News, which is fun.
November 12, 2025 at 7:41 PM
OpenAI's announcement of GPT-5.1 immediately triggered the flame war detector on Hacker News, which is fun.
Reposted by Max Woolf
i am begging every skeptic to find some middle ground, any middle ground at all, between rejecting speculation as sci-fi and refusing to speculate even slightly about what happens next
A relatively small number of people in certain jobs say that ChatGPT and other LLMs have made them more productive at work. But in the overall economy, it does not look like net productivity is up.
Most of the supposed value is in sci-fi speculation. “Imagine a machine that cures cancer.”
Most of the supposed value is in sci-fi speculation. “Imagine a machine that cures cancer.”
I honestly don’t get the value of this company. They hoover up energy and water. Their product constantly gets things wrong and, in extreme cases, coaches people into suicide.
And it’s all built on what seems to be malicious and vast intellectual property theft.
What does OpenAI offer the world?
And it’s all built on what seems to be malicious and vast intellectual property theft.
What does OpenAI offer the world?
November 9, 2025 at 4:52 PM
i am begging every skeptic to find some middle ground, any middle ground at all, between rejecting speculation as sci-fi and refusing to speculate even slightly about what happens next
i kinda want to try "this is your last chance to dodge a block" as a prompt engineering technique.
November 7, 2025 at 4:19 PM
i kinda want to try "this is your last chance to dodge a block" as a prompt engineering technique.