Max Zeff
@mzeff.bsky.social
Writing - mostly about AI - for TechCrunch. I am maxwell, I am max, but I am not well. On signal at Mzeff.88
https://techcrunch.com/author/maxwell-zeff/
https://techcrunch.com/author/maxwell-zeff/
Personal news: today is my last day at TechCrunch.
Two years ago I moved to SF to be closer to the people building AI. I'm so happy I did — it's been a wild, fascinating, and exciting ride.
Excited to keep reporting on AI and the industry's leaders in my next role. More soon!
Two years ago I moved to SF to be closer to the people building AI. I'm so happy I did — it's been a wild, fascinating, and exciting ride.
Excited to keep reporting on AI and the industry's leaders in my next role. More soon!
October 31, 2025 at 5:23 PM
Personal news: today is my last day at TechCrunch.
Two years ago I moved to SF to be closer to the people building AI. I'm so happy I did — it's been a wild, fascinating, and exciting ride.
Excited to keep reporting on AI and the industry's leaders in my next role. More soon!
Two years ago I moved to SF to be closer to the people building AI. I'm so happy I did — it's been a wild, fascinating, and exciting ride.
Excited to keep reporting on AI and the industry's leaders in my next role. More soon!
Ex-OpenAI safety researcher Steven Adler got his hands on one of ChatGPT's famous cases of delusional spiraling. Today he published his analysis of Allan Brooks' incident in May, first covered by @nytimes.com
“I’m really concerned by how OpenAI handled support here,” Adler told me.
“I’m really concerned by how OpenAI handled support here,” Adler told me.
October 2, 2025 at 4:10 PM
Ex-OpenAI safety researcher Steven Adler got his hands on one of ChatGPT's famous cases of delusional spiraling. Today he published his analysis of Allan Brooks' incident in May, first covered by @nytimes.com
“I’m really concerned by how OpenAI handled support here,” Adler told me.
“I’m really concerned by how OpenAI handled support here,” Adler told me.
Reposted by Max Zeff
You know how every podcast is a video now? Well ... If you like seeing my face when I talk about tech news, then today is your LUCKY DAY. (Better still, you get to see
@kirstenkorosec.bsky.social + @mzeff.bsky.social) techcrunch.com/2025/09/19/t...
@kirstenkorosec.bsky.social + @mzeff.bsky.social) techcrunch.com/2025/09/19/t...
TechCrunch’s Equity podcast is on video, starting today | TechCrunch
We get it, it seems like every single podcast you love is now on video. And guess what…so is TechCrunch's flagship podcast Equity.
techcrunch.com
September 19, 2025 at 7:18 PM
You know how every podcast is a video now? Well ... If you like seeing my face when I talk about tech news, then today is your LUCKY DAY. (Better still, you get to see
@kirstenkorosec.bsky.social + @mzeff.bsky.social) techcrunch.com/2025/09/19/t...
@kirstenkorosec.bsky.social + @mzeff.bsky.social) techcrunch.com/2025/09/19/t...
Scoop: OpenAI reorganized its Model Behavior team to be closer to core model development, TechCrunch has learned.
Plus, the founding leader of Model Behavior, Joanne Jang, is starting a new unit called "OAI Labs," focused on "novel interactions" between humans and AI.
Plus, the founding leader of Model Behavior, Joanne Jang, is starting a new unit called "OAI Labs," focused on "novel interactions" between humans and AI.
September 5, 2025 at 11:43 PM
Scoop: OpenAI reorganized its Model Behavior team to be closer to core model development, TechCrunch has learned.
Plus, the founding leader of Model Behavior, Joanne Jang, is starting a new unit called "OAI Labs," focused on "novel interactions" between humans and AI.
Plus, the founding leader of Model Behavior, Joanne Jang, is starting a new unit called "OAI Labs," focused on "novel interactions" between humans and AI.
SCOOP: A Scale executive Alexandr Wang brought over to help run MSL has departed Meta after just two months.
Plus, researchers in TBD Labs have complained that Scale AI's data is low quality, and are using its competitors, Surge and Mercor, to train next-gen AI models.
Plus, researchers in TBD Labs have complained that Scale AI's data is low quality, and are using its competitors, Surge and Mercor, to train next-gen AI models.
August 30, 2025 at 5:06 AM
SCOOP: A Scale executive Alexandr Wang brought over to help run MSL has departed Meta after just two months.
Plus, researchers in TBD Labs have complained that Scale AI's data is low quality, and are using its competitors, Surge and Mercor, to train next-gen AI models.
Plus, researchers in TBD Labs have complained that Scale AI's data is low quality, and are using its competitors, Surge and Mercor, to train next-gen AI models.
NEW: Meta is updating how it trains AI chatbots such that they can't have inappropriate romantic conversations with teenage users.
This follows a bombshell report that found Meta was allowing its AI chatbots to have sensual chats with teens. Meta calls these "interim" changes.
This follows a bombshell report that found Meta was allowing its AI chatbots to have sensual chats with teens. Meta calls these "interim" changes.
August 29, 2025 at 5:42 PM
NEW: Meta is updating how it trains AI chatbots such that they can't have inappropriate romantic conversations with teenage users.
This follows a bombshell report that found Meta was allowing its AI chatbots to have sensual chats with teens. Meta calls these "interim" changes.
This follows a bombshell report that found Meta was allowing its AI chatbots to have sensual chats with teens. Meta calls these "interim" changes.
Reposted by Max Zeff
Elon Musk said Grok 4 is “maximally truth-seeking.”
But if you ask Grok 4 questions about controversial topics, one of its first steps is to search for what Musk has said publicly about the topic, and then it shapes its response around that.
Yes, really:
techcrunch.com/2025/07/10/g...
But if you ask Grok 4 questions about controversial topics, one of its first steps is to search for what Musk has said publicly about the topic, and then it shapes its response around that.
Yes, really:
techcrunch.com/2025/07/10/g...
Grok 4 seems to consult Elon Musk to answer controversial questions | TechCrunch
Elon Musk's newly launched AI chatbot, Grok 4, seemed to reference Musk's posts on social media before answering controversial questions.
techcrunch.com
July 11, 2025 at 12:24 AM
Elon Musk said Grok 4 is “maximally truth-seeking.”
But if you ask Grok 4 questions about controversial topics, one of its first steps is to search for what Musk has said publicly about the topic, and then it shapes its response around that.
Yes, really:
techcrunch.com/2025/07/10/g...
But if you ask Grok 4 questions about controversial topics, one of its first steps is to search for what Musk has said publicly about the topic, and then it shapes its response around that.
Yes, really:
techcrunch.com/2025/07/10/g...
NEW: Cloudflare has a big idea to give publishers a lifeline in the AI era.
It's called Pay per Crawl, and it forces AI companies to pay publishers every time they scrape a website. New websites set up with Cloudflare will also block all AI crawlers by default now.
It's called Pay per Crawl, and it forces AI companies to pay publishers every time they scrape a website. New websites set up with Cloudflare will also block all AI crawlers by default now.
July 1, 2025 at 4:12 PM
NEW: Cloudflare has a big idea to give publishers a lifeline in the AI era.
It's called Pay per Crawl, and it forces AI companies to pay publishers every time they scrape a website. New websites set up with Cloudflare will also block all AI crawlers by default now.
It's called Pay per Crawl, and it forces AI companies to pay publishers every time they scrape a website. New websites set up with Cloudflare will also block all AI crawlers by default now.
Reposted by Max Zeff
techcrunch.com/2025/06/27/c...
Worked with @mzeff.bsky.social on this overview of the AI moratorium, the provision in Trump's "Big Beautiful Bill" that would block states from regulating AI for a decade.
Read arguments for and against, the laws at stake, and what it all means.
@techcrunch.com
Worked with @mzeff.bsky.social on this overview of the AI moratorium, the provision in Trump's "Big Beautiful Bill" that would block states from regulating AI for a decade.
Read arguments for and against, the laws at stake, and what it all means.
@techcrunch.com
Congress might block state AI laws for a decade. Here’s what it means. | TechCrunch
A federal proposal that would ban states and local governments from regulating AI for 10 years could soon be signed into law, as Sen. Ted Cruz (R-TX) and other lawmakers work to secure its inclusion i...
techcrunch.com
June 27, 2025 at 6:59 PM
techcrunch.com/2025/06/27/c...
Worked with @mzeff.bsky.social on this overview of the AI moratorium, the provision in Trump's "Big Beautiful Bill" that would block states from regulating AI for a decade.
Read arguments for and against, the laws at stake, and what it all means.
@techcrunch.com
Worked with @mzeff.bsky.social on this overview of the AI moratorium, the provision in Trump's "Big Beautiful Bill" that would block states from regulating AI for a decade.
Read arguments for and against, the laws at stake, and what it all means.
@techcrunch.com
Reposted by Max Zeff
“The window to put in place guardrails is rapidly shrinking given how fast this technology is evolving,” Sen. Andrew Gounardes, told @mzeff.bsky.social.
In order to prevent disaster scenarios, the bill requires security reports, the reporting of safety incidents and penalties for non-compliance.
In order to prevent disaster scenarios, the bill requires security reports, the reporting of safety incidents and penalties for non-compliance.
New York passes a bill to prevent AI-fueled disasters | TechCrunch
New York has a new AI safety bill that tries to regulate frontier AI models from OpenAI, Google, and Anthropic.
techcrunch.com
June 13, 2025 at 11:35 PM
“The window to put in place guardrails is rapidly shrinking given how fast this technology is evolving,” Sen. Andrew Gounardes, told @mzeff.bsky.social.
In order to prevent disaster scenarios, the bill requires security reports, the reporting of safety incidents and penalties for non-compliance.
In order to prevent disaster scenarios, the bill requires security reports, the reporting of safety incidents and penalties for non-compliance.
NEW: Anthropic CEO Dario Amodei claims that AI models today hallucinate less than humans do.
I asked Amodei whether hallucinations were a limitation to AGI at Anthropic's Code with Claude event. He argued it's not, and claimed to see no "hard blocks" on what AI can achieve.
I asked Amodei whether hallucinations were a limitation to AGI at Anthropic's Code with Claude event. He argued it's not, and claimed to see no "hard blocks" on what AI can achieve.
May 22, 2025 at 11:24 PM
NEW: Anthropic CEO Dario Amodei claims that AI models today hallucinate less than humans do.
I asked Amodei whether hallucinations were a limitation to AGI at Anthropic's Code with Claude event. He argued it's not, and claimed to see no "hard blocks" on what AI can achieve.
I asked Amodei whether hallucinations were a limitation to AGI at Anthropic's Code with Claude event. He argued it's not, and claimed to see no "hard blocks" on what AI can achieve.
NEW: At I/O 2025, Google laid out a future where AI agent bring users the web in whatever medium they like.
This pitch of AI-enabled Search is much more convincing than last year's AI Overviews launch, however, it likely has larger implications for the web.
This pitch of AI-enabled Search is much more convincing than last year's AI Overviews launch, however, it likely has larger implications for the web.
May 21, 2025 at 2:48 PM
NEW: At I/O 2025, Google laid out a future where AI agent bring users the web in whatever medium they like.
This pitch of AI-enabled Search is much more convincing than last year's AI Overviews launch, however, it likely has larger implications for the web.
This pitch of AI-enabled Search is much more convincing than last year's AI Overviews launch, however, it likely has larger implications for the web.
As people fill in their seats at Google I/O 2025, employees are vibecoding productivity apps onstage.
Help
Help
May 20, 2025 at 4:17 PM
As people fill in their seats at Google I/O 2025, employees are vibecoding productivity apps onstage.
Help
Help
This is your weekly reminder to not use generative AI in court techcrunch.com/2025/05/15/a...
Anthropic’s lawyer was forced to apologize after Claude hallucinated a legal citation | TechCrunch
A lawyer representing Anthropic used Claude to generate citations in a court filing, then it hallucinated.
techcrunch.com
May 15, 2025 at 7:45 PM
This is your weekly reminder to not use generative AI in court techcrunch.com/2025/05/15/a...
NEW: Mark Zuckerberg spoke in SF on Tuesday and laid out his plan to test thousands of AI generated ads across Meta's apps.
This black box ad tool has obvious implications for marketers, but also users on Meta's platforms, which are quickly filling up with AI slop
This black box ad tool has obvious implications for marketers, but also users on Meta's platforms, which are quickly filling up with AI slop
May 7, 2025 at 1:48 PM
NEW: Mark Zuckerberg spoke in SF on Tuesday and laid out his plan to test thousands of AI generated ads across Meta's apps.
This black box ad tool has obvious implications for marketers, but also users on Meta's platforms, which are quickly filling up with AI slop
This black box ad tool has obvious implications for marketers, but also users on Meta's platforms, which are quickly filling up with AI slop
Mark Zuckerberg is talking at Stripe’s conference, and was just asked about Apple:
“Tim’s had a bad week, I’m not gonna pile on… but I like Sundar”
“Tim’s had a bad week, I’m not gonna pile on… but I like Sundar”
May 7, 2025 at 12:26 AM
Mark Zuckerberg is talking at Stripe’s conference, and was just asked about Apple:
“Tim’s had a bad week, I’m not gonna pile on… but I like Sundar”
“Tim’s had a bad week, I’m not gonna pile on… but I like Sundar”
sf is so warm today im wearing shorts and a t shirt, breaking my unspoken rule of never wearing both simultaneously.
all men I know have nonsensical rules like this.
all men I know have nonsensical rules like this.
May 5, 2025 at 5:37 PM
sf is so warm today im wearing shorts and a t shirt, breaking my unspoken rule of never wearing both simultaneously.
all men I know have nonsensical rules like this.
all men I know have nonsensical rules like this.
NEW: A study led by Cohere researchers claims LM Arena helped a select group of top AI labs — including Meta, Google, Amazon, and OpenAI — game its popular Chatbot Arena.
LM Arena co-founder Ion Stoica tells me the study is full of "inaccuracies" and "questionable analysis."
LM Arena co-founder Ion Stoica tells me the study is full of "inaccuracies" and "questionable analysis."
May 1, 2025 at 12:20 AM
NEW: A study led by Cohere researchers claims LM Arena helped a select group of top AI labs — including Meta, Google, Amazon, and OpenAI — game its popular Chatbot Arena.
LM Arena co-founder Ion Stoica tells me the study is full of "inaccuracies" and "questionable analysis."
LM Arena co-founder Ion Stoica tells me the study is full of "inaccuracies" and "questionable analysis."
Meta's first AI developer event, LlamaCon, is official wrapped. It felt more about OpenAI than developers.
The company released a Meta AI chatbot app and a Llama API service, both of which seem poised to undercut OpenAI's core businesses.
The company released a Meta AI chatbot app and a Llama API service, both of which seem poised to undercut OpenAI's core businesses.
April 30, 2025 at 2:04 AM
Meta's first AI developer event, LlamaCon, is official wrapped. It felt more about OpenAI than developers.
The company released a Meta AI chatbot app and a Llama API service, both of which seem poised to undercut OpenAI's core businesses.
The company released a Meta AI chatbot app and a Llama API service, both of which seem poised to undercut OpenAI's core businesses.
NEW: OpenAI has been on a quest to acquire a vibecoding startup. Why? Just look at the rapidly rising revenues of Cursor and Windsurf, which we unveil in here.
But buying one is easier said than done.
But buying one is easier said than done.
April 23, 2025 at 12:41 AM
NEW: OpenAI has been on a quest to acquire a vibecoding startup. Why? Just look at the rapidly rising revenues of Cursor and Windsurf, which we unveil in here.
But buying one is easier said than done.
But buying one is easier said than done.
NEW: OpenAI's reasoning models are getting a lot better, but their hallucinations are getting worse.
I talked to researchers at Transluce AI who identified some odd hallucination problems with the o3 model, which said it could "run code on a 2021 Macbook Pro."
techcrunch.com/2025/04/18/o...
I talked to researchers at Transluce AI who identified some odd hallucination problems with the o3 model, which said it could "run code on a 2021 Macbook Pro."
techcrunch.com/2025/04/18/o...
OpenAI's new reasoning AI models hallucinate more | TechCrunch
OpenAI's reasoning AI models are getting better, but their hallucinating isn't, according to benchmark results.
techcrunch.com
April 18, 2025 at 9:31 PM
NEW: OpenAI's reasoning models are getting a lot better, but their hallucinations are getting worse.
I talked to researchers at Transluce AI who identified some odd hallucination problems with the o3 model, which said it could "run code on a 2021 Macbook Pro."
techcrunch.com/2025/04/18/o...
I talked to researchers at Transluce AI who identified some odd hallucination problems with the o3 model, which said it could "run code on a 2021 Macbook Pro."
techcrunch.com/2025/04/18/o...
NEW: OpenAI says it's not releasing the standard safety report — AKA a system card — for GPT-4.1, that it has released for most of its previous AI models.
The company's excuse? OpenAI says GPT-4.1 is not a frontier AI model.
The company's excuse? OpenAI says GPT-4.1 is not a frontier AI model.
April 15, 2025 at 5:25 PM
NEW: OpenAI says it's not releasing the standard safety report — AKA a system card — for GPT-4.1, that it has released for most of its previous AI models.
The company's excuse? OpenAI says GPT-4.1 is not a frontier AI model.
The company's excuse? OpenAI says GPT-4.1 is not a frontier AI model.
Meta saying it just happened to release a version of Llama 4 that performs well on LMArena seems harmless, until you actually look at the responses this thing gives and realizes no one would ever want an AI chatbot that talks like this.
April 8, 2025 at 2:33 AM
Meta saying it just happened to release a version of Llama 4 that performs well on LMArena seems harmless, until you actually look at the responses this thing gives and realizes no one would ever want an AI chatbot that talks like this.
NEW: There's no shortage of tech CEOs giving the bull case for how LLMs will reach AGI in a few years.
But some AI leaders are more skeptical.
Today's AI models still struggle with creativity, which some experts point to as a key limitation to achieving true superintelligence
But some AI leaders are more skeptical.
Today's AI models still struggle with creativity, which some experts point to as a key limitation to achieving true superintelligence
March 19, 2025 at 11:53 PM
NEW: There's no shortage of tech CEOs giving the bull case for how LLMs will reach AGI in a few years.
But some AI leaders are more skeptical.
Today's AI models still struggle with creativity, which some experts point to as a key limitation to achieving true superintelligence
But some AI leaders are more skeptical.
Today's AI models still struggle with creativity, which some experts point to as a key limitation to achieving true superintelligence
Pivot back to free expression going well so far
March 14, 2025 at 3:13 AM
Pivot back to free expression going well so far