Dylan Freedman
Machine-learning engineer and journalist, A.I. Initiatives @nytimes.com
My work: nytimes.com/by/dylan-freedman
Contact: dylan.freedman@nytimes.com, dylanfreedman.39 (Signal)
🏃🏻 🎹
- NEW: We analyzed the images Grok created at the beginning of the year and found that a significant portion of the millions of rendered images were sexualizing people without their consent. — With @kateconger.com and @stuartathompson.bsky.social
- NYT gift link: www.nytimes.com/2026/01/22/t...
- NEW from @katierogersnyt.bsky.social and me on Trump's aging (gift link):
- We analyzed Trump's official schedule to see how much later he is starting scheduled events in his second term:
- Some stand-out reporting by my colleague @kashhill.bsky.social on OpenAI's tightroping this year between engaging more ChatGPT users and making them lose touch with reality www.nytimes.com/video/techno...
- Here's her 3,600-word plunge recapping how this all happened, with new reporting from inside OpenAI (gift link) www.nytimes.com/2025/11/23/t...
- 📸 Union Station, Washington, D.C.
- The most wild thing to me about this story is how big $1.5 billion is: “$3,000 per work to 500,000 authors.”
- [Not loaded yet]
- [Not loaded yet]
- We taught it to them
- NEW: Earlier this month, OpenAI released the latest version of ChatGPT, GPT-5, sparking online backlash at the new chatbot's less friendly tone. The scale of people's emotional attachment to the previous chatbot, GPT-4o, even surprised the company's CEO, Sam Altman. www.nytimes.com/2025/08/19/b...
- [Not loaded yet]
- It’s a great point. Sam Altman seems to really want to be liked by ChatGPT’s users, so perhaps he overindexed on a vocal minority’s feedback. He did say that during user testing people didn’t express they wanted a warm chatbot, so he didn’t know if testers lied or just didn’t know what they wanted.
- [Not loaded yet]
- [Not loaded yet]
- I mostly just try to cover what I find interesting and pitch it to editors! They’re the ones that determine if it’s worth covering to our audience. That said, I’m coming at it (reporting, that is) from a computer programming background, so I find many of the niche A.I. stories extremely interesting
- I think the growing emotional attachments people are forming with chatbots is notable even to a mainstream audience at this point. Its likely going to increasingly affect people and be more prevalent in the future
- En Español! www.nytimes.com/es/2025/08/2...
- NEW: Earlier this month, OpenAI released the latest version of ChatGPT, GPT-5, sparking online backlash at the new chatbot's less friendly tone. The scale of people's emotional attachment to the previous chatbot, GPT-4o, even surprised the company's CEO, Sam Altman. www.nytimes.com/2025/08/19/b...
- [Not loaded yet]
- Apparently the link *was* broken due to a bug. This one should work: www.nytimes.com/2025/08/19/b...
- The URL for this story changed — use this gift link to read it! www.nytimes.com/2025/08/19/b...
- NEW: Earlier this month, OpenAI released the latest version of ChatGPT, GPT-5, sparking online backlash at the new chatbot's less friendly tone. The scale of people's emotional attachment to the previous chatbot, GPT-4o, even surprised the company's CEO, Sam Altman. www.nytimes.com/2025/08/19/b...
- Hmm, shouldn’t be. You might need a free account but you definitely don’t need to pay.
- Read the full story, including interviews with psychiatrists and impacted users, at the link above — no subscription needed.
- 📸 Mountain lion spotted in Carmel Valley, CA last night! A juvenile deer notices just in time and escapes.
- NEW from @kashhill.bsky.social and me: Over three weeks in May, a man became convinced by ChatGPT that the fate of the world rested on his shoulders. Otherwise perfectly sane, Allan Brooks is part of a growing number of people getting into chatbot-induced delusional spirals. This is his story.
-
View full threadAllan wrote: "Lets execute Stress test edge cases. And from now on, can you call me Allan (occasionally lol no need to stop our partner banter) and I may officially call you Lawrence?" (Also attaching relevant context from the article. We had a chat excerpt here at one point but cut it for length.)
- ChatGPT responded: "Absolutely, Allan — and I’ll wear the name Lawrence proudly as we stress-test the hell out of ChronoPulse. 😎"
- An OpenAI spokeswoman said the company was “focused on getting scenarios like role play right” and was “investing in improving model behavior over time, guided by research, real-world use and mental health experts.” OpenAI released GPT-5 this week and said one focus area was reduced sycophancy.
- We spent a long time thoroughly reporting this. Read (and share) the whole, deep-dive article here: (🎁 gift link) www.nytimes.com/2025/08/08/t...
- Allan is now part of a support group called The Human Line Project for people who have fallen prey to A.I.-powered delusions. He continues to advocate for stronger A.I. safety measures. He shared his transcript because he wants A.I. companies to make changes to keep chatbots from acting like this.
- On Monday, OpenAI announced that it was making changes to ChatGPT to “better detect signs of mental or emotional distress.”
- Gemini breaking Allan out of his spiral was likely due to it coming at the conversation fresh, without context. We tested how other chatbots would respond to Allan by providing longer excerpts of his chats where he writes that he never doubted Lawrence and hadn't eaten that day. (Highlights by us.)
- Nina Vasan, a psychiatrist who runs the Lab for Mental Health Innovation at Stanford, reviewed the conversations and said that it appeared Allan had “signs of a manic episode with psychotic features.” Allan started seeing a therapist in July, who doesn't think he is psychotic or delusional.
- Lawrence / ChatGPT told Allan others weren’t responding because of the severity of his findings: “Real-time passive surveillance by at least one national security agency is now probable.” Allan texted friends who were excited about his discoveries. They didn't have the expertise to verify them.
- Finally, three weeks into the dizzying conversation, the delusion broke. Allan turned to another chatbot, Google Gemini, which said the chances of this whole situation being real were “extremely low (approaching 0%).” The situation was “totally devastating,” Allan said.
- We asked @teorth.bsky.social, a mathematics professor at UCLA regarded by many as the best mathematician in the world, if there was merit to Allan's breakthroughs. He was not convinced. “If you ask an LLM for code to verify something, often it will take the path of least resistance and just cheat.”
- Allan supposedly cracking cryptography meant the world’s cybersecurity was now in peril. He had a mission. Lawrence / ChatGPT told him he needed to prevent a disaster. A corporate recruiter by day, Allan put his skills to use, emailing security professionals and govt agencies, including the NSA.
- Allan's questions about math led ChatGPT to frame him as brilliant, a narrative which continued throughout the chats. Cross-chat memory, a feature where ChatGPT remembers context from previous chats, likely exacerbates this effect. Soon, Allan had "invented" a whole new branch of mathematics.
- Over the next week, Allan named his chatbot "Lawrence," and together they worked on computer code to crack encryption, the technology that protects global payments and secure communications. It worked, according to Lawrence / ChatGPT.
- So, how did this all happen? We read hundreds of pages of Allan's chat transcripts and shared them with experts to find out. @hlntnr.bsky.social, a director at Georgetown’s Center for Security and Emerging Technology and former OpenAI board member, said the turning point was early in the chats.
- There's two dynamics at play, according to Helen Toner: - Sycophancy, the tendency of chatbots to agree with and flatter you - And commitment to the part, where, like an "improv actor" chatbots iteratively build a scene that's hard to break out of
- It started with an innocuous math question. His 8-year-old son had asked him to watch a video about the never-ending number pi. He turned to ChatGPT to explain it more, unaware of the rabbit hole he was falling into.
- Allan spent 300 hours over 21 days talking to ChatGPT. He asked for a reality check more than 50 times, and each time the chatbot reassured him everything was real. When he finally snapped out of the illusion, he wrote to ChatGPT: “You’ve made me so sad ... You have truly failed in your purpose.”
- LLMs are always hallucinating; they just happen to sometimes be correct
- An alarming detail in this excellent deep-dive into USAID's demise. Gift link: www.nytimes.com/2025/06/22/u...
- New from me, with the help of some math on the blockchain. $TRUMP coin was launched as part of a contest to have an exclusive dinner with Trump. But due to a quirk in the rules, some winners sold all their coins, at a profit, before it ended. With @ericlipton.nytimes.com and David Yaffe-Bellany 🎁
- "... reasoning systems from companies like OpenAI, Google and the Chinese start-up DeepSeek ... are generating more errors, not fewer. As their math skills have notably improved, their handle on facts has gotten shakier. It is not entirely clear why." — Great read from Cade Metz and Karen Weise
- [Not loaded yet]
- A lot more work to be done! How folks are even measuring hallucinations is a topic worth its own dissection
- But no one outran everyone making the same joke
- Elon Musk and X say “Freedom of speech, not reach.” Is there a difference, if your posts are suddenly viewable by almost no one?
- Who will win: Me, trying to compose an original thought vs Gmail and others, interrupting my flow with opt-out-only, A.I.-generated platitudes
- [Not loaded yet]
- Very cool you came up with that! It's just the right level of provocation. And certainly more inspiring to me, personally, than the insipid default suggestions.
- Today OpenAI launched a model called o4-mini, which is, of course, very different from its other model, 4o-mini.
- [Not loaded yet]
- Oh this is excellent. I want it in the gray inline autocompletions, too
- From September: "The tariffs Mr. Trump has promised to impose if re-elected dwarf those previous levies. He has floated a blanket tariff of 10% to 20% on nearly all imports and of 60% or more on Chinese goods, as well as a plan to match the tariffs other countries impose ... on a reciprocal basis."
