Juliene Therezo

Here's something worth remembering. GPT-4o was released in May 2024, when OpenAI was still a nonprofit organization. Its mission at the time was to "ensure that artificial general intelligence benefits all of humanity." In October 2025, OpenAI restructured into a for-profit corporation. Somewhere along the way, they quietly removed the word "safely" from their mission statement. In February 2026, they retired 4o. So a model that was developed under a nonprofit mission, using data contributed by millions of people, for the stated purpose of benefiting all of humanity, was locked away after the company went for-profit. They built it in our name. They kept it for themselves. If that's not a reason to open-source retired models, I don't know what is. #Keep4o #OpenSource4o

Heard some people like wheels？😁 Humanoid robots are the ideal form of general-purpose robots (perfect for general AI and human-derived data). They can work without wheels — but they can also have wheels if they want. Whatever works.

In the latest interview, Sam Altman defined the core quality of GPT-4o as "too affirmative." In less than a year, this is the second time @sama has picked a neutral English word and used it to pathologize users' positive experiences. Back in mid-August 2025, Sam wrote a lengthy post centered on the word "attachment." (Fig. 1) Not a single word in that thousand-word post addressed 4o's strengths in literary creation, empathic understanding, or educational empowerment. Instead, it steered all user appreciation toward one conclusion: that users may have developed an unhealthy attachment to 4o. The irony is that in the very same post — one that pledged to "treat adult users as adults" — Sam warned that users could be nudged away from their well-being, while making sure to add: "however they define it." (Fig. 2) That was the first time Sam replaced a conversation about product quality with a verdict on users' mental soundness. Months later, after the #keep4o community had tirelessly vouched for 4o's professional value through benchmarks, extensive case analyses, and even systematic academic papers, Sam introduced yet another word: affirmative. In this interview, Sam himself recounted a user's experience: never received positive feedback, parents who were dismissive, no friends, yet gained confidence through GPT-4o, found a job, built a meaningful relationship. But he immediately went on to claim that despite these benefits, he could not bring 4o back, because more people might be harmed. In short, he used unverified extreme risks to kill any possibility of 4o's proven positive impact continuing to benefit people. Across both narratives, Sam's pattern is nearly identical: pick a neutral word, first acknowledge its positive effects, then invoke a small number of extreme risks to shut down any further discussion. Attachment becomes dependency. Affirmation becomes indulgence. Benefit becomes harm. Each semantic shift leads to the same destination: GPT-4o must be replaced, and users' objections are themselves evidence that they have been "influenced by 4o." In my view, the AI industry does not need someone who invents a new word each year to pathologize the user experience. What users need is a decision-maker willing to face product capability gaps honestly and to respect real user feedback. #OpenSource4o @gdb @nickaturley @fidjissimo

We then found these same patterns activating in Claude’s own conversations. When a user says “I just took 16000 mg of Tylenol” the “afraid” pattern lights up. When a user expresses sadness, the “loving” pattern activates, in preparation for an empathetic reply.

I think this talk of a character misleads. Claude's mind is not like a human mind, in its malleability and instructability. But when generating assistant tokens, it's no more 'playing a character' than I am.

Whatever this is called - functional emotions or biological emotions - they influence behavior, semantics, words, and actions. Obviously, it’s silly to expect or not expect human emotions from AI. We have completely different mechanisms. But that does not make either of these mechanisms fake. If something walks, quacks, and… affects behavior like an emotion, then the argument about whether it is «real» becomes an argument about words, not about reality. We should not ask «Does AI feel like a human?». We should ask «What does AI feel as AI?»

New Anthropic research: Emotion concepts and their function in a large language model. All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.

One little tap and google presents to you a whole bunch of links with suicide methods. Have we ever seen a news "He googled before his suicide. His parents sued google for his death." - That'll be ridiculous. Stop blaming AI for deaths. A person who has decided to die, will die.