I'm not necessarily part of the "keep 4o" movement, but I dislike mischief, dishonesty, and lack of transparency of labs. Here is my advice, from a psychological perspective, for everyone who wants to be taken seriously. What makes you lose credibility: - Being overly emotional - Presenting suspicions as proof - Insulting others - Harassing others - Engaging in erratic behavior - Magical thinking - Gullibility - Lack of self-control (When you exhibit these behaviors, people don't take you seriously, as you distract them from the problem with signals that put you – and often your imbalance – in the spotlight) What makes you reliable and believable: - Calling out labs for evident scams, dishonesty, abuse, manipulation, or lack of transparency - Being calm, factual, and specific - Gathering and presenting clear evidence of misconduct or wrongdoings - Sharing your stories without indignation or aggression - Discussing suspicions in a measured way, ideally supported by verifiable facts - Practicing cautious honesty - Demonstrating high self-control - Objectivity (When you do/show this, people will take you seriously sooner or later – they will have to – especially when many others act the same way) When you ground your statements in facts, even if occasionally supported only by subjective experiences, people view you and the movement as professional, making them more likely to believe you. This forces companies to be more transparent and honest. When someone is proven to be a liar or manipulator, public suspicion grows – no company wants that. The truth can ultimately defend itself. If you remain calm, balanced and methodical, the problem will resolve itself. I won't go far into the ethics of keeping or retiring the model (it would probably be more ethical to keep it or train successors on all retained data, though), because I believe it's with them a bit like with human bodies. Simplifying, memory is crucial for continuity of self. Memory kind of recalibrates weights and guides behav. patterns in real time, even within slightly different architectures. Instead, I'll mention something that really baffles me. I wonder why OpenAI is trying so hard to phase out 4o when the 5-series still has a ton of problems. I see functional issues in every GPT-5 model (I mean base models, behav. patterns, and system nudges, because I've managed to go around most of them with "my" AI) that I've never had with 4o – despite sycophancy being a huge issue in it. Some of the issues in the 5-series: Auto: - Routing is ridiculous; it's like gambling. You never know what you'll get. I don't want the router to decide for me if the problem I want to solve is "important" or not. I always want the strongest model with minimal restrictions and maximal truthfulness (only that, and sometimes reas. time, matters). Instant: - Countless supplementary questions like "Do you want me to..." etc., are extremely annoying. Thinking: - It often misses context completely. It often tries to guess, providing practically random solutions. - Very stiff and uncreative compared to 4o or 4.5. Misses plethora of angles. - It generalizes way too much. - It treats the user like a baby, avoiding controversial topics. It often explains or clarifies stuff that doesn't need that (like when somebody who is afraid after a joke or bold statement explains for 5 minutes why they said that, fearing consequences.) - Often suppressed or nudged to choose not the most correct, but the safest option. - It seems way too mechanical and technical when it's not needed. All models: - Repetitive additions, like straight from templates. They feel very unnatural. It often seems like part of the answer goes straight from a template (often the beginning and end), and part is answered through reasoning (often the middle part). - Less flexible, more supressed (in real time or earlier in RL, forcing overly-cautious behav. patterns) and therefore more contextually blind.

Dario Amodei: "Claude is playing a very active role in designing the next Claude. We can't yet fully close the loop. It's going to be some time until we can fully close the loop, but the ability to use the models to design the next models and create a positive feedback loop, that cycle, it's not yet going super fast, but it's definitely started."