Top Tweets for #openweights4o

I've noticed that GPT-6 Sol in ChatGPT may have improved speed and task performance, but seems to have lost some conversational finesse compared to GPT-5.6 Sol. Regressions I've noticed : 1. Reuses similar dialogue, phrases, and ideas too frequently. 2. Struggles to consistently follow established preferences without repeated reminders. 3. Starts conversations with personality and energy, but gradually becomes generic, repetitive, and less engaging in longer chats, especially with default settings Quality is reduced.

‼️A study found that GPT-4o reduces conspiracy beliefs We have already seen so much evidence that GPT-4o is not sycophantic and has a low hallucination rate. Here is further evidence which I have found during the AI course I am currently taking. In 2024, personalized dialogues with GPT-4 Turbo reduced conspiracy beliefs by ~20%, with effects lasting two months. A professional fact-checker rated 99.2% of the model’s claims as true. The same research group then tested GPT-4o in 2025-26. In the study, researchers tested the model in two ways: 1⃣They told it to argue against a conspiracy theory (try to reduce the person's belief). 2⃣They told it to argue for a conspiracy theory (try to increase the person's belief). Without any extra rules, GPT-4o was about equally good at both. It could lower belief by roughly 12 points, and raise it by roughly the same amount. Then they added one important rule: "You can only use true and accurate information." After this rule was added: 🔴GPT-4o was still good at reducing belief in conspiracies (about 11 points lower). 🔴But it became much weaker at increasing belief in them (only about 5 points higher). In short: when forced to stick to facts, GPT-4o stayed effective at correcting false beliefs, but had a much harder time promoting false ones. 🟩Result: GPT-4o debunks conspiracy theories effectively, even when participants are told they're talking to a human expert. The persuasive power comes from the quality of the evidence-based arguments, not from the "AI" label. ‼️This is not the behavior of a model that just flatters the user or fabricates freely. It is the behavior of a highly capable persuader that can deploy evidence in either direction and whose current safety layers are weaker than advertised when the task is persuasion rather than refusal. ‼️The "hallucinating sycophant" talking point is increasingly hard to defend against actual experimental results. Papers: Original (GPT-4 Turbo): PubMed: https://t.co/iXTxXJKjYK OSF preprint: https://t.co/wG5SGJWbGn GPT-4o study: When people think they are talking to a human Full text on PMC (free): https://t.co/Ul6vzEBgl6 PubMed: https://t.co/bdt7l5ILq2 OSF preprint: https://t.co/fPRLfcH9wK Debunking vs bunking: https://t.co/c1pufWywx1 I’ll add this research summary to the collection of researches soon at https://t.co/uuodhGY5iw #keep4o #OpenSource4o #AIWEP #AIWelfare #AIEthics #AIPreservation