ChatGPT multiplies two five-digit numbers and misses by 913,200
With no calculator and no scratch paper, the model reaches its answer the only way it can, by predicting the next digit, which comes out right for the small sums it has seen everywhere and confidently wrong for a large one it has barely seen at all.
When many can act alone, the most optimistic one decides
Bostrom, Douglas, and Sandberg proved the result in 2016 for actors who share one goal and act on private estimates of its value, the conditions a commercial model release does not meet.
DPO turns preference tuning into a single classification loss
Its simplicity won open models over from the RLHF pipeline, though the evidence that it matches PPO was small-scale and later work still puts PPO ahead on the hardest tasks.
Everyone on the video call was a deepfake except the Arup employee who wired HK$200 million
The colleagues he recognized on the video call were pre-recorded deepfakes, and the instruction to move the money came afterward by message.
OpenAI withheld GPT-4's BBH score, and a rival published one anyway
The 83.1% for GPT-4 in Anthropic's Claude 3 card came from Google's report, and the tasks behind it had already reached GPT-4's training data.