- GPT-4 lands and demonstrates a capabilities chasm between proprietary and next-best open source alternatives, while also validating the power of reinforcement learning from human feedback.
- It’s unclear how long human-generated data can sustain AI scaling trends (some estimate that data will be exhausted by LLMs by 2025) and what the effects of adding synthetic data are.Videos and data locked up in enterprises are likely up next.
- LLMs and diffusion models continue to offer gifts to the life science community by producing new breakthroughs for molecular biology and drug discovery.
-LLM和扩散模型通过在分子生物学和药物发现方面取得新突破,继续为生命科学界提供礼物
- Multimodality becomes the new frontier and excitement around agents of all flavors grows substantially.
-多模式成为新的前沿,围绕所有flAvors代理的兴奋大幅增长
Industry
行业
- NVIDIA rips into the $1T market cap club with voracious demand for its GPUs from nation states, startups, big tech and researchers alike.
- Export controls rate limit advanced chip sales to China, but major chip vendors create export control-proof alternatives.
-出口管制限制了对中国的高级芯片销售,但主要芯片供应商创造了防止出口管制的替代方案
- Led by ChatGPT, GenAI apps have a breakout year across image, video, coding, voice or CoPilots for everyone, driving $18B of VC and corporate investments.
- The world has divided into clear regulatory camps, but progress on global governance remains slower.
The largest AI labs are stepping in to fill the vacuum
.-世界已分成明确的监管阵营,但全球治理进展仍然缓慢,最大的人工智能实验室正在填补真空
- The chip wars continue unabated, with the US mobilising its allies, and the Chinese response remaining patchy.
-芯片战争继续有增无减,美国动员其盟友,而中国的回应仍然参差不齐
- AI is forecast to affect a series of sensitive areas, including elections and employment, but we’re yet to see a significant effect.
-人工智能预计将影响一系列敏感领域,包括选举和就业,但我们还没有看到显著的效应
Safety
安全问题
- The existential risk debate has reached the mainstream for the first time and intensified significantly.
-存在风险的辩论首次成为主流并强烈地表明
- Many high-performing models are easy to ‘jailbreak’.To remedy RLHF challenges, researchers are exploring alternatives, e.g. self-alignment and pretraining with human preferences.