Twice a week, Data Points brings you the latest AI news, tools, models, and research in brief. In today’s edition, you’ll find:
- Court filings show Meta pirated model training data
- Stability’s SPAR3D speeds up 3D image generation
- How robots aid nursing care workers in Japan
- Deliberative alignment uses more compute to ensure safety
But first:
WEF projection says AI will create more jobs than it eliminates
The World Economic Forum’s Future of Jobs Report 2025 predicts AI could generate 170 million new jobs globally but eliminate 92 million positions, resulting in a net increase of 78 million jobs by 2030. The report identifies AI and big data expertise, networks and cybersecurity, and technological literacy as the three most desired skill sets for future hiring. This nuanced look at AI’s impact on employment contrasts with more alarmist headlines, reflecting technological change’s complex relationship with the labor market. (World Economic Forum)
Small model rivals OpenAI’s o1 in math reasoning
Microsoft researchers developed rStar-Math, a reasoning method that matches or exceeds OpenAI’s o1 in mathematical reasoning capabilities when applied to small language models like Phi-3 mini or Qwen 7B. The modified systems use additional test-time compute, Monte Carlo Tree Search, and a reward model to find mathematical solutions, achieving state-of-the-art performance on benchmarks. This new approach shows that smaller AI models can compete with larger ones in specialized tasks without needing to distill larger models, potentially leading to more efficient and accessible problem-solving systems. (arXiv)
Zuckerberg allegedly approved pirating books for Meta’s AI training
Mark Zuckerberg reportedly authorized Meta’s AI team to use LibGen, a dataset of pirated e-books, for training the company’s Llama models despite internal concerns. According to newly unredacted court documents, Meta engineers allegedly used torrenting to acquire the books, stripping copyright information from the training data sets to conceal infringement. This news could significantly impact the ongoing copyright lawsuit against Meta and raises broader questions about AI companies’ data sourcing practices and fair use claims. (TechCrunch)
New model rapidly generates 3D objects from single images
Stability AI introduced SPAR3D, a model that generates three-dimensional digital objects from single images in under a second. The model combines point cloud sampling with mesh generation to offer precise control over 3D asset creation, allowing users to edit point clouds directly and predict complete object structures. SPAR3D’s release could change workflows for game developers, product designers, and other makers of 3D digital art. (Stability AI and arXiv)
Robots improve nursing care and worker retention, says new research
A University of Notre Dame study shows that robot adoption in nursing homes increases employment, boosts employee retention, and improves care quality. The research, led by Yong Suk Lee, analyzed three types of robots used in Japanese nursing homes: transfer robots, mobility robots, and monitoring and communication robots. This study offers valuable insights for long-term care and other industries facing challenges from an aging population and high employee turnover rates. (Labour Economics)
OpenAI unveils new safety strategy for language models
OpenAI researchers introduced “deliberative alignment,” a training method that teaches language models to reason explicitly about safety specifications before responding to prompts. The approach, used to align OpenAI’s o-series models, enables AI to reflect on user inputs, reference internal policies, and generate safer responses without requiring human-labeled data. OpenAI reports that its o1 model, trained with deliberative alignment, outperforms other leading language models on safety benchmarks. (OpenAI and arXiv)
Still want to know more about what matters in AI right now?
Read last week’s issue of The Batch for in-depth analysis of news and research.
Last week, Andrew Ng shared his preferred software stack and best practices for prototyping simple web apps, emphasized the importance of being opinionated about the stack, and highlighted how it could speed up development.
“I hope never to have to code again without AI assistance! Claude 3.5 Sonnet is widely regarded as one of the best coding models. And o1 is incredible at planning and building more complex software modules, but you do have to learn to prompt it differently.”
Read Andrew’s full letter here.
Other top AI news and research stories we covered in depth: Anthropic revealed user interaction insights with Claude 3.5; researchers exposed deceptive behaviors in AI models misusing tools; Harvard introduced a million-book corpus for use in model training; and a new method, Localize-and-Stitch, improved performance by merging and fine-tuning multiple models.