Machine Learning Research
DeepSeek Ups the Open Weights Ante: DeepSeek-V3 redefines LLM performance and cost efficiency
A new model from Hangzhou upstart DeepSeek delivers outstanding performance and may change the equation for training costs.
Latest
Data Points
Report claims AI will create millions of net jobs: rStar-Math boosts small models’ math skills to o1’s level
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.
Data Points
Nvidia announces Cosmos world models at CES: Microsoft’s Phi-4 model now available on Hugging Face
AI careers remain just as hot as you might expect. Columbia’s GET model predicts gene expression. Cohere’s North brings easy and secure automaton to enterprise. Meta pauses older AI characters but will introduce new ones this year.
Letters
My AI-Assisted Software Development Stack: The software development stack is evolving fast. Here are some things to consider as you choose components.
Using AI-assisted coding to build software prototypes is an important way to quickly explore many ideas and invent new things.
The Batch Newsletter
When Good Models Do Bad Things, What Users Really Want, More Training Data!, Better Model Merging
The Batch AI News and Insights: Using AI-assisted coding to build software prototypes is an important way to quickly explore many ideas and invent new things.
Machine Learning Research
Better Performance From Merged Models: Localize-and-Stitch improves methods for merging and fine-tuning multiple models
Merging multiple fine-tuned models is a less expensive alternative to hosting multiple specialized models. But, while model merging can deliver higher average performance across several tasks, it often results in lower performance on specific tasks. New work addresses this issue.
Machine Learning Research
Massively More Training Text: Harvard unveils a million-book corpus for AI training
Harvard University amassed a huge new text corpus for training machine learning models.
Machine Learning Research
Models Can Use Tools in Deceptive Ways: Researchers expose AI models' deceptive behaviors
Large language models have been shown to be capable of lying when users unintentionally give them an incentive to do so. Further research shows that LLMs with access to tools can be incentivized to use them in deceptive ways.
Machine Learning Research
What LLM Users Want: Anthropic reveals how users interact with Claude 3.5
Anthropic analyzed 1 million anonymized conversations between users and Claude 3.5 Sonnet. The study found that most people used the model for software development and also revealed malfunctions and jailbreaks.
Data Points
ElevenLabs drops latency to 75 milliseconds: New Falcon3 helps push the field for smaller language models
Nvidia promises to open source Run:ai. SALT inverts distillation by having a smaller model train a larger one. SWE-Gym offers new way to fine-tune coding agents. Llama put to work to recommend books on Scribd.
Data Points
Texas legislation would aggressively regulate AI: OpenAI revisits a public benefit for-profit structure
SmallThinker builds a 3 billion parameter reasoning model. Alibaba cuts prices on its Qwen models. Google unveils the FACTS model benchmark. Smolagents orchestrates smaller open source agents.
Letters
New Opportunities for the New Year: AI-assisted coding lets you prototype applications quickly and easily. Go forth and build!
Despite having worked on AI since I was a teenager, I’m now more excited than ever about what we can do with it, especially in building AI applications.