June 2021: True to our mission of advancing human-centered AI research, we piloted a new Ethics and Society Review process for awarding grants to encourage researchers to consider the wider ramifications of their work in AI. #MondayMilestones 10/n
AI
-
Stanford Launches Embedded Ethics Initiative for Computer Science Students
By
–
October 2020: Together with @stanfordethics and @stanfordeng
, we launched the Embedded Ethics initiative to ensure that computer science students understand the importance of ethics in a technological context. #MondayMilestones 8/n https://
hai.stanford.edu/news/building-
ethical-computational-mindset
… -
Users Want AI Tools to Follow Instructions Reliably
By
–
Gotta admit I'd resent it I am a deeply curious person, but not "find an Etruscan scholar" curious for something like this I understand the limitations of these tools, and I'd be fine with the disclaimer, but I also like my computers to do the things I tell them to do!
-

Scaling Mamba: Addressing Training Instability with Layer Adjustments
By
–
We found that scaling Mamba to substantial sizes isn’t trivial and causes unstable training. We had to make some adjustments to the Mamba layers to stabilize them, as can be seen in the graph below (training loss before & after the change) 6/6
-

Jamba combines Transformer, Mamba, MoE for efficient scaling
By
–
Combining Transformer, Mamba & MoE allows flexibility in balancing low memory usage, high throughput, and high quality. Jamba’s KV cache – which becomes a limiting factor when scaling context in pure Transformers – is 8x smaller compared to a pure Transformer. 5/6
-

Jamba’s Impressive Long-Context Performance with Minimal Attention Layers
By
–
Jamba was trained to handle contexts of up to 256K. Jamba has excellent performance in the needle-in-a-haystack evaluation, which is especially interesting given its use of only 4 attention layers. It also outperforms Mixtral on most long-context benchmarks. 4/6
-

Mamba Models Struggle with In-Context Learning Compared to Attention
By
–
We noticed that pure Mamba models struggle to develop in-context learning capabilities. E.g., they performed substantially worse than the pure attention model in 3 common benchmarks while the attention–Mamba exhibits similar results to just Transformers. 3/6
-

Jamba Hybrid Model Outperforms Pure Attention and Mamba Architectures
By
–
We see that the hybrid Jamba model outperforms both pure attention and pure Mamba models. The ratio of attention-to-Mamba layers of either 1:3 or 1:7 performs comparably, but given that a 1:7 ratio is more compute-efficient, we opt for it in our model. 2/6
-

Jamba Whitepaper Released: Hybrid SSM-Transformer Architecture Details
By
–
Jamba whitepaper is out!
The whitepaper details our in-depth ablations on this novel hybrid SSM-Transformer architecture, and how we chose to interleave Mamba, Transformer and MoE. https://
arxiv.org/abs/2403.19887 Here are some highlights from the paper 1/6 -
OpenAI Releases AGI Later Today
By
–
You heard it here first folks, @OpenAI is releasing AGI later today!