• AI Rundown
  • Posts
  • The AI Rundown by Lightscape Partners – 04/08/24

The AI Rundown by Lightscape Partners – 04/08/24

Apple released a new foundational model capable of understanding context, Princeton researchers developed a new SWE-agent, and Microsoft may be building the most expensive computer ever.

Author

Ted Wagner
April 08, 2024

Image generated by Ted Wagner using DALL-E 3

Good morning and welcome to this week’s AI Rundown.

Last week in AI,

  • Apple released a paper announcing a new model called ReALM, claiming it understands spoken and on-screen context better than any other model and outperforming GPT-4 with ReALM’s largest models.

  • Researchers at Princeton developed a framework to convert any language model into a software engineer capable of debugging real issues in a codebase.

  • Microsoft is rumored to build a $100 billion data center to house OpenAI’s next AI supercomputer.

Check out the AI conferences happening this month here.

If you’d like a quick refresher on industry terms and the current landscape of AI, skip to the bottom.

Top AI Stories of the Week

Apple reveals new model with context, outperforms GPT-4. Link.

  • The new model understands context, like unclear references in human speech and visual context on a screen, to increase interaction performance with Siri.

  • The study claims that ReALM-3B substantially outperforms GPT-4.

  • The paper was released ahead of WWDC 2024, further building hype for Apple’s annual event many believe will be centered around their upcoming AI strategy.

Researchers at Princeton NLP developed SWE-agent, an open-sourced project that converts GPT-4 into a software engineer. Link.

  • The framework converts language models like GPT-4 into software engineering agents that “can fix bugs and issues in real GitHub repositories.”

  • SWE-agent achieved a SOTA performance achievement on the SWE-bench benchmark after resolving 12.29% of issues.

Microsoft and OpenAI begin work on a $100 billion data center project called project “Stargate.” Link.

  • The project would have Microsoft spend $100 billion building a supercomputing cluster to support OpenAI’s future models.

  • Microsoft has already spent several hundred million dollars building the clusters to train GPT-4.

Hardware + Software

Researchers at Stanford announced Octopus v2, an advanced open-source LLM engineered to operate on Android devices. Link.

  • The new framework allows for AI agents to run more efficiently directly on devices.

  • The framework is currently only available for Android devices.

  • It’s 35x faster and more accurate than RAG and has <1s latency when running Llama-7B.

Stability AI announced Stable Audio 2.0. Link.

  • Stable Audio 2.0 is Stability’s product in AI-generated audio, able to produce high-quality, full tracks with coherent musical structure.

  • The tracks are up to 3 minutes in length.

  • The model is capable of audio-to-audio generation, allowing for users to upload samples.

  • The training data was licensed through the AudioSparx music library, ensuring fair compensation.

Google DeepMind researchers introduced Mixture-of-Depths, a method to optimize compute in transformer-based models. Link.

  • Transformer-based language models are the primary model architecture for almost all modern generative AI models.

  • The paper claims they can train transformers to “dynamically allocate FLOPS (or compute) to specific positions in a sequence, optimizing the allocation.”

  • The technique makes models up to 50% faster to step during post-training sampling.

Ethics

Anthropic discovered “many-shot jailbreaking,” a new technique to avoid the safety measures built in to LLMs. Link.

  • Many-shot jailbreaking evades the safety guardrails in LLMs and works on generally all available models.

  • The hack utilizes LLM advancements in growing context windows, the amount of information an LLM can process at once.

  • Anthropic published the research paper to shine light on the issue and ensure proper attention is given.

Venture

SiMa.ai announced a $70 million funding round with participation from Dell Technologies Capital. Link.

  • The startup is developing a software-focused edge AI solution via a purpose built chip and ML software.

  • The investment from Dell is the firm’s first in hardware, marking a shift in investment intent in the AI race.

Upcoming AI Conferences

  • San Jose, CA

  • April 16 – 17, 2024

Get up to speed on the current landscape

Image credit: SONYA HUANG, PAT GRADY, Sequoia Capital

Thank you for reading the AI Rundown by Lightscape Partners. Please send any questions, comments, or suggestions to [email protected].