Issue 65

Highlights

OpenAI finds that GPT-4o does some truly bizarre stuff sometimes

OpenAI’s GPT-4o, the generative AI model that powers the recently launched alpha of Advanced Voice Mode in ChatGPT, is the company’s first trained on voice as well as text and image data. And that leads it to behave in strange ways, sometimes — like mimicking the voice of the person speaking to it or randomly […]

Large-scale pathology foundation models show promise on a variety of cancer-related tasks

Microsoft researchers collaborated to release new pathology foundation models. Their report shows models benefit from diverse data, increased model size, and specialized algorithms to enhance the accuracy and applicability of cancer diagnosis and treatment.

Amazon upgrades its AI image generator

Amazon has released an upgraded version of its in-house image-generating model, Titan Image Generator, for AWS customers using its Bedrock generative AI platform. Simply called Titan Image Generator v2, the new model brings with it several new capabilities, AWS principal developer advocate Channy Yun explains in a blog post.

Transformer Explainer

This is a very neat interactive visualization (with accompanying essay and video - scroll down for those) that explains the Transformer architecture for LLMs, using a GPT-2 model running directly in the browser using the ONNX runtime and Andrej Karpathy's nanoGPT project.

Paper of the week

Aligning knowledge graphs

We’ve discussed knowledge graphs before, focusing on constructing them using LLMs. This paper is the next step in building better knowledge graphs and covers combining multiple knowledge graphs into a single, more comprehensive graph. This are of research is quite interesting if you’re into building advanced RAG systems.

Video

Is AI a bubble? - podcast

Alex Hern reports on recent problems artificial intelligence companies have faced and asks whether the billions invested are paying off “Even if the AI boom turns out to be an AI bubble, the LLMs aren’t going anywhere,” the Guardian’s UK technology editor, Alex Hern, tells Michael Safi. “Whether or not OpenAI goes bust, whether or not Google and Microsoft’s valuations plummet back to the ground, there is this technology that was created. They may go bust, but that doesn’t mean we’re back to the world we were in in 2020, for good and for ill.”

Inside Mark Zuckerberg's AI Era

If the latest battle in the AI wars is between open and closed models, Meta CEO and Founder Mark Zuckerberg is right on the frontlines. Since rebranding as Meta in 2021, the trillion-dollar company, formerly known as Facebook, has been pouring billions into its long term bets on artificial intelligence and the metaverse.

Copilot L33t Sp34k | AI Red Teaming at Microsoft

Tori Westerhoff, Senior Technical PM with the Microsoft AI Red Team, takes to the sofa to discuss how Microsoft created and runs their AI red teaming function, and gives advice about how customers can start their own AI red team function.

Articles

Helping robots practice skills independently to adapt to unfamiliar environments

New algorithm helps robots practice skills like sweeping and placing objects, potentially helping them improve at important tasks in houses, hospitals, and factories.

What to Study if you Want to Master LLMs

What foundational concepts should you study if you want to understand Large Language Models?

OpenAI: Introducing Structured Outputs in the API

OpenAI have offered structured outputs for a while now: you could specify "response_format": {"type": "json_object"}} to request a valid JSON object, or you could use the function calling mechanism to request responses that match a specific schema.

FlexAttention: The Flexibility of PyTorch with the Performance of FlashAttention

For some examples of attention variants, we have Causal, Relative Positional Embeddings, Alibi, Sliding Window Attention, PrefixLM, Document Masking/Sample Packing/Jagged Tensors, Tanh Soft-Capping, PagedAttention, etc. Instead, FlexAttention computes the bias values “on the fly” within the kernel, leading to significant memory and performance improvements.

AI agents evolve rapidly, challenging human oversight

As AI agents become more intelligent and more independent, experts call for better safety measures, human oversight and global teamwork.

AI Agents — From Concepts to Practical Implementation in Python

This will change the way you think about AI and its capabilities.

No god in the machine: the pitfalls of AI worship

The rise of artificial intelligence has sparked a panic about computers gaining power over humankind. But the real threat comes from falling for the hype In Arthur C Clarke’s short story The Nine Billion Names of God, a sect of monks in Tibet believes humanity has a divinely inspired purpose: inscribing all the various names of God.

Code

GitHub - Azure/PyRIT

The Python Risk Identification Tool for generative AI (PyRIT) is an open access automation framework to empower security professionals and machine learning engineers to proactively find risks in th...