TLDR AI 2023-11-17

Ready to be part of the AI evolution? There’s a significant shortage of Machine Learning Engineers, with companies offering salaries upwards of $300,000 - $500,000.

Interview Kickstart’s ML SwitchUp Program is your bridge to an exhilarating career in ML. Here’s why:

• Curriculum designed and taught by FAANG Machine Learning engineers.
• For developers, data engineers, and software professionals with no background in AI/ML
• Individual coaching and 1:1 help
• Capstone project for hands-on experience
• Up to 15 mock interviews with FAANG+ ML engineers
• Results: Alumni who consistently bag > $300K job offers. Highest compensation received: $1.28 Million

Don’t get left behind. Make the switch today to an exciting ML career with Interview Kickstart.

Register for the free webinar to learn more

DeepMind has been working on music synthesis for a number of years and now just announced a powerful new system. Interestingly, much of the boost came from a data partnership with revenue sharing. Meaning it trained on artists’ music for a better performing model while ensuring that the artists were compensated. The model will be available in a number of forms, one of which is via YouTube Shorts Studio.

Oftentimes when you generate an output image with a generative model, it isn’t quite what you were looking for. However, editing that image with the same model is extremely challenging. Meta had a key insight that treating all generations as instructions allows editing capabilities to emerge. This, coupled with the new simplicity of the model architecture, is quite a nice step forward.

Microsoft launched Azure AI Speech text-to-speech avatar at the Microsoft Ignite 2023 event, allowing users to create photorealistic avatars that can speak scripted text in various languages using text-to-speech technology.

Researchers found that Large Language Models sometimes make errors in Video Question Answering (VideoQA) by depending too much on the language and ignoring the actual video content. To solve this, researchers introduced a new approach called Flipped-VQA, which makes these models better understand the relationship between videos, questions, and answers, leading to more accurate results.

Researchers have expanded the SCB-ST-Dataset4, which captures activities like hand-raising, reading, and writing to better understand and detect students' classroom behaviors using deep learning.

SentAlign is a new tool for aligning sentences in large parallel documents, capable of handling thousands to tens of thousands of sentences efficiently.

Most companies are still figuring out how to charge for AI and LLM tools. Packages? Credits? Tokens? Whichever model you choose, Orb makes it as easy as can be to implement — just choose your pricing model and billable metric and you’re done! Track consumption, prevent fraud, and align pricing to value (including GPU runtime). Try the free sandbox

We can speed up Segment Anything 8x by using Torch compile, Sparsity, customer kernels with Triton, and a number of other PyTorch performance features.

Building tools with language models is an emerging engineering discipline that spans high performance computing, GPU orchestration, and monitoring.

A collection of real-world AI/ML exploits for responsibly disclosed vulnerabilities.

ControlNet was a novel way to give fine-grained control to image synthetics models. Now there is a somewhat analogous model for music generation that lets you control a number of features like speech and pitch.

OpenAI's call for proposals on implementing democratic processes in AI decision-making appears restrictive and seems to prefer handling sensitive political issues without taking responsibility, potentially limiting the scope and effectiveness of democracy in AI governance.

The Copilot AI business model enhances existing workflows for efficiency without creating new markets or disrupting lower ends, but its true disruptive potential lies in reimagining workflows, a challenge that could unlock significantly larger market opportunities.

Let AI automate your meeting notes using GPT-4 Turbo.

Google DeepMind's AI Lyria model will use SynthID to watermark audio tracks generated using its technology, making them identifiable as AI-created without affecting the listening experience, even after compression or other modifications.

YouTube is implementing new guidelines requiring creators to disclose the use of AI in their videos, including both AI-altered and entirely synthetic content.