“We're on the cusp of a great recalibration in tech. The winners will be those who can navigate the complex interplay of AI, open-source, and security while addressing the looming productivity paradox.”
Hive Five Newsletter

Today’s edition includes :

• Dioptra - A tool for testing AI model risk

• OpenAI steps into the AI voices industry

• DeepMind’s AlphaProof can get a silver medal at the IMO

• and more…

Read time: 5 min.

AI and Cybersecurity

Dioptra - A tool for testing AI model risk

Dioptra, an open-source tool developed by NIST, has been re-released to assess AI model vulnerabilities, particularly against data poisoning attacks.

This web-based platform allows various organizations to evaluate AI systems' performance under adverse conditions, providing a standardized approach to AI risk assessment.

While Dioptra offers significant potential for AI safety testing, it has limitations. It currently only works with models that can be downloaded and used locally, excluding popular API-gated models. Additionally, it cannot completely eliminate risks in AI models and faces broader challenges in AI benchmarking.

Dioptra's primary use cases include model testing, research, evaluations, and red-teaming.

Its key motivation is to address the growing challenge of evaluating and securing Machine Learning systems against diverse adversarial attacks, filling a crucial gap in standardized testing approaches.

The tool is built on a flexible microservices architecture, allowing for scalable deployment across multiple machines or on a single laptop. It utilizes a central API, data storage component, and a Redis queue for job management, all built using open-source resources for enhanced extensibility. Learn more about this with the documentation.

AI and Voices

OpenAI steps into the AI voices industry

OpenAI is rolling out ChatGPT's Advanced Voice Mode, starting with a small group of ChatGPT Plus users. The feature will be available to all Plus users by fall 2024. This new voice mode is powered by GPT-4o, a multimodal model capable of processing audio tasks without auxiliary models.

The initial demo of GPT-4o's voice sparked controversy due to its similarity to Scarlett Johansson's voice. OpenAI denied using her voice and later removed it from the demo. The company delayed the release to improve safety measures.

The current alpha release includes voice capabilities but not video or screensharing features. OpenAI claims GPT-4o can detect emotional intonations in users' voices, including sadness, excitement, and singing.

OpenAI has conducted extensive testing with external red teamers speaking 45 languages. The company is limiting Advanced Voice Mode to four preset voices and has implemented measures to prevent impersonation of real individuals or public figures.

To avoid potential copyright issues, OpenAI has introduced filters to block requests for generating music or copyrighted audio. This comes in response to recent legal challenges faced by AI companies over copyright infringement.

AI voices is a niche that is witnessing a tremendous boost with a great number of companies and tool in the market like Elevenlabs, Elai and OpenAI.

AI and Mathematics

DeepMind’s AlphaProof can get a silver medal at the IMO

DeepMind has developed two AI systems, AlphaProof and AlphaGeometry 2, which together solved four out of six problems from the 2024 International Mathematical Olympiad (IMO). This achievement is equivalent to a silver medal performance, marking a significant breakthrough in AI's mathematical reasoning capabilities.

AlphaProof uses reinforcement learning and a pre-trained language model to prove mathematical statements in the formal language Lean. It generates and verifies solution candidates, learning from each successful proof. AlphaGeometry 2 is an improved version of its predecessor, utilizing a Gemini-based language model and a faster symbolic engine to solve geometry problems.

These systems demonstrate advanced problem-solving skills in algebra, number theory, and geometry. They can tackle complex mathematical challenges, including the hardest problem in this year's IMO, which only five human contestants solved. This progress opens up new possibilities for AI-assisted mathematical research and problem-solving.

DeepMind is also exploring a natural language reasoning system based on Gemini for mathematical problem-solving. The company plans to release more technical details on AlphaProof and continues to investigate various AI approaches to advance mathematical reasoning capabilities.

“The fact that the program can come up with a non-obvious construction like this is very impressive, and well beyond what I thought was state of the art.”
Prof Sir Timothy Gowers

Amongst other things

• Mistral launches Mistral Large 2, making significant improvements in code generation, mathematics and reasoning while keeping the model’s size smaller than it’s competitors.

• MIT researchers turn to AI to understand how LLMs work.

• Researchers at the University of Cambridge found that feeding LLMs too much AI generated content can lead to “model collapse”.

• Canva has acquired Leonardo AI, the Australian Image generation startup.

Tools to make you cool

• Murf: AI enables real people’s voices

• OpenLit: Monitor LLMs and GenAI

• Speech to note: A must-have sppech to note tool

• Truffle Security: Open-source secret scanning engine

Super links

AI Generated Images

prompt: Alan Turing as a part of a lead hacker in a red team. He using colossus, the giant computer, at Bletchly park in the 1940's during world war 2

Thanks for reading.

Hamza from AI and Security.