💡 What are the nerds up to?
➜ My friend Krystian Bergmann is doing a webinar tomorrow about the 10 biggest AI implementation mistakes that he’s seen among the many companies he’s done AI workshops with.
LinkedIn
➜ The first quantum-enhanced generative AI model to surpass classical models in drug discovery for cancer therapy – genAI models ran on classical and quantum hardware to generate one million drug candidates, leading to the synthesis and testing of 15 molecules.
Zapata AI
➜ If, like our team, you enjoy using Azure services and want to get more value from their ML offerings, here’s a 4-week ML roadmap created “to empower developers and data scientists alike to build, deploy, and manage high-quality models faster and with confidence”.
Microsoft
➜ Microsoft released the Python Risk Identification Tool for genAI (PyRIT) – a library developed by their AI Red Team “for researchers and engineers to help them assess the robustness of their LLM endpoints against different harm categories such as fabrication/ungrounded content (hallucination), misuse (bias), and prohibited content (harassment).”
Microsoft, GitHub
➜ The ChatGrid frontend gives you access to the Exascale Grid Optimization (ExaGO) model, which can simulate the USA’s power grid in real time, allowing grid planners to analyze the ripple effects of any disruptions.
GitHub
➜ Gemma is Google’s new family of open models designed for responsible AI development – they’re lightweight, inspired by the Gemini models, with 2B and 7B sizes, and they surpass larger models on key benchmarks while running on standard developer laptops or desktop computers.
Google
➜ Meta’s TestGen-LLM uses ‘Assured LLM-based Software Engineering’ (Assured LLMSE) with private, internal LLMs that are probably fine-tuned with Meta’s codebase to test and improve code – it uses an ensemble approach to generate code improvements, choosing from the outputs of multiple LLMs and different prompts, and it’s designed to improve existing human-written tests, and not generate them from scratch.
arXiv, explanation on Engineer’s Codex
➜ Managers that want to drive AI adoption should first use leading tools and models themselves – once they understand the capabilities, it’s time to ask these four questions:
–What useful thing you do is no longer valuable?
–What impossible thing can you do now?
–What can you move to a wider market or democratize?
–What can you move upmarket or personalize?
One Useful Thing
➜ The killer app of Gemini Pro 1.5 is video – it seems to do amazingly well at understanding video and solving problems based on video input, as evidenced by a few of the industry’s influencers.
Simon Willison, Mckay Wrigley on X(Twitter)
➜ When Andrej Karpathy posts a video, it’s always a must-watch – in this one he builds the GPT Tokenizer.
YouTube
➜ Groq built their own chip (LPU Inference Engine), compiler and software, systems, and cloud for AI inference – try GroqChat yourself and see if you notice a performance difference.
Groq
➜ The new image generation model from Stability is here, Stable Diffusion 3, boasting significant advancements in creating images from text descriptions – particularly with prompts involving multiple subjects, and improved spelling accuracy in text-to-image translations.
Stability
➜ New AI model can prevent plasma from escaping a nuclear fusion reactor’s magnetic field – it’s the first method to predict and prevent reactor instabilities before they occur rather than after the fact. It’s just a proof-of-concept, but it could be a huge step towards boundless clean fusion energy
Nature
➜ Mistral has released Mistral Large, supposedly a cheaper alternative to GPT-4 Turbo, and Le Chat, an alternative to ChatGPT which uses various models.
Mistral
➜ Phind is kinda like Perplexity but specifically for developers – it’s an “intelligent answer engine” for devs that uses genAI to help you solve challenging problems in minutes, also available as VS Code extension.
Phind
➜ GenAI image generation models can actually understand and map out the details of a scene, like lighting and shapes, all on their own.
GitHub