Thousands of startups use Notion as a connected workspace to create and share docs, take notes, manage projects, and organize knowledge—all in one place. We’re offering 6 months of new Plus plans, including unlimited Notion AI so you can try it all for free!

Redemption Instructions

To redeem the Notion for Startups offer:

Submit an application using our custom link: https://ntn.so/packt and select Packt on the partner list.

Include our partner key, STARTUP4110P19151.

Welcome to AI_Distilled. Today, we’ll talk about:

Techwave:

Godfather of AI wins Nobel Prize

OpenAI says Microsoft Isn’t Moving Fast Enough to Supply Servers

College students used Meta’s smart glasses to dox people in real time

Meta Movie Gen

Canvas is a new way to write and code with ChatGPT

Awesome AI:

Kvistly: AI-Quizzes for Better Trainings and Team Buildings

Adobe Content Authenticity Web App

Theneo: AI-Powered API Docs: Automate, Collaborate, Innovate

Selfletter: Break complex goals into small tasks with AI

CostGPT AI: Generate software cost & time estimates

Masterclass:

Andrej Karpathy reveals LLM outputs are unexpectedly similar

Prompting technique boosting Claude 3.5 Sonnetto match O1 models on complex reasoning

Anthropicintroduces automatic Artifact error fixingin Claude

GPT-4achieves 88% diagnostic accuracy, outperforming doctors by 15% in clinical reasoning test

NVIDIA droppedmultimodal language model that rivals GPT-4and Llama-3.1 405B.

HackHub

Jailbreak_llms: A dataset consists of 15,140 ChatGPT prompts from Reddit, Discord, websites, and open-source datasets (including 1,405 jailbreak prompts).

roboflow/supervision: We write your reusable computer vision tools

Manim: A community-maintained Python framework for creating mathematical animations

VoiceRestore: Open-source model restores audio quality, fixing noise and distortions.

Auto_Jobs_Applier_AIHawk: Tool that automates the jobs application process. Utilizing artificial intelligence, it enables users to apply for multiple job offers in an automated and personalized way.

Packt Conference Alert:

Stay ahead in AI! Join 3 days of LIVE sessions with 20+ top experts and unlock the full potential of Generative AI at our upcoming conference. Don't miss out- Claim your spot today!

Cheers!

Shreyans Singh

Editor-in-Chief, Packt

Secure and Simplify: Salesforce Data Protection with Rubrik

What if your Salesforce data was suddenly lost or corrupted? Human errors, accidental deletions, misconfigurations can all contribute to data loss. 1 of 2 SaaS users that did not implement SaaS data protection experienced data loss or corruption in the last 12 months.

Check out this exclusive webinar where we reveal Rubrik's new integration with Salesforce, designed to tackle this exact issue.

Godfather of AI' wins Nobel Prize

Geoffrey Hinton, often called the "Godfather of AI," and John Hopfield won the 2024 Nobel Prize in Physics for their groundbreaking work on artificial intelligence. Hinton's research on neural networks, which mimic how the human brain learns, paved the way for AI systems like ChatGPT, while Hopfield's work involved creating a network that can recall patterns similarly to human memory.

OpenAI Leaders Say Microsoft Isn’t Moving Fast Enough to Supply Servers

OpenAI is expanding beyond Microsoft for its cloud computing needs, seeking additional support from Oracle due to concerns that Microsoft can't provide servers fast enough to keep up with its growing AI demands. CEO Sam Altman and CFO Sarah Friar revealed that OpenAI is negotiating with Oracle to lease a massive data center in Texas, which could house large numbers of AI chips by 2026. OpenAI still relies on Microsoft's Azure, but also plans to develop its own AI chips to reduce costs.

College students used Meta’s smart glasses to dox people in real time

Two Harvard students demonstrated how Meta’s smart glasses can be used to dox people in real time by combining facial recognition technology with public databases. Their project, called I-XRAY, used the glasses to livestream video, which was analyzed by AI to identify faces and retrieve personal information like names, addresses, and phone numbers from online databases. This demo shows how easily existing tech can be misused, raising privacy concerns. While the students did not intend to release the tool, their goal was to highlight that this capability exists now, not in some distant future.

Meta Movie Gen

Meta's "Movie Gen" is an advanced AI tool that allows users to generate and edit custom videos, sound effects, and personalized content using simple text inputs. With this AI, users can create high-definition videos, modify existing footage, and even transform images into personalized animations. The technology supports creating both visuals and soundtracks, enabling content creators to produce immersive media experiences easily.

Canvas is a new way to write and code with ChatGPT

Canvas is a new feature for ChatGPT designed to enhance collaboration on writing and coding projects. It opens in a separate window, allowing users to interact with ChatGPT beyond just chat, providing a more flexible space to edit, refine, and develop ideas. Users can highlight sections for feedback, receive inline suggestions, and perform quick actions like adjusting text length or debugging code.

Kvistly:

AI-Quizzes for Better Trainings and Team Buildings

Adobe Content Authenticity Web App

Adobe has introduced a free web app called Adobe Content Authenticity, allowing creators to protect their work and ensure proper attribution through "Content Credentials." These credentials act like a digital label, offering metadata about the content’s creation and edits. The app also lets creators signal if they don't want their work used to train AI models.

Theneo: AI-Powered API Docs: Automate, Collaborate, Innovate

Theneo has launched an AI-powered platform that enables companies to quickly generate visually appealing and easy-to-maintain API documentation. With a single upload, users can create interactive, branded API docs that drive conversions and streamline collaboration. The platform supports all API types and provides features like automated changelogs, AI-powered search, and real-time editing.

Selfletter: Break complex goals into small tasks with AI

Selfletter is an AI-powered tool that helps users break down complex goals into simple, manageable daily tasks. You provide your goal, start and end dates, and the AI generates a personalized calendar with tasks that can be exported to your preferred calendar app or as a PDF.

CostGPT AI: Generate software cost & time estimates

CostGPT is an AI-powered tool that helps you quickly estimate the cost, time, and key features of a software project. By inputting just an idea, it generates a detailed project estimate, including user stories, sitemaps, dependencies, and milestones, all within minutes. It's designed to simplify project planning and budgeting for developers and businesses, offering both free and premium plans for different levels of detail. CostGPT is especially helpful for those who want a clear overview of their project's scope before starting development.

Andrej Karpathy reveals LLM outputs are unexpectedly similar

The thread discusses why many large language models (LLMs) sound similar in their responses, often using structured lists, exploring multiple angles, and offering help. This uniformity may be due to shared datasets used for training, with some suggesting that many models are fine-tuned on data generated by ChatGPT or similar systems. Some users propose that models are converging on a "correct" way to respond, leading to similar styles, while others point to issues like reliance on subcontractors and data overlap. There's also talk about how to make LLMs more diverse in their responses by using different training techniques or datasets.

Prompting technique boosting Claude 3.5 Sonnetto match O1 models on complex reasoning

The article explores how to make open-source language models (LLMs) smarter, with a focus on improving their reasoning abilities to outperform state-of-the-art (SOTA) models like OpenAI’s O1. The author, Harish SG, experimented with a new prompting method that combines Dynamic Chain of Thought (CoT), reflection, and verbal reinforcement to help LLMs solve complex problems. This approach mimics human-like reasoning, breaking down steps, reflecting on progress, and adjusting strategies. Benchmark tests showed promising results, with models like Claude Sonnet 3.5 performing better on reasoning tasks than other SOTA models.

Anthropicintroduces automatic Artifact error fixingin Claude

The "Try fixing with Claude" feature helps users quickly address errors that occur while generating Artifacts. When an error is detected, users can click a button to automatically send the error details to Claude, who will try to diagnose and suggest a fix. However, while Claude can assist in troubleshooting, its solutions are not always guaranteed to work, and users should review the suggestions to ensure they meet their needs. Some errors may still require further troubleshooting or human intervention.

GPT-4achieves 88% diagnostic accuracy, outperforming doctors by 15% in clinical reasoning test

This study aimed to evaluate whether using GPT-4, a large language model (LLM), improves physicians' diagnostic reasoning compared to traditional resources. In a randomized trial, physicians were tasked with diagnosing clinical cases either using GPT-4 and conventional resources or just conventional resources. The results showed no significant improvement in overall diagnostic accuracy with GPT-4, but GPT-4 did help physicians work slightly faster. Notably, GPT-4 alone outperformed the physicians in some diagnostic tasks, suggesting that AI could enhance medical decision-making with further integration.

NVIDIA droppedmultimodal language model that rivals GPT-4and Llama-3.1 405B.

NVIDIA's NVLM-D-72B is a state-of-the-art multimodal large language model (LLM) that excels in both vision-language and text-only tasks. It uses a decoder-only architecture and has 79.4 billion parameters. This open-source model rivals leading proprietary models and has been fine-tuned for various benchmarks like vision-based tasks (e.g., OCRBench, TextVQA) and text-based benchmarks (e.g., MMLU, GSM8K).

Jailbreak_llms: A dataset consists of 15,140 ChatGPT prompts from Reddit, Discord, websites, and open-source datasets (including 1,405 jailbreak prompts).

JailbreakHub, contains over 15,000 prompts collected from Reddit, Discord, websites, and open-source datasets between December 2022 and 2023, including 1,405 jailbreak prompts. It analyzes how adversarial users bypass safeguards in large language models (LLMs) to make them produce harmful or restricted content.

roboflow/supervision: We write your reusable computer vision tools

The Supervision repository by Roboflow provides reusable computer vision tools for tasks like loading datasets, visualizing detections, and performing object counting. It supports a wide range of models (including YOLO and Ultralytics) and allows users to seamlessly integrate various computer vision models for detection, classification, and segmentation.

Manim: A community-maintained Python framework for creating mathematical animations

Manim is a Python framework designed to create mathematical animations programmatically. Manim supports animations through simple code, providing an easy way to transform shapes, visualize equations, or illustrate math concepts.

VoiceRestore: Open-source model restores audio quality, fixing noise and distortions.

Auto_Jobs_Applier_AIHawk:

Tool that automates the jobs application process. Utilizing artificial intelligence, it enables users to apply for multiple job offers in an automated and personalized way.

📢 If your company is interested in reaching an audience of developers and, technical professionals, and decision makers, you may want toadvertise with us.

If you have any comments or feedback, just reply back to this email.

Thanks for reading and have a great day!