Good morning. It’s Friday, June 21st.

Did you know: On this day in 2003, the Wikimedia Foundation was founded?

In today’s email:

• Claude 3.5 Sonnet

• Ilya’s AI Safety Company

• 10 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

Anthropic Launches Claude 3.5, Potentially The Most Capable AI Model Yet

Anthropic introduces Claude 3.5 Sonnet, a demonstrably superior AI surpassing its predecessor, Claude 3 Opus in reasoning, coding, and content generation. Sonnet 3.5 operates at double the speed with no cost inflation, setting new benchmarks in AI performance. It excels in tasks requiring graduate-level reasoning, undergraduate-level knowledge, and complex coding assignments, achieving significant improvements over its predecessor in internal coding assessments.

Anthropic emphasizes Sonnet's improved grasp of nuance, humor, and intricate instructions, leading to more refined natural language outputs. Claude 3.5 Sonnet excels in image processing, transcribing from imperfect sources vital for retail, logistics, and finance.

Sonnet also introduces Artifacts, facilitating real-time editing of AI outputs like code or designs within chat threads. Artifact allows collaborative refinement of prototypes and workflows, evolving Claude from chat tool to design/development ally. In web preview, Artifacts aim to enhance AI integration and boost productivity in creative industries.

Available through the Anthropic API, Claude.ai, and major cloud platforms, it supports a broad range of applications with pricing at $3 per million input tokens and $15 per million output tokens, alongside expanded user interaction features. Read more.

Related image: Benchmark scores of Claude, ChatGPT and Gemini over time 

Former OpenAI Chief Scientist Ilya Sutskever Launches Company for Safe AI

Ilya Sutskever, the brain behind OpenAI and its former chief scientist, has embarked on a new venture, Safe Superintelligence Inc. (SSI), alongside investor Daniel Gross and ex-OpenAI engineer Daniel Levy. Headquartered strategically in Palo Alto and Tel Aviv, SSI is on a mission to tackle what they dub the 'most pressing technical conundrum of our era'—creating safe superintelligent AI.

SSI’s goal is to advance AI capabilities while ensuring safety remains paramount, by recruiting a select team of top engineers and researchers. Sutskever's departure from OpenAI in May 2024 followed disagreements with CEO Sam Altman regarding the rapid commercialization and associated safety risks of AI. SSI's mission, as articulated in their announcement, is singularly focused on creating reliable and safe superintelligent AI, aligning their team, investors, and business model towards this objective. Read more.

Microsoft drops Florence-2, a unified model to handle a variety of vision tasks: Microsoft's Azure AI division released Florence-2, a versatile foundation model for vision tasks. Available on Hugging Face under an MIT license, Florence-2 tackles a wide range of tasks through a unified prompt-based approach. Offered in two sizes (232M and 771M parameters), it excels in captioning, object detection, visual grounding, and segmentation, even surpassing some larger models.

Luma AI's Dream Machine can now generate over a minute of AI video: Luma AI has introduced a significant update to its Dream Machine, an AI video generator, enabling the creation of videos over a minute long. Initially, video length was limited to five seconds, but the new Extend Video feature allows for extending clips up to 1:20 based on prompts, considering the context for seamless continuation. This tool, available to Standard, Pro, and Premier users, now supports watermark removal and promises forthcoming editing features with intuitive controls.

OpenAI upgrades DALL-E 3 instead of rolling out GPT-4o's (much better) imaging capabilities: OpenAI's DALL-E 3 gets a tune-up, despite the introduction of the more advanced multimodal GPT-4o. Pioneering the field with DALL-E 2 in 2022, OpenAI faced competition from Midjourney and Adobe Firefly with DALL-E 3, which initially struggled with photorealism. However, OpenAI has since upgraded DALL-E 3, particularly its ability to accurately render text within generated images.

Hugging Face CEO sees a surge in AI startup founders looking to sell: According to Clément Delangue, CEO of Hugging Face, there has been a notable increase in AI startup founders looking to sell their companies. Delangue reports receiving inquiries from about ten founders weekly interested in acquisition opportunities. Hugging Face, valued at $4.5 billion, recently acquired Argilla for $10 million. This marks the fourth acquisition for the New York-based startup, which raised $235 million last year.

Futureverse Launches Jen, An AI Music Model Focused on 'Transparency' Futureverse has launched Jen, an AI music model developed by Shara Senderoff and Mike Caren, trained on 40 licensed music catalogs. Jen’s alpha release allows users to generate 10-45 second song snippets from text prompts, with a “continuation” feature to extend tracks up to 3:30. It utilizes latent diffusion, similar to technologies in Stable Diffusion and DALL-E 2, and provides transparency on its workings through research papers.

New medical LLM, PathChat 2, can talk to pathologists about tumors, offer diagnoses: Mahmood Lab's PathChat 2, an advanced large language model (LLM), demonstrates exceptional performance in pathology tasks. Combining vision processing and pre-trained language capabilities, PathChat 2 surpasses benchmarks set by previous models (LLaVA, GPT-4V) in image analysis (78% accuracy standalone, 89.5% with clinical context). PathChat handles various diagnostic tasks such as differential diagnosis and tumor grading, typically requiring extensive labeled data. It supports pathologists interactively by generating clinically relevant responses and facilitating human-in-the-loop diagnostics.

Etcetera: Stories you may have missed

AI Decodes Emotions in Tennis Players

France leads the pack for Generative AI funding in Europe  

AI boom makes European universities the new Harvard and Stanford for tech talent  

'Marathon not a sprint': Apple's AI push faces big challenges in China

GrayMatter scores $45M for robots that speed-up manufacturing with 'physics-informed AI' 

Elon Musk claims Tesla's new AI supercluster will grow to over 500 MW, record AI chip

AI chipmaker Cerebras Systems, a competitor to Nvidia, reportedly files for an IPO

Sierra's new benchmark reveals how well AI agents perform at real work

DataSquirrel.ai provides non-tech managers with fast, guided data analytics and BI, featuring auto-cleaning, insights, visualizations, and dashboard reports.

AI Logo Reveals animates logos with AI-powered tools directly in your browser. Upload logos, select presets, and export stunning videos featuring effects like smoke and lightning.

Plansom is an AI SaaS platform that generates optimized, prioritized plans for organizations and individuals, simplifying complex planning tasks instantly.

AI Content Mate generates step-by-step user journey screens and contextual text directly in Figma, using a free Groqcloud API.

NeuraLead is an AI-driven B2B tool for discovering and nurturing leads, integrating seamlessly with your CRM and communication systems.

Socap.ai accelerates startup fundraising by matching founders with investors using AI-driven community support, machine learning, and agentic workflows.

Document AI by Playmaker automates document processing by extracting and validating data, then integrating with over 300 systems to streamline workflows.

TrustLoop utilizes AI to enhance feedback capture and reputation management for products, driving high-quality reviews and actionable insights for improvement.

Dropbase AI facilitates rapid development of custom internal tools and backend operations software using AI, Python-based architecture, and self-hosted deployment.

SQLPilot innovates as an AI-first SQL editor, enabling natural language query writing by contextualizing database sources directly in prompts.

📄 Depth Anywhere: Enhancing 360 Monocular Depth Estimation via Perspective Distillation and Unlabeled Data Augmentation

📄 From RAGs to rich parameters: Probing how language models utilize external knowledge over parametric information for factual queries

📄 ​​Multimodal Needle in a Haystack: Benchmarking Long-Context Capability of Multimodal Large Language Models

📄 Bootstrapping Language Models with DPO Implicit Rewards  ​​

📄 Long Code Arena: a Set of Benchmarks for Long-Context Code Models

📄 MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding  

📄 PIN: A Knowledge-Intensive Dataset for Paired and Interleaved Multimodal Documents 

📄 Prism: A Framework for Decoupling and Assessing the Capabilities of VLM

📄 ExVideo: Extending Video Diffusion Models via Parameter-Efficient Post-Tuning  

📄 Whiteboard-of-Thought: Thinking Step-by-Step Across Modalities

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, apply here.