Today:

• GPT-4 Turbo Unveiled: Reclaims Top Chatbot Spot

• Cohere Launches Rerank 3

• OpenAI Fires Researchers for Leaks

• Meta AI Expands to WhatsApp, Instagram

• OpenAI's $86B Offer Includes Ex-Staff

• Buildbox 4's No-Code AI game creation

Sam Altman "gpt-4 now significantly smarter" | OpenAI Updates GPT-4 and Reveals Open Source Evals

OpenAI introduces GPT-4 Turbo for improved chat experiences, boasting smarter, more natural responses. The release coincides with open-sourcing a library for transparent model evaluation. 

Results show GPT-4's significant improvement in various tests, reclaiming its top position in the chatbot arena. However, debates persist over test relevance and model performance. 

WATCH THE VIDEO ON YOUTUBE

Introducing Rerank 3: A New Foundation Model for Efficient Enterprise Search & Retrieval

Cohere introduces Rerank 3, a new model tailored for enterprise search and Retrieval Augmented Generation (RAG) systems. Rerank 3 enhances search performance and reduces costs with minimal impact on latency, compatible with any database or search index. It excels in searching multi-aspect and semi-structured data, such as emails and JSON documents, in over 100 languages. 

With improved latency and accuracy, Rerank 3 optimizes document retrieval, particularly for long-context scenarios. It integrates seamlessly with Elastic's Inference API and significantly lowers total cost of ownership while boosting the efficiency of RAG systems.

COHERE

OpenAI Researchers, Including Ally of Sutskever, Fired for Alleged Leaking

OpenAI has terminated two researchers, Leopold Aschenbrenner and Pavel Izmailov, over alleged information leaks. Aschenbrenner, an ally of Chief Scientist Ilya Sutskever, was part of a team focused on AI safety. Izmailov, who also worked on reasoning, had a stint in the safety division. The reasons for their dismissal and the leaked information remain undisclosed. 

These firings mark one of the first public staffing changes since CEO Sam Altman's reinstatement in March, following a board-led investigation that cleared him of past actions leading to his brief dismissal in November.

THE INFORMATION

Meta trials its AI chatbot across WhatsApp, Instagram and Messenger in India and Africa

Meta is testing its AI chatbot, Meta AI, in India and parts of Africa on WhatsApp, Instagram, and Messenger. This move aims to tap into large user bases to expand AI offerings. India, with over 500 million users, is a key market. Meta aims to engage users in developing markets where smartphone usage is rising. 

Meta AI, unveiled in 2023, can answer queries and generate images from text. The late entry into AI tools stemmed from doubts about readiness. Now, Meta prioritizes user retention, investor confidence, and maintaining its technological edge. Integrating Meta AI into WhatsApp and Instagram targets billions of users, potentially surpassing competitors.

TECHCRUNCH

OpenAI Expands $86 Billion Tender Offer, Inviting Former Employees

OpenAI, valued at $86 billion, expands its share sale to former employees, offering them a chance to cash in on their holdings. Previously excluded, ex-staffers now have until April 15 to decide on participating in the offer. 

This move marks a significant shift in the company's approach, potentially benefiting those with a stake in the AI developer's success.

THE INFORMATION

Buildbox 4 launches with no-code AI game creation engine

Buildbox 4 revolutionizes game development with AI-driven, no-code solutions. Users can bring game ideas to life in minutes, as the platform generates assets and scenes based on prompts. It caters to amateurs, semi-professionals, and professionals alike, democratizing game creation. 

With AI assistance, developers can prototype faster and explore innovative designs. Studios benefit from streamlined development processes, enabling scalable game production. Buildbox's evolution from complex coding barriers to AI-first development signifies a new era in game creation, empowering creators of all skill levels.

VENTUREBEAT

🧠RESEARCH

BRAVE: Broadening the visual encoding of vision-language models

"BRAVE: Broadening the visual encoding of vision-language models" delves into enhancing vision-language models (VLMs) by improving their visual encoders. VLMs like CLIP, while advanced, struggle with certain image features and visual hallucination. BRAVE, a new method, merges features from multiple encoders to boost performance across tasks, reducing VLM issues with fewer parameters.

Urban Architect: Steerable 3D Urban Scene Generation with Layout Prior

"Urban Architect: Steerable 3D Urban Scene Generation with Layout Prior" introduces a method for generating large-scale urban scenes from textual descriptions. By incorporating a 3D layout representation as a prior, the model can better interpret ambiguous text and enable controllable generation. It employs Layout-Guided Variational Score Distillation for optimization and a Scalable Hash Grid structure to handle the vast scale of urban scenes. The framework can generate scenes covering over 1000m driving distance, facilitating scene editing demonstrations. 

CodecLM: Aligning Language Models with Tailored Synthetic Data

CodecLM, a framework for fine-tuning large language models (LLMs) to specific tasks by generating tailored synthetic data. By leveraging LLMs as codecs, it encodes seed instructions into concise metadata and decodes them to create customized instructions. This approach reduces labor and time in data collection while improving instruction-following abilities across various distributions and LLMs.

MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies

MiniCPM, advocating for Small Language Models (SLMs) as resource-efficient alternatives to Large Language Models (LLMs). MiniCPM, including 1.2B and 2.4B non-embedding parameter variants, showcases capabilities comparable to larger LLMs. It emphasizes scalability in model and data dimensions, with innovations in model scaling and the introduction of a Warmup-Stable-Decay (WSD) learning rate scheduler for efficient training and domain adaptation. MiniCPM's performance across diverse SLM applications solidifies its significance in the field.

Hash3D: Training-free Acceleration for 3D Generation

Hash3D, a novel acceleration technique for 3D generation that eliminates the need for model training. By identifying and reusing redundant feature maps across neighboring timesteps and camera angles, Hash3D significantly accelerates the inference process while improving smoothness and view consistency of generated 3D objects. Experimentation across text-to-3D and image-to-3D models demonstrates Hash3D's versatility and efficiency gains of 1.3 to 4 times, with processing times reduced to approximately 10 minutes for text-to-3D and 30 seconds for image-to-3D conversions.

🛠️TOP TOOLS

Personno Slash research costs and accelerate insights with Personno’s AI-driven respondents.

Superpipe - optimized LLM pipelines for structured data

Lilys - Summaries built for learning, without context loss.

Tidepool - analyzes chat conversations, user feedback, LLM prompts,
and more, helping you make better decisions for your business.

Lutra - Create AI workflows just from English instructions without the need for coding or drag-and-drop visual programming.

📲SOCIAL MEDIA

🗞️MORE NEWS

European car manufacturer will pilot Sanctuary AI’s humanoid robot

Sanctuary AI, a Canadian robotics firm, is sending its humanoid robot to a Magna auto plant. Magna, a major car manufacturer with facilities in Austria, invests in Sanctuary AI. This move mirrors similar deals with BMW and Mercedes, marking a trend in automating manufacturing processes with humanoid robots. TECHCRUNCH

Italy considers tougher penalties for AI-related crimes

Italy plans stricter penalties for AI-related crimes, including market manipulation and money laundering, in a proposed law. The draft emphasizes AI's use in healthcare and the judiciary, proposing fines and jail time for offenses like copyright infringement and deepfake creation. Italy aims to lead AI governance during its G7 presidency. REUTERS

Anthropic says its AI models are as persuasive as humans

Anthropic, an AI startup, claims its language models match human persuasiveness, raising concerns about disinformation. Research reveals successive AI model generations are increasingly persuasive, with the latest rivaling human arguments. Amid debates over AI's supremacy, this prompts questions about its societal impact and potential for misinformation. AXIOS

Asus’ new OLED laptops run on the latest AI-capable chips

Asus unveils new laptops, including the Vivobook S series and Zenbook 14 OLED (UM3406), featuring OLED displays and advanced AMD and Intel processors. These chips enhance AI capabilities, power efficiency, and performance for tasks like gaming and multitasking. Prices start at $749 for the Vivobook S series and $999.99 for the Zenbook 14 OLED. THE VERGE

Texas is replacing thousands of human exam graders with AI

The Texas Education Agency (TEA) is implementing an "automated scoring engine" utilizing natural language processing for grading open-ended questions on exams, aiming to reduce costs and improve efficiency. While promising savings, concerns arise over accuracy and potential biases, highlighting ongoing debates about AI in education. THE VERGE

‘Like Wikipedia And ChatGPT Had A Kid’: Inside The Buzzy AI Startup Coming For Google’s Lunch

Perplexity, a search engine resembling a blend of Wikipedia and ChatGPT, garnered early attention and investments from tech bigwigs like Jeff Bezos by offering personalized AI-powered searches. Although facing Google's dominance, its novel approach attracts users seeking quick, concise answers, positioning it as a complementary rather than competitive tool in the search landscape. FORBES

AI race heats up as OpenAI, Google and Mistral release new models

OpenAI, Google, and Mistral have all recently released new versions of their frontier AI models, marking an intense period of activity in the AI sector. The launch of these models, including OpenAI's anticipated next major version of GPT, indicates a significant advancement in AI technology and hints at further developments in the industry over the summer. THE GUARDIAN

MPs warn over lack of AI rules for UK’s creative industries

A group of cross-party MPs in the UK have raised concerns about the lack of robust rules governing how artificial intelligence (AI) platforms interact with the creative industries, warning that this is benefiting large tech companies. The MPs found that many UK artists have seen a decline in their royalties, partly due to the adoption of digital distribution models that pay creators less than physical distribution systems. The committee has called on the government to ensure that creators have mechanisms in place to enforce their consent and receive fair compensation when their work is used by AI systems. FINANCIAL TIMES