Generative AI Funding at an All time High 📈

Plus: OpenAI Function Calling and API Updates, Google's Virtual Try-On Tools, Pirated GPT-4 and more.

Hey there 👋

We’re starting this week’s edition with Salesforce jumping in the generative AI bandwagon to join the flag-bearers, with the below announcements:

• Salesforce AI Cloud that leverages generative AI to help enterprises enhance customer experiences and productivity across sales, service, marketing, commerce, IT, and development workflows.

• Salesforce Generative AI Fund doubled to $500 million, adding Humane and Tribble to its AI startup ecosystem.

• Salesforce Accelerator - AI for Impact, a philanthropic initiative to provide purpose-driven organizations with equitable access to generative AI technologies.

Also, this week was brimming with a palpable sense of “rivalry” as Meta released and open-sourced an AI music generator MusicGen in a bout to Google’s guarded MusicLM, and AMD launched a new AI chip which is apparently the “world’s most advanced AI accelerator”, contesting Nvidia’s dominance. Adding to the spice was GPT-4 that has surpassed human capabilities in crafting pitch decks across multiple industries to become a favoured choice for securing funding.

If this made you curious, keep scrolling for more juicy details because we’ve covered it all!

This issue covers:

• Latest Developments 🌍

• News from the Industry 🧑‍🏫

• Tools of the Trade ⚒️

• AI Meme of the Week 🤡

Latest Developments 🌍

Our Pick 👌

Tracking Everything Everywhere All at Once: OmniMotion, a globally consistent motion representation, that allows for accurate, full-length motion estimation of every pixel in a video.

• Weakly supervised information extraction from inscrutable handwritten document images: Addressing the limitations of existing information extraction methods when dealing with handwritten documents.

• Video-ChatGPT: A multimodal model that combines visual and language understanding to generate human-like conversations about videos.

• FasterViT: Combines the benefits of CNNs and ViT for high image throughput in computer vision applications, using a Hierarchical Attention approach.

• Transformers learn through gradual rank increase: Transformers exhibit incremental learning dynamics with increasing rank difference between trained and initial weights.

• Face0: Enables instant conditioning of a text-to-image model on a face, for prompt-based image generation and control.

• STUDY: Socially-aware recommender system that utilizes a modified transformer decoder network for joint inference over user groups in a social network.

• Judging LLM-as-a-judge with MT-Bench and Chatbot Arena: GPT-4 as a judge for evaluating chat assistants shows over 80% agreement with human preferences.

• Scalable 3D Captioning with Pretrained Models: Cap3D, an automatic approach to generating descriptive text for 3D objects, that leverages pretrained models to consolidate captions from multiple views of a 3D asset.

• Image Captioners Are Scalable Vision Learners Too: Plain image captioning is a more powerful pretraining strategy for vision encoders than contrastive pretraining on image-text pairs.

• Speech-to-Text Adapter and Speech-to-Entity Retriever Augmented LLMs for Speech Understanding: Significantly improves dialog state tracking performance and reduced word error rate in automatic speech recognition.

• Galactic: A high-speed simulation and reinforcement learning framework for training robotic mobile manipulation skills.

• ChatGPT is fun, but it is not funny!: ChatGPT struggles with generating diverse and original jokes, repeating the same 25 Jokes over 90% times.

• Retrieval-Enhanced Contrastive Vision-Text Models: Utilize external memory to improve fine-grained knowledge retrieval and boost CLIP performance.

• SayTap: A method to control quadrupedal robots using natural language commands and foot contact patterns.

• TART: A plug-and-play transformer module that enhances reasoning abilities in LLMs, improving performance across various tasks, models, and modalities.

• Mind2Web: A dataset for developing and evaluating language-based generalist agents to perform complex tasks on real-world websites.

• WebGLM: An efficient web-enhanced Q&A system that augments pre-trained LLMs with web search and retrieval capabilities, improves upon WebGPT.

• GPT-Calls: Using GPT model for efficient and accurate call segmentation and topic extraction without the need for labeled data.

---

News from the Industry 🧑‍🏫

Our Pick 👌

• OpenAI has announced the following Function calling and API updates:

  • Developers can now describe functions to GPT-4 and GPT-3.5-turbo, allowing the model to output a JSON object containing arguments to call those functions.

  • The existing GPT-4 and GPT-3.5-turbo models have been improved, and new models have been introduced with extended context length.

  • The cost of text-embedding-ada-002 is being reduced by 75% to $0.0001 per 1K tokens.

  • The cost of GPT-3.5-turbo’s input tokens is reduced by 25%, to $0.0015 per 1K input tokens and $0.002 per 1K output tokens.

  • Deprecation timelines have been announced for older versions of the models.

• A four-weeks-old French startup Mistral AI secures a $113 million seed funding round at $260 million valuation, to challenge OpenAI in LLM development with a focus on open-source solutions.

• Google has released Virtual Try-On for apparel that shows how clothes look like on real models with different body shapes and sizes, with accurate details like draping, stretching and wrinkles.

• Meta has released and open-sourced AI-powered music generator MusicGen that turns text descriptions into short audio clips.

• Meta has developed a human-like AI I-JEPA that learns by creating an internal model of the outside world, delivering strong performance on computer vision tasks while being computationally efficient and versatile.

• Meta plans to make LLaMa, its open-source LLM, open for commercial use. It is is currently available only to researchers.

• Meta is rolling out an in-house AI chatbot called 'Metamate' to its employees, trained on the company data, instead of partnering with Microsoft or OpenAI.

• Meta is introducing AI-generated stickers in Messenger, with plans to expand generative AI in its social apps like WhatsApp, Facebook and Instagram.

• AMD has developed a new AI chip MI300X, claiming it to be the “world’s most advanced accelerator for generative AI”, challenging the status quo of Nvidia.

• According to a survey, GPT-4-generated pitches have a 3x higher likelihood of securing funding compared to human ones, across all industries.

• Microsoft and OpenAI's partnership is rumoured to be in conflict as Microsoft disregarded OpenAI's warning and rushed to integrate GPT-4 into Bing search, causing resentment and strained relations between the companies.

• OpenAI, Google DeepMind and Anthropic have committed to providing the UK priority access to their AI models to support research into evaluation and safety, with the UK onset to become a global hub for AI safety.

• OpenAI CEO Sam Altman calls for U.S.-China collaboration in regulating AI development, potentially for the company’s interest while OpenAI's products are not currently available in China.

• People are pirating GPT-4 by scraping exposed API keys, allowing them to access and use the model without paying, potentially leading to unauthorized charges on stolen accounts.

• Alphabet has cautioned its employees on entering confidential information into chatbots, including its own chatbot Bard, due to privacy concerns.

• Cerebras has released SlimPajama, a cleaned and deduplicated version of RedPajama, trimming the dataset from 1210B to 627B tokens, offering the largest, multi-corpora, high-quality, open-source dataset for training LLMs.

• Hugging Face and AMD are partnering to accelerate state-of-the-art models for CPU and GPU platforms, providing improved performance and cost-effective training and inference.

• Cohere raises $270 million in Series C round, at $2.1 billion valuation, to further develop its cloud-agnostic AI platform for enterprises.

• Accenture plans to invest $3 billion in AI over the next three years, to offer new industry solutions, launch an AI Navigator for Enterprise platform, and double its AI talent to 80,000 people.

---

Tools of the Trade ⚒️

Our Pick 👌

Framer AI: Create and publish a website in seconds using simple text prompts, provides many editing options and an in-built copywriter.

• Airplane Autopilot: Develop internal tools and dashboard to simplify engineering operations on the Airplane platform, using text prompts, without coding.

• Juri Flow: Get instant legal assistance from expert AI lawyer well-versed in various legal domains.

• Greenifs AI: Ensures compliance with green marketing guidelines, detects greenwashing errors, and helps improve marketing communications.

• Composer: Build trading algorithms using AI, backtest strategies, and execute trades, all without coding.

• JobWizard: AI-powered job hunting tool that automates job applications, provides personalized answers and tracks applications in real-time.

• Credal: Secure AI solution for enterprises that integrates with existing data sources, provides secure chat UI and APIs, enforces access policies, generates audit logs and redacts sensitive data.

• AIAgent: An intelligent web app that empowers users to automate workflows, runs multiple AI Agents concurrently, powered by GPT-4, no API keys.

• Bothatch: Transform your data into conversations, create and train AI-powered chatbots that engage in personalized interactions and automate tasks.

• RestoGPT: AI that generates free online ordering storefront with integrated POS and delivery, enables autopilot order acceptance and fulfillment without fees.

• Deeto: Connect your prospects with top customers to provide trustworthy insights, facilitate dialogue, and close deals faster.

• Whisper Web: Offers ML-powered speech recognition directly in your browser, enabling audio-to-text conversion in real-time.

• Perplexity AI Profile: Create your own AI profile by setting your bio, choosing language and location, and get answers tailored to your preferences.

• Sentelo: Efficient learning through paraphrasing, code explanations, summaries, control questions, expanded content and more.

---

AI Meme of the Week 🤡

Source

That’s all for this week!

Will see you next Saturday with more such content. Don’t forget to subscribe and give your feedback below.

BONUS 🎉

Share this newsletter with three other friends and stand a chance to win my book GPT-3: The Ultimate Guide to build NLP Products with OpenAI API. Winners will be selected on a monthly basis.