The first two weeks of May were packed with information and exciting developments:
- Google finally made a comeback, and the I/O conference regained its liveliness.
- The pursuit of enabling large models to "understand longer content" continued. MosaicML announced the support for 65k token inputs with their open-source MPT-7B model. Exploration solutions like Unlimiformer and InfiniteGPT, which support infinite input lengths, were also released. Additionally, Anthropic's Claude started supporting inputs of 100k tokens, impressing with its effectiveness.
- In terms of content generation, while discussions on various video solutions remained hot, there was a sudden surge in interest for 3D content and sound generation.
- Policy developments also continued their progress. The European Union took the lead in implementing stringent AI restriction measures on the other side of the Atlantic.
The True Global Ventures team compiled a list of 25 interesting AI products from early May. Feel free to share your thoughts and engage in discussions with us in the comments section!
In this issue, we bring you:
- 20 useful and fun new products: In the useful category, we have organized 5 data processing products, 2 text-to-3D content products, 2 AI proxy products, 1 product from the miscellaneous category, and 3 Chrome extensions. In the fun category, we have curated 4 new products. For Chinese developers, we share 2 interesting new products.
- New products from 3 startups: Anthropic, Hugging Face, and MosaicML.
- New products from 2 big companies: This includes Google's list of new products announced at the I/O conference and a new game release from Nintendo.
Enjoy exploring these exciting offerings!
The tool that enables Pandas DataFrame to "speak human language" was officially released on April 28th and has received 5.4k+ stars on GitHub ✨
First of all, what is Pandas? Pandas is an open-source third-party Python library and a vital advanced tool for data analysis in Python. It is built on top of Numpy and Matplotlib, earning the reputation as one of the "Three Musketeers of Data Analysis" alongside NumPy and Matplotlib.
So, what does Pandas AI specifically do? As mentioned earlier, PandasAI is not a replacement for Pandas but rather transforms Pandas into a "chatbot". Users can pose data-related questions in the form of Pandas DataFrames, and PandasAI responds in natural language, tables, or charts. Currently, it only supports the GPT model, and users need to provide their own OpenAI API key.
Additionally, since the author is an Italian living in Germany, the tool has robust data protection measures in place!
Here's a small example of generating a bar chart based on data:
pandas_ai.run(report_table, prompt="Make a barplot that shows how much money each gender spends on each product line")
The tool automatically performs data cleaning and visualizes the execution process, helping users quickly transform raw data into usable data analysis tools for analysis/reporting without the need for formulas, macros, or code. The current platform provides use cases that cover industries such as B2B e-commerce, human resources, financial accounting, and survey data analysis.
Additionally, DataSquirrel emphasizes that the platform is compliant with GDPR (General Data Protection Regulation) and PDPA (Personal Data Protection Act) standards.
Link - https://datasquirrel.ai/
AIHelperBot is a tool that automatically generates SQL queries and supports one-click database connection or import. In previous articles, we have briefly introduced some SQL query generation tools such as AI2SQL, Text2SQL, AIQuery, and so on. However, AIHelperBot appears to have the most comprehensive support for database structures, formats, addition methods, and processing capabilities. It is currently priced at $5 per month and offers a 7-day free trial.
ChartGPT by CadLabs
CadLabs is a tool developed by CadLabs that is based on GPT-3.5. It has the ability to generate charts based on data and answer questions related to the data. In a previous newsletter, we shared a similar tool with the same name, developed by Kate Reznykova, a member of the investment team at Next47. Both tools have a similar drawback, which is that they are prone to errors.
The functionality is just as the name suggests - an all-in-one Embedding Marketplace. It supports public, private, and third-party data for discovering, evaluating, and accessing relevant embeddings. The product has not been launched yet.
Link - https://www.embedding.store/
Skybox by Blockade Labs
It claims to be a 3D content generation tool, but it's more like a 360-degree panorama generation tool. By inputting text (at least one word), it generates a 360-degree panoramic view, which produces a very cool effect!
To see how well it understands specific text, let's take a look at my example: Moon.
Nyric by Lovelace Studio
This is a highly anticipated demo video of a text-to-3D content generation tool. It starts with the command "/create winter winter alien forest," followed by a segment with very low video quality. We cannot confirm whether this video is actually generated by AI.
Lovelace Studio describes Nyric as "an AI world-generation platform for digital communities, build the world of your dreams in seconds." However, there is a clear inconsistency between the company's Twitter and website descriptions. From the former, it appears to be an AI for Gaming company, while the latter suggests it is a VR + Web 3.0 company.
Link - https://lovelacestudio.com/
SmartGPT, built by KOL AI Explainer, is a modular AI agent that offers better reasoning capabilities and easier configuration. According to the creator, SmartGPT aims to break down tasks into smaller sub-problems and leverage the internet and other external sources to enable LLMs (currently GPT-3.5 & 4) to accomplish complex tasks without user input.
When compared to Auto-GPT and BabyAGI, SmartGPT has the following advantages and disadvantages. Its strengths lie in its modularity, easy configuration, and improved reasoning performance. However, one notable drawback is the lack of an integrated memory management module.
Regarding the achievement of "better reasoning performance," there are two main aspects:
1. Agent Division: Tasks are divided among different levels of agents (Manager, Boss, Employee, Minion), each assigned with different inference-related tasks. This division allows the model to perform reasoning effectively:
- Manager: Breaks down the main task into several high-level sub-tasks and sequentially assigns these sub-tasks to Boss.
- Boss: Receives the task, creates a loose plan, and further breaks it down into individual sub-tasks to be assigned to Employees.
- Employee: Receives the task, writes pseudocode, and passes it on to Minion.
- Minion: Refines the pseudocode into LUA scripts and executes them.
2. Context Segmentation: Users segment the context between different agents, saving significant token costs and enabling the use of smaller models with Llama support.
These features enhance SmartGPT's ability to handle complex reasoning tasks by breaking them down and leveraging contextual segmentation for improved efficiency.
No-code AI agent building tool - Currently, you can create one for free. Come and give it a try!
Link - https://fine-tuner.ai/
A questionnaire generation tool essentially allows AI to automatically generate related questions based on a description. In the field of education, there have been many such attempts, such as Wooflash's Quiz Wizard.
Link - https://ai.feathery.io/
SciSpace Copilot, developed by the Indian research platform SciSpace (formerly known as Typeset), is primarily used to explain the text, numbers, and tables in scientific literature (such as research papers, technical blog articles, or reports) in a question-and-answer format.
Typeset is a company founded in 2015 and has accumulated a database of over 200 million papers and more than 50 million full-text open-access PDFs. They have experience in handling academic paper data and have fine-tuned their tool based on data from mainstream academic publishing platforms. As a result, the output of the tool is more accurate compared to other tools like ChatPaper that we are familiar with.
In the updated version released on May 7th, the tool supports explaining specialized vocabulary, conducting in-depth research based on topics (with provided template questions), translation among 13 languages, and adding URL bookmarks. In the future, it will also support annotations and highlighting key points with different colors.
Link - https://typeset.io/
AIPRM for ChatGPT
In simple terms, AIPRM is an SEO prompt template plugin. The developer positions it as a prompt tool for ChatGPT, which also supports Midjourney prompts.
The plugin was launched on January 8th and has already gained over a million users. The Prompt Template has been used over fifty million times. The pricing for B2B clients, including Disney, Adobe, Intel, Microsoft, and Condé Nast, is as follows:
Link - https://www.aiprm.com/
Teamsmart is an interesting document assistant that offers a range of different robots with varying capabilities based on different professions or skills. While its features may be similar to other text-processing plugins, Teamsmart allows users to summarize content, generate code, draft tweets, and more directly within the browser. The distinguishing factor is its collection of robots tailored to different professional or skill sets, each offering unique abilities and functionalities.
Link - https://www.teamsmart.ai/
The Artifact of Clickbait is a tool designed to combat sensationalized headlines by removing exaggerated expressions from articles, leaving only objective facts. While it falls under the "For Fun" category, it is actually quite useful in promoting accurate and unbiased content.
AI-generated drum beats are available in several styles: Default, Hip Hop, Techno, Trap, and Drum n'Bass. There is also an option for random beats, and the results are quite good!
As of May 15th, here are the statistics:
The author Nick is also the creator of the popular AI hairstyle generation product called HairstyleAI.
Link - https://drumloopai.com/
health-related apps and environmental data from smart home apps to provide personalized and professional sleep guidance. It analyzes the data comprehensively and offers individualized recommendations to improve sleep quality. It sounds like a service that would be highly beneficial to individuals in need of better sleep.
Link - https://www.centenarian.co/
Ask, imagine, guess
Playing "You Draw, I Guess" with AI—except, it's the users who ask AI! Can you guess what AI is thinking?
n AI dream interpretation tool developed by my favorite YouTuber and content creator, "不说话的白菜“ (Runde Yang). It supports four modes: Freudian, Zhou Gong's Dream Dictionary, Scientific Dream Analysis, and Shiva. The tool aims to be entertaining.
Recently, people's exploration of dreams seems to have become increasingly diverse. Do you remember the super dreamy dream generation tool The Dreamkeeper (https://thedreamkeeper.co/) developed by Altassian that we shared in Newsletter 008?
Link - https://dreamore.app/zh
Automatically clipping long videos into 10 short videos and seamlessly publishing them on platforms like YouTube Shorts, TikTok, and Instagram Reels, the founder of this tool is Young Zhao, a Chinese entrepreneur who graduated from UC Berkeley.
The role of AI is mainly reflected in several aspects:
- Automatic analysis of video highlights and clipping.
- Automatic scoring of generated short clips based on their virality difficulty, along with providing reasons.
- Automatic face tracking and capturing of the most shareable facial images.
- Automatic generation of subtitles with emphasis on key points.
- Automatic addition of emojis to subtitles.
Link - https://www.opus.pro/
claude-x-100k by Anthropic
Claude's ability to support 100k tokens, which is equivalent to approximately 75,000 words or a 200-page English book, makes it my favorite model. Currently, Poe users with a subscription can use the claude-instant-100k model up to 100 times per month.
As for the effectiveness, let's take a look at the demonstration by the blackboard teacher. However, it's important to note that while Claude has made significant progress on a literary (or philosophical) level, it still produces many factual errors. Therefore, caution is advised when using it for tasks that require high accuracy or even scientific research.
Attached: List of Anthropic Models
Transformers Agents by Hugging Face
「This removes the barrier of entry to machine learning」
In simple terms, Transformers Agents are designed to facilitate the creation of AI agents. Users can leverage them to quickly create various types of AI agents, incorporating LLMs for multimodal capabilities.
Transformers Agents provide a wide range of practical tools, including technical document question answering, text-to-speech conversion, text-to-image synthesis, text-to-video synthesis, website content summarization, image segmentation, and more.
Foundation Series by MosaicML
MosaicML has released its own open-source LLM, MPT-7B, which can be summarized with the following features:
- Commercially available with model quality comparable to Llama. It was trained for approximately 2 weeks at a cost of $200,000.
- Trained on a commercially available 1T tokens dataset. For comparison, Llama was also trained on 1T tokens, while Pythia and OpenLLaMA used 300B tokens, and StableLM used 800B tokens.
- Accepts inputs of up to 65k tokens (estimated maximum of 84k), whereas other open-source models typically support 2-4k tokens.
- Offers a significantly lower price, even as low as 1/15th of competitors like OpenAI or Anthropic.
HF Demo - https://huggingface.co/mosaicml/mpt-7b
Official Announcement - https://www.mosaicml.com/blog/mpt-7b
This year marks the 15th year of Google's I/O conference, and AI has been a prominent focus, especially this year, with nearly 200 mentions of "AI" throughout the event. So, what are the new products from Google worth paying attention to? Here's a brief introduction to each product:
- Google's next-generation LLM, serving as the foundational model for most of the AI products announced at the I/O conference.
- Built on Google's latest JAX and TPU v4 infrastructure, but no further technical details were disclosed.
- Improved in common-sense reasoning, mathematics, and logic: According to the official technical report, PaLM2 outperforms GPT-4 in certain tasks like mathematics. In @Mark Tenenholtz's evaluation, its programming capabilities are superior to ChatGPT, whether it's GPT-3.5-turbo or GPT-4.
2. Gemini - A multimodal large-scale model being developed by Google DeepBrain.
- Built on top of PaLM2.
- Open to all users and will gradually support more languages.
- Allows exporting results to Google Docs, Replit, and Gmail.
- Offers tools similar to ChatGPT Plugins, such as image recognition, displaying locations on maps, and image generation in collaboration with Adobe Firefly.
4. Duet AI - In the previous issue, we introduced AI for Workplace, which now has a name: Duet AI. It is a competitor to Office Copilot but currently lacks standout features that truly impress.
5. SGE（Search Generative Experience） - An "AI-first" search experience where, when the Google algorithm determines that the AI-generated result is superior to the standard answer, the AI-generated answer is displayed at the top of the search results.
Application link - labs.google.com/search
6. Magic Editor for Google Photos - Selective photo editing.
7. MusicLM - Text-to-music generation — currently available for testing at AI Test Kitchen with Google. Worth a try!
8. Gen AI for Android
- Magic Compose - Automated message replies.
- Gen AI for Wallpaper - Mobile wallpaper generation.
9. ML Hub for Developers - End-to-end ML model training and deployment service.
10. Immersive View for Google Map - Simulated 3D immersive maps.
11. 「Help me write」 in Gmail - Email writing assistant (text generation).
Peridot by Niantic
AI-driven AR virtual pet, or you can say, a modern version of Tamagotchi.
Link - https://playperidot.com/
Will AI really replace editors?
The first magazine entirely written by AI, produced by the American non-profit publishing company McSwenney's—Jonathan Zeller wrote: "You may be skeptical about machine-written work at first, but once you see the software rearranging familiar-seeming paragraphs into different orders and changing a few words, you’ll realize it’s a suitable replacement for your favorite authors, who can now rest and starve."
—— but it's a joke :)