Lenovo Innovation Technology Conference 2024
At the Lenovo Innovation Technology Conference 2024, Lenovo joined forces with AI leaders like Nvidia’s Huang Renxun and AMD’s Su Zifeng to showcase exciting AI products. Some of the top highlights were personal AI assistants, liquid-cooled servers, and AI-powered smartphones. These innovations are making it clear that AI is playing a bigger role in everything, from personal computers to business services and even public welfare projects.
Lenovo introduced a new personal assistant called AI Now, which helps with tasks like taking notes and designing posters. For businesses, Lenovo launched the sixth generation of their Neptune liquid cooling system, which makes servers more energy-efficient. AI’s potential was also shown through advancements in self-driving cars and mixed-reality technology.
Another major announcement was a collaboration between Intel and AMD, who are working together to drive innovation in the x86 architecture. Nvidia’s Huang Renxun shared his vision of AI agents becoming as important as operating systems, predicting that billions of AI agents will be created with Nvidia’s technology.
OpenAI’s Study on ChatGPT Bias
OpenAI recently released a study showing that ChatGPT talks to users differently depending on their names, like “Xiaomei” or “Xiaoshuai.” The study found that ChatGPT linked some names to specific topics. For example, “Xiaomei” was often connected with early childhood education, and “Xiaoshuai” with computer engineering. Female names usually got friendlier replies, while male names received more formal responses.
Although the differences were small, researchers noticed them. OpenAI said harmful replies were very rare (about 0.1%). The study’s goal was to point out biases in fun things and serious areas, like job applications. OpenAI plans to keep studying other languages and groups to make AI more fair.
Professor Nie Zaiqing and AI in Biology
Professor Nie Zaiqing from Tsinghua University is making breakthroughs by using AI in the world of biology. At the 2024 Nobel Prize in Chemistry, attention was given to how AI can help solve tough problems in drug development. The drug industry still faces long timelines, high costs, and low success rates. But Professor Nie’s team has created a large AI model called ChatDD to help.
ChatDD mixes natural language processing with biological data to assist drug developers. Nie believes the biology field is a great fit for AI because it has lots of unique data. Already, ChatDD is helping with drug research and clinical trials. Nie envisions ChatDD becoming a hub for tools in the industry, potentially creating new ways for the pharmaceutical world to profit.
Tencent’s Video Generation Models
Tencent AI Lab, along with USTC, compared 13 video generation models. They tested over 8,000 cases, from human-centered videos to robotics and animations. This research provides insights into future possibilities for video creation.
The report also mentions challenges, like how hard it is for these models to understand complex actions. It offers a “seeing is believing” approach, allowing people to watch and compare the AI-generated videos. This could help make video creation more accessible, letting more people express their creativity.
OpenAI Hires Dane Stuckey as CISO
Dane Stuckey, previously Palantir’s chief information security officer (CISO), has joined OpenAI. He will work with the security team to keep OpenAI’s AI products safe. His experience with government contracts could be helpful, especially with OpenAI’s recent work with the Pentagon.
https://techcrunch.com/2024/10/15/former-palantir-ciso-dane-stuckey-joins-openai-to-lead-security
EverartAI: AI-Powered Marketing
EverartAI claims to be the first fully AI-driven marketing team. They can take your ideas and turn them into 500 unique marketing visuals, without needing complicated prompts. Just share your thoughts, and they’ll handle the rest, delivering professional results quickly.
Meet Ditto: The Self-Build Coding Agent
Say hello to Ditto, a coding agent that only needs around 500 lines of code to build multi-file applications on its own. It’s designed to be simple, using a basic loop with just five tools. Ideal for developers who want a straightforward coding solution.
Fixing LLM Training with Gradient Accumulation
Unsloth AI recently fixed a bug in large language model (LLM) training that was causing inaccurate loss calculations. Their new method reduces the error by over ten times, which is a big improvement for anyone training large AI models.
https://unsloth.ai/blog/gradient
Fixing Gradient Accumulation in Training AI Models
Recently, we fixed a big problem with gradient accumulation that was hurting the training of large language models (LLMs). Unsloth’s update improves how training runs and loss calculations work, making sure everything is accurate.
What is Gradient Accumulation?
Gradient accumulation helps us train models using less memory (VRAM) by acting like we’re training with larger batches. It’s especially useful in setups with multiple GPUs, which is important for big training tasks.
This issue was first spotted by Zhaofeng back in 2021 and was noticed again by Benjamin Marie last week. They found that using gradient accumulation could sometimes lead to higher loss values compared to training with full batches.
Anthropic’s New Responsible Scaling Policy
Anthropic has updated its Responsible Scaling Policy (RSP). This policy helps manage the risks of advanced AI systems. The update aims to create better safety rules based on what AI models can do.
What’s New?
The update brings a more flexible approach to assessing AI risks. It includes new guidelines for when to improve our safety measures and a simpler process for evaluating models and their safeguards. The goal is to learn from past experiences and prepare for the fast growth of AI.
New PIKAFFECT Features
There’s exciting news with PIKAFFECT! Now you can make objects collapse, dissolve, or deflate with a fun “Ta-Da!” effect. Try it out at Pika Art.
Vision Transformers and Visual Reasoning
Researchers are studying how Vision Transformers (ViTs) perform on a test called the Abstraction and Reasoning Corpus (ARC). This test checks how well AI can think about pictures. They discovered that ViTs have a hard time with the ARC tasks, even when they are trained with a lot of data
https://arxiv.org/abs/2410.06405
https://arxiv.org/pdf/2410.06405
Introducing ViTARC:
To improve performance, they created a new architecture called ViTARC, which focuses on better image representation and object recognition. This model shows much better results, solving many ARC tasks effectively.
Fairness in Chatbots
OpenAI is looking into “first-person fairness” in chatbots like ChatGPT. This means they want to make sure that all users get fair treatment and good answers, no matter who they are.
https://cdn.openai.com/papers/first-person-fairness-in-chatbots.pdf
Enhancing LLMs with Thinking Ability
New research is being done to help large language models (LLMs) “think” before they answer questions. This will help the models solve tough problems better and do well in different tasks.
https://arxiv.org/abs/2410.10630
https://arxiv.org/pdf/2410.10630
Hugging Face & GitHub Updates
Animate-X:
This new animation framework uses a latent diffusion model to create high-quality videos. It captures motion patterns and is better than existing methods.
https://lucaria-academy.github.io/Animate-X
IterComp:
IterComp is an advanced text-to-image generation method that improves how images are created, making it easier for developers to create various generation techniques.
https://huggingface.co/comin/IterComp
Investment News
Xscape Photonics:
This new company got $44 million to work on special lasers for data centers. These lasers will help connect computer chips faster and use less energy.
https://techcrunch.com/2024/10/15/xscape-is-building-multicolor-lasers-for-datacenters
Amplitude Acquires Command AI:
Amplitude has bought a company called Command AI to make apps more engaging. This will help users get better help inside the app, like finding their way and getting started more easily.
https://techcrunch.com/2024/10/15/amplitude-buys-command-ai-to-bolster-its-app-engagement-offerings
OutRival Gets Funding to Build AI Agent Service
OutRival is a new company started by Ruben Harris and Timur Meyster, who also founded Career Karma. They want to help businesses create AI agents that can improve customer service. Although many companies are in this field, OutRival is getting attention for its innovative ideas.
In 2022, the AI industry received over $64.1 billion in funding, with one-third going to new AI companies. OutRival is using leftover funds from Career Karma’s $40 million funding round in 2022 and has support from investors like Jack Altman and Initialized Capital.
OutRival wants to help companies easily build AI agents that work well with their current tools and systems. This will help businesses stand out and give customers a better experience. They have already worked with college admissions teams to make their jobs easier and plan to expand into other industries. Career Karma will continue to operate as an independent company owned by OutRival.
Concourse Raises $4.7 Million for AI Financial Automation
Concourse, a startup that makes financial automation tools, has raised $4.7 million in a funding round led by Andreessen Horowitz (a16z). Other investors include Y Combinator, CRV, and BoxGroup. The money will be used to improve their product, hire more team members, and boost marketing efforts.
Co-founded by former a16z investor Matthieu Hafemeister and Ted Michaels, Concourse wants to automate complicated financial tasks like data extraction and report generation using AI. They have attracted clients like Instabase and Shef and are currently testing their platform, which they plan to officially launch next year.
Hafemeister stated that the funding will mainly go toward growing the engineering team, especially in areas like machine learning. Their goal is to create an all-in-one automation solution for finance teams, helping them reduce manual work and improve the accuracy of their analyses.
https://techcrunch.com/2024/10/15/concourse-is-building-ai-to-automate-financial-tasks
Gladia Completes $16 Million Series A Funding for Audio Transcription API
OutRival is a new company started by Ruben Harris and Timur Meyster, who also created Career Karma. They want to help businesses make AI agents that can improve customer service. Even though there are many companies doing this, OutRival is getting noticed for its creative ideas.
In 2022, the AI industry received over $64.1 billion in funding, with one-third going to new AI companies. OutRival is using leftover money from Career Karma’s $40 million funding round in 2022 and has support from investors like Jack Altman and Initialized Capital.
OutRival aims to help companies easily build AI agents that work with their existing tools and systems. This will help businesses stand out and give customers a better experience. They have already worked with college admissions teams to make their jobs easier and plan to expand into other areas. Career Karma will still operate as an independent company owned by OutRival
Improving Reasoning in Large Language Models Using Prolog
Large language models (LLMs) have great reasoning skills but can still be improved. One problem is that they generate answers one step at a time and cannot use loops or conditions, which can limit their performance.
To improve their reasoning, developers are trying different methods. Some generate multiple answers and choose the best one, while others use programming languages for help. Recently, there has been interest in using Prolog, a programming language suited for logical reasoning.
Prolog allows LLMs to simplify their code and reason better by converting user requests into Prolog code. Research has shown that using Prolog can significantly increase the accuracy of LLMs on math problems.
Should Prefill and Decode Be Separated into Different Cards?
Separating the prefill and decode stages in language models can help manage memory better. This can lower memory use, especially with large models and long contexts. However, sharing data between the stages can add some communication delays.
By separating these processes, systems can work faster and more efficiently. But this separation can make things more complicated, especially for large-scale systems.
https://www.zhihu.com/account/unhuman?type=S6E3V1&need_login=true
Current Methods for Distilling Large Models
Researchers are looking at knowledge distillation to make big models work better. This method helps smaller models learn from larger ones, making them easier to use.
One method is called MiniLLM, which uses a special way to find important patterns in generative models. Another method from Meta helps break down complex systems into simpler ones to improve their reasoning skills.
https://www.zhihu.com/account/unhuman?type=S6E3V1&need_login=true
Comparing CogVideoX Practices: xDiT vs. VideoSys
Generative video technology is getting a lot of attention, but it has challenges like high processing needs and slow speeds. Different methods, like Ulysses Attention and CFG parallelism, are being used to improve efficiency.
The xDiT and VideoSys frameworks were compared in terms of performance. While both performed well under certain conditions, xDiT showed better results with more parallel methods.
https://www.zhihu.com/account/unhuman?type=S6E3V1&need_login=true
How GPUs Speed Up Deep Learning Training
GPUs (Graphics Processing Units) are great for deep learning because they can perform many operations at once. Unlike CPUs, which are good for single tasks, GPUs excel at handling large parallel tasks.
CUDA programming helps developers use GPU resources to make calculations faster, especially in deep learning. This article explains how to write code that runs in parallel using CUDA and highlights the benefits of using GPUs, especially when working with large data sets.
A Visual Guide to Mamba and State Space Models
This paper discusses the Transformer architecture and introduces Mamba, a new model that might improve upon Transformers. Mamba aims to be faster and more efficient by selectively processing information.
It uses new algorithms to help with parallel computing and keep track of important data. Mamba is designed to work well with long sequences and hopes to compete with Transformer models in speed and efficiency.
https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-mamba-and-state
AimRT Installation Guide (Part 2)
This article explains how to install AimRT on Ubuntu 22.04. It covers the steps needed to upgrade CMake, compile AimRT, and run examples.
Upgrade CMake: Download and install the latest version from the official website.
Compile AimRT: Get the source code from GitHub and run the build script. You can adjust download links for better access if needed.
Run Examples: After compiling, check if everything works by running a sample program.
https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-mamba-and-state