Is OpenAI Ignoring AI Safety for Shiny Products?
PLUS: $10 million in free compute by Hugging Face, Nvidia ChatQA outperforms GPT-4 Turbo
Today’s top AI Highlights:
OpenAI’s Superalignment Team Head calls out Sam Altman on ignoring AI safety for “shiny products”
Hugging Face offers $10 Million in free GPUs to support AI development
Nvidia’s ChatQA models outperform GPT-4 Turbo on conversation QA and RAG
AI Copilot for MacOS that offers more than the ChatGPT desktop app
& so much more!
Read time: 3 mins
Latest Developments 🌍
OpenAI’s Transparency is on Question Again 🔍
It seems OpenAI has landed into controversy, yet again. Last week Ilya Sutskever, Co-founder & Chief Scientist, resigned from OpenAI. This was followed by another resignation from Jan Leike who was involved with the development of ChatGPT and the head of OpenAI’s Superalignment team.
While Sutskever’s resignation spurred more conversations about OpenAI not focusing on making AI safe and focusing on releasing products, Leike went upfront in his recent tweet on how “OpenAI is shouldering an enormous responsibility on behalf of all of humanity.”
Key Highlights:
Superalignment Team Issues: OpenAI formed the Superalignment team to control superintelligent AI, promising it 20% of the company’s compute resources. However, requests for even a fraction of that compute were often denied, significantly hindering the team’s work and causing frustration among its members.
High-Profile Resignations: Several team members, including Leike and Sutskever, resigned due to disagreements with OpenAI’s leadership over the company’s priorities. They felt that OpenAI was not sufficiently focused on safety and preparedness for future AI models.
Restrictive Off-Boarding Agreements: Departing employees were subjected to extremely restrictive off-boarding agreements with lifetime nondisclosure and non-disparagement clauses. Violating these agreements could result in the loss of vested equity worth millions of dollars.
Disaster Management: Following criticism, Sam Altman has acknowledged this unfair clause and says he’s “embarrassed”. OpenAI is now changing this policy to remove the “potential equity cancellation” provision.
No Transparency: OpenAI continues to maintain its stance on not opening its model due to safety concerns. This along with the departure of senior safety team members and restrictive NDAs, is just increasing the skepticism about the company’s commitment to transparency and accountability.
Chai Research is a leading Generative AI Company that serves 15 trillion+ tokens every month on par with industry giants like Anthropic AI with just a team of 12.
If AI research excites you, this is your chance to be part of something big. Chai Research is offering AI/ML roles with competitive salaries ranging from $100k to $500k.
Apply Now: https://www.chai-research.com/jobs
ZeroGPU Makes AI Resources More Accessible 🤗
Hugging Face is one of the leading forces in democratizing AI development. They’ve pledged $10 million worth of free compute power to help developers create new AI technologies. This generous move is a direct response to how AI advancements are concentrated within the walls of large tech companies like Microsoft, Google and Tesla. Hugging Face is launching a new program called ZeroGPU, which provides shared GPU resources to developers, empowering them to build powerful AI models without the usual financial and technical barriers.
Key Highlights:
Shared GPUs for All: ZeroGPU allows multiple users and applications to share access to GPUs, making them more affordable and accessible to individuals and small teams. This is a stark contrast to the current model where developers typically have to commit to expensive, long-term contracts with cloud providers.
Level Playing Field: By removing the financial burden of acquiring powerful GPUs, ZeroGPU empowers startups, researchers, and independent developers to compete with tech giants. This could lead to exciting new innovations across various fields.
Powered by Nvidia GPUs: ZeroGPU relies on Nvidia’s A100 GPUs, giving a boost to AI development without breaking the bank. This means developers can access powerful computational resources without large budgets or extensive infrastructure.
Technical Research 👇
Opensource ChatQA Models Surpass GPT-4 in Key Benchmarks
Building conversational AI models that can understand context, integrate information, and answer complex questions like GPT-4 is a challenging task. Access to the same data and resources used to train powerful models like GPT-4 is the biggest challenge. Nvidia’s new open-source project ChatQA tackles this challenge. ChatQA, a suite of models, significantly improves conversational Q&A and RAG tasks, surpassing even GPT-4 Turbo, using open-source data and resources.
Key Highlights:
Two-stage instruction tuning: A two-stage instruction tuning method was developed. The first stage involves supervised fine-tuning on a blend of high-quality datasets. The second stage, context-enhanced instruction tuning, integrates contextualized QA datasets to enhance the model’s ability to leverage provided or retrieved context..
Efficient Retrieval: For retrieval, a dense retriever optimized for conversational QA is fine-tuned on multi-turn QA data. This method yields performance comparable to state-of-the-art query rewriting models while reducing deployment costs.
Comprehensive benchmark: ChatQA introduces CHATRAG benchmark consisting of 10 conversational QA datasets, covering diverse question types and document complexities. These include datasets that require retrieval, table understanding, arithmetic calculations, and even handling scenarios with unanswerable questions.
Results: ChatQA-1.0-70B model, built on Llama 2, slightly outperforms both GPT-4 and GPT-4-Turbo on the CHATRAG bench without relying on any OpenAI synthetic data. The team also developed the Llama 3-ChatQA-1.5-70B model, which surpasses the latest version of GPT-4-Turbo.
😍 Enjoying so far, share it with your friends!
Tools of the Trade ⚒️
Invisibility: Your AI copilot for MacOS that is always available to help you with whatever you’re doing on your computer. It appears with just a shortcut (Option + Space). It can see your screen so you can ask it questions without switching tabs or apps. It might sound just like the ChatGPT desktop app but it offers more: You can choose from the various models available like GPT-4O, Gemini, Llama 3, Claude 3, Mixtral, etc.
Interactive Graph by Columns AI: Transform simple prompts into visually captivating graphs. Columns AI uses GPT-4O to turn simple text prompts into interactive and appealing graphs and charts that can be directly used in presentations. You can further edit those graphs to align with your brand colors and style.
Userevaluation’s AI Interviews: AI sidekick to conduct, analyze, and summarize customer interviews. You can set up a research study, define participant criteria, and craft detailed interview plans. The AI handles the interviews, provides transcripts and summaries, and generates detailed reports with key insights and visual data presentations.
Awesome LLM Apps: Build awesome LLM apps using RAG for interacting with data sources like GitHub, Gmail, PDFs, and YouTube videos through simple texts. These apps will let you retrieve information, engage in chat, and extract insights directly from content on these platforms.
Hot Takes 🔥
OpenAI - we care for safety a lot, are extremely prudent and will hold back release for an indefinite time until we are completely satisfied about safety!
Also OpenAI - we stayed up all weekend trying io ship 4o just before I/O. We envision a future where everyone will get a slice of GPT-7 compute to survive because our AGI will be taking over!!
🤯🤯🤯 ~
Bindu ReddyWhen you put sufficiently many people in a room together with such a distorted view of reality that they perceive a impending Great Evil, they often fall victim to a spiral of purity that makes them hold more and more extreme beliefs.
Pretty soon, they become toxic to the organization that hosts and funds them.
They become marginalized and eventually leave.
This scenario is quite common. ~
Yann LeCun
Meme of the Day 🤡
openai lawyers cooking up the most complicated corporate structure and employee stock option plans
Sorry, one more!
That’s all for today! See you tomorrow with more such AI-filled content.
Real-time AI Updates 🚨
⚡️ Follow me on Twitter @Saboo_Shubham for lightning-fast AI updates and never miss what’s trending!
PS: I curate this AI newsletter every day for FREE, your support is what keeps me going. If you find value in what you read, share it with your friends by clicking the share button below!