A Daily chronicle of AI Innovations June 10 2024: ⚖️Meta faces legal complaints over AI data plans 🚀Alibaba’s Qwen2 AI models outperform GPT-4 & Llama-3 🧠SAP & Nvidia are developing applications with AI & digital twins🕵️♂️Chinese tech giants exploit Nvidia AI chip loophole 👑AI beauty pageant
A Daily chronicle of AI Innovations June 10th 2024:
⚖️ Meta faces legal complaints over AI data plans
🚀 Alibaba’s Qwen2 AI models outperform GPT-4 & Llama-3
🧠 SAP & Nvidia are developing applications with AI & digital twins
🕵️♂️ Chinese tech giants exploit Nvidia AI chip loophole
🍎 Apple launches “Apple Intelligence” at WWDC 2024 for iPhone, iPad, and Mac
🚀 TCS launches TCS AI WisdomNext™, an industry-first GenAI aggregation platform
🤝 Human Native AI is building a marketplace for AI training licensing deals
🤖 Hugging Face and Pollen Robotics launched an open-source robot for household chores
👑 The world’s first AI beauty pageant
🧠 Microsoft Recall gets safety changes
🔊 AI TRAINING: Using ‘Background Conversations’ in ChatGPT
🧠 AI RESEARCH: Concise prompting cuts AI costs by 20%
Enjoying these daily updates, listen to our podcast at: https://podcasts.apple.com/ca/podcast/ai-unraveled-latest-ai-news-trends-gpt-gemini-generative/id1684415169
Apple unveils massive AI push:
Apple announced a series of long-awaited AI features at #WWDC, its annual developer conference today. Some things that stood out to me:
🔴 The sheer footprint of Apple’s AI push: Apple Intelligence’s generative models would be available across its ecosystem — iOS, iPadOS and macOS.
🔴 The focus on personalization: Experts say that no company knows you better than Apple, and the fact that the company can now draw upon all that personal context to power AI experiences could be a game-changer in boosting productivity and efficiency. Further, it plans to do this through “Private Cloud Compute” using Apple Silicon, so guarantees privacy too.
🔴 The first mainstream use of AI agents? With “cross-application tasking,” Apple touted how it could delve into your apps and execute tasks on your behalf. This is essentially the premise behind AI agents, or AI software that can asynchronously take over tasks (More here: https://lnkd.in/gExb9KuN), a behavior that Apple could help lead the adoption of with this move.
💡 Apple Intelligence: This AI will draw from your personal data to provide answers, ensuring privacy.🔍 Smarter Siri: Siri, enhanced by Apple Intelligence, will now pull information from multiple apps to answer your questions. For example, it can prep you for a meeting by providing details from your calendar, summarizing prep documents, and even giving you the weather forecast.
😊 AI-Generated Emojis: Users can create custom emojis using text prompts in iMessage. Additionally, non-photorealistic images can be generated for conversations.
🤖 OpenAI Partnership: In certain cases, users can opt to use OpenAI’s ChatGPT to answer questions directly from their Apple device, even without a ChatGPT account.
🔒 Hidden Apps and Privacy: New privacy features for iPhone allow users to lock apps with Face ID, Touch ID, or a passcode. Users can also hide apps from the home screen, ensuring media from these apps does not appear elsewhere.
📞 Real-Time Call Transcripts: iPhone users can now record and create transcripts of calls directly from the phone app. All parties will be notified when recording is active.
🎧 Gesture Controls for AirPods: AirPods users can now answer or decline calls with simple head gestures.
⌚ Apple Watch Health Monitoring: New health features on Apple Watch can track vital signs and notify users if they may be getting sick, based on body temperature and heart rate.
Source: https://www.apple.com/apple-intelligence/
🚀 Alibaba’s Qwen2 AI models outperform GPT-4 & Llama-3
Alibaba launched Qwen2 with five sizes ranging from 0.5B to 72B parameters. These models are trained in 27 additional languages besides English and Chinese, showcasing state-of-the-art performance in benchmarks. The models deliver significantly improved performance in coding and mathematics and extended context length support up to 128K tokens. Despite having fewer parameters, qwen2–72 B outperforms leading models like Llama-3–70B and its predecessor Qwen1.5–110B.
Qwen2–72B-Instruct performs comparably to GPT-4 in terms of safety and significantly outperforms Mistral-8x22B. The models are released under Apache 2.0 and Qianwen License on Hugging Face and ModelScope.
Why does it matter?
Qwen2 beats Meta’s model despite being trained on relatively fewer tokens. The researchers attribute it to more efforts put into data cleaning and training, implying innovative approaches on their end.
However, it also signals the slow shift in how LLMs are developed– from solely relying on quantity of data to prioritizing the quality of data and training techniques.
Source: https://qwenlm.github.io/blog/qwen2
🧠 SAP & Nvidia are developing applications with AI & digital twins
At SAP’s Sapphire event in Orlando, Florida, SAP and NVIDIA announced their collaboration to enhance SAP’s generative AI copilot, Joule, with two new capabilities: SAP Consulting and ABAP Developer. These new features are powered by NVIDIA AI Enterprise software.
Additionally, SAP is integrating NVIDIA Omniverse Cloud APIs into its Intelligent Product Recommendation solution to simplify the buying and selling process for complex products. This integration will allow salespeople to visualize 3D product digital twins directly within the SAP Intelligent Product Recommendation interface, making it easier to understand the products.
Why does it matter?
Using NVIDIA Omniverse Cloud APIs in SAP’s Intelligent Product Recommendation solution accelerates the quote generation process and increases sales and customer satisfaction by enabling sales representatives to provide more accurate, tailored recommendations.
Source: https://blogs.nvidia.com/blog/sap-sapphire-ai-omniverse
🕵️♂️ Chinese tech giants exploit Nvidia AI chip loophole
The U.S. government prohibits Nvidia from selling A.I. chips directly to Chinese companies due to national security concerns. Still, ByteDance is accessing Nvidia’s A.I. chips for its U.S. operations by leasing them from Oracle, as the current U.S. rules do not explicitly prohibit Chinese companies from accessing the chips if used within the U.S.
Other Chinese tech giants like Alibaba, Tencent, and China Telecom seek similar arrangements with U.S. cloud providers. The U.S. Commerce Department proposed a rule to tighten controls, but it faced opposition from cloud providers and remains in limbo.
Why does it matter?
Even if the loophole is closed, Alibaba and Tencent have discussed obtaining Nvidia chips for their U.S.-based data centers. It could further escalate the AI “arms race” and rivalry between the USA and China as both nations seek to outpace each other in developing advanced AI systems for economic and military advantages.
What Else Is Happening in AI on June 10th 2024
Apple will launch “Apple Intelligence” at WWDC 2024 for iPhone, iPad, and Mac
Leaks suggest Apple will reveal “Apple Intelligence”, aka AI, at the WWDC event this week. These AI features will focus on broad appeal and privacy, with opt-in not mandatory. Apple will use its own tech and OpenAI tools to power the new AI features.
Source: https://www.theverge.com/2024/6/7/24173528/apple-intelligence-ai-features-openai-chatbot
TCS launches TCS AI WisdomNext™, an industry-first GenAI aggregation platform
The platform allows organizations to compare and experiment with GenAI models across cloud services in a single interface. It offers ready-to-deploy business solution blueprints with built-in guardrails for quick adoption.
A study by Harvard, MIT, and Wharton reveals junior staff is not reliable for AI training
Junior consultants who participated in a GPT-4 experiment struggled with AI risk mitigation, with their tactics lacking a deep understanding of the technology and focusing on changing human behavior rather than AI system design. The findings highlight the need for top-down AI governance, expert input, and upskilling across all levels of the organization.
Human Native AI is building a marketplace for AI training licensing deals
The platform helps AI companies find data to train their models while ensuring rights holders are compensated. Rights holders upload content for free and connect with AI companies for revenue share or subscription deals. Human Native AI helps prepare and price content, monitors for copyright infringements, and takes a cut of each deal.
Hugging Face and Pollen Robotics launched an open-source robot for household chores
The humanoid Reachy2 was initially controlled by a human wearing a VR headset. Then, a machine learning algorithm studied the teleoperation sessions to learn how to perform the tasks independently. The dataset and trained model used for the demo are open-sourced on Hugging Face, allowing anyone to replicate the process on smaller robots at home.
The world’s first AI beauty pageant
The World AI Creator Awards and creator platform FanVue are currently running the inaugural ‘Miss AI’ contest, with over 1,500 AI-generated models competing for the world’s first AI beauty pageant.
- 10 finalists were selected from a pool of 1,500 AI-generated contestants, with the winner set to be announced at the end of June.
- The AI models ‘hail’ from countries across the globe, showcasing various causes and personalities in addition to their photorealistic images.
- Judges will also evaluate the AI tech behind the avatars, including prompts/image outputs and the creator’s ability to engage audiences on social media.
- The prize pool includes $20,000 in awards and access to PR and mentorship programs.
With how good AI image generation capabilities have become, we’re likely already being exposed to an influx of generated brand ambassadors and models without even realizing it. While the idea of an AI beauty pageant may sound strange… it’s probably only going to get weirder from here.
Microsoft Recall gets safety changes
Microsoft is making changes to its new Recall on-device AI feature following security concerns — with the tool now starting off by default and gaining new encryption protections ahead of the release on June 18.
- The Copilot + Recall feature was unveiled at the Microsoft Build event in May, which will run locally and take constant screenshots to remember user actions.
- The company faced backlash after experts warned of the tool’s ‘privacy nightmare’ with the potential to expose screenshots and activity to hackers.
- The feature will now be off by default, with users having to opt in manually — with new encryption and authentication being added for extra safety measures.
- The tool will roll out in ‘preview’ on the new Copilot + PCs set to launch on June 18.
While the Recall feature was one of the flashiest features revealed at Build — it also is one of the most controversial. While the screenshots enable memory and get us a step closer to the AI agent dream, its also a tough hurdle to get past for increasingly privacy-focused consumers.
Source: https://www.therundown.ai/p/microsofts-total-recall
AI TRAINING: Using ‘Background Conversations’ in ChatGPT
ChatGPT’s latest update introduced the “Background Conversations” feature, allowing you to continue your chat even when using other apps or when your screen is off.
- Install the latest ChatGPT app update on your phone.
- Open the app and go to Settings.
- Scroll to “Voice Mode” and toggle “Background Conversations” on.
- Use ChatGPT hands-free while multitasking or with your screen off.
Note: If you still don’t see the option, the feature may not have rolled out to your account yet.
AI RESEARCH: Concise prompting cuts AI costs by 20%
Researchers just found that adding a simple ‘be concise’ instruction to chain-of-thought prompts can reduce the length of AI responses by 50% with minimal impact on accuracy — leading to over 20% cost savings on API calls.
- Chain-of-thought (CoT) prompting has the AI explain its reasoning step-by-step before giving a final answer, improving accuracy on complex problems.
- Researchers tested adding ‘be concise’ to CoT prompts on GPT 3.5 and 4, using 1,000 multiple-choice questions across 10 topics.
- The instruction made responses about 50% shorter on average vs. normal CoT prompting, with no significant accuracy impact.
- The approach also reduced per-query API costs by over 20% for both models, with the more concise prompts generating fewer tokens.
Next time your AI chatbot is getting a little too wordy, this one simple prompt trick is all you need! Despite the massive complexities and uncertainties surrounding how LLMs work, sometimes these strange, seemingly basic quirks make all the difference in getting the optimal outputs.
Source: https://arxiv.org/pdf/2401.05618
New AI Job Opportunities on June 10th 2024:
- Meta — Research Engineer, Conversational AI (Reality Labs): https://jobs.therundown.ai/jobs/62419977-research-engineer-conversational-ai-reality-labs
- Palantir Technologies — Product Designer, US Government: https://jobs.therundown.ai/jobs/62398950-product-designer-us-government
- Databricks — Strategic Sourcing Manager, Technology: https://jobs.therundown.ai/jobs/62399340-strategic-sourcing-manager-technology?
- Anyscale — Marketing Operations Manager: https://jobs.therundown.ai/jobs/62367786-marketing-operations-manager
