Einstein in your Basement

Good morning.

In a great video (that I urge you to watch), Henrik Kniberg likens ChatGPT4o to having Einstein in your basement. Where GPT2 was a high schooler and GPT3 was someone at a prestigious university, you now have access to Albert Einstein. He knows almost everything about anything.

But most of us are not engaging in conversation with him; they ask him silly questions and turn away when he doesn’t nail it on the first try.

The only limitation is YOU and your imagination. As Kniberg says in the video, this is like having a world-class chef in your kitchen and asking him to peel the onions.

A lot of people are quick to dismiss Generative AI, rushing to find something so specific to their jobs that Einstein fails to answer.

“Ha!”, they say. “I told you not to be worried.”. But the fact is, as this progresses, it will soon be able to answer those specific questions.

Back to the restaurant metaphor: it’s a fact there is now a world class chef (imagine Paul Bocuse or Adrián Ferra) in your kitchen.

I talk to a lot of people about AI and there’s three approaches I see people take:

1. They welcome him in the kitchen and step back; they let him cook and perhaps do the final review before they send out the dish. They give full credit to him.

2. They’re convinced they are still the better chef; they ignore the fact that Paul Bocuse is standing there and assign him to the dishwasher station

3. They let Paul Bocuse cook in a backroom and present the dish as your own.

Point 1 is where we all need to go. Point 3 is what I suspect many people are doing - it’s taboo to say you use GPT in your work (which is a whole other topic)

And when you choose point 2 - you’ll be blindsided, I’m afraid.

Here’s the video mentioned above - that I found buried under thousands of cat videos. Enjoy (but first read the newsletter below :))

Welcome to the Blacklynx Brief.

AI News

• At its WWDC event, Apple introduced its ‘Apple Intelligence’ AI strategy, featuring a partnership with OpenAI and new AI features for iOS 18, iPadOS, and macOS 15. Siri will now converse more naturally, remember context, and perform complex tasks, while new AI tools in apps like Mail and Notes will help auto-generate and edit text. Privacy remains a priority with on-device processing and Private Cloud Compute, and the AI features will be optional for users.

• Chinese tech firm Kuaishou has launched KLING, a new AI model that generates high-quality videos up to 2 minutes long, surpassing OpenAI's still-unreleased Sora. KLING produces 1080p resolution videos, showcasing both realistic and surreal clips, and uses a 3D space-time attention system for more lifelike motion and interactions.

• TikTok's parent company ByteDance is renting advanced Nvidia AI chips in the U.S., exploiting a loophole to bypass restrictions on AI chip exports to China. Although the U.S. bans direct sales of Nvidia's AI chips to Chinese firms, ByteDance is reportedly leasing servers with these chips from Oracle, accessing over 1,500 H100 and several thousand A100 chips. Other Chinese tech giants like Alibaba and Tencent are considering similar strategies, highlighting the intensifying AI race between the U.S. and China despite regulatory efforts.

• Microsoft is updating its “Recall” on-device AI feature due to security concerns, making it off by default and adding new encryption protections before its release on June 18. The feature, which takes constant screenshots to remember user actions, faced backlash over potential privacy risks. Users will now need to opt in manually, and the tool will launch in preview on new Copilot + PCs.

• Researchers found that adding a "be concise" instruction to AI prompts can cut response length by 50% without significantly affecting accuracy, saving over 20% on API costs. This method was tested on GPT-3.5 and GPT-4 with 1,000 multiple-choice questions, showing shorter and equally accurate responses compared to normal prompts. This simple adjustment can make AI chatbots more efficient and cost-effective

• Elon Musk has threatened to ban Apple devices from his companies, criticizing Apple’s new partnership with OpenAI as a "security violation" and "creepy spyware." Musk tweeted that visitors to Tesla, SpaceX, and xAI will need to store their Apple devices in a Faraday cage. This conflict highlights Musk's ongoing issues with OpenAI, a company he co-founded, and suggests he might create a competing "Grok phone" if Apple proceeds with the integration.

• Weirdly enough , two days later, Elon Musk withdrew his lawsuit against OpenAI, which had accused the organization of shifting its focus from benefiting humanity to making profits. Musk co-founded OpenAI in 2015 but claimed the mission changed after his departure in 2017, especially with its partnership with Microsoft. He’s up to something, we don’t know yet what.

• A University of Michigan study used AI models, originally designed for human speech, to interpret dog noises, identifying breed, age, gender, and emotional state with 70% accuracy. Researchers gathered vocalizations from 74 dogs and fine-tuned the AI to analyze these sounds. The AI successfully distinguished individual dogs, breeds, and emotional contexts like play and aggression

• Microsoft is ending the custom GPT creation feature in its Copilot Pro subscription service, shifting its focus to enterprise and commercial applications. This feature, introduced just months ago, allowed users to create custom chatbots similar to OpenAI’s GPT Store. While this functionality will be removed for Pro users on July 10, it will remain available for enterprise customers.

• Luma Labs has introduced Dream Machine, a new AI model capable of generating high-quality, realistic 5-second video clips from text and image prompts. Unlike OpenAI’s yet-to-be-released Sora, Dream Machine is already available to the public. It offers a free plan for 30 video generations and a paid tier for up to 2,000 outputs per month. Go try it !
  

Quickfire News

Another week - another flurry of activity. Things are moving fast. This section tells you everything that happened outside of our main points.

• Nvidia’s market value surpassed $3T, with the AI chip leader soaring past Apple to become the world’s second-most valuable public company behind only Microsoft.

• Meta launched a new AI assistant for businesses in WhatsApp, along with other AI tools aimed at supporting businesses with customer interactions and sales.

• Social platform Cara has grown from 40,000 to 650,000 users in the past week, with the artist-focused, anti-AI social media app seeing an inflow of traffic amid backlash against Meta’s AI policies.

• Jasper AI released Flash Diffusion, a new method for accelerating diffusion models, producing faster, higher-quality image generations while reducing computing costs.

• The U.S. government opened antitrust investigations into Microsoft, OpenAI, and Nvidia, with the FTC and DOJ reportedly focusing on monopolistic practices and investment scrutiny.

• Alibaba’s open-source Qwen 2-72B model moved into the top spot on Hugging Face Open LLM Leaderboard, surpassing rivals Mixtral and Llama-3 across a range of benchmarks.

• Robot Era showcased a demo of its XBot-L AI-powered humanoid robot successfully traversing the Great Wall of China, demonstrating advanced movement and balance capabilities.

• Michael Kors is partnering with Mastercard to offer customers the company’s AI-powered ‘Shopping Muse’ for personalized product recommendations.

• Anthropic posted a new blog detailing the process behind crafting its Claude AI model’s personality, using ‘Character Training’ to help instill curiosity, thoughtfulness, and diverse viewpoints.

• The U.S. Air Force and Space Force launched NIPRGPT, a new AI tool designed to assist with tasks like report summarizing and coding to improve information access across operations.

• The U.S. government is reportedly drafting additional restrictions on China’s access to advanced AI chips, looking to target a powerful ‘GAA’ architecture currently being introduced by chipmakers.

• Canadian PM Justin Trudeau spoke about AI’s potential, saying he believes development should not be slowed and instead focused on building for good to counter the bad.

• Adobe updated its terms of service following backlash over language stating that users' work is used to train AI, with new policies going into effect on June 18th.

• Google announced its June Pixel update, introducing its Gemini Nano AI model for Pixel 8 and 8a users, enabling new features like enhanced transcription capabilities.

• François Chollet and Mike Knoop launched ARC Prize 2024, a $1M coding competition challenging AI developers to create systems that solve new reasoning problems and advance towards AGI.

• Google DeepMind researchers published a new paper arguing that open-endedness — the ability to create new, learnable ‘artifacts’ — is the key to achieving artificial superhuman intelligence (ASI).

• OpenAI announced the addition of two new executives, with former Nextdoor CEO Sarah Friar joining as CFO and Kevin Weil stepping in as CPO.

• Perplexity is facing backlash following the launch of its new Pages feature, with accusations of generated content republishing nearly identical stories to major news outlets without attribution.

• Meta is planning to use European user data to train its AI models while navigating EU data protection laws, aiming to create tech that better reflects diverse languages and cultures.

• Morgan Stanley CEO Ted Pick said that AI could save financial advisors 10-15 hours a week, with the firm currently testing AI chatbot integrations from OpenAI.

• xAI’s Grok may be gaining new image capabilities and the ability to access web search results, according to new code found within the system.

• Flyhomes debuted its AI-powered home search portal, aiming to transform real estate searches with advanced data integration and conversational AI.

• Apple released new details on its on-device AI models, which perform slightly above other small models like Phi-3-mini, Mistral-7b, and Gemma-7b in testing.

• Mistral secured a new $640M Series B funding round, bringing the French AI startup’s valuation to $6B.

• Brazil’s government is hiring OpenAI to leverage AI for analyzing and flagging lawsuits, aiming to reduce court-related costs and reduce that impact on the federal budget.

• Pika Labs rolled out a new model update to its video generation platform, showcasing improved quality.

• Stability AI released the open model weights for Stable Diffusion 3 Medium, with the 2B parameter text-to-image model offering advanced photorealism, prompt understanding, and typography capabilities.

• OpenAI has reportedly doubled its annual revenue to $3.4B over the last six months, according to The Information.

• Midjourney unveiled model personalization, allowing users to better tune outputs to their own styles and preferences by rating images.

• Microsoft released a new research paper presenting DenseAV, an architecture that learns the meaning of words and sounds by watching unlabeled videos.

• Oracle announced new cloud infrastructure deals with OpenAI and Microsoft to ‘extend Azure AI’s capacity’ and scale deployment of AI models.

Closing Thoughts

That’s it for us this week.

If you want a version of this newsletter with pictures etc. Be sure to subscribe below:

https://www.blacklynxbrief.com/subscribe