Nvidia reports this Wednesday evening, August 28 at 5 PM EST. Then we’ll know if my enthusiasm paid off. And Nvidia’s stock continue s to climb. Fingers crossed.
Sign up for the Nvidia webcast here.
My continuing AI research focuses on three areas:
+ Is AI useful to you, me and our friends in the corporate world? The answer is a resounding YES. But not everybody will adopt it at the same rate — just as they didn’t adopt electricity at the same rate. Try this: The first commercial alternating current central station in the United States started in 1886. In 1925, only half of American houses had electrical power. It took until 1960 before virtually all the homes in the U.S. had electricity. That’s 74 years. AI is happening faster. Much faster. Thank the Lord.
+ Which are the companies who are benefiting the most? Which ones should we invest in? Nvidia is number one. So far, I haven’t found a sound second one — though all my friends (and me) are desperate for the magic AI beneficiary. Likely candidates include Dell, SMCI, Apple, Oracle, Microsoft, Facebook (Meta), Amazon, Google, Broadcom, Lenovo and Taiwan Semiconductor. But all these have other flourishing businesses. AI is a blip, albeit nice blip. But still a blip.
+ Since AI is so powerful, how should we, the unwashed, approach it personally — without being overwhelmed by the complexity and the cost (in time and money)?
We have to start with one question: What’s the one thing I could improve that would make the biggest positive difference to my life and my business?
If I were running a business selling stuff online, I’d want to fix my chatbot, which are universally awful and annoy far more of our customers than they make happy.
I don’t run a business. I do research to find the next great thing/company to invest in? I like Perplexity. Frankly, I’m still learning the questions to ask. But Perplexity is good. Everyone’s getting better, even Google.
If were a travel agent, I’d use ChatGPT, Perplexity and other AI engines to help me plan my clients’ next trips.
If I were inside a corporation managing a group of people, I’d use Microsoft’s Teams/Copilot. It arranges video meetings, records the meetings, summarizes what was said, sends out a report and comes up with some ideas of what should be done next.
My best technology researcher is Richard “Zippy” Grigonis. I have worked with Zippy for over 30 years. I asked him,” Where now with AI?” Here’s his excellent response:
Where Do We Stand Now with AI?
The AI field is currently experiencing rapid and at times bewildering developments across several key areas:
- Generative AI: Smaller Models, Open-Source Models: The trend continues towards smaller, more efficient versions of Large Language Models (LLMs) that can run on less powerful hardware, making AI more accessible and cost-effective for a broader range of applications. This shift is partly driven by increasing GPU shortages and rising cloud computing costs. The smaller AI models tend to be open source, so minor developers and even hobbyists with beefed-up PCs are tinkering with chatbots, looking for business opportunities.
The “holy grail” for small versions of AI models is to run them entirely on consumer-end hardware, particularly smartphones, with no Internet connection. This is difficult, since there are issues with power, thermals and system memory. Moreover, smartphones usually don’t have a dedicated graphics card, but all phones do have an integrated GPU.
Even so, Apple researchers managed to run large AI models with highly limited system memory in their so-called “LM In A Flash” study. The Falcon 7B AI model, with its 7 billion parameters, previously took 14GB of memory, but an iPhone 15 Pro with 8GB was able to run it, and it ran 90% faster than before thanks to Apple’s new techniques.
In the Android world, you can download MLC LLM, a program that deploys and load models and then run them inside of the app using MLC Chat. MLC has an iOS version too, available on the App Store. Hundreds of models are available to run on the Hugging Face community site. Downside: because it relies solely on the CPU the performance is slow—about three tokens per second on older devices. Fortunately, newer devices like the Samsung Galaxy S23 Ultra (powered by the Snapdragon 8 Gen 2 chipset) are optimized to run the MLC LLM / MLC Chat app.
Meta is also working on a solution called MobileLLM, a compact LLM model designed for phones. Their Meta Reality Labs say their research is spurred by their estimate that users may soon be spending more than an hour daily either in direct conversation with chatbots or having LLM processes running in the background.
- Free, Free, Free: On July 23, 2024 Meta released Llama 3.1 LLM, the largest free open-source AI system to date with 405 billion parameters. Llama 3.1 is Meta’s most advanced model which is comparable to market leaders ChatGPT-4o and Claude 3.5 Sonnet, which can be accessed for free over at https://claude.ai/new. As if that isn’t enough, the renowned Midjourney image generator is allowing anyone visiting their dedicated website to create up to 25 images for free through a free trial. Ideogram 2.0 also debuted, their most advanced text-to-image model, also available to all users for free.
- Responsible AI and Ethical Concerns: With AI becoming more embedded in critical sectors like healthcare, finance, and public services, there’s a growing emphasis on responsible AI. This includes ensuring transparency, fairness, and safety in AI systems. Concerns about the misuse of AI, such as deepfakes and misinformation, are also driving discussions around better governance and regulation.
For example, BBC Panorama recently discovered dozens of deepfakes portraying black people as supporting Donald Trump for president. Trump also posted “I accept!” on his Truth Social account, along with a whole slew of Taylor Swift images, many of which were AI-generated. One shows Swift as Uncle Sam with the text, “Taylor wants you to vote for Donald Trump.” The other photos depict fans of Swift wearing “Swifties for Trump” T-shirts. Earlier, Trump had shared an AI-generated image depicting Kamala Harris holding a communist military rally at the Democratic national convention, as well as a deepfake video of him dancing with X owner Elon Musk, who has endorsed him.
- Copyright and Legal Challenges: The legal landscape for AI is evolving, particularly around the use of copyrighted material for training AI models. High-profile lawsuits like The New York Times against OpenAI have overshadowed those of Getty Images, Universal Music Group and several class action suits representing book authors, visual artists, and software developers filed against Alphabet, Anthropic, Meta, Microsoft, OpenAI and Stability AI. Everybody is challenging the legality of using copyrighted works without explicit permission, which could have significant implications for the future development of generative AI systems.
- Automation and Job Impact: AI continues to automate many aspects of white-collar work, particularly in repetitive tasks like data entry and customer support. However, rather than replacing entire job functions, most experts think that, at least in the short-term, AI is more likely to augment human work, allowing people to focus on higher-impact tasks.
- AI Scientist Learns the Scientific Method: In June 2024, ex-OpenAI researcher Leopold Aschenbrenner posted a 50,000-word, 165-page paper, entitled “Situational Awareness: The Decade Ahead.” In it, Aschenbrenner claims that at around 2027 we will achieve “automated AI research” where AI models will be able to improve and expand their knowledge and capabilities on their own, without human help or limitations. When this happens, an “intelligence explosion” will occur, rapidly leading us to the world of superintelligent AI.
Well, guess what. On August 13, 2024, a famous AI lab in Japan, Sakana AI, published a paper entitled, “The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery.”
The AI Scientist software, developed by Sakana AI in collaboration with the University of Oxford and the University of British Columbia, is an amazing advancement in fully automated scientific discovery. This system leverages LLMs to automate the entire research lifecycle, from generating research ideas and writing code to conducting experiments, analyzing results, and producing full scientific manuscripts. It even includes an automated peer review process that evaluates and improves the generated papers with near-human accuracy.
The AI Scientist can continuously iterate and refine ideas, building upon previous work in an open-ended fashion, much like the human scientific community. It’s been successfully applied to machine learning research, discovering novel contributions in AI image generation and transformer-based LLMs. Despite some flaws in the papers produced, the system demonstrates the potential to democratize research by significantly reducing the cost and time required to produce scientific papers—approximately $15 per paper.
And so, we now enter a new era of scientific discovery, where AI can independently conduct research, accelerating scientific progress and creativity in addressing complex scientific and global challenges, as well as improving AI itself… Which eventually leads us to the scary world of superintelligent AI.
Aside from the AI Scientist, AI’s role in scientific discovery, especially in fields like material science and biomedicine, has been expanding all along. These applications are not only advancing our understanding but are also expected to lead to breakthroughs that could have immense societal impacts. For example, an article published April 25, 2024 in Science, describes a collaboration between Pfizer and the Research Center for Molecular Medicine of the Austrian Academy of Sciences (CeMM) that has led to a new AI-driven drug discovery method that could make it faster and easier to identify small molecules with therapeutic potential.
- Intelligent Search: Intelligent search platforms like SearchGPT, Perplexity, You.com, Elicit, Knowledgie and similar AI-powered search tools represent a potentially big advance in the search engine market, offering more personalized, contextual, and precise results compared to traditional search engines like Google or Bing. These platforms use LLMs and advanced natural language processing (NLP) to understand and generate human-like text, making them capable of answering complex queries, summarizing information, and providing insights beyond basic keyword-based search.
Unlike traditional search engines that primarily rank results based on SEO (Search Engine Optimization) metrics, intelligent search platforms focus on understanding the intent behind queries, which can lead to more accurate and contextually appropriate results. This capability positions them as valuable tools for users who need more than just a list of links, such as researchers, professionals, and those in complex fields.
These intelligent search platforms are well-positioned to dominate niche markets where specialized knowledge and detailed answers are needed. For example, industries like healthcare, legal, finance, and academia could greatly benefit from AI-driven search tools that quickly provide precise and trustworthy information.
Intelligent search could completely disrupt the traditional search engine market by offering superior user experiences, which would probably lead to a shift in how people access information online.
- The Fear that AI is Hitting a Ceiling: On his YouTube channel that follows developments in the AI world, Matt Wolfe examined nearly all of the current AI image generators: MidJourney 6.1, Idogram 2.0, Mystic, Phoenix, Flux.1 (Grok), DALL-E 3, SD3, Firefly 3, Meta Emu, Imagen 3, and Playground v3. He gave them similar prompts and put the results in a big chart in Figma. Commenting on the various modes, he found that, “…to be honest, they’ve all kind of caught up to each other.”
This statement in a roundabout way reflects concerns about the capabilities of LLMs potentially hitting a ceiling and the associated fear that the substantial investments in them may not yield sufficient return on investment (ROI).
As LLMs like ChatGPT, Claude and Llama have grown in size, improvements in performance have been less dramatic compared to earlier iterations. For example, while moving from GPT-2 to GPT-3 showed substantial gains, the leap from GPT-3 to GPT-4 was less pronounced, leading to concerns that simply increasing the number of parameters may not continue to yield significant advancements.
Medium did an AI piece
Generating New Yorker-Style Cartoons with AI
Harnessing the power of OpenAI’s ChatGPT to Transform Cartoon ideas into Reality
For the Medium piece (with other fun cartoons), click here.
Lyme
Friends have Lyme disease. You can suffer Lyme disease for 30 years and more — unless you catch it in the first few days.
My advice: Stay out of the woods. Stay out of your garden.
No matter how much you check, you’ll never find all the tiny little critters that carry Lyme.
If it’s hard to read today’s blog on your phone, go to the web site. Click here.
That’s it for now. — Harry Newton
