AI Continues Its Rapid Advance
VR advances will transform education, life, and work
*Rapid AI adoption (way faster than any technology in history, and it's a general-purpose technology).
*(Way) better AR/VR and what that means for education
*Better AI applications
*Better models (with expanding capabilities)
*Better training data
*Better robots
*Beyond just predicting words
In one year, we've moved way beyond your "parents’ AI."
https://open.substack.com/pub/stefanbauschard/p/ai-continues-its-rapid-advance?r=2flwzd&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
This is a look at some of the major advances in AI over the last two weeks.
The goal is to give educators a sense of the progression. Remember, we are just in the first inning of AI technologies, and these will seem like toys in two years.
Fast integration and use
Note that the last chart is only ChatGPT and only goes through June.
This is not surprising, as we already have the technologies we need to use AI (phones, laptops, desktops, Meta glasses, and VR systems).
Electricity took 37 years because cities had to be wired. PCs had to be bought. Then we had to figure out how to make the internet work in phone lines and rewire in many instances. Smartphones had to be bought.
AI is just delivered through all of the above.
We cover this in more detail in our report.
Increased Access
Generative AI is simply available in more places.
Free. Through heycami, users can directly access Perplexity.ai on Whats App, which enables real-time semantic web search. You can communicate with it through text or voice. For example, I asked it about the weather in my hometown, and it answered correctly.
Expensive but cool and transformative. Apple released their Apple Vision Pro, which integrates both Open AI’s ChatGPT and perplexity.ai directly into the augmented reality headset. This enables individuals to directly interact with both of these tools in this space.
This Vision Pro is currently very expensive ($3500), but as affordability grows and more students have access, you will start to see the development of more immersive tutoring and teaching systems that will quicken the learning pace.
The Vision Pro is obviously state-of-the art, which allows you to see what is coming, but even current VR technology and platforms are having a huge impact.
Just think about how what is in this list above is going to transform teaching and learning.
Midjourney (the company that generates some of the most photorealistic images) just hired one of the Apple Vision Pro developers to be its Head of Harwdware. So, it seems they will be building completely immersive experiences that are “real life.”
Imagine what it will be like when MidJourney is on V9 and has its own “Vision Pro” and can completely immerse you in your own movie, classroom, or historical setting (Imagine your students as participants in a “real” Constitution Convention or participating in a debate tournament in an immersive environment).
Ease of Use
ChatGPT. Directly from the home screen, you can now call one of the thousands of GPT assistants to help you complete your task. And, of course, you can easily create your own GPT assistant.
These do require a paid ChatGPT account, but you can also make your own free bots at Poe, and now GitHub.
It’s pretty easy for someone to create their own virtual assistant to do whatever they want.
Robots
AI is moving fast from software applications like ChatGPT to physical applications like robots. According to a new report from Bloomberg, humanoid robot maker Figure AI is in talks with Microsoft and OpenAI to lead a funding round that could raise as much as $500 million at a valuation of $1.9 billion.
Many experts anticipate 2024 as a pivotal year for robotics, with expectations of major advancements akin to the impact of ChatGPT. Jim Fan, a Senior Research Scientist, and AI Agents Lead at Nvidia, suggests we're nearing a significant robotics innovation, propelled by advances in hardware and AI technology.
Figure AI positions itself as a pioneering AI robotics firm, aiming to introduce a versatile humanoid robot. The company envisions its robots boosting productivity, alleviating labor shortages, and reducing hazardous employment. It aims to serve sectors like manual labor, domestic assistance, and space exploration, as detailed on its website.
The robotics field is witnessing swift developments elsewhere too. Tesla is developing its humanoid robot, Optimus, and Norwegian firm 1X Technologies has recently secured a $100 million investment, indicating a growing interest and investment in robotic technologies.
See more below.
Better Tools
AI is getting better every day at what we want to use it for.
Real-time web search. The AI-driven search application Perplexity has climbed into the top 10 rankings of productivity apps on the App Store, surpassing popular applications such as Google Calendar and Microsoft Word.
A better browser. After Perplexity made a splash recently with its AI-powered search engine, The Browser Company has just announced another AI-powered alternative that claims to provide a better online search experience.
It’s called Arc Search and it combines a search engine, a browser, and AI into something new and different. Whereas traditional search engines return a list of links you have to click through to find what you’re looking for, Arc Search looks through these links itself and creates a unique webpage with more intuitive answers made just for you.
Essentially, Arc Search saves you time and effort by searching on your behalf and organizing all the relevant information into a new webpage. CEO Josh Miller claims that Arc Search provides a 2x faster search experience than Safari and Chrome.
Better Google AI. Google’s Bard has been updated to create images from basic text descriptions. Additionally, it operates on Gemini Pro, Google's intermediate model, significantly enhancing response quality, and is now available in at least 40 languages. Google Maps is also receiving an update to include conversational search, offering users AI-driven suggestions for exploring new places. Bard will also summarize YouTube videos. Simply go to bard.google.com and type summarize [URL of the video].
There are rumors that Gemini Ultra will be out this week.
Better models. A mysterious AI model called ‘Miqu‘ was leaked on data science platform HuggingFace last week. AI models are increasingly common, but a standout detail has caught widespread attention: early evaluations placed this model fourth on the EQ-Bench, a benchmark assessing AI emotional intelligence, outperforming all but OpenAI's GPT-4. This marks a significant achievement for the open-source AI community, contrasting with the proprietary models from major entities like OpenAI. Although its origins are under discussion, it's widely believed to be a leaked model from Mistral, a French AI startup known for its contribution to open-source AI.
Mistral’s CEO seemed to confirm these assumptions in a recent post and told readers to “stay tuned,” implying that Mistral is releasing an official version of the model soon, potentially with performance upgrades that bring it closer to, or exceed, the performance of OpenAI’s GPT-4.
Student Use
The Higher Education Policy Institute in the UK's study reveals a significant trend: 53% of students are using generative AI for assignment preparation, demonstrating its normalization in higher education. Interestingly, only 5% have used AI to cheat, defined as submitting AI-generated text without modification. This suggests a predominant use of AI as a learning aid rather than a means to circumvent academic effort, though we don’t know how much they modified it. Moreover, the projection that 73% of students expect to continue using AI post-graduation highlights its perceived value in the professional world. This points to a broader recognition of AI's utility, necessitating educational and policy adjustments to ensure responsible use and integration of AI skills into future workplaces.
Models and moving beyond just predicting words.
The foundation of LLMs is next work/token prediction, and there is a lot of criticism of LLMs as only being able to predict next words/tokens (most famously the 2020 and 2021 articles by Emily Bender et al.).
While the foundation of LLMs is next word/token prediction, language models have now been trained for basic inference reasoning, and more advanced reasoning abilities are starting to be demonstrated due to expanded training methods. It’s reasonable to expect material improvements in this in GPT5 and Gemini Ultra.
Beyond that, these models are starting to be able to work with symbolic models and other models (ChatGPT4 is considered to be a Mixture of Experts (probably at least 8 models)). There are also entirely new models being worked on (objective-driven world models) that can reason and plan. Those will either end up working with or replacing language models. And the models do other things, such as work with RAG databases that are improving.
Even other leading scientists who are critics of LLMs (Yejun Choi, Yann LeCun, Melanie Mitchell) agree that LLMs have at least very limited reasoning abilities, and LeCun constantly argues that AI tools that are not LLMs will develop advanced reasoning and planning abilities in the not-too-distant future.
Anyhow, changes in education take a long time. Educators need to think about preparing students for a world where AIs can reason and climb toward human-level intelligence, not just next-word predictors.
The idea that current AIs only predict the next word and will only continue to be able to predict the next word, should stop being shared.
Moving Beyond ““Garbage” Data
There is also this idea that the language models have been trained on bad data, such as Reddit. This is true; it was trained on some screened Reddit data (not every random thing a person said on Reddit) and I’m sure there's still some bad data in the models (as there is in the articles people find through Google searches), but now the most advanced models have been trained on the entire public internet (around 10 trillion words), and developers are focusing on training the models on high-quality, often proprietary data. This training + RAG advances + other methods (RLAIF, for example) mean the training data will keep getting better.
Anyhow, these models are getting better and better every day and all this talk about what they can’t do at the moment, especially when it’s no longer true, is holding appropriate educational adoption back and not moving it forward.
AI in More Places
Amazon is testing a new AI shopping assistant, “Rufus" that answers shopping questions, recommends products, and compares items.
Of Note
OpenAI released a study on GPT-4’s effectiveness in potentially creating a bioweapon and found that the risks aren’t that high.
Robots.
Anyhow, no matter how intelligent it is, it’s going to change the world, and we’ve barely scratched the surface of what we can do with even the current technologies.
For example:
