What’s next for LLMs like GPT-4 and chatbots like Bing Chat, Google Bard and OpenAI’s ChatGPT?
ChatGPT and GPT-4 remain the dominant themes in the world of generative artificial intelligence (AI) this week, and they will continue to do so for some time to come. Even India may soon launch a ChatGPT-like AI-powered chatbot, going by the recent hint of Union minister for electronics and information technology, Ashwini Vaishnaw. Here are five other important developments in this space.
1. AGI or not? “Strikingly close” to human-level performance?
In a 154-page document titled ‘Sparks of Artificial General Intelligence: Early experiments with GPT-4’, Microsoft researchers begin with the contention that the fourth version of the generative pre-trained transformer model (GPT-4) is part of a new cohort of large language models (LLMs), including ChatGPT and Google’s PaLM, that exhibit more general intelligence than previous AI models. They also suggest that “beyond its mastery of language, GPT-4 can solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more, without needing any special prompting”. And in all of these tasks, GPT-4’s performance is “strikingly close to human-level performance...”, which makes the Microsoft researchers believe that GPT-4 “could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system”. You may read the entire paper here.
I believe that even as such developments cause a stir because these AI models excel humans in specific tasks, the current developments do not necessarily indicate that a true AGI system is anywhere close to the ones that are shown in movies such as Skynet, Terminator, Transformers, iRobot, HER, etc.
I have written in detail about this topic so you may read my views here:
I have argued earlier that it’s tempting to label such developments as the first steps towards AGI (as Microsoft researchers have done in the paper cited above) because machines are indeed becoming extremely adept at narrow AI (handling specialized tasks) and give us the impression that they can think like humans -- GPT-4 being a very good case in point. To be sure, over the past few years, we have seen that AI excels at controlling your spam; improving the images and photos you shoot from your cameras; translating languages and converting text into speech and vice versa on the fly; helping doctors diagnose diseases and finding use in drug discovery; helping astronomers to look for exoplanets while simultaneously assist farmers in predicting floods; and now even helping in spotting some early signs of earthquakes.
However, while such tasks may tempt us to ascribe human-like intelligence to machines, we must remember that even driverless cars and trucks, however impressive they look or sound, are still higher manifestations of “weak or narrow AI” -- in other words, they do not think for themselves but prove intelligent when making predictions from data (LLMs, of course, have to be pre-trained). You may read more of this here.
Meanwhile, hundreds of notable people, including Elon Musk, Gary Marcus and Andrew Yang, have called for a six-month moratorium on training systems that are “more powerful than GPT-4”. In an open letter, they conclude that “...Powerful AI systems should be developed only once we are confident that their effects will be positive, and their risks will be manageable...”
That said, the developments in the field of AI are too rapid for anyone to stick to any one point of view. But my humble submission to CXOs is not to get distracted by AGI for now at least, and instead focus on how the power of LLMs like GPT-4 and LLM-powered chatbots like ChatGPT can be leveraged by individuals and companies with the right policy frameworks to prevent its abuse. That said, the claims of AGI are something that policymakers and governments must take very seriously.
2. Google Bard is live in the US and UK
On 21 March, Google announced that “We’re starting to open access to Bard, an early experiment that lets you collaborate with generative AI. We’re beginning with the US and the UK and will expand to more countries and languages over time.” Bard is powered by a “lightweight and optimized” version of LaMDA (Google’s research LLM), and will be updated with newer, more capable models over time.
Google wants us to think of an LLM as a prediction engine. When given a prompt, it generates a response by selecting one word at a time from words that are likely to come next. Picking the most probable choice every time wouldn’t lead to very creative responses, so there’s some flexibility factored in. Google acknowledges that while LLMs are an exciting technology, “they’re not without their faults,” including biases and stereotypes and inaccurate, misleading or false information while presenting it confidently. And as Google puts it, Bard is a doorway to Google Search. To join the waitlist (presuming you’re in the US or UK), you have to use your personal Google account (run by you and not a parent, guardian or admin); be 18 years of age or older; and have supported web browser (Chrome, Chromium-based Edge, Firefox, Opera or Safari).
3. Google Bard Vs Bing Chat Vs ChatGPT: Early impressions
It’s early days for Bard and Bing since ChatGPT clearly had the first mover’s advantage and now has a little over 100 million users. But unlike ChatGPT (which still has October 2021 as the data cut-off date unless you use the new plugins), Bard and Bing Chat have access to current data. Bard is not available in India, so you can check out some comparisons here. I have been using Bing Chat, on the other hand, for a while. The interface is clean, and Bing Chat’s own prompts are smart too. The best part is that the data is provided with references so that you can verify the sources yourself if you doubt the information. ChatGPT, on its part, does not provide references unless you use it with the new plugins. Bard can access Google’s search engine. ChatGPT has no internet access, so Bard competes more with Bing Chat--Microsoft’s new AI-powered Bing that uses GPT-4.
4. ChatGPT Plugins
Plugins are chat extensions designed specifically for language models like ChatGPT, enabling them to access up-to-date information, run computations, or interact with third-party services responding to a user’s request. Developers can create a plugin by exposing an API through their website and providing a standardized manifest file that describes the API. ChatGPT consumes these files and allows the AI models to make calls to the API defined by the developer.
OpenAI says plugins can be the “eyes and ears” for language models by giving them access to information that is recent, personal, or specific (information that is not typically included in the training data). ChatGPT, for instance, now has a browser plugin that helps it retrieve current data from the web. OpenAI claims its “code interpreter” plugin is especially useful in solving mathematical problems, doing data analysis and visualization, and converting files between formats.
OpenAI also has an open-source “Retrieval” plugin that enables ChatGPT to access personal or organizational information sources (with permission) from their data sources, such as files, notes, emails or public documentation, by asking questions or expressing needs in natural language. As an open-source and self-hosted solution, developers can deploy their own version of the plugin and register it with ChatGPT.
OpenAI’s first ChatGPT plugins have been created by some companies, including Expedia (travel itinerary, places to see, things to do), FiscalNote (Provides and enables access to select market-leading, real-time data sets for legal, political, and regulatory data and information), KAYAK (Search for flights, stays and rental cars. Get recommendations for all the places you can go within your budget) and Wolfram (Access computation, math, curated knowledge & real-time data through Wolfram|Alpha and Wolfram Language).
OpenAI also has a “third-party” plugin. We will see many more plugins as OpenAI gets feedback from developers and users.
Source: Company data. Picture courtesy of Credit Suisse’ ChatGPT: Unlocking the Potential of Large Language Models’
5. AI deception for sale: FTC’s word of caution
The US Federal Trade Commission (FTC) cautioned companies that before they develop or offer a synthetic media (colloquial term for chatbots developed using LLMs) or generative AI product, they should account for its misuse or fraud or any other potential harm at the design stage itself. “Then ask yourself whether such risks are high enough that you shouldn’t offer the product at all”. If companies still decide to make or offer such a product, FTC recommends that they should take all “reasonable precautions” before it hits the market. According to the FTC blog, “...Evidence already exists that fraudsters can use these tools to generate realistic but fake content quickly and cheaply, disseminating it to large groups or targeting certain communities or specific individuals. They can use chatbots to generate spear-phishing emails, fake websites, fake posts, fake profiles, and fake consumer reviews or to help create malware, ransomware, and prompt injection attacks. They can use deep fakes and voice clones to facilitate imposter scams, extortion, and financial fraud. And that’s very much a non-exhaustive list.” The FTC is watching, and so should policymakers.
QUOTE OF THE WEEK
Jim Fan @DrJimFan (Nvidia AI scientist)
“When I was 20, the barrier of entry to AI was sky-high. Had to learn a ton of overly complicated methods and work with badly shaped tools. Now the barrier is so low that English is the new programming language. If this isn’t the time to jump in, I don’t know when it is.”
Pocket-friendly device to detect milk adulteration in 30 seconds
IIT-Madras says it has developed a 3D paper-based portable device that could be adapted for use in homes as an instant remedy to prevent the consumption of adulterated milk. The device, according to the premier institution, can detect multiple substances commonly used as adulterating agents, including urea, detergents, soap, starch, hydrogen peroxide, sodium-hydrogen-carbonate and salt. The research paper has been published in the journal Nature. Milk is one of the most vital foods important to lead a healthy lifestyle, and yet it is the most adulterated food item in the world.
Picture courtesy of Livemint
The adulteration of milk is a growing menace, especially in developing countries like India, Pakistan, China, and Brazil. Consumption of adulterated milk could cause medical complications such as kidney problems, infant death, gastrointestinal complications, Diarrhoea, and even cancer.
Answering repeated questions can put you at risk online
People reveal more personal information when you repeat the same questions, according to new research from the University of East Anglia. The research team asked 27 study participants for a range of personal information online, including their height, weight and phone number as well as their opinions on topics including immigration, abortion, and politics. The participants then structured the questions from the least to most intrusive and were asked how much of their personal information they would ‘sell’ to be made available on a specific website for two weeks. They then asked them again how much information they would sell -- to appear for a further two weeks -- for a chance of even more money. In a second larger online study, 132 participants were asked how much information they would sell at two-time points, as well as a range of personality questions.
According to the researchers, the first study revealed that asking for real personal data led to increased information disclosure when repeated. The second study replicated this effect and found no change in people’s associated concerns about their privacy -– people change their behaviour but not their view. “This demonstrates that simple repetition can make people over-disclose, compared to their existing and unchanged concern,” the researchers said in a press statement. The study titled ‘Tell Me More, Tell Me More: Repeated Personal Data Requests Increase Disclosure’ has been published in the Journal of Cybersecurity.