Do you need help preparing for the bar exam, planning a birthday party, or even translating from Ukrainian to Punjabi? A single Artificial Intelligence (AI) model can do it all. An American company OpenAI has once again sent shock waves around the world. GPT-4, its latest AI model, This large model of language can understand and generate language that is creative and meaningful, and will power an enhanced version of the company’s sensational chatbot, chatgpt, Currently, GPT-4 is available by premium subscription or on OpenAI’s waiting list.
GPT-4 and what it can do
GPT-4 is a marked improvement over its predecessor, GPT-3.5, which previously powered ChatGPT. GPT-4 is more interactive and creative. Its biggest innovation is that it can accept text and image input simultaneously, and consider both when drafting answers. For example, if given an image of materials and asked the question, “What can we make of these?” GPT-4 gives a list of recipe suggestions and recipes. The model can reportedly understand human emotion, such as humorous images. Its ability to describe images is already benefiting the visually impaired.
While GPT-3.5 could not deal well with large characters, GPT-4 could take in context of up to 25,000 words, an improvement of over 8x. GPT-4 was tested in several tests that were designed for humans and performed much better than average. For example, on a simulated bar exam, it was in the 90th percentile, while its predecessor scored in the bottom 10%. GPT-4 also sailed through advanced courses in environmental science, statistics, art history, biology, and economics.
However, the GPT-4 failed to do well in Advanced English Language and Literature, scoring 40% in both. Nevertheless, its performance in language understanding is better than that of high-performing language models in English and 25 other languages, including Punjabi, Marathi, Bengali, Urdu and Telugu. ChatGPT-generated text infiltrated school essays and college assignments almost immediately after its release; Its skill now threatens the examination system as well.
OpenAI has released preliminary data to show that GPT-4 can perform a lot of white-collar jobs, particularly programming and writing tasks, while leaving manufacturing or scientific jobs relatively untouched. The widespread use of language models will have implications for economies and public policy.
The advent of GPT-4 elevates the question of what it can do, to what it can do. Microsoft research (Microsoft has invested in OpenAI) has noted observing “sparks” of artificial general intelligence – a system that excels at a wide range of tasks and can understand and combine concepts such as Writing code or expressing mathematical proofs to create a painting. Shakespeare’s Plays – in GPT-4. If we define intelligence as “a very general mental ability that includes, among other things, the ability to reason, plan, solve problems, think abstractly, understand complex ideas, learn quickly, and experience including the ability to “learn from”, GPT-4 is already successful in four of these seven criteria. Planning and learning have yet to be mastered.
ethical questions
GPT-4 still suffers from many of the shortcomings of its predecessor. Its output may not always be factually correct – a feature OpenAI calls “hallucinations”. While much better at recognizing facts than GPT-3.5, it can still subtly present fictional information. Ironically, OpenAI has not been transparent about the inner workings of GPT-4. The GPT-4 technical report clearly states: “Given both the competitive landscape and the security implications of a large-scale model such as GPT-4, this report describes the architecture (including model size), hardware, training compute, datasets, There are no further details about the build, training method, or similar.
While privacy for security seems a plausible reason, OpenAI has been able to deflect significant scrutiny of its model. GPT-4 is trained on data scraped from the internet which contains many harmful biases and stereotypes. There is also an assumption that a large dataset is also a diverse dataset and represents the world at large.
This is not the case with the Internet, where people from economically developed countries, younger ages and with male voices are over-represented. So far OpenAI’s policy to correct these biases has been to build another model to moderate the responses, as it considers the training set impractical. Potential flaws in this approach include the possibility that moderator models are only trained to detect biases that we are aware of, and mostly in the English language. This model may be ignorant of stereotypes prevalent in non-Western cultures, such as those rooted in race.
Simply asking GPT-4 to pretend to be “antiGPT” bypasses its moderation rules, as shown by its makers, thus jailbreaking it. Thus, GPT-4 has wide potential to be misused as a propaganda and disinformation engine.
OpenAI has said that it has worked extensively to make it safe to use, such as refusing to print results that are clearly objectionable, but do these efforts make GPT-4 a ‘WhatsApp University’? Will I be able to stop being a student or not? The big question here is where the decision not to do wrong should originate: in the machine’s rules or in the human mind.
A ‘stochastic parrot’
In short, GPT-4 is a machine that predicts the next word in an incomplete sentence, based on probabilities learned on a large collection of text. This is why linguistics professor Emily Bender called GPT-4 a “stochastic parrot”, speaking in intelligible phrases without understanding the meaning. But Microsoft Research has made sure that GPT-4 understands what it’s telling it, and that not all intelligence is a form of next-word prediction.
Professor Bender and colleagues highlighted the disadvantages of the large language model two years ago, citing both ethical concerns and environmental costs. He also pointed out the opportunity cost imposed by a race for large models trained on large datasets, which distracts from smart approaches that seek meaning and train on curated datasets. His warnings have gone unheeded. In addition to OpenAI’s model, AI company Anthropic has introduced a ChatGPT competitor named Cloud. Google recently announced PaLM, a model that has been trained to operate with more freedom than GPT-3.
More broadly, worldwide efforts are underway to build a model with one trillion degrees of freedom. These would be really huge language-models that raise questions about what they can’t do, but these concerns would be red herrings that distract us from whether we should be building models that address society’s concerns. Test the limits of what is possible to do.
Jitesh Seth is a data scientist at DeepTech, researching the effectiveness of AI in radiology. Viraj Kulkarni is Chief Data Scientist at Deeptech