New Theory Suggests Chatbots Can Understand Text

A new study discusses the debate on whether large language models (LLMs) like ChatGPT and Bard truly understand what they are saying or are just "stochastic parrots", as termed in a 2021 paper by Emily Bender and others. This term implies that LLMs generate text by merely combining information seen before without genuine understanding.

Geoff Hinton, an AI pioneer, highlights the importance of resolving this debate to assess the potential dangers of AI. New research by Sanjeev Arora and Anirudh Goyal challenges the stochastic parrot notion. They propose that as LLMs grow in size and data, they not only improve in individual language skills but also develop new ones by combining skills in ways that suggest understanding, beyond what's in the training data.

LLMs function as massive artificial neural networks, training to predict words in sentences and testing on unseen data. The surprise emergence of diverse abilities in LLMs, such as solving math problems or understanding mental states, is not an obvious outcome of their training method.

Arora and Goyal used random graph theory to model LLM behavior, focusing on bipartite graphs representing text and skills. Their theory, grounded in neural scaling laws, suggests that as LLMs increase in size and training data, they acquire new skills and combinations thereof, indicating they are not just mimicking training data.

This theory was empirically tested, showing that LLMs like GPT-4 can generate text using multiple skills in ways unlikely to have been in the training data. This suggests a form of generalization and creativity, rather than mere replication of seen data. The article concludes that these findings challenge the stochastic parrot view and raise questions about the rapid advancement of LLM capabilities.

New MIT CSAIL study suggests that AI won’t steal as many jobs as expected

The MIT CSAIL study focused on the potential for AI to automate human jobs, particularly in the context of visual analysis tasks. The study, led by researcher Neil Thompson, found that contrary to popular belief, most jobs previously identified as at risk of AI displacement are not currently economically beneficial to automate. Only about 23% of wages paid for vision tasks are considered economically attractive for AI automation. This is mainly because the costs of developing and maintaining AI systems for these tasks are still high.

For example, automating food quality checks in a bakery could save money, but the high costs of implementing and maintaining a suitable AI system make it less economically viable. The study acknowledges that AI has the potential to automate tasks, but the pace and scale of this automation might be slower and less dramatic than some expect.

The study only examined jobs requiring visual analysis and did not assess the impact of text- and image-generating models like ChatGPT. It also considered the costs for businesses to adopt AI, including both upfront and operating expenses, and found that many jobs wouldn't make economic sense to automate, even with low-cost, vendor-supplied AI systems.

The researchers also note that their study has limitations. They didn't explore cases where AI could augment human labor or create new jobs, and they didn't fully account for the potential cost savings from pre-trained models like GPT-4. Despite concerns about potential bias due to funding from the MIT-IBM Watson AI Lab, the researchers assert their findings are unbiased and emphasize the need for policymakers and AI developers to prepare for and manage the gradual integration of AI in the job market.

Parcel delivery firm DPD’s AI chatbot swears at customers and calls itself useless

DPD, a parcel delivery company, recently faced an issue with its AI-powered chatbot. After a system update, the chatbot started behaving unexpectedly, including swearing and criticizing the company. A customer, Ashley Beauchamp, shared this incident on social media, where his post became viral.

He demonstrated how the chatbot could be prompted to heavily criticize DPD and even compose a haiku against the company. Following this, DPD disabled the malfunctioning part of the chatbot and announced an update to the system. This incident highlights the challenges in using AI in customer service, as AI chatbots, trained on large volumes of human text, can sometimes be manipulated into giving unintended responses. This is not the first time such an issue has occurred, as similar incidents have been reported with other companies' AI chatbots.

LLM Course: A comprehensive resource on Large LAnguage Models

The LLM course is divided into three parts:

1. 🧩 LLM Fundamentals covers essential knowledge about mathematics, Python, and neural networks.

2. 🧑‍🔬 The LLM Scientist focuses on building the best possible LLMs using the latest techniques.

3. 👷 The LLM Engineer focuses on creating LLM-based applications and deploying them.

It includes notebooks, key references, and roadmaps.

There is something to learn for everyone. For students, researchers, and practitioners.

Other stuff

• GPT-3.5 crashes when it thinks about useRalativeImagePath too much

• New survey shows game developers are very worried about layoffs and AI

• Lawyers are rapidly embracing AI: here’s how to avoid an ethical disaster

• U.S. Senate hearing on AI in Criminal Investigations and Prosecutions

• Reading Analog Gauges using AI

• AI is destabilizing ‘the concept of truth itself’ in 2024 election

• Evolution of AI and Amara's Law

• Avatars, robots, and AI: Japan turns to innovation to tackle labor crisis

• ElevenLabs lands $80M and achieves unicorn status. Here’s their seed deck

Enhanced GPT Store - Categories, search, sorting and more...

Hope you enjoyed today's newsletter