What Happened

In 2020, Sathvik Nair and I wrote that, while it was an impressive breakthrough, GPT-3 was not coming for your job. Indeed, GPT-3 did enable impressive work without causing major disruption, but ChatGPT and its ilk feel qualitatively different from their predecessors and, in the opinion of many people watching recent developments, augur badly for workers. 

Read more

What is In-Context Learning (ICL)?

AI has come a long way in recent years (and months!), with systems like ChatGPT demonstrating impressive abilities in solving a wide variety of language-based tasks. Before LLMs, most AI models were still limited by the data they were trained on - they can only perform tasks they have explicitly been optimized for through training. GPT-3 and subsequent LLMs have been able to do something more powerful: learn new tasks and skills simply from new examples in the input, without any gradient updates or changes to the pretrained model. This ability is known as in-context learning (ICL).

Why is ICL exciting?

In-context learning has enormous potential to unlock more flexible, general, and human-like intelligence in AI systems. Some reasons it is generating so much interest:

• Versatility - With ICL, a single model can learn a wide variety of skills at the same time, instead of needing separate training for each one.

• Generalization - ICL allows models to learn underlying rules and patterns from just a few examples, and generalize them to new situations.

• Efficiency - No lengthy or costly re-training of models is needed. Skills can be acquired instantly.

• Accessibility - ICL enables AI systems that can be taught by everyday users through simple demonstrations of the task.

In short, ICL enables LLMs to become powerful systems that can continually learn, reason, and adapt to new tasks. But how does ICL work and where does it come from?

How Does In-Context Learning Work?

Recent research has revealed 3 key factors that enable and enhance in-context learning abilities in large language models, and we’ll go through each one.

Read more

TL;DR

• The Writers Guild of America (WGA), consisting of two labor unions representing over 11,000 writers in various media industries, went on strike on May 2nd. As stated in the WGA’s proposals, the strike aims to achieve higher pay, a dependable payment structure, and more fairness in deals, contracts, and regulations related to artificial intelligence.

• Writers have been picketing the headquarters of major studios including Netflix, Amazon, Warner Bros., Universal Studios, and others, brandishing signs and demanding higher wages. 

• Amongst the many demands regarding streaming residual, the document also discusses concerns stemming from the threat that AI poses to their profession. The writers want regulations in place to safeguard their future in the industry. 

• Advancements in AI, specifically large language models, are quite massive. While its exact impact on various domains and professions may yet be unclear, it is important to discuss AI regulations in the context of providing job security and mitigating negative impacts. 

What Happened

Roughly 11,500 TV and movie writers belonging to the Writers Guild of America (WGA) went on strike on May 2nd for the first time in 15 years after their negotiations with the Alliance of Motion Picture and Television Producers (AMPTP) failed to reach a new contract. AMPTP represents eight major studios: Amazon, Apple, Disney, Warner Bros. Discovery, NBC Universal, Netflix, Paramount, and Sony along with many other producers and distributors.

Read more

Overview

• AI algorithms like DeepArt and DeepDream pioneered and ushered in new artistic possibilities, blurring the lines between human and machine creativity.

• Now, modern image and video-generation models like DALLE 2, Stable Diffusion, Midjourney, Imagen, …

Read more

We have come a long way from the days when senators described the internet as a “series of tubes”. At last month’s Senate Judiciary Committee hearing on AI, Senator Richard Blumenthal kicked off the opening statements with a deepfake of his voice and a statement written by ChatGPT, and senators spent three hours talking with top AI leaders about issues of transparency, privacy, and more. 

The past half-year of AI frenzy seems to have resulted in a genuine desire for lawmakers to come together with industry to understand this new technology and develop standards to promote its safe development. Some, in fact, were quick to point out that the interaction seemed almost too friendly. The three key witnesses — OpenAI CEO Sam Altman, IBM Chief Privacy & Trust Officer Christina Montgomery, and New York University Professor Gary Marcus — faced a much warmer environment compared to the more antagonistic congressional hearings of past tech executives like Mark Zuckerberg. While it’s possible that this unprecedented friendliness signals better cooperation between government and industry, many analysts warn that it indicates a high likelihood of regulatory capture. 

Yet if there’s one thing known to be true in American politics, it’s that the path from a policy discussion to a newly adopted bill is a long one. With top tech companies pouring nearly $70 million into congressional lobbying last year, and the vast majority of bills never making it past committee, the latest hearings are only the beginning of the long road to come.  

To see where AI regulation may be headed, it’s helpful to look into the subtle aspects of this hearing and see what questions are still left unanswered. Finding points of agreement is valuable, but it’s in the areas of uncertainty where we have the greatest opportunity to actively shape the discourse — ensuring that it’s well-informed, productive, and focused on the issues that matter to all of us. 

Here’s what still remains to be debated: 

1. Which concerns will be lawmakers’ primary focus?

While developing standards for the responsible use of technology appears to be a common goal, the scope of proposed regulations will vary significantly depending on which problems they aim to target. The seven-minute questioning rounds converged towards many common themes, showing hope for future government action on fronts like misinformation and system transparency. 

Read more

TL;DR: Transformer models, and therefore ChatGPT, Bing/Sydney, and their ilk still have fundamental limitations; I think these are important barriers to something we might call “AGI”—but I also think we might consider jettisoning the term and just saying what we mean. We should be impressed by what’s possible now, but be aware of limitations, think clearly about what these limitations mean, and better conceptualize our own flourishing. 

Intro

On February 24, OpenAI’s CEO Sam Altman released a post titled “Planning for AGI and Beyond.” Given OpenAI’s charter to bring safe general intelligence to the world, it’s not much of a surprise that we’d see Altman write something like this. The article first lays out three principles they care about:

1. AGI should “empower humanity to maximally flourish in the universe”

2. Benefits of, access to, and governance of AGI should be widely shared

3. Successfully navigating massive risks will require iterative learning and adaptation by deploying less powerful versions of the technology [ “the technology” here is presumably “AGI” ]

Altman then discusses how we should prepare for AGI in the short term: this section re-iterates principle 3 above and expands on the notion that we should learn to navigate AI deployment challenges and that a gradual approach to releasing more powerful technologies will give people, policymakers, and institutions time to understand and adapt to new advances. First-hand experience with these technologies will allow the greater public to understand their downsides, “adapt our economy,” and put regulations in place. 

In planning for the long term, Altman thinks that “the future of humanity should be determined by humanity” and that there should be great scrutiny of any effort attempting to build AGI; accordingly, major decisions should require public consultation. 

In this article, I want to do a few things—broadly, I want to answer the question “Should we be ‘planning for AGI’ right now?” Answering that question requires two things first:

1. Articulating what we mean by “AGI.” And, in turn,

2. Articulating what we mean by “intelligence.” 

Along the way, we should realize that the precise nature of what we’re talking about matters a lot more than the terms we use. Overloading terms without much explanation often allows us to make vague statements that sound important, but say little. I will center much of this discussion around terms like AGI because we use them so frequently, but we should try to remember that actions like “achieving AGI” can mean very different things to different people.

Loaded Terms

TL;DR: This section attempts to clarify the terms “intelligence” and “AGI,” drawing from a motley of perspectives on what those might mean. I consider a notion of intelligence, and of AGI, that defines a threshold for intelligence based on the capacity to act and perceive/associate contexts for action, and a gradation of intelligence based on the efficiency with which one learns new skills. 

Read more

It’s been a couple of months since OpenAI released ChatGPT, the large language model chatbot that not only blew the minds of many AI researchers but also introduced the power of AI to the general public. Briefly, ChatGPT is a chatbot that can respond to human instructions to do tasks from writing essays and poems to explaining and debugging code. The chatbot displays impressive reasoning capabilities and performs significantly better than prior language models.

In this editorial, I will discuss my personal takes on ChatGPT’s impact on three groups of people: AI researchers, Tech developers, and the general public. Throughout the piece, I will speculate on the implications of technologies like ChatGPT and outline some scenarios that I believe are likely to happen. This article leans more toward an op-ed than a fact-based report, so take these views with a grain of salt. With that, let’s dive in…

ChatGPT for AI Researchers

To me, a fellow AI researcher, the most important lesson of ChatGPT is that curating human feedback is very important for improving the performance of large language models (LLMs). ChatGPT changed my, and I suspect, many researchers’ views on the problem of AI alignment for LLMs, which I’ll explain now.

Before ChatGPT I intuitively thought that we had two different problems when it comes to LLMs: 1) improving LLMs’ performance on certain language-based tasks (e.g. summarization, question and answering, multi-step reasoning) while 2) avoiding harmful/toxic/biased text generations. I thought of these two objectives as related but separate and called the second problem the alignment problem. I learned from ChatGPT that alignment and task performance are really the same problem, and aligning LLMs’ outputs with human intent both reduced harmful content as well as improved task performance.

For a bit of context, on a very high level, one can think of separating modern LLM training into 2 steps:

• Step 1: Self-Supervised Learning (SSL) of a neural network model to predict the next word (token) given a sequence of previous words (tokens) - this is trained on a very large, Internet-scale dataset.

• Step 2: Aligning the LLMs’ generations with human preferences through various techniques, like fine-tuning the LLM on a small dataset of high-quality instruction-following texts and using Reinforcement Learning to fine-tune the LLM with a learned reward model that predicts human preferences.

For ChatGPT, OpenAI likely used many different techniques in tandem with each other to produce the final model. It also seemed like OpenAI was able to quickly respond to online complaints of misaligned behaviors (e.g. generating harmful texts), sometimes in days if not hours, so the company must also have ways of modifying/filtering model generations without actually retraining/fine-tuning the model.

ChatGPT marks a quiet come-back for Reinforcement Learning (RL). Briefly, Reinforcement Learning with Human Feedback (RHLF) first trains a reward model that predicts how high a human would score a particular LLM generation, then it uses this reward model to improve the LLM through RL.

I won’t go into too much in detail about RL here, but OpenAI has traditionally been known for its RL prowess, having authored OpenAI gym which jump-started RL research, trained RL agents to play DoTA, and famously trained robots to play the Rubik’s cube using RL on millions of years of simulation data. After OpenAI disbanded its robotics team, it seemed like RL was fading into the background for OpenAI, as its achievements in generative models came mostly from Self-Supervised Learning. The success of ChatGPT, which hinges upon RLHF, is bringing new attention to RL as a practical method for improving LLMs.

Read more

Introduction

You’re reading text right now–it’s serving as a medium for me to communicate a sequence of thoughts to you. Ever since humanity became a band of degenerates that actually wrote things down instead of using their memories, we’ve been using sets of signs to transmit information. Under some definitions, you might call all of this “text.” 

Today, and over the past centuries, we have encoded our knowledge of the world, our ideas, our fantasies, into writing. That is to say, much of human knowledge is now available in the form of text. We communicate in other ways too–with body language, images, sounds. But text is the most abundant medium we have of recorded communications, thoughts, and ideas because of the ease with which we can produce it. 

When GPT-3 was fed the internet, it consumed our observations about the world around us, our vapid drama, our insane arguments with one another, and much more. It learned to predict next words in sequences of the tokenized chaos of human expression. In learning how we form sequences of words to communicate, a large language model learns to mimic (or “parrot”) how we joke, commiserate, command. GPT-3 kicked off something of a “revolution” by being extremely good at “text-to-text”: prompted with examples of a task (like finishing an analogy) or the beginning of a conversation, the generative model can (often) competently learn the task or continue the conversation. 

There is almost a “universality” to the ways we use text, and we have only recently reached a point where AI systems can be put together in order to exploit how we use language to describe other modalities. The progress that enabled powerful text generation also enabled text-conditioned multimodal generation. “Text-to-text” became “text-to-X.” 

In “text-to-text,” you could ask a model to riff on a description of a dog. In text-to-image, you could turn that description into its visual counterpart. Text-to-image models afforded a new ability not present in existing image generation systems. Existing models like GANs were trained to, given noise inputs (and class information for class-conditional image generation) generate realistic images. But these models did not offer the level of controllability that DALL-E 2, Imagen, and their ilk provide users: you could ask for a photo of a kangaroo with sunglasses, standing in front of a particular building, holding a sign bearing a particular phrase. Your wish was the algorithm’s command. 

Read more

TL;DR

Beyond generic code completion, code-writing language models can also write domain-specific code according to natural language instructions. In our recent work, Code as Policies (CaP), we explore this idea in the context of robotics and show how we can prompt language models to directly write code that controls robots to perform tasks according to language instructions. Code allows the language model to perform precise arithmetic, spatial-geometric reasoning, and express a degree of behavioral common sense. We deployed CaP on various robot platforms and showed it can perform diverse tasks, from tabletop object manipulation to 2D drawing, all without any additional model training. CaP represents a new way of programming robots and points to a promising future of using language models to write code to do tasks. For more, see our website, coverage by TechCrunch, and Twitter thread:

Background

Large language models (LLMs) have shown impressive capabilities not just in natural language understanding, but also in “reasoning” tasks, from reading comprehension to answering difficult math questions. Recent advances in robotics also leverage this capability to use LLMs for robot planning (see video above and gif below). Essentially, we want to build a system that takes as input high-level natural language instructions of a task (e.g. cleaning up a coffee spill) and outputs actions the robot can execute. In PaLM-Saycan (video above and gif below), this is done by having the LLM generate a sequence of low-level actions, described by natural language, that the robot knows how to do (e.g. go to the trashcan, pick up the sponge, close cabinet drawer, etc).

This is a powerful paradigm that lets us build robot systems that 1) interface with users through language, something non-experts can do, 2) leverage LLMs’ reasoning capabilities for task planning, and 3) enable the robot to plan new tasks without task-specific model training (everything is done through prompting).

While using natural language as the input is an incredibly expressive way to specify robot tasks, there are some limitations to using natural language as the action output for robot task planners. It’s difficult for LLMs to reliably reason about spatial-geometric relationships, perform vector arithmetic, and use logic structures like if conditions and loops. It’s also hard to provide visual feedback to the language model directly through natural language (e.g. describing object coordinates, bounding boxes, and segmentation masks), and feedback is crucial for robot decision-making. It turns out these types of operations are very naturally expressed in code, and this is the question I explored in my internship this past summer at Google - what if we just get language models to write robot code? This seems like a promising direction, but how do we exactly do that, and how well does this actually work in practice?

Language Model Programs

Read more

• At the recent AI Day 2022, Tesla unveiled its Optimus robot, presented the inner workings of its self-driving and AI technologies, and briefly captured the attention of everyone working in AI and robotics.

• Reactions from media and experts poured in during the aftermath of the event; some praised the speed of the humanoid hardware development, while others criticized the lack of anything truly novel.

• Regardless of actual product development, this AI Day worked wonders for Tesla in terms of publicity, which the company hopes will help them recruit the necessary talent for its AI ambitions.

What Happened

Tesla recently hosted its second AI Day to present the company’s progress in AI and to recruit engineers in robotics, AI, and hardware. Perhaps what the public and media looked forward to the most is Tesla’s work-in-progress humanoid robot, Optimus, which at last year’s AI day was demo’ed as a person dressed up in a spandex suit. We’ll give a quick summary of what went down at AI day and some commentary, some reactions from the press and experts, and our take on the event.

Optimus Bot

Tesla first showed a working prototype that was able to walk and wave its arms on stage. Afterward, they also showed the newer in-development prototype design of the robot, which they say is designed for mass-manufacturing. This prototype couldn’t walk on its own, so Tesla engineers rolled out the robot on a platform:

Read more

On June 11, Google engineer Blake Lemoine published an “interview” he and a collaborator at Google conducted with the language model-powered chatbot LaMDA. Following his publication of the interview transcript, Lemoine was put on administrative leave, with the stated reason being breach of confidentiality. 

According to Lemoine, he was asked in th…

Read more

Language is complicated. There’s a lot of complexity and nuance to how people string words and sentences together, but AI systems have gained a fairly impressive grasp on it. Natural language processing (NLP) is everywhere today, from autocorrect to smart assistants like Siri to voice-to-text features, and some individuals wonder if similar technology could let them communicate with animals.

That task may seem impossible at first, but so did robot-human communication at one point. The world’s first chatbot emerged in 1966, and IBM’s Watson won “Jeopardy!” less than 50 years later. Technology grows exponentially, so it’s not outlandish to think that AI systems could eventually help people talk with animals.

These tools may not look like what some people imagine, but AI-powered human-animal communication could be possible. Research is already underway.

Read more

Read more

Read more

Read on to find out why that’s the case, or skip to “TLDR” at the bottom for the short version.

Note: I will be using ‘sentience’ throughout in the loose sense that it is used in Blake Lemoine’s claims, i.e. LaMDA has some sort of human-like consciousness that merits considering it with some level of legal personhood. A more detailed discussion of what t…

Read more

The Russian invasion of Ukraine goes on, and much continues to be made of its implications. France’s re-elected President Emmanuel Macron vowed to “act first to avoid any escalation following the Russian aggression in Ukraine” during his inauguration speech. That aggression itself, as we’ve already commented, has possibly involved AI-powered weapon…

Read more

Read more

Russia's shocking war against Ukraine – the biggest war in Eur…

Read more

Read more

Last Week in AI

Generating AI Art from Text with Google Colab

Generating AI Art from Text with Google Colab With Dream, using AI tools to generate art has officially gone mainstream in a big way, powered by the viral popularity it gained on TikTok. Amazingly, the technique that powers this app was introduced less than a year before the app itself was released, when OpenAI announced…

Read more

4 years ago · Andrey Kurenkov

Read more