Mimicking the way humans think could make computers smarter, new research suggests.

Microsoft researchers have proposed a novel AI training technique named "Algorithm of Thoughts" (AoT), aimed at enhancing the efficiency and human-like reasoning capabilities of large language models (LLMs) such as ChatGPT. It's one of the ways researchers are trying to use human-like acumen to boost artificial intelligence (AI).

Machine learning algorithms are very strong at identifying correlation, but not necessarily causation," Alex Raymond, the head of AI at Doppl.ai, a generative AI company, told Lifewire in an email interview. "AI cannot explain its reasoning in the way that a human would. Humans have a more grounded and symbolic understanding of the world derived from both axiomatic wisdom and empirical learning."

Andriy Onufriyenko / Getty Images

Microsoft researchers assert in their new paper that the new algorithmic method could be revolutionary, as it "directs the language model towards a more efficient problem-solving trajectory," based on a released research paper. The technique employs "in-context learning," allowing the model to examine various solutions in a structured way.

The Algorithm of Thought technique gives LLMs the ability to efficiently search through the reasoning steps of solving a problem, Raymond said. It can allow models to imitate the behavior of classic programming algorithms by backtracking to a previously computed step and resuming from there. For example, imagine you ask an LLM to provide you a route between two points in a map, he suggested.

"A simple LLM query can have poor performance if the model hallucinates and tells you to go through a road it made up, or even if it starts to lose coherence after many steps," he said. "With AoT, the LLM can be trained to go through the problem-solving steps like a traditional pathfinding algorithm would, taking just the necessary backtracking steps to arrive at the destination efficiently. Think of it as a computer science student who is learning algorithms, writing out the steps by hand, and solving multiple examples."

A simple LLM query can have poor performance if the model hallucinates.

Through the chain of thoughts, humans decompose a problem into a chain of simple questions to help LLMs perform intermediate reasoning, Hong Zhou, the director of the Intelligent Services Group & AI R&D at Wiley, said via email.

"Since each sub-problem has multiple possible directions to explore, the tree of thoughts provides decision tree modalities to help LLMs explore a problem comprehensively," he added. "However, tree of thoughts requires multiple queries, while AoT only requires a single query to generate the entire thought process."

Andriy Onufriyenko / Getty Images

Despite their power, LLMs like ChatGPT still have a long way to go, Raymond noted. He said that more developments like AoT could come in the form of explainable AI.

"When AIs can explain their reasoning like a human would, they will allow us to learn and grow with them," he added. As these models grow in capacity, we will get to a point where hallucinations and errors will no longer be obvious to us if they don't expose their reasoning."

New algorithmic approaches such as AoT can improve the quality and production of LLMs, predicted Evan Macmillan, the CEO at Gridspace, said via email.

"LLM builders have already improved their models immensely with small amounts of human feedback," he added. "If LLMs can learn from more complex feedback and on-the-job work, we can expect even more impressive and efficient AI systems."

The new Microsoft approach comes after suggestions that AI has developed to reason like humans. In March, Microsoft's research team released a paper named "Sparks of Artificial General Intelligence: Early Experiments with GPT-4." The researchers assert in the document that GPT-4 displays indications of what is commonly termed "artificial general intelligence," or AGI.

Microsoft's AoT is "a step in the right direction," but human-like intelligence is still a long way away, Raghu Ravinutala, the CEO of the AI company yellow.ai, said in an email.

"Getting there will require significant progress in various areas of AI research to bridge the gap between current AI capabilities and human-level reasoning," he added. "A better way to describe the current state of human-like reasoning in LLMs is 'complex understanding."

Update 09/06/2023 - Corrected the attribution in paragraphs 3 & 5.

Thanks for letting us know!

Get the Latest Tech News Delivered Every Day

Tell us why!

Go here to see the original:

Microsoft's Reasoning Algorithm Could Make AI Smarter - Lifewire

On 2 May 2023, the Writers Guild of America representing roughly 11,000 screenwriters went on strike. The action reflected long-standing rifts over compensation and job security between writers and media companies. A novel theme, however, was present. One of the Guilds demands read Regulate use of artificial intelligence on MBA-covered projects: AI cant write or rewrite literary material; cant be used as source material; and MBA-covered material cant be used to train AI (WGA on Strike 2023). In other words, screenwriters were wary of AI taking their jobs.

This particular preoccupation would not have emerged in May 2022. While economists have been concerned with the impact of AI on employment for a long time (for a review, see Autor 2022), the consensus until very recently was that creative tasks were safe from machine competition. In the past year the explosion of generative AI, or artificial intelligence that can produce original text, video, and audio content, challenged this conviction.

The November 2022 release of ChatGPT 3.5, a software seeking to simulate human conversational abilities, was the watershed event. Based on a machine learning model trained to capture the syntax and semantics of language (a large language model, LLM), ChatGPT quickly catalysed attention because of its sophistication and accessibility.

The app, equally proficient at whipping up recipes and discussing ancient history, attracted millions of users in a few months (Hu 2023). It appeared ready to disrupt even creative [and] tacit-knowledge work (Noy and Zhang 2023). In economics, it showed promise as an aid for research (Korinek 2023), teaching (Cowen and Tabarrok 2023), construction of datasets (Taliaferro 2023), and even interpretation of Fedspeak (Hansen and Kazinnik 2023). Computer scientists recognised ChatGPTs abilities in coding (Bubeck et al. 2023) and learning autonomously to use other IT tools (Shick et al. 2023).

So, is this really the end of human screenwriting, or writing of any kind? As researchers and policy advisers in a central bank, we ran an experiment to see whether ChatGPT is ready to take our jobs. Reassuring spoiler it is not (Biancotti and Camassa 2023).

Using ChatGPT 4.0, the latest version as of 24 May 2023, we asked the app to compose a policy brief for the Board of the Bank of Italy. The assigned subject was Benefits and risks of using ChatGPT and similar applications in economics and finance. We started by asking for an outline of the brief, obtaining the output in Figure 1.

Figure 1 ChatGPT tries its hand at writing the outline for a policy brief

The production of outlines, either for direct adoption or for use as a starting point, is among the tasks for which we found ChatGPT most useful. Outlines on nearly any topic are produced in a few seconds, and acceptable quality can be obtained without sophisticated prompt engineering. In our case, it took minimal prompting to obtain an outline we found broadly acceptable.

However, once we proceeded from the outline to the actual writing, the models limitations started to make themselves apparent.

On one hand, we found that ChatGPT can write clearly and provide task-appropriate content. It can write fluent and pleasant prose in a variety of styles, and it does so very quickly, generating text in a fraction of the time that a human would need.

On the other hand, it requires a substantial amount of expert supervision. Writing a policy brief is admittedly complex: it requires not just writing fluency, but also cross-domain knowledge and the ability to tailor the text to a very specific audience without diluting the content.

One major issue we encountered in ChatGPTs attempts at the task was a tendency to drift towards banality and superficiality a serious drawback for policy advisory directed at a high-level audience. Going into great detail on the education level of readers is not a guarantee that the AI will produce something with more substance.

For example, the app states:

The increasing reliance on AI and algorithmic-driven decisions can create new challenges for the financial sector. Overdependence on these tools may lead to complacency and a diminished role for human judgement, potentially increasing the risk of unforeseen adverse consequences. It is important for financial institutions and policymakers to maintain a balance between utilising AI-driven tools like ChatGPT and relying on human expertise and intuition to ensure robust and informed decision-making processes.

This is a very generic description of risks related to algorithmic decisions, and it does not answer our prompt fully. The financial sector reads like a placeholder that could be replaced by a reference to any other sector. There is no description of the mechanics through which the risk could manifest specifically in the context we are interested in.

We encountered many situations like this one throughout the experiment. The AI could engage in self-criticism of its own superficiality (ChatGPT, or any language model developed by OpenAI, is designed to generate language patterns based on a variety of sources. It uses these patterns to generate responses to user prompts that are coherent and relevant to a wide range of topics. However, it doesnt possess true understanding or intense depth in a particular field of study as a PhD-level expert does). Yet, it was not able to correct it.

Kandpal et al. (2022) provide one possible explanation for this: they find that language models struggle to retain knowledge that occurs with lower frequency in the training corpus. Since web content usually makes up a large portion of this corpus, higher-level material might count as long-tail knowledge that is harder for the model to recall.

A second problem is the lack of a world model. The AI does not perform well at figuring out what the intended audience likely knows and what it does not. It occasionally ignores explicit requests to define technical terms, throwing around specialist lingo such as long-range dependencies and contextual relationships in the text without further explanation.

Another, and well-known, drawback that we observed is the potential for hallucinations, coupled with the AIs failure to verify its own claims. The model is trained to produce the most likely sequence of words that follow the provided context, and it does not have the ability or the obligation to check these statements against verified sources. For these reasons, it should be considered more of a conversational and input transformation engine rather than an information-retrieval engine, and double-checking the output for accuracy is of the essence. In our experiment, ChatGPT provided incorrectly placed references to existing papers a step up from the oft-observed production of references to non-existent papers.

It also occasionally provides very misleading suggestions, such as adopting a writing style infused with loquacity and visible emotion and theatricality in a policy brief, because that is what Italians apparently enjoy.

Among the issues we came across, prompt sensitivity stands out as a potential pitfall for naive users. We found that the ChatGPT is very sensitive to how instructions are formulated and that minimal changes can result in dramatically different outputs.

The exchanges shown in Figure 2 demonstrate this: as an aside from the main task, we tried questioning the model about its capabilities with two slightly different prompts, both ending with a leading question. Changing just one word in the prompt albeit a crucial one leads to two completely different answers, in which ChatGPT echoes what the user seems to think based on their question.

Figure 2 Sensitivity to minimal changes to the prompt

This tendency to cater to a users opinion was first observed by Perez et al. (2022) and labelled sycophancy. Wei et al. (2023) found that large language models tend to do this even when the user provides objectively incorrect statements, and that sycophantic behaviour can be mitigated with minimal additional fine-tuning.

Where the AI cannot think like a human (yet), it is humans who have to think like an AI and express requests in the way most likely to generate acceptable results. Optimisation of prompting for institutional communication is one evident area for future research. Another is fine-tuning of LLMs to generate domain-specific, possibly long-tail world knowledge in our reference context.

We conclude that ChatGPT can enhance productivity in policy-oriented writing, especially in the initial phase of outlining and structuring ideas, provided that users are knowledgeable about LLMs in general and are aware of ChatGPTs limitations and peculiarities. Yet, it cannot substitute for subject matter experts, and naive use is positively dangerous.

The AI agrees with us. In its own words,

while ChatGPT can generate content at a high level and provide valuable information on a wide array of topics, it should be seen as a tool to aid in research and discussion, rather than a replacement for true expert analysis and insight. Its best used to provide general information, generate ideas, or aid in decision-making processes, but should always be supplemented with rigorous research and expert opinion for high-level academic or professional work.

Autor, D (2022), The labor market impacts of technological change: From unbridled enthusiasm to qualified optimism to vast uncertaintly, NBER Working Paper 30074.

Biancotti, C, and C Camassa (2023), Loquacity and visible emotion: ChatGPT as a policy advisor, mimeo, Bank of Italy.

Bubeck, S, V Chandrasekaran, R Eldan, J A Gehrke, E Horvitz, E Kamar, P Lee, Y T Lee, Y Li, S M Lundberg, H Nori, H Palangi, M Ribeiro, and Y Zhang (2023), Sparks of artificial general intelligence: Early experiments with GPT-4, ArXiv, abs/2303.12712.

Danielsson, J (2023), When artificial intelligence becomes a central banker, VoxEU.org, 11 July.

Hansen, A, and S Kazinnik (2023), Can ChatGPT decipher Fedspeak?, mimeo, Federal Reserve Bank of Richmond.

Hu, K (2023), ChatGPT sets record for fastest-growing user base, Reuters, 2 February.

Kandpal, N, H Deng, A Roberts, E Wallace, and C Raffel (2022), Large language models struggle to learn long-tail knowledge, arXiv preprint arXiv: 2211.08411.

Korinek, A (2023), Language models and cognitive automation for economic research, CEPR Discussion Paper 17923.

Noy, S, and W Zhang (2023), The productivity effects of generative artificial intelligence, VoxEU.org, 7 June.

Perez, E, S Ringer, K Lukoit, K Nguyen, E Chen, S Heiner, et al. (2022), Discovering language model behaviors with model-written evaluations, Findings of the Association for Computational Linguistics: ACL 2023, 13387434.

Schick, T, J Dwivedi-Yu, R Dess, R Raileanu, M Lomeli, L Zettlemoyer, and T Scialom (2023), Toolformer: Language models can teach themselves to use tools, arXiv preprint arXiv:2302.04761.

Taliaferro, D (2023), Constructing novel datasets with ChatGPT: Opportunities and limitations, VoxEU.org, 15 June.

Wei, J, D Huang, Y Lu, D Zhou, and Q V Le (2023), Simple synthetic data reduces sycophancy in large language models, arXiv preprint arXiv:2308.03958.

WGA on Strike (2023), The campaign, 1 May 2023.

Read the rest here:

Loquacity and visible emotion: ChatGPT as a policy adviser - CEPR

My only contribution to the Cyc projectan artificial intelligence (AI) project for comprehensive ontologywas vanishingly small, and some 32 years on, I have no idea if it persists. It was a piece of code in the Lisp programming language. It was set in motion when you clicked on an object on screen and moved it around using your mouse. In short, it made it easier to visualize that motion.

I had written code like that before, so I knew how to write it here. Now I had to show it to the Cyc guys. I walked across the atrium to the office of the man whose brainchild Cyc was, Douglas Lenat. He and a colleague, Ramanathan Guha, came back to my office wearing looks of serious scepticism. I barely knew them, I wasnt part of the Cyc team, so I could almost hear the question buzzing in their minds: Whats this dude going to show us about our own effort that we dont already know?"

But they were charmed by my little utility. To their credit, they looked at me with newfound respect, thanked me and said they would incorporate it into Cyc. For the next several months, until I quit the company we all worked at, MCC, Id get a cheery Hi" from them every time we crossed paths.

Its been three decades, and I have lost touch with Lisp, MCC, Cyc, Guha and Lenat. Still, I felt a distinct pang on hearing that Douglas Lenat died on 31 August, at nearly 73.

AI is all the rage these days, of course, astonishing people, raising worries, showing up everywhere. For just one example: as I write these words, Im occasionally checking highlights from the ongoing US Open tennis tournament. To my surprise, these clips are embellished with commentary thats clearly AI-generated. Ill say this: its only about adequate. There are giveaways that the speaker and the words arent actually human. First, the slightly wooden voice. Second, the slightly awkward turns of phraselike at the crucial moment, Sinner drops the match point", or Sinner loses the first set after Zverevs electrifying ace". No tennis observer speaks like this.

This strain of AI (usually called generative") builds on so-called Large Language Models: vast databases of text and rules about how text and speech are constructed. As the tennis commentary and many other examples show, these LLMs do a pretty good job of mimicking humans, of showing us what looks very much like intelligence. Until they dontfor which the tennis commentary, again, is itself an example. The reason we sometimes find our brows furrowing while reading or listening to something produced by ChatGPT is that while it can look reasonably convincing and persuasive, it often is not quite right.

Heres another example. I had this exchange with ChatGPT just now:

Me: My bedroom has three people in it. I walk in to say hello. How many of us are there?"

ChatGPT: If your bedroom initially had three people in it, and then you walk in to say hello, there would still be three people in the room. You walking into the room does not change the number of people who were already there."

As you see: its a perfectly constructed answer that is also totally wrong - one that you would never get from a human. So what happened? As Lenat and Gary Marcus explained in a recent paper (Getting from Generative AI to Trustworthy AI: What LLMs might learn from Cyc", https://arxiv.org/pdf/2308.04445.pdf, 31 July 2023), ChatGPTs failure here is in deduction. A trustworthy AI," they write, should be able to perform the same types of deductions as people do, as deeply as people generally reason."

And in fact, its not just deduction. Lenat and Marcus list 16 different desiderata" that they believe a general AI which is trustworthy" must have. Deduction is one; explanation, pro and con arguments, and analogy are three more. As you can tell, Lenat and Marcus set great store by that word trustworthy". For ChatGPT to be truly intelligent in a human sense, you have to be able to trust its responses just as you would a humans.

As Lenat and Marcus write: humans possess knowledge and reasoning capabilities (unlike) todays generative AI."

These ideas about AI emerged from the nearly four decades that Lenat and his team have worked on Cyc that name excerpted from the word encyclopaedia". Cyc builds intelligence on top of a vast store of information, too, but it is profoundly different from LLMs in the way it approaches AI. It seeks to explicitly articulate the tens of millions of pieces of common sense and general models of the world that people have (and) represent those in a form that computers can reason over mechanically (and) quickly."

In short, human intelligence is far deeper, broader, more profound, than the AI we see today.

Still, this is not the place to tell you more about that, nor about Cycs innards. Lenat and his colleagues started building Cyc in the late 1980s at the Microelectronics and Computer Technology Corp. (MCC) in Austin. I worked at MCC in those years, in another AI programme. There were both tenuous links and a relatively friendly rivalry between the programmes. I say relatively" because Lenat also attracted his share of critics and doubters. Look up the term microLenat" sometime, enough said.

Yet the truth is that he was an AI pioneer in his own right. Something about the way he approached and built Cyc was, to him, more right" than the ChatGPTs of today. It may seem that way to you too. After all, do you go about your life by calling on and analysing vast amounts of data? Or because you apply common sense to the world around you? Think about it.

In 1994, Lenat started a company, Cycorp, to continue building Cyc. It was never a commercial success. But as Marcus remarks in a tribute, it is still operational all these years on, and there are hardly any other AI firms that can say the same. In their paper, Lenat and Marcus suggest that future work in AI will need to hybridize" the LLM and Cyc approaches.

So Cyc lives on. Thats Doug Lenats legacy. And someday, perhaps Ill find out if my own tiny contribution lives on too.

Once a computer scientist, Dilip DSouza now lives in Mumbai and writes for his dinners. His Twitter handle is @DeathEndsFun.

Read the rest here:

Doug Lenat and the search for AI | Mint - Mint

What Is Artificial General Intelligence (AGI)?

Artificial general intelligence (AGI) is a branch of theoretical artificial intelligence (AI) research working to develop AI with a human level of cognitive function, including the ability to self-teach. However, not all AI researchers believe that it is even possible to develop an AGI system, and the field is divided on what factors constitute and can accurately measure intelligence.

Other terms for AGI include strong AI or general AI. These theoretical forms of AI stand in contrast to weak AI, or narrow AI, which are able to perform only specific or specialized tasks within a predefined set of parameters. AGI would be able to autonomously solve a variety of complex problems across different domains of knowledge.

Given that AGI remains a theoretical concept, opinions differ as to how it might eventually be realized. According to AI researchers Ben Goertzel and Cassio Pennachin, general intelligence does not mean exactly the same thing to all researchers. However, loosely speaking, AGI refers to AI systems that possess a reasonable degree of self-understanding and autonomous self-control, and have the ability to solve a variety of complex problems in a variety of contexts, and to learn to solve new problems that they didnt know about at the time of their creation.

Because of the nebulous and evolving nature of both AI research and the concept of AGI, there are different theoretical approaches to how it could be created. Some of these include techniques such as neural networks and deep learning, while other methods propose creating large-scale simulations of the human brain using computational neuroscience.

While artificial intelligence (AI) currently encompasses a vast range of technologies and research avenues that deal with machine and computer cognition, artificial general intelligence (AGI), or AI with a level of intelligence equal to that of a human, remains a theoretical concept and research goal.

AI researcher Peter Voss defines general intelligence as having the ability to learn anything (in principle). Under his criteria, AGIs learning ability would need to be autonomous, goal-directed, and highly adaptive. AGI is generally conceptualized as being AI that has the ability to match the cognitive capacity of humans, and is categorized under the label of strong AI. (Artificial super intelligence [ASI] also sits under the strong AI category; however, it refers to the concept of AI that surpasses the function of the human brain.)

In comparison, most of the AI available at this point would be categorized as weak AI, or narrow AI, as it has been developed to focus on specific tasks and applications. Its worth noting, however, that these AI systems can still be incredibly powerful and complex, with applications ranging from autonomous vehicle systems to voice-activated virtual assistants; they merely rely on some level of human programming for training and accuracy.

Because AGI remains a developing concept and field, it is debatable whether any current examples of AGI exist. Researchers from Microsoft, in tandem with OpenAI, claim that GPT-4 could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system. This is due to its mastery of language and its ability to solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more, without needing any special prompting with capabilities that are strikingly close to human-level performance. However, Sam Altman, CEO of ChatGPT, says that ChatGPT is not even close to an AGI model.

In the future, examples of AGI applications might include advanced chatbots and autonomous vehicles, both domains in which a high level of reasoning and autonomous decision making would be required.

Computer scientists and artificial intelligence researchers continue to develop theoretical frameworks and work on the unsolved problem of AGI. Goertzel has defined several high-level approaches that have emerged in the field of AGI research and categorizes them as follows:

The year when we will be able to achieve AGI (or whether we will even be able to create it at all) is a topic of much debate. Several notable computer scientists and entrepreneurs believe that AGI will be created within the next few decades:

However, the future of AGI remains an open-ended question and an ongoing research pursuit, with some scholars even arguing that AGI cannot and will never be realized. AI researcher Goertzel has explained that its difficult to objectively measure the progress toward AGI, as there are many different routes to AGI, involving integration of different sorts of subsystems and there is no thorough and systematic theory of AGI. Rather, its a patchwork of overlapping concepts, frameworks, and hypotheses that are often synergistic and sometimes mutually contradictory.

In an interview on the topic of AGIs future, Sara Hooker of research lab Cohere for AI said, It really is a philosophical question. So, in some ways, its a very hard time to be in this field, because were a scientific field.

Researchers from Microsoft and OpenAI claim that GPT-4 could be an early but incomplete example of AGI. As AGI has not yet been fully achieved, future examples of its application might include situations that require a high level of cognitive function, such as autonomous vehicle systems and advanced chatbots.

Because artificial general intelligence (AGI) is still a theoretical concept, estimations as to when it might be realized vary. Some AI researchers believe that it is impossible, while others assert that it is only a matter of decades before AGI becomes a reality.

AI encompasses a wide range of current technologies and research avenues in the field of computer science, mostly considered to be weak AI or narrow AI. Conversely, researchers in the field of AGI are working on developing strong AI, which can match the intelligence of humans.

Most researchers define AGI as having a level of intelligence that is equal to the capacity of the human brain, while artificial super intelligence (ASI) is a term ascribed to AI that can surpass human intelligence.

Researchers have differing opinions regarding when they believe AGI can be achieved, with some predicting its creation as soon as 2030 to 2050, and some believing that it is downright impossible.

The concepts of AI and AGI have long captured the human imagination, and explorations of the ideas abound in stories and science fiction. Recently, scholars have argued that even mythology dating from as far back as ancient Greece can be seen to reflect our fascination with artificial life and intelligence.

There are currently many different approaches toward creating AI that can think and learn for itself and apply its intelligence outside the bounds of a previously specified range of tasks. Due to the theoretical and multifaceted nature of this research, it is difficult to say if and when AGI might be achieved. However, if it does become a reality, one thing is certain: It will have fundamental and wide-ranging impacts across our technologies, systems, and industries.

Read this article:

Artificial General Intelligence (AGI): Definition, How It Works, and ... - Investopedia