An age-old adage, often introduced to us during our formative years, is designed to nudge us beyond our self-centered, nascent minds: "Two heads are better than one." This proverb encourages collaborative thinking and highlights the potency of shared intellect.

Fast forward to 2023, and we find that this wisdom holds true even in the realm of artificial intelligence: Multiple language models, working in harmony, are better than one.

Recently, a team from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) embodied this ancient wisdom within the frontier of modern technology. They introduced a strategy that leverages multiple AI systems to discuss and argue with each other to converge on a best-possible answer to a given question. This method empowers these expansive language models to heighten their adherence to factual data and refine their decision-making.

The crux of the problem with large language models (LLMs) lies in the inconsistency of their generated responses, leading to potential inaccuracies and flawed reasoning. This new approach lets each agent actively assess every other agents responses, and uses this collective feedback to refine its own answer. In technical terms, the process consists of multiple rounds of response generation and critique. Each language model generates an answer to the given question, and then incorporates the feedback from all other agents to update its own response. This iterative cycle culminates in a final output from a majority vote across the models' solutions. It somewhat mirrors the dynamics of a group discussion where individuals contribute to reach a unified and well-reasoned conclusion.

One real strength of the approach lies in its seamless application to existing black-box models. As the methodology revolves around generating text, it can also be implemented across various LLMs without needing access to their internal workings. This simplicity, the team says, could help researchers and developers use the tool to improve the consistency and factual accuracy of language model outputs across the board.

Employing a novel approach, we dont simply rely on a single AI model for answers. Instead, our process enlists a multitude of AI models, each bringing unique insights to tackle a question. Although their initial responses may seem truncated or may contain errors, these models can sharpen and improve their own answers by scrutinizing the responses offered by their counterparts," says Yilun Du, an MIT PhD student in electrical engineering and computer science, affiliate of MIT CSAIL, and lead author on a new paper about the work. "As these AI models engage in discourse and deliberation, they're better equipped to recognize and rectify issues, enhance their problem-solving abilities, and better verify the precision of their responses. Essentially, we're cultivating an environment that compels them to delve deeper into the crux of a problem. This stands in contrast to a single, solitary AI model, which often parrots content found on the internet. Our method, however, actively stimulates the AI models to craft more accurate and comprehensive solutions."

The research looked at mathematical problem-solving, including grade school and middle/high school math problems, and saw a significant boost in performance through the multi-agent debate process. Additionally, the language models showed off enhanced abilities to generate accurate arithmetic evaluations, illustrating potential across different domains.

The method can also help address the issue of "hallucinations" that often plague language models. By designing an environment where agents critique each other's responses, they were more incentivized to avoid spitting out random information and prioritize factual accuracy.

Beyond its application to language models, the approach could also be used for integrating diverse models with specialized capabilities. By establishing a decentralized system where multiple agents interact and debate, they could potentially use these comprehensive and efficient problem-solving abilities across various modalities like speech, video, or text.

While the methodology yielded encouraging results, the researchers say that existing language models may face challenges with processing very long contexts, and the critique abilities may not be as refined as desired. Furthermore,the multi-agent debate format, inspired by human group interaction, has yet to incorporate the more complex forms of discussion that contribute to intelligent collective decision-making a crucial area for future exploration, the team says. Advancing the technique could involve a deeper understanding of the computational foundations behind human debates and discussions, and using those models to enhance or complement existing LLMs.

"Not only does this approach offer a pathway to elevate the performance of existing language models, but it also presents an automatic means of self-improvement. By utilizing the debate process as supervised data, language models can enhance their factuality and reasoning autonomously, reducing reliance on human feedback and offering a scalable approach to self-improvement," says Du. "As researchers continue to refine and explore this approach, we can get closer to a future where language models not only mimic human-like language but also exhibit more systematic and reliable thinking, forging a new era of language understanding and application."

"It makes so much sense to use a deliberative process to improve the model's overall output, and it's a big step forward from chain-of-thought prompting," says Anca Dragan, associate professor at the University of California at Berkeleys Department of Electrical Engineering and Computer Sciences, who was not involved in the work. "I'm excited about where this can go next. Can people better judge the answers coming out of LLMs when they see the deliberation, whether or not it converges? Can people arrive at better answers by themselves deliberating with an LLM? Can a similar idea be used to help a user probe a LLM's answer in order to arrive at a better one?"

Du wrote the paper with three CSAIL affiliates: Shuang Li SM '20, PhD '23; MIT professor of electrical engineering and computer science Antonio Torralba; and MIT professor of computational cognitive science and Center for Brains, Minds, and Machines member Joshua Tenenbaum. Google DeepMind researcher Igor Mordatch was also a co-author.

The rest is here:
Multi-AI collaboration helps reasoning and factual accuracy in large ... - MIT News

A New York Times article this morning, titled How to Tell if Your AI Is Conscious, says that in a new report, scientists offer a list of measurable qualities based on a brand-new science of consciousness.

The article immediately jumped out at me, as it was published just a few days after I had a long chat with Thomas Krendl Gilbert, a machine ethicist who, among other things, has long studied the intersection of science and politics. Gilbert recently launched a new podcast, called The Retort, along with Hugging Face researcher Nathan Lambert, with an inaugural episode that pushes back on the idea of todays AI as a truly scientific endeavor.

Gilbert maintains that much of todays AI research cannot reasonably be called science at all. Instead, it can be viewed as a new form of alchemy that is, the medieval forerunner of chemistry, that can also be defined as a seemingly magical process of transformation.

Many critics of deep learning and of large language models, including those who built them, sometimes refer to AI as a form of alchemy, Gilbert told me on a video call. What they mean by that, he explained, is that its not scientific, in the sense that its not rigorous or experimental. But he added that he actually means something more literal when he says that AI is alchemy.

The people building it actually think that what theyre doing is magical, he said. And thats rooted in a lot of metaphors, ideas that have now filtered into public discourse over the past several months, like AGI and super intelligence. The prevailing idea, he explained, is that intelligence itself is scalar depending only on the amount of data thrown at a model and the computational limits of the model itself.

But, he emphasized, like alchemy, much of todays AI research is not necessarily trying to bewhat we know as science, either. The practice of alchemy historically had no peer review or public sharing of results, for example. Much of todays closed AI research does not, either.

It was very secretive, and frankly, thats how AI works right now, he said. Its largely a matter of assuming magical properties about the amount of intelligence that is implicit in the structure of the internet and then building computation and structuring it such that you can distill that web of knowledge that weve all been building for decades now, and then seeing what comes out.

I was particularly interested in Gilberts thoughts on alchemy given the current AI discourse, which seems to me to include some doozies of cognitive dissonance: There was the Senates closed-door AI Insight Forum, where Elon Musk called for AI regulators to serve as a referee to keep AI safe, while actively working on using AI to put microchips in human brains and making humans a multiplanetary species. There was the EU parliament saying that AI extinction risk should be a global priority, while at the same time, OpenAI CEO Sam Altman said hallucinations can be seen as positive part of the magic of generative AI and that superintelligence is simply an engineering problem.

And there was DeepMind co-founder Mustafa Suleyman, who would not explain to MIT Technology Review how his company Inflections Pi manages to refrain from toxic output Im not going to go into too many details because its sensitive, he said while calling on governments to regulate AI and appoint cabinet-level tech ministers.

Its enough to make my head spin but Gilberts take on AI as alchemy put these seemingly opposing ideas into perspective.

Gilbert clarified that he isnt saying that the notion of AI as alchemy is wrong but that its lack of scientific rigor needs to be called what it really is.

Theyre building systems that are arbitrarily intelligent, not intelligent in the way that humans are whatever that means but just arbitrarily intelligent, he explained. Thats not a well-framed problem, because its assuming something about intelligence that we have very little or no evidence of, that is an inherently mystical or supernatural claim.

AI builders, he continued, dont need to know what the mechanisms are that make the technology work, but they are interested enough and motivated enough and frankly, also have the resources enough to just play with it.

The magic of generative AI, he added, doesnt come from the model. The magic comes from the way the model is matched to the interface. The magic people like so much is that I feel like Im talking to a machine when I play with ChatGPT. Thats not a property of the model, thats a property of ChatGPT of the interface.

In support of this idea, researchers at Alphabets AI division DeepMind recently published work showing that AI can optimize its own prompts and performs better when prompted to take a deep breath and work on this problem step-by-step, though the researchers are unclear exactly why this incantation works as well as it does (especially given the fact that an AI model does not actually breathe at all.)

One of the major consequences of the alchemy of AI is when it intersects with politics as it is now with discussions around AI regulation in the US and the EU, said Gilbert.

In politics, what were trying to do is articulate a notion of what is good to do, to establish the grounds for consensus that is fundamentally whats at stake in the hearings right now, he said. We have a very rarefied world of AI builders and engineers, who are engaged in the stance of articulating what theyre doing and why it matters to the people that we have elected to represent our political interests.

The problem is that we can only guess at the work of Big Tech AI builders, he said. Were living in a weird moment, he explained, where the metaphors that compare AI to human intelligence are still being used, but the mechanisms are not remotely well understood.

In AI, we dont really know what the mechanisms are for these models, but we still talk about them like theyre intelligent. We still talk about them liketheres some kind of anthropological ground that is being uncovered and theres truly no basis for that.

But while there is no rigorous scientific evidence backing for many of the claims to existential risk from AI, that doesnt mean they arent worthy of investigation, he cautioned. In fact, I would argue that theyre highly worthy of investigation scientifically [but] when those things start to be framed as a political project or a political priority, thats a different realm of significance.

Meanwhile, the open source generative AI movement led by the likes of Meta Platforms with its Llama models, along other smaller startups such as Anyscale and Deci is offering researchers, technologists, policymakers and prospective customers a clearer window onto the inner workings of the technology. But translating the research into non-technical terminology that laypeople including lawmakers can understand, remains a significant challenge.

That is the key problem with the fact that AI, as alchemy and not science, has become a political project, Gilbert explained.

Its a laxity of public rigor, combined with a certain kind of willingness to keep your cards close to your chest, but then say whatever you want about your cards in public with no robust interface for interrelating the two, he said.

Ultimately, he said, the current alchemy of AI can be seen as tragic.

There is a kind of brilliance in the prognostication, but its not clearly matched to a regime of accountability, he said. And without accountability, you get neither good politics nor good science.

VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.

Read more here:
Why generative AI is 'alchemy,' not science - VentureBeat

RT-2 utilizes chain-of-thought reasoning, allowing it to make multi-stage decisions, such as selecting alternative tools or beverages based on specific situations.

Google DeepMind unveiled Robotic Transformer 2 (RT-2), a vision-language-action (VLA) model designed to enhance robotic control through plain language instructions. Harnessing data from the Internet, RT-2 aims to foster robots that can adeptly navigate human environments, akin to well-known fictional robot companions from science fiction.

RT-2, drawing inspiration from how humans learn by reading and observing, relies on a vast language model akin to ChatGPT, which is trained using online text and images. This allows RT-2 to achieve the feat of generalization, enabling it to recognize patterns and perform untrained tasks.

Google showcased RT-2s proficiency by demonstrating its ability to identify and discard trash without prior training. This includes recognizing potentially ambiguous items like food packaging as trash. A separate test had a robot powered by RT-2 successfully pinpoint a dinosaur figurine when instructed to Pick up the extinct animal. These capabilities are transformative as, traditionally, robotic training has been labor-intensive, relying on extensive manual data acquisition.

See also: AI and Robotics Research Continues to Accelerate

RT-2s prowess can be attributed to Google DeepMinds adoption of transformer AI models, celebrated for their generalization capabilities. The technology is built on Googles prior AI innovations, such as the Pathways Language and Image model (PaLI-X) and the Pathways Language model Embodied (PaLM-E). Moreover, RT-2 was co-trained using data from its precursor, RT-1, gathered over 17 months.

The RT-2 framework refines a pre-trained VLM model with robotics and web data, leading to a model that processes camera images from robots and predicts subsequent actions. Interestingly, actions are represented as tokens, akin to word fragments, aiding in the robots control. This method, applied to RT-1, was also employed for RT-2, converting actions into symbolic string representations to facilitate new skill acquisition.

Additionally, RT-2 utilizes chain-of-thought reasoning, allowing it to make multi-stage decisions, such as selecting alternative tools or beverages based on specific situations. Comparative tests revealed RT-2s stellar performance in new situations, recording a 62% success rate against RT-1s 32%.

However, the model has its limitations. Although web data enhances generalization over concepts, it cannot bestow the robot with new physical skills it hasnt practiced. Google acknowledges these constraints and the considerable research journey ahead but remains optimistic, viewing RT-2 as a significant stride towards achieving general-purpose robots.

View original post here:
Enhanced Robotic Control with DeepMind RT-2 - RTInsights

Connor Coley, 29, developed open-source software that uses artificial intelligence to help discover and synthesize new molecules. The suite of tools, called ASKCOS, is used in production by over a dozen pharmaceutical companies, and tens of thousands of chemists, to create new medicines, new materials, and more efficient industrial processes.

One of the largest bottlenecks in developing new molecules has long been identifying interesting candidates to test. This process has played out in more or less the same way for decades: make a small change to a known molecule, and then test the novel creation for its biological, chemical, or physical properties.

Coleys approach includes a form of generative AI for chemistry. A chemist flags which properties are of interest, and AI-driven algorithms suggest new molecules with the greatest potential to have those properties. The system does this by analyzing known molecules and their current properties, and then predicting how small structural changes are likely to result in new behaviors.

As a result, chemists should spend less time testing candidates that never pan out. The types of methods that we work on have led to factors of maybe two, three, maybe 10 [times] reduction in the number of different shots on goal you need to find something that works well, says Coley, who is now an assistant professor of chemical engineering and computer science at MIT.

Once it identifies the candidate molecules, Coleys software comes up with the best way to produce them. Even if chemists imagine or dream up a molecule, he says, figuring out how to synthesize something isnt trivial: We still have to make it.

To that end, the system gives chemists a recipe of steps to follow that are likely to result in the highest yields. Coleys future work includes figuring out how to add laboratory robots to the mix, so that even more automated systems will be able test and refine the proposed recipes by actually following them.

Original post:
Alhussein Fawzi - MIT Technology Review

The Rara AVIS Reason Behind Their Intentions

If you cant beat it, join it.

Thats how the saying goes, and thats precisely what our dear friend Google is doing regarding AI.

But, weirdly enough, Google is taking it to the extreme, purposely contributing to the demise of its ad-base revenue model, one of the most successful businesses in the history of capitalism and the cash cow behind its trillion-dollar business.

However, you shouldnt fear for Googles integrity, this is the plan all along.

This article was originally published days ago in my free weekly newsletter, TheTechOasis.

If you want to be up-to-date with the frenetic world of AI while also feeling inspired to take action or, at the very least, to be well-prepared for the future ahead of us, this is for you.

Subscribe below to become an AI leader among your peers and receive content not present in any other platform, including Medium:

Anyone who follows the AI industry will probably agree on the fact that were approaching the death of Internet search as we know it.

And Google knows it too.

Undeniably, using ChatGPT or Claude is much quicker and more convenient than doing link-based searches.

Thus, Google had two options:

As Googles execs arent dumb, theyve naturally gone for the second option.

In fact, theres no company in the world right now more heavily focused on disrupting search with AI than, ironically, Google.

See the original post here:
This is Why and How Google Will Kill its Business Model - Medium