new For all of the bravado of AI, before we become too dazzled or dismayed (depending on which side of this issue you reside), at AI’s potential impact on projects, it is worthwhile to sit back with a skeptical bird’s eye view. After all, as we humans evolved throughout the information age of the mid to late twentieth century, AI can just be considered the latest development in a long series of technological advances driven by the engine of human information input. Even if every one of the improvements stated above are tangible, how well does it translate to the ultimate outcome? Organizations around the globe have been trying to digitalize and computerize, some as early as the 1960s and 1970s. Digitization and computerization promised a revolution in productivity through automation, streamlining, and making humans redundant. But the road to a more significant productivity boost has been arduous at best. One of the primary reasons is captured in the Theory of Constraints (TOC) developed by Eliyahu M. Goldratt, an Israeli physicist turned business management guru. In his landmark book The Goal, which sold over 6 million copies, he explains the importance of focusing on the right constraints to improve the overall system. For projects, the constraints are generally not within the technology and tools, but within humans. — Healthcare Business Today, 1d ago

new Incorporating bioethics principles into medical AI algorithms is undoubtedly a crucial aspect of creating trustworthy technology that serves to advance society. While the paper highlights multiple critical topics that should be considered and offers robust solutions, gaps and questions remain. More research needs to be done regarding when these AI tools are considered to be conscious of their own actions and, by extension when liability is placed on humans or the machines. Additionally, while the authors’ proposed solution of instituting liability fees and other systems between public and private parties is an interesting point to consider, it may be difficult to establish in countries such as the United States, where the healthcare system is incredibly disjointed. Additionally, in places with many competing private healthcare companies, it should be considered that these parties do not necessarily have patients’ best interests at heart. Instead, they tend to prioritize profits over patient well-being—thus adding another obstacle to ensuring ethical medical AI is instituted. — Montreal AI Ethics Institute, 1d ago

new Our paper shows many examples of AI systems that have learned to deceive humans. Reinforcement Learning (RL) agents trained to play strategic games have learned to bluff and feint, while large language models(LLMs) will output falsehoods in creative ways that help achieve their goals. One particularly concerning example of AI deception is provided by Meta’s AI, CICERO, which was trained to play the alliance-building world-conquest game Diplomacy. Meta put a lot of effort into training CICERO to be “largely honest and helpful,” claiming that CICERO would “never intentionally backstab” its allies. But when we investigated Meta’s rosy claims by studying games that CICERO had played, we found that Meta had unwittingly trained CICERO to be quite effective in its deception. — Montreal AI Ethics Institute, 1d ago

new Diabetic retinopathy (DR) is one area where AI is being used as a screening tool. According to Dr Gopal Pillai, professor and head of department of ophthalmology at Amrita Institute of Medical Sciences in Kochi, approximately one third of people with diabetes will have DR in India, and about 10-12 per cent of those will have vision-threatening DR, indicating that the development of DR has seriously affected the patient’s vision and failure to treat it in a timely manner will result in irreversible vision loss. One thing that was hampering early detection of DR was that it is asymptomatic, until, that is, the person completely loses his vision. “One might be driving, reading, watching TV and doing their activities without even knowing that a time bomb is ticking inside the eye. And because the patient would often come in late, there was no way of early diagnosis,”says Pillai, who is leading a government-sponsored clinical trial network. — theweek.in, 1d ago

new The validation results reveal a remarkable alignment between LLMCarbon’s projections and real-world data for diverse LLMs, surpassing the performance of existing tools like mlco2. LLMCarbon’s adaptability to various data center specifications and its capability to pinpoint optimal parallelism settings enhance overall operational efficiency. This adaptability, combined with its ability to accurately gauge the environmental impact of LLMs, positions LLMCarbon as a pragmatic tool in assessing and mitigating the carbon footprint associated with LLMs, offering an indispensable resource for the future of sustainable AI development. — Montreal AI Ethics Institute, 1d ago

new Predictive work systems and automated testing equipment are already saving health systems time and money, and the benefits are sure to grow as the technologies advance. However, human oversight is still needed. A technician should still use their expertise to determine if a computer-generated solution is indeed appropriate, rather than trusting the proposed solution without critical thinking. It’s also important for health systems to engage with technicians when deploying AI tools—to gather feedback and ensure new procedures & technologies are enhancing their work experience rather than adding to their list of responsibilities. — Healthcare Business Today, 2d ago

new In brains, online learning (editing weights, not just context window) is part of problem-solving. If I ask a smart human a hard science question, their brain may chug along from time t=0 to t=10 minutes, as they stare into space, and then out comes an answer. After that 10 minutes, their brain is permanently different than it was before (i.e., different weights)—they’ve figured things out about science that they didn’t previously know. Not only that, but the online-learning (weight editing) that they did during time 0<t<5 minutes is absolutely critical for the further processing that happens during time 5<t<10 minutes. This is not how today’s LLMs work—LLMs don’t edit weights in the course of “thinking”. I think this is safety-relevant for a number of reasons, including whether we can expect future AI to get rapidly smarter in an open-ended way without new human-provided training data (related discussion). — alignmentforum.org, 2d ago

new Now these are really hard conversations to have. I'm using very human words to describe an AI which don't fit, but the shorthand is unavoidable. Now, doesn't AI have a value? No. Does it pretend? Yes. Can it mimic values? Yes. Can it do it? Well, not yet. All these things are going to change. A lot of this is in flux, but I think this is really interesting as these AIs become capable of doing things that used to be the exclusive purview of humans. So now I want to give the big problem here. All of the AIs that we have are built, are designed, are trained, and are controlled by for-profit corporations. — harvard.edu, 2d ago

new California’s report also usefully lays out the potential risks associated with using these new tools, making clear that while there are some new potential harms, in many cases many of the risks are common to the use of any technology. Governments need to be conscious of the fact that tools that enable the easy generation of high-quality content could be misused to dupe consumers and residents.Perhaps because 35 of the 50 leading AI businesses are in California, as the state's report points out at the outset, it is silent on the risks to governments and those they serve of relying excessively on technologies developed and governed by unaccountable companies, especially when those technologies are procured by public servants without a deep knowledge of the tech. — GovTech, 2d ago

new ROBERT CHAPMAN-SMITH: In the lead up to, you know, the release of some of OpenAI's models, there, there's been sort of like a speaking tour of folks in going to Washington talking to legislators about AI and there was a, the worry at that time was about regulatory capture. Like are they, are folks going to essentially gate the technology in such a way that smaller players are not gonna be able to play ball? And we've seen regulatory capture happen a lot within the political realm within Washington. But there's also this question of like effectiveness in terms of regulation. Like just because the regulation has passed doesn't mean it's actually a good regulation or if this body of Congress is actually able to regulate this fast-moving technology well, like they can't even pass a budget, like how are they gonna keep up with the pace of AI change? So I'm curious about that as a tool for dealing with AI safety because it in some sense it feels like one the the legislative body or processes or capable to be captured by interested parties and two, even when they do regulate sometimes they just do a poor job, they just miss the thing that is the key regulatory factor. So I'm curious about your conception there, and how to deal with some of the messiness that comes with those types of approaches to dealing with technological safety. — Big Think, 2d ago