“Study estimates Texas NOx emissions could be 2 to 10 times higher if data center demand is met with on-site combustion rather than grid power paired with solar and wind.” Somewhat a worst-case analysis, but points to potential impacts. | https://bit.ly/4z926SL 

Increases in emissions for these and similar businesses should be fairly straightforward to model and then compare risks and benefits. While some are trying to limit public input and permitting requirements, we could streamline analytical requirements and data disclosures. Air quality issues are not produced or solved by any one facility.  

Maintenance engineering or the sustaining of technology and structures after their design and installation is often not seen as the fun, sexy part of engineering with tasks doled out to inexperienced or less qualified colleagues in some organizations or occasionally not involving an engineer at all, however, it is just as critical to the health and safety of the public as any initial design is. | https://cnn.it/4hP5v2S 

I have no special information as to what causes this investigation is likely to uncover, but it would not be surprising to find insufficient engineering attention paid to failure or anomaly reports, the design of maintenance schedules in light of that information, or changes or modifications to the initial design. Risk management depends on giving appropriately qualified people the time to consider the implications of the information available to them. We will likely conclude in this case that we did a poor job of that and it cost us a man’s life. We are getting away with that same inattention right now in dozens or hundreds of similar cases by good luck that won’t last forever. 

Ephemeral Saboteurs. Much to the surprise of many people, even new engineers, we rarely do anything in the design of technology to prevent intentional misuse. We rely on post hoc enforcement actions (criminal law, civil law), social stigma, and the potential saboteur’s own fear of harm to deter that misuse. A lot of the anxiety around AI control of technology is probably related to the potential loss of those mechanisms. What do we do to a complex, electronic mind that cannot really be punished and lacks predictable human self-interest?

Running data centers on power from natural gas that would otherwise be flared would seem to be a low net environmental impact <https://bit.ly/3TW9EbZ>. Still, though, the “small-scale” designation can seem to be hiding the true impact from the public’s perspective. The costs to do air quality modeling and install monitoring devices is falling, so maybe we should just plan to do this for all industry to instill confidence and ensure the cumulative impact is being managed appropriately.

“For world leaders, one key question about AI is this: Who should control life-and-death decisions — human or machine?” | https://to.pbs.org/4yVo7Un | Public figures seem to be focused on killer robots and drones that may decide on their own what targets are valid and what means are justified. This is absolutely a decision that we need to clearly decide as a society and enforce, but there are many more lurking behind this one.

Who decides when risk to human life is at stake: when the AI surgeon choses one technique during an operation that reduces the probability of patient disability but increases the probability of patient death; when the AI operated vehicle takes action to decrease the probability of a pedestrian death and increases the probability of severe passenger injuries. There are so many potential questions here, where system design is making decisions that we could never before expect humans to be able to make.

The framing of this article is from the perspective of financial analytics, but I think it applies more broadly: “Why AI agents are like staff you can’t trust” | https://bit.ly/4hJJJ06. 

In engineering and technology development, losing credibility can be the end of one’s career, or at least the end of promotions. The risks are too great. We don’t put public and worker safety and the company’s viability in the hands of an intern, though interns can be valuable team members and can grow into that responsibility with time. Maybe AI models will, too, or maybe they never will, but for now, everything we let them do requires scrutiny.  

I guess the idea is that if air pollution data is not disclosed during the permit phase that there will be less opposition to the development and a greater chance of making the installation a fait accompli? | https://bit.ly/4wSBtiU | not sure it really works like that | transparency is more congruent with sustained political support; companies should be clear on how they are mitigating the negative impacts, and ensuring that sufficient benefits are distributed locally as well 

Reading over the NY RAISE Act (Responsible AI Safety and Education), I see it includes some good components such as requiring public publication of a “safety and security protocol” with requisite testing and reporting of certain “safety incidents”, but its definition of “critical harm” is an incident that causes “death or serious injury to 100 or more people or one billion dollars of damage”.  Most risk management standards would call any human death “catastrophic” and require documented mitigations. And, many future specialized AI models may stay below the law’s threshold. So, this is fine as long as it is the start of conversation around AI regulation and not the end of it. | https://bit.ly/3V6rVUg 

The objective for environmental and EHS engineers is to find ways to make our technology less harmful and safer even in cases of failure like this F-16 crash. Hydrazine is toxic, but energy dense, and so is used in spacecraft thrusters and emergency power systems on military aircraft. Are there other, safer options? Probably, but they await development and testing that hydrazine already has or they have their own negative tradeoffs that we need to figure out how to mitigate. We need new engineers to dive in and help solve these problems. | https://bit.ly/476YPH0 

“[Some] business leaders are accelerating the deployment of agentic artificial intelligence with ‘no idea how to manage risk.'”

The tools for risk management of any technology including new ones like agentic AI are similar: hazard analysis, failure analysis, probabilistic modeling, testing and validation, and quality control. This may be a whole new career pathway, and maybe we should be offering degrees for it, but there are skilled professionals out there already with the background and experience to assist businesses with managing their risks with agentic AI: safety and reliability engineers. | https://bit.ly/4xTVWoB 

This effort to write a safety bill for AI really illustrates the problem of treating AI use as one issue. There are the issues associated with AI-provided content for underage users; there are the issues of using AI models in energy systems; there are the issues associated with AI-suggested harmful medical advice; there are the issues associated with AI use in transportation systems and vehicles; there are the issues associated with AI misuse in education. We don’t regulate human decisions in these diverse areas with one set of guidelines, or one set of risk management or training structures, so we don’t need to do that with AI either. But that doesn’t mean doing nothing. Congress needs to make sure the regulatory agencies have the tools they need to address this new technology. | https://bit.ly/4yIfA7j 

So, the main policy question for congress is what agency or agencies should regulate AI models and their uses in the US. My opinion is that agencies’ jurisdiction should cover AI use in their domains. FAA for aviation, NHTSA for automobiles, CPSC for consumer products, etc. No new AI regulator is needed. Just make sure they each have the necessary authority to develop rules, testing and validation protocols for their industry. Some kind of update to each agencies authorizing legislation may be needed plus direction and funding from congress to be sure this is initiated sooner rather than later. | https://bit.ly/4h4kp3l 

“AI Safety” to me looks like 6+ different problems wearing a trenchcoat. Some of these risks are likely ready for regulation, but others need a lot more investigation and soon. | https://reut.rs/4yxPiUX 

I guess the idea is that if air pollution data is not disclosed during the permit phase that there will be less opposition to the development and a greater chance of making the installation a fait accompli? | https://bit.ly/4wSBtiU | not sure it really works like that | transparency is more congruent with sustained political support; companies should be clear on how they are mitigating the negative impacts, and ensuring that sufficient benefits are distributed locally as well 

“That means even NHTSA’s report about agency lateness is late.” | https://bit.ly/4A9SBDM 

Ever wonder why, if wildfires cause so much negative air quality impact, prescribed burning, which also emits a fair amount of pollution, is mentioned as a primary mitigation for wildfires? Aren’t they doing the same thing? The answer is related to magnitude. This 2017 paper collected data on this subject: https://bit.ly/4ctavXS. Wildfires, beyond their risks to life and damage to property, burn hotter and longer (as if years of potential prescribed burns happened all at once, with greater intensity, and more complex chemistry) than prescribed burns generating more pollution per unit of land area, lofting that pollution higher in the atmosphere, and affecting a wider area. Mitigating the human health impact of a prescribed burn is easier, more local, and less disruptive.