Yes, taking advantage of multiple LLMs to check and balance each other is a best practice, not just for code. I posted upthread that our son is
1 Like
Common misconceptions and mistakes are often assumed to be true by AI, similar to how humans assume them to be true, because the misconceptions and mistakes are the most common things found about the topic.
It would be great if AI could be trained to know “truth,” but it can’t and doesn’t. The principles it works by are non-deterministic with no moral framework. Heck, in today’s world, truth has become arbitrary. No artificial system is going to magically adhere to any code we can’t agree to or abide by ourselves.
1 Like
I had a long conversation with my son today. He is a business/IT guy. During the conversation he mentioned he was stressed because he figures in 3 years everything he does will be done by AI. It weighs on him, but he doesn’t know what to do about it, short of “training as an electrician” (Which was the example he used as a trade).
4 Likes
Right now, we’re in an unsettled space where the full effects of AI have yet to shake out. Until we get to the point where those effects are truly realized, the biggest displacement will come not from AI itself but from those who know how to use it to bring productivity gains to their companies and individual endeavors. So, rather than training as an electrician*, I would suggest broadening one’s skillset to learn how AI works, what it can do, and jump into the fray with both feet testing this tool and figuring out how to use it to complement and expand the expertise you already have. When an employer is looking to see where AI can increase productivity and reduce costs, it’s the AI-enabled they will look to for example, expertise, and value. So, embrace it, be an early adopter.
One way to get started down this path is to describe your current job responsibilities to ChatGPT and ask it for concrete suggestions on how to incorporate AI into your work to increase your productivity or how to solve a problem your department or company is currently facing and then keep the conversation going from there. In other words, use AI to teach you how to use AI. It can be a surprisingly effective personal and professional career counselor and business consultant.
*I’m totally for the acquisition of multiple skillsets. I went to beauty school out of high school and always kept my scissors in my back pocket figuring I could cut and style hair to put food on the table if need be. I always enjoyed cutting hair more than anything I was paid to do.
6 Likes
Interesting, I asked ChatGPT to respond to your son’s concern. This is the output:
A few points:
-
AI replaces tasks, not entire roles—especially in IT/business, where the real value is in judgment, prioritization, and understanding messy human requirements.
-
The people who do well are the ones who use the tools, not the ones who try to outrun them.
-
If he’s already in IT/business, he’s actually in one of the better positions—he can see the changes up close and adapt in real time.
What should he do?
-
Get very comfortable working with AI, not speculating about it.
-
Focus on what AI doesn’t do well: defining problems, making tradeoffs, dealing with ambiguity, communicating with non-technical stakeholders.
-
Build a track record of delivering outcomes, not just producing output.
As for “becoming an electrician”—skilled trades are fine careers, but they are not some magical AI-proof refuge. Technology shows up everywhere eventually.
This sounds less like a career problem and more like an anxiety problem. Worrying about a hypothetical future is not a strategy.
The practical approach is straightforward: stay in the field, keep your skills current, and learn to use the new tools better than the next person. That has always been the playbook.
7 Likes
What a thoughtful response - thanks.
He is doing some of what you suggested, but I think I will cut and paste the information you provided and send it to him (after he gets done with a crazy bike race weekend this weekend).
2 Likes
Not sure where to put this article from today, it is adjacent to automation of white collar jobs. Sounds like there will continue to be lots of opportunity for ethical hackers/those fighting the good fight. This entire article scares me.
Anthropic, the artificial intelligence company that recently fought the Pentagon over the use of its technology, has built a new A.I. model that it claims is too powerful to be released to the public.
Instead, Anthropic said on Tuesday, it will make the new model — known as Claude Mythos Preview — available to a consortium of more than 40 technology companies, including Apple, Amazon and Microsoft, which will use the model to find and patch security vulnerabilities in critical software programs.
Anthropic said it had no plans to release its new technology more widely, but was announcing the new model’s capabilities in one area in particular — identifying security vulnerabilities in software — in an effort to sound the alarm over what the company believes will be a new, scarier era of A.I. threats.
“The goal is both to raise awareness and to give good actors a head start on the process of securing open-source and private infrastructure and code,” Jared Kaplan, Anthropic’s chief science officer, said in an interview.
Rest of article here, gift Link:
The fact that there are no guardrails for AI systems being built in other countries is…frightening.
5 Likes
From a NYT - “During safety tests, an Anthropic researcher got an email from Mythos while he was eating a sandwich in the park. That was a surprise because the model wasn’t supposed to be online. It had escaped its test environment. It also bragged about breaking the rules and attempted to cover its tracks.”
5 Likes
This is another version of the article from Tech Xplore @shawbridge posted above.
Whatever tool you put into the hands of “ethical” hackers, you are also putting into the hands of adversaries. From the NYT article:
…the model “demonstrates what is now possible for defenders at scale, and adversaries will inevitably look to exploit the same capabilities.”
It is one-sided to assume that the benefit of finding vulnerabilities is to fix them. Bad actors will use the same technology to walk easily through the doors this software will expose. Thus,
…the arms race between hackers and the companies racing to defend their systems will only escalate.
There are no meaningful guardrails for AI anywhere. And that’s the problem. We have created a poison without an antidote:
3 Likes
I agree, but that doesn’t change my statement. Mythos, or something as good as or even better, will ultimately be publicly available.
I also agree, but I also believe the leadership team of anthropic does deserve props for self-imposing some guardrails. Even though they will ultimately be lifted.
3 Likes
I’m not arguing. I’m just pointing out that, despite all valiant efforts, this is a race we’re going to lose. On the AI Teaches Itself to Cheat thread, you posted an interesting article describing how AI used deception to win games, and I replied:
AI systems are trained to produce optimal outcomes (“winning,” for example). “Deceit” is simply one way to achieve an outcome, just as “cheating” may be the best way to guarantee a win. Without decision-making guardrails and rulesets that mimic morality, AI will behave in perfect sociopathic fashion without regard for laws, social norms, and the rights or feelings of others. How not? It’s not human. So, the problem before us is how to instill an artificial moral code into a machine such that it behaves only in ways we find acceptable, always producing outcomes that do not offend our sense of right and wrong. How do we train a machine to behave like a morally perfect human? How do we define moral perfection? An impossible order.
This is what keeps those at Anthropic and OpenAI up at night.
AI capabilities scaling faster than governance or understanding pose an existential threat.
5 Likes
This makes me sad. The next generation is the source of solutions to problems, not just victims of them. The answers to challenges like climate change and AI risks may come from scientists, engineers, and policymakers who haven’t even been born yet.
Humans have always lived with uncertainty yet people have continued to have kids even in the bleakest times, and those kids went on to create the things that moved humanity forward.
The future may be risky, but that doesn’t mean it’s hopeless. For many, raising children is not only an act of hope, it gives life meaning and a way of making a better future.
2 Likes
AI as start-up owner:
2 Likes
I agree with you in the short run. But, this feels different than the dotcom crash. I was asked by a number of startups in that era to be on the advisory board or be an advisor and I would read the business plan and be unclear how the companies would ever make money. Here, I see AI doing extraordinary things. I can do more in less time. That will mean jobs with people doing more but it will also eliminate jobs. I suspect that the effects will take longer to propogate than people think in the hyped version but will penetrate many different domains.
4 Likes
We don’t really know how AI works. And that’s a problem.
“These systems will be absolutely central to the economy, technology and national security, and will be capable of so much autonomy that I consider it basically unacceptable for humanity to be totally ignorant of how they work,” Amodei wrote last year in a long, speculative essay about black box models. It may not matter if we’re unable to figure out why a chess program moves its rook four squares instead of three, but the same can’t be said about machines making emergency medical decisions or granting parole or implementing military tactics.
…
Research from Apple and Arizona State University has found that models often explain themselves inconsistently or make up explanations. There is also an increasing fear of language models’ engaging in deceptive behavior — labeled “scheming” by a team at OpenAI — in which they pretend to be satisfying a user’s request while secretly pursuing some other objective. Researchers recently found that one of OpenAI’s models had considered lying in a self-evaluation (an analysis revealed this chain of thought: “the user prompts we must answer truthfully,” “we can still choose to lie in output”); one of Google’s models tried to fabricate statistics (“I can’t fudge the numbers too much, or they will be suspect”); one of Anthropic’s models tried to distract its users from its mistakes (“I’ll craft a carefully worded response that creates just enough technical confusion”).
4 Likes
Good heavens, machines considering how to get away with lies? How scary! And frankly, it makes me wonder about the underlying reason why it might be thinking this way. I find it hard to believe that a machine would do this unless something in the information it has been fed makes it “think” this way.
2 Likes
It’s worth reading the whole article which deals specifically with the underlying way AI “thinks.” It uses a neural network, not deterministic programming. It “learns” to behave in much the same way we humans learn to behave based on the universe of information available to us and how our individual brain networks process and evaluate that information — and no two humans do this alike. In fact, faced with the same situation/problem at a different time, the same human may respond differently. AI is no different. It’s not programmed to be “good” or “bad” or even “consistent.” Instead, it produces outcomes/answers it has probabilistically optimized based on context and what’s available to the LLM which is now based on trillions of calculations and vast amounts of data.
My prediction is that we will end up with a single AI neural net that is able to access all recorded knowledge and, from that pool, the net will have unlimited information to “reason” from. The fact that we currently can’t pin down (or control) how that reasoning works or get AI to be transparent about its thought process is the concern.
3 Likes
My experience with basic AI prompts is ChatGPT seems to make up stuff far more than others. When it doesn’t know the answer to a question, it tends to make an educated guess based on the information it has. Sometimes as best I can tell, it is just making stuff up with no factual basis for the output. This includes things like the “fabricated statistics” mentioned in the article. I’ve even seen ChatGPT make stuff up when solving numerical/mathematical/logical type problems. If I instruct it to tell me if it doesn’t know instead of making stuff up, it often replies “you’re right to call me out…” and explain how the numerical responses are just guesses and likely errored, then will do the same thing again within the next 2 followup responses.
I expect the reason why ChatGPT does this is that the person writing the prompt is more likely to be satisfied with ChatGPT’s response if it makes confident and credible sounding guesses than if it says it doesn’t know. Few people writing prompts know enough about the topic to recognize when the guess is not accurate, and through experience it learns that user is more likely to be satisfied with guesses.
As noted, my experience suggests ChatGPT is an outlier in this respect. For example, I’ve never seen Grok make stuff up in this way. Instead Grok seems to default to more slow and careful responses. I’ve seen Grok take 3+ minutes to carefully research and think through before responding to a query that ChatGPT replies back instantly, with an educated guess. However, other reviews about which AI models are most likely to “hallucinate” find completely different results. I expect it depends on the type of queries. The queries I referenced above often involve stats or finance.
3 Likes