I work in AI safety and Human Data Ops research, so I can loosely answer most of these from my experience as somebody who is not privy to top level projects. Obviously, this is one person's perspective rather than a definitive view of the industry and will keep it broad in line with NDA.
1)Is there a way to switch off all the electricity that AI runs on - if it goes rogue and starts to unleash nuclear weapons for example?
A) "AI" isn't run from one central power source, so not in that sense. In real terms it would require the specific system to be taken offline, its credentials revoked, isolation from a network etc. There are many layers to the process. We do test for this eventuality and close-down is possible because human oversight is currently, retained. The AISI has also recently demonstrated that it can terminate and isolate systems during safety testing following a security incident. That shows shutdown and containment are feasible in a controlled environment, although it would be wrong to extrapolate that into a guarantee that every conceivable future AI system could always be stopped.
2)Are AI experts attempting to teach AI to behave in a moral and ethical way?
A) Definitely yes, there are huge numbers of people working in the domain who are very concerned about this and many job roles and departments dedicated to safety and ethics. I have yet to work on a project without a safety team, ethics consult, or significant testing of all throughput and output for safety, risk, bias or harm across all metrics and rubrics. We test through algorithm and human oversight across all types of output and actively try to jailbreak models to expose flaws. This has been the case regardless of which company I've worked with on a model. I've worked on practical and research products around ethics governance and its very real and ongoing but hasn't reached an end point at this stage. AKA we cannot guarantee with certainty that a failure case is not possible - but this applies to many industries and circumstances where harm can be done. Of course there will also be players in the game who are less concerned with this - just know there are lots of players in the game who are and who push for it.
3)Are tech experts putting enough safeguards in place - to protect us from plausible dangers eg bad agent’s engineering deadly pathogens?
A) Bad Actors are a critical risk, yes. Are there many safeguards, yes, are there enough safeguards - that's difficult to quantify and we just don't know.
- Is the Government prepping for an AI induced breakdown on infrastructure on a massive scale?
A) Yes, they are and more than ever. Digital resilience is high priority and the National Risk Register has been updated to include this risk across our infrastructure. Taking a non-sensationalist view I would say that the government expects there to be incidents of significant digital disruption and is equipping itself for these, not that it believes or suggests we are all about to be killed by a rogue model or bad actor with access to technology.
5)Can you think of ways we can help ourselves?
I would recommend avoiding a reliance on digital technology. If you are concerned about the environmental and societal implications, consider whether every use is actually necessary. Consider seriously (without a panic mindset) how you would cope and what you would do if you found yourself in a situation where you needed to go up to a few days without access to digital wallets, power, documents, contact details, medication etc. Actively canvass for sensible regulation via your MP, respond to government consultations, vote with AI policies as a consideration and keep up to date with the work and advice of independent safety organisations such as AISI and don't amplify sensationalist claims if you don't fully understand whether or not they are factually accurate.
All those aside I would advise people to educate yourself on what "AI" is and isn't. There are hundreds of thousands of different and developing models and applications. It is a technology with many capabilities and a vast range of failure modes.
On Jacob Coxon referencing a 10% chance this will actually come to pass,I think the important distinction is that he is expressing a subjective probability, not establishing that there is objectively a 10% chance of human extinction. He is rightly concerned about the push forward of technology at a rate not sufficiently supported by a check and balance. He has seen enough in-role to support that stance, Anthropics own research supports that possibility and it's worth considering that there are also others in comparable roles who disagree with those findings. They are all referring to future models, not ones currently in the public domain.
One can only hope that flagging the issue as loudly as he has done achieves what he appears to want it to achieve: slowing the progression of increasingly autonomous and potentially self-improving systems enough to ensure that we develop the ability to control and govern them appropriately.
Ideally, the relevant labs and governments would act collaboratively and within an effective regulatory framework towards a safer outcome, rather than creating a race in which everyone feels compelled to be the first to reach increasingly powerful systems regardless of whether the necessary safeguards are in place, which is what he is objecting to and why he has quit and made a public statement.