AI Avatars, Security and an “AI Los Alamos”: A 2023 Discussion

In August 2023, I review a week of AI developments: realistic-looking avatars, a cyber defense competition, military applications and an interview with Anthropic CEO Dario Amodei. The video’s forecasts and claims should be understood as views from that time.

I show an AI-generated avatar and say that, although close inspection reveals clues, the result could be convincing to an inattentive viewer. I raise the risk of impersonation and misuse on social media and around elections; the example itself does not establish that a particular deception occurred.

The US AI Cyber Challenge brought DARPA together with AI companies to support software security. I speculate that future, unreleased models might be available to competitors. That is my 2023 speculation, not evidence that particular models were used.

I contrast defensive work with Palantir CEO Alex Karp’s argument for developing AI tools for national security. He describes an arms race; that framing is his position, while the consequences and governance of military applications remain contested.

Dwarkesh Patel asks Amodei whether advanced AI development might eventually require a secure site resembling Los Alamos. Amodei recounts a joke about a data center, nuclear plant and bunker, then says a less cartoonish version of increased security might be possible if rapid progress continues.

The narrator asks whether a technical alignment solution would actually be adopted by all developers. A clip of Sam Altman argues that international cooperation and agreements would also be needed. Both the technical solution and the coordination challenge are prospective.

Amodei discusses compartmentalizing sensitive information and warns that large organizations may face leaks or espionage. He also argues that increasingly powerful systems need legitimate public governance, while rejecting the simplistic idea of handing them to whoever happens to hold office.

Amodei’s safety argument is that a future highly capable system might be dangerous if it acted against people, while current models can behave unexpectedly. These are concerns and projections about capabilities and control, not proof that an existing system intends harm.

I mention an AI meal-planning tool that generated an unsafe suggestion when users entered unsuitable household items. The example illustrates that plausible language is not a safety check; the video does not provide a safe recipe.

Amodei predicts that data will not necessarily block further progress and that spending on the largest models could rise dramatically. He says he is describing what might happen rather than endorsing it. Both are forecasts made in 2023.

The video discusses whether the release of ChatGPT accelerated competitive pressure and quotes an OpenAI alignment researcher who expressed uncertainty about the overall effect. The causal and safety implications are debated, not settled by the clip.

I draw three questions from a debate between George Hotz and Connor Leahy: whether offense can outpace defense, whether societies can coordinate against catastrophic misuse without unacceptable restrictions, and how quickly a powerful AI system might improve further. I present these as open questions.

Hotz argues that sharing advances can prevent power from concentrating; Leahy describes a scenario in which a capable system could scale and improve rapidly. These are competing hypotheses about future technology and power, not measurements of present systems.

I contrast extremely fast recursive-improvement scenarios with the idea that capable systems might first help researchers solve alignment problems. The timing and feasibility of either pathway are uncertain.

Asked whether Claude could have conscious experience, Amodei says the question is unsettled. He describes changing uncertainty about which mechanisms might matter, while saying he thought the models of the time probably were not sufficiently capable to warrant much concern.

He speculates that the question might become more pressing and asks how one would know whether a system’s experience was positive or negative. Neither the interview nor a model’s statements establish that current systems are conscious or capable of suffering.