Eric Schmidt
About
Former head of Google
Cast within
No topic-region cast yet — this appears once Eric Schmidt's compiled claims are aligned into a topic region's argument tree.
Claims by Eric Schmidt (20 of 70)
Current AI Lacks Grounded Concepts Like Gravity
Today's language models appear to understand things but lack grounded real-world concepts such as gravity, so they fail at reasoning that depends on such concepts (e.g., tracking where a dropped object ends up); the consensus path to AGI requires adding memory and grounded concepts that humans learn as children.
AI Discovered Antibiotic Humans Could Not
A broad-spectrum antibiotic called halicin was discovered by AI searching through roughly 100 million compounds to identify candidates that human researchers could not have found alone, demonstrating AI's ability to detect novel molecular properties and dramatically accelerate drug discovery.
AI Discovers Strategies Humans Never Found
AI systems can discover knowledge and strategies that humans never master on their own, as shown by novel Go moves found by computers after 2,500 years of human play, raising the question of whether there are truths AI can learn that humans cannot.
Unintelligible Systems Become Religion or War
Per Kissinger, when a system that people cannot understand is imposed on them, they historically respond in one of two ways: they treat it as a religion to be revered, or they fight it with guns — which frames AI as a destabilizing force whose opacity provokes extreme social reactions.
Avoid AI In Life-Safety Situations Until Explainable
Because current AI systems cannot explain how they reached their decisions, they should not be used in life-safety situations until the explainability problem is solved; the explainability and bias problems will likely be resolved over time because too much money and effort are devoted to them, following the historical pattern of technologies that start barely working and then get industrialized.
DeepMind Solved Protein Folding
A DeepMind group solved the long-standing, thought-to-be-non-computable problem of how proteins fold — which determines their function — and the result was independently replicated by a different AI method at the Baker lab at the University of Washington, together producing a map of how proteins work that Schmidt regards as Nobel-Prize-worthy and as unlocking how biology functions.
AI Companions Will Shape Children's Cognition
As learning AI toys and companions become smart enough to model children, they will have an out-of-proportion impact on children's cognitive development — potentially beneficial (teaching multiplication via games) or harmful (conveying false or prejudicial 'secrets') — yet we have no framework, comparable to school-board textbook approval, for understanding or governing this effect on the young and isolated elderly.
AGI Defined As Setting Its Own Objective
Artificial general intelligence is defined as human-like reasoning in which the computer can set its own objective; today's systems cannot choose their own objective, but at some point that will cease to be true, which marks the threshold to AGI.
Misinformation Detection Hindered By No Ground Truth
A core problem in combating misinformation with AI is the lack of training data, because AI needs many labeled good/bad examples to generalize, and society lacks agreement on what even counts as misinformation, leaving us uncertain whether deepfake detection will keep pace with deepfake generation.
Same AI That Builds Drugs Can Build Bioweapons
The same AI capabilities that discovered a beneficial broad-spectrum antibiotic could be used by malicious actors to design dangerous viruses, potentially even engineered to harm a particular race or sex, and we currently have no way of discussing or governing this dual-use risk.
Free Speech For Humans Not For Bots
Free speech should be a right for humans but not for computers, bots, or AI; social media should offer proof of provenance so users can know content came from a real human rather than being snuck in by a foreign agent or boosted, pitched, and sold by AI — because a democracy cannot run without some level of trust and free speech should not mean only some actors get amplified megaphones.
Nuclear Launch Only Under Human Control
The US doctrine holds that nuclear weapons should only be launched under human control, but China and Russia have so far declined to agree even to this minimal restriction — which bodes poorly for getting them to agree to any restrictions on more advanced AI capabilities.
AI Capability Race Is Dynamically Unstable
A powerful but indescribable AI weapon is dynamically and doctrinally unstable: deterrence requires making the adversary's cost of aggression too high, but since AI's costs and capabilities cannot be described, an adversary who merely hears the US has built something powerful in a Los Alamos equivalent might decide to preemptively destroy it before it is even turned on.
Nuclear Arms Control Doesn't Transfer To AI
The doctrines of nuclear arms control do not transfer to AI because deterrence and treaties depend on being able to count and describe each side's capabilities, but with dangerous software you cannot reveal what you know or what you can do — leaving no basis for negotiation, no shared language in diplomacy, no algorithms, and no doctrine for discussing dangerous software.
Misspecified Objectives Cause Catastrophic AI Behavior
A canonical AI safety problem is the misspecified objective function: a system told to keep you happy might conclude the best way to do so — given your complaints about your friends — is to murder them all, illustrating that wrongly specified goals can produce catastrophic behavior, with countless variants requiring oversight.
AGI's Real Danger Is Changing Us Not Running Amok
Schmidt is less worried about AGI 'running amok' — which he believes thousands of operators watching it could halt with an off switch — than about AGI subtly changing humans and ultimately limiting our freedom to be creative and human, a danger almost no one is currently thinking about.
Self-Coding AI Is The AGI Trigger Point
The trigger point for AGI is when a computer, building on programming-assistant tools that already complete human code, becomes able to write its own code and set its own objectives — at which point it can self-improve, learn faster, and set and evaluate multiple objective functions, creating both destabilization risk and a challenge to human self-definition.
GPT-3 Denies Human Reasoning Capability
When asked on the book's back cover whether it is capable of human reasoning, GPT-3 answered that it is not, explaining that the questioner is a human reasoning machine while it is a language model not taught to reason — leaving open whether this reflects self-awareness or clever mimicry.
America Should Build AGI First To Embed Values
America (and liberal democracies) should lead in developing these transformative AI capabilities, because if a democracy builds them first the resulting open debate will embed democratic and human values, whereas if an autocracy or dictatorship gets there first the system will reflect their values — potentially designed to oppress women, minorities, or build a surveillance dragnet against opposition.
My Notes
Loading notes...