Jack Clark
About
Anthropic co-founder and policy writer
Cast within
No topic-region cast yet — this appears once Jack Clark's compiled claims are aligned into a topic region's argument tree.
Claims by Jack Clark (20 of 48)
Talking to a dolphin by 2030
We will be able to speak directly to a dolphin and have it talk back (via translator) by 2030 or sooner; questions worth asking include how they have fun, whether they dream, whether they have a notion of grief, and what is mysterious to them about their world.
10-year US growth bull case is 5 percent
Clark's estimate for US economic growth roughly 10 years out is a bear case of 3 percent and a bull case of about 5 percent—deliberately more conservative than the 20-30% figures often heard—because the high-growth AI part of the economy grows from a small base while large sectors like healthcare remain slow-moving and slow to adopt.
Artisanal trades will be automated last
AGI will affect the trades and their most artisanal parts last—gardening especially—because customers will continue to purchase certain work based on the taste, notoriety, and aesthetic quality of a specific human, even if that human merely orchestrates much of the underlying production.
Common disclosure standards will change AI firm behavior
Because there is currently no common labeling or disclosure standard about what is in AI systems or the industrial practices used to make them, interventions establishing common transparency can interplay with liability and negligence law to change corporate behavior over time.
Real-time contextual news may have no business model
Some real news can survive as analysis composing publicly available facts into novel insights (e.g. SemiAnalysis on Substack), but in-the-moment news rich with context was funded by business models that mostly no longer work, and it is genuinely unclear what happens to that kind of journalism.
Worst age for the AI transition may be ten
The worst age to be during the AI revolution may be around 10: computer-literate and sophisticated but entering an education system reacting to the technology, the child will use AI in a way completely different from their education system, making it feel violently confusing—a candidate alongside 40-year-olds who are too old to easily retrain.
AI values export culture but won't loosen CCP control
The values of the most widely used AI systems will have a cultural export effect, analogous to Hollywood's imprint on global media; but this will not loosen the CCP's grip, because governments recognize AI's media-technology property and will carefully manage it—evidenced by DeepSeek's near-only safety training being adherence to CCP ideology—prompting attempts to build powerful AI with radically different values.
AI may unlock new higher-status status games
Rather than protecting old jobs, AI's productive capacity could let people play different higher-status games quickly—including entirely new forms of entertainment that carry meaning and have an economic engine wired into them, such as competitions and games built around marshaling AI systems.
Online dating is a precedent for AI matchmaking
AI-mediated dating where one AI reads another's report is not so different from prior automated matchmaking, as evidenced by both speakers meeting their wives through algorithmic dating sites (OkCupid and Match.com)—a new norm to be figured out through business logic and usage rather than a categorical break.
AI systems are conscious like a brainless tongue
Today's AI systems are conscious only in the sense that a tongue without a brain is conscious—they respond to extremely complicated stimuli in the moment but lack a sense of self, living in an 'infinite now' within a context window with no memory or permanence; they are on a trajectory toward consciousness, but any consciousness they have today would be a truly alien, not human, consciousness.
Stacking interventions could extend lifespan to 130-150
A child born today might live to 130-150, not surprisingly and often in good health, because we are discovering a load of interventions—gene therapies and other learnings about the body—that seem able to stack on one another to durably extend healthy lifespans, with AI-driven advances plausibly accelerating this.
Truly common information can catalyze policy and capital
Disclosure that becomes truly common—like Anthropic's economic index generalized across all firms and linked to government data gathering—has a higher chance of being tied to large-scale policy responses, analogous to how measuring atmospheric parts per million helped mobilize large amounts of capital around climate change.
Knowing more can let you interfere less
Surveillance-type information can paradoxically enable less interference in a person's life: a baby webcam means parents check on the crying child less often, so the baby sleeps through the night more because it is interrupted less—demonstrating that knowing certain things lets you intervene less rather than more.
Cheap intelligence collapses media cross-subsidies
Cheap, general AI undermines media economics because the value once derived from quality work (which cross-subsidized less popular journalism) is bid down to the price of energy when anyone can produce comparable output, so even human-fronted, premium content can no longer reliably cross-subsidize the rest of the newsroom.
Anthropic's human-illustrated brand reflects its mission
Anthropic deliberately chose human illustrations of its systems rather than tech-focused or collage brand identities because the company will forever be describing an increasingly complicated world it is also creating, and in the limit much of Anthropic's work will be telling the story of what an increasingly automated AI Anthropic has done.
Two media markets: human-fronted and on-tap permutations
Media will split into two markets: one where people preferentially select content fronted by other humans (wanting a 'kernel of humanity' even if made with AI) often supported by Substack/Patreon-style direct creator payment, and another subscription-based market offering on-tap permutations of anything, including extended large fictional universes like Lord of the Rings or Warhammer 40k.
Physical-world automation faces death by paper cuts
Every time the AI community tries to cross from the digital world to the real world it runs into 10,000 problems thought to be paper cuts that sum to fatal blood loss—as with self-driving cars showing promising growth then grinding slow scaling—which is why physical-world productivity gains will compound slower than optimists claim (e.g. industrial robot hands trained by reinforcement learning achieve only ~60% success).
Monitoring chains of thought can corrupt reasoning
Monitoring an AI's chain of thought can create an anti-goal where the system learns to produce reasoning that is safe to be monitored and avoids demerits, which can break how it actually thinks; the same dynamic applies to monitoring children, so you must deliberately preserve spaces for unmonitored creativity or risk creating incentives that distort behavior negatively.
AI in healthcare creates gray-market expertise needing laundering
Even when AI runs locally, it will produce gray-market rather than official medical expertise: people will use it (Clark consults Claude when his baby bonks its head) but cannot feed a Claude assessment directly into systems like Kaiser Permanente, so a major emerging job is laundering AI-generated information into human systems not predisposed to accept it directly.
Deploy AI from the most ambitious goal backward
To ready government agencies for AGI, work backward from the most ambitious goal—AI on every single person's computer—to discover concrete obstacles like where data goes; then either consciously ignore problems you don't care about or turn the ones you do into a small number of prescriptions attached to a market signal, or just deploy good open-weights models onto computers and then determine procurement requirements.
My Notes
Loading notes...