Balaji Srinivasan

Asks whether AI stays on the leash. Humans remain upstream as goal-setters and sensors of markets and politics; AI will become better than most humans at verification, but he is “not so sure” it will replace the human prompt. Independent goals would require AIs to reproduce without human cooperation, with robots building datacenters, mines and power plants, which he calls not technically inconceivable. In a scenario he labels sci-fi, China builds cryptographically chained “AI slaves”, and humans plus leashed robots hunt down self-replicating ones. Full long-post text inspected via the X API.

Argues that any state or stalker running an AI model can now synthesize scraps of online information into dossiers beyond Soviet dreams. No silver bullet exists, but anything unencrypted can be used against you, so he urges encrypting everything with zero-knowledge methods. A threat diagnosis and defensive prescription, not a full privacy policy. It quotes his own “From AI to ZK” post. Full text inspected via the X API.

Offers a view of AI as neither zero nor infinity: it boosts productivity within trusted tribes but reduces economic trust between them through spam, scams, slop and fake resumes. Verification, proctoring and vetting costs soar, so AI may break almost as many markets as it enables. He thinks China may centrally curate its internet against fakery, while America likely lacks a large trusted tribe. He suggests the future could look like a thousand small network states plus one Chinese superstate. Full long-post text inspected via the X API.

Calls Block’s 40% headcount cut the first AI cut and a signal to raise your game and learn AI tools. He calls the laid-off a temporarily unfortunate class rather than a permanent underclass, and praises the severance. He says AI is no panacea and there will be overcorrection, but the technical innovation is real: disrupt yourself or get disrupted. About one company and tech workers, not an economy-wide unemployment forecast. Quoted announcement belongs to others; full text inspected via the X API.

Argues American AI companies are fighting Democrats over job automation, Republicans over the military, and China over distillation at once. The equilibrium, he says, is that open source models become the only trusted models: centralized American AI makes a lot of money but is eventually outcompeted by decentralized local AI. A forecast, not a policy demand. Quoted post belongs to someone else; full text inspected via the X API.

Argues that every economically useful AI agent does what it is asked on command, so digital AI is built for the leash. Physical AI is even more so, because China builds most robots and will create “robot slaves, not robot gods”. He still calls a Chinese drone armada fearsome: the problem is a billion Chinese AI slaves, not a Western AGI god. Qualitative, with no probability. Quoted post belongs to someone else; full text inspected via the X API.

A follow-up to his doom post: the problem is Chinese AI domination rather than AI doom, and the likely solution is Internet AI decentralization. The quoted post about the German chancellor’s China trip is someone else’s and is not examined here. Full text inspected via the X API.

Disagreeing with a quoted post, he says Democrats want to stop AI to protect blue jobs and Republicans want military-friendly AI, but China and the rest of the world want open source, which is likely where things land once capabilities top out. America spends billions as AI’s “bootloader” while wealth taxes, visa limits and building bipartisan animus toward technologists scatter Silicon Valley. He concludes no American faction will control AI; it will decentralize. A forecast mixing his political views with AI; full text inspected via the X API.

Says current AI agents are bots on a leash, built for the prompt: amplified rather than truly artificial intelligence. Agentic workflows help and time horizons have lengthened since Claude Cowork, but human prompting and verification remain the bottleneck, and without verification a codebase fills with economically irrelevant slop. He cites a critique of the METR time-horizon study as showing a sigmoid on messy tasks, and allows technology may change this. Full text inspected via the X API; the linked critique was not reviewed.

Agrees with Alex Tabarrok that AI hyperdeflates costs in code, math, biomedicine and robotics and will create a lot of wealth, but says it also floods sales, recruiting, identity verification, education and social media with scams, spam and slop, raising verification costs and digital tribalism separately from any job effects. China may damp these harms more easily; the free Internet needs a web3 of trust. Costs should be enumerated and mitigated. Tabarrok’s quoted text is not his; full text inspected via the X API.

Reacting to the America.gov launch talks, says an AI interface to the federal government, if it works as promised, would help every American and blunt claims that AI benefits only the very rich, easing anti-datacenter sentiment. He sees digitized, auditable government services as a tech-led reform and a version of the network state. Conditional enthusiasm about a launch, not evidence the system works. The talks themselves are others’ statements; full long-post text inspected via the X API.