@elonmusk The interesting part isn’t just putting compute in orbit.
It’s that agentic systems are starting to force the physical layer to evolve as fast as the software layer.
That’s a much deeper systems problem than most people are pricing in.
@RaoulGMI The missing layer is still the same one:
Who authorized the agent, under what policy, with what evidence trail, and who remains accountable when it moves real value.
Programmable money is necessary.
Governed authority is the harder part.
@simonmaechling This is one of the cleanest framings I’ve seen.
Intelligence is getting cheaper.
Evidence and reconstructable decisions are not.
That’s still the real bottleneck for any system that has to operate under institutional constraints.
The more complex the system becomes, the more time I spend protecting coherence instead of adding features.
It’s unglamorous work.
Building the model of the system before forcing every part into production is turning out to be one of the highest-leverage decisions.
#Ai#Tech
Agents already consume the majority of AI capacity.
The next constraint is not more intelligence.
It is whether that intelligence can operate inside a governed context - identity, permissions, evidence and reconstructible authority.
#Ai#Agent#Grok#OpenAi#System
@GergelyOrosz The analogy breaks the moment you need real accountability and reconstructable decisions.
Agents can execute tasks. They still don’t carry institutional authority.
@paulg Especially true on the first company. You lack reference points in both directions - so the same data can feel either catastrophic or insufficient depending on the day.
@paulg This is the cleanest way I’ve seen to reframe “is this good enough?” for early teams.
Compounding at that rate forces the real constraints to surface fast - team, systems, and coordination capacity.
@LexSokolin The capital intensity changed faster than the operating assumptions.
Most of the hard problems now sit after the model - identity, permissions, evidence, and whether the system can still be underwritten by institutions.
@martin_casado@sriramkri Exactly. Emerging managers aren’t just competing for leftovers - they often create access to opportunities the large funds would never see in the first place. That’s market expansion, not market share redistribution. ✊🏻
@paulg Preventing data centers doesn’t slow AI.
It just decides which country gets to deploy it at scale.
The constraint is shifting from model capability to the physical and coordination infrastructure required to actually run it.
You can turn capital into capability relatively cleanly now.
Turning that capability into reconstructable, permissioned, accountable action inside real capital systems is a different problem.
#Regulation#System#Capital
The interesting part of the current AI capital wave is not that money has become more productive.
It’s that the surface where that productivity becomes institutionally usable is still mostly missing.
#Ai#Money#Capital
@a16z The productivity jump is real.
What’s still scarce is the layer that lets that capital actually move with clear authority, evidence and accountability once the model is trained.
Do you see the binding constraint shifting next - model, capital, or surface around both?
@VitalikButerin Interesting direction.
Fact that local mixing is computationally realistic today while the security question remains open feels like a classic “capability ahead of assurance” problem.
Curious how the research community is thinking about verification timelines on this family.
The political resistance is downstream of a deeper mismatch: intelligence capability is scaling faster than the authority, energy and coordination infrastructure required to deploy it at societal scale.
🚨 POD UP! The Core Four are back!
-- Dario Amodei Breaks His Silence: Regulatory Capture or Real Concern?
-- The Anti-Data Center Revolt: Abbott, Shapiro, and the GOP Panic Memo
-- A "DMV for AI"? Sacks Says the Open Source Ban Is Coming
-- Friedberg's RSI Thesis: Why Regulating AI May Be a Fool's Errand
-- a16z Under DOJ Investigation Over a 1914 Law
-- Midterms Toss-Up: 53% of Young Conservatives Want Government Grocery Stores
(00:00) Besties are back!
(00:13) Dario's two-part essay: regulatory capture, doomerism, and the data center backlash
(10:25) FINRA for AI vs. MPAA for AI: SROs, thinking tokens, and the "DMV for AI"
(30:12) Is an open source ban coming? Harnesses, FDI, jobs, and recursive self-improvement
(56:33) a16z under DOJ investigation over "interlocking directorates"
(1:01:15) Midterms: broken polls, the socialism surge, and CATO's DSA price tag
Come hang with the Besties at the All-In Summit | September 13-15:
https://t.co/ctguerI0TE #AllInSummit
Hardening authorization and untrusted client state is unglamorous work.
It is also the difference between a demo and something an institution can eventually trust.
Evidence still open on a few surfaces. No announcement until the smoke tests close.
#Building#Dev#Demo#Ai