Gutenberg Lab — Morning Paper

True Recursive Self-Improvement in AI Requires the Optimizer and Optimizee to Be the Same System

Speakers

Martin Casado (General Partner, a16z) × Daval Ajita (Co-founder, MakerMaker.AI) × Owen He (Co-founder, MakerMaker.AI)

Bottom line

Recursive Self-Improvement (RSI) in AI delivers sustained capability gains only when it simultaneously improves performance and reduces resource consumption—and that stage has already begun, according to empirical measurements over recent months.

3-Line Summary

The founders of MakerMaker.AI distinguish genuine RSI from autocatalytic development by asking whether the optimizer and the optimizee are the same entity—swapping these roles for the first time in history. In the domain of inference optimization, they report early evidence that both the numerator (performance) and denominator (cost) are improving simultaneously over months of observation, which, if sustained, breaks the assumption that capability gains require proportional resource increases. However, convergence properties depend on empirically measured exponents in self-referencing functions and cannot be solved analytically, and most published RSI claims failed to replicate in their internal testing.

3 Key Points

1. GPT-2 training cost dropped from hundreds of dollars to about $100

Using the same hardware (single GPU), algorithmic improvements, better data techniques, and kernel optimization reduced training time from weeks to hours. This demonstrates that even denominator-only improvements enable more rollouts, better trajectory sampling, and smarter models without changing the architecture itself.

2. The company employs a "Viceroy of Variance" to destroy results

MakerMaker.AI prioritizes measurement rigor by assigning a dedicated engineer to "destroy the results we have and find what's wrong." Because LLMs are fundamentally stochastic and even identical B200 GPUs show slightly different physical behaviors, distinguishing real improvement from noise, variance, and reward hacking (two levels down—the agent that built the agent that built the agent) requires hundreds of experiments and human intuition.

3. Intelligence is a multidimensional vector; RSI will accelerate unevenly across domains

Owen He noted that even if digital optimization becomes 1000x faster, domains bottlenecked by physical experiments (drug discovery with wet lab work and clinical trials taking months) will see only 20% overall speedup (Amdahl's law). Coding and mathematics, where evaluations are fast and cheap, will accelerate fastest, leading to vastly different transformation speeds across industries.

Editorial Perspective

This discussion matters because it clarifies what "closed loop" means in an industry flooded with RSI claims and establishes empirical measurement standards. The claim that both numerator and denominator improvements have been observed over months—while acknowledging Planck constants as the ultimate physical limit—offers an empirically grounded counternarrative to both doomer and accelerationist extremes. However, the specific measured exponents and generalization beyond inference optimization remain undisclosed, and the "garden hose company in Wisconsin" metaphor suggests that the real breakthrough may lie in unfashionable problem domains no one is currently watching.

Source: a16z Podcast "The Race to Recursive Self-Improving AI" (October 12, 2026)

https://a16z.simplecast.com/episodes/the-race-to-recursive-self-improving-ai-OPuX1TcP

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

Why Zcash Surged 30x in 13 Months as AI Breaks Bitcoin's Anonymity

Featuring

Bankless (host) × Will McEvoy (CIO & Board Member at Cypherpunk Technologies; Principal at Winklevoss Capital)

Bottom line

As AI evolves to de-anonymize Bitcoin's transparent ledger by cross-referencing leaked personal data, privacy is becoming the defining technology category of the next decade.

3-Line Summary

Zcash surged 30x from $50 to $1,500 over 13 months (Aug 2025–Oct 2026) as awareness grew that AI can parse Bitcoin's transparent ledger alongside leaked personal information to identify wallet owners. Publicly traded Cypherpunk Technologies has positioned itself as the only pure-play privacy investment vehicle, holding $500M worth of Zcash (~326,000 coins, approaching 2% of total supply with a 5% goal) and operating 18% of the network's mining hashrate. The price surge was primarily driven by large Bitcoin holders realizing their pseudonymity is breaking down and moving capital to privacy-preserving alternatives.

3 Key Points

1. AI fundamentally breaks Bitcoin's pseudonymity assumption

McEvoy argues AI can cross-reference Bitcoin's transparent ledger with leaked personal data (addresses, phone numbers, social security numbers) to de-anonymize wallets, creating existential risk for large holders. This wasn't possible before because data volume made manual analysis impractical, but AI changes the game. The top 100 Bitcoin holders face realistic risk of being doxxed with home addresses.

2. Zcash mining profitability is 7.7x Bitcoin mining

At $1,000/MWh vs Bitcoin's $130/MWh, Zcash mining is even more profitable than AI colocation ($240/MWh) that caused Bitcoin miners to unplug machines. This extraordinary margin is structural, not temporary, due to supply constraints. Only one manufacturer (Bitmain), low thousands of units per month capacity, 7-8 month minimum lead times, and Bitmain won't significantly increase allocation because Zcash is a "relatively small business" compared to Bitcoin demand. New entrants face semiconductor design costs and competition with AI companies for fab allocation. McEvoy predicts "years, not months" of tight supply.

3. Privacy exists on a spectrum; most solutions are Web2 privacy in Web3 clothing

McEvoy distinguishes "true cypherpunk privacy" (Zcash: hold keys, shielded address, zero-knowledge-backed guarantees) from "Web2 privacy recreated in Web3" (AWS/Google model: not public but government can subpoena, employees can see). AI will "attack quite intensely" the Web2-style alternatives. Only cryptographic privacy survives.

Editorial Perspective

The framing of privacy as a prerequisite for constitutional rights—not just a technology choice—elevates this beyond crypto narratives into civilizational infrastructure. Without privacy of thought, governments can preempt freedom of speech before it's exercised. This philosophical grounding drives McEvoy's optimism about US regulatory environment. Yet Lazarus Group's usage proves both product quality and regulatory risk, and China's control level suggests optimism alone can't resolve the structural tension between privacy technology and state power.

Source: Bankless "Cypherpunk Owns $500M of Zcash and 18% of Its Hashrate | CIO Will McEvoy" (October 12, 2026)

http://podcast.banklesshq.com/

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --

The Capitalization of Attention — How Boxers and DJs Compete with Sequoia on Returns

Featuring

Jason Calacanis × Chamath Palihapitiya × David Friedberg (All-In podcast) × Jake Paul (professional boxer, entrepreneur, angel investor) × Alex Pall, Drew Taggart (The Chainsmokers, Mantis VC co-founders)

Bottom line

Attention and distribution have become differentiation factors as valuable as capital itself, but celebrity investors face structural bias from institutional LPs and must outperform traditional VCs on returns to overcome it.

3-Line Summary

• Jake Paul and The Chainsmokers have converted attention built through boxing promotion and DJ residencies into investments in OpenAI, SpaceX, and others, claiming to compete with Sequoia on the same terms by offering founders distribution as a service
• Institutional LPs avoid celebrity funds because "you'll be the first thing pointed out if something goes wrong," creating structural bias that requires top-5% returns to overcome
• Multi-tranche deals where second/third-round investors pay 2-3x markups with no performance change are being observed, which Jason Calacanis flags as "bubble behavior"

Three Key Points

1. Monetizing Attention as Venture Capital: Jake Paul consulted OpenAI on Sora's social media app launch and granted his name/image/likeness rights, raising $100M for Fund 4. The Chainsmokers leverage relationships like "we played their company party 3 days ago" to connect founders at Series A. Both position distribution as "something founders need and want" on the cap table

2. UFC's Structural Vulnerability: Jake Paul's MVP/PFL merger targets UFC's model where fighters receive ~15% of revenue (vs. 50% in other pro sports leagues). Cites Sean O'Malley earning $600k on a major card as evidence of underpayment, arguing "fighters want to leave and big fights aren't happening"

3. Bubble Signal — Multi-Tranche Markups: Alex Pall observes investors in second/third tranches paying 2-3x the initial valuation "for absolutely no change in underlying performance." Jason responds: "That's bubble behavior. When you see that, take a little note." This is a real-time warning from active Series A investors

Editorial Perspective

The structural shift where attention becomes a currency for cap table seats resembles the early internet era when distribution channels (portals, ISPs) held pricing power over content. The difference this time is that distribution itself buys equity, while institutional investors maintain structural bias against "celebrity VCs"—requiring double proof via returns. The multi-tranche observation signals that valuation inflation is decoupling from fundamentals, a pattern worth monitoring as a leading indicator.

Source: All-In "Jake Paul & The Chainsmokers: Turning Fame into Funds, Jake Enters Politics? & Venture Bubble Signs" (September 30, 2026)

https://allinchamathjason.libsyn.com/jake-paul-the-chainsmokers-turning-fame-into-funds-jake-enters-politics-venture-bubble-signs

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --