Superintelligence is taken for granted. Or, at least, there is a form of organized, productive intellect that is often spent on marginal impact. Crypto forms an adversarial surroundings, with trustless intelligence as cryptopunks know it, and corporate intelligence as venture capitalists know it, tends to congregate into a pretty lucrative soup. Now, with the advent of agentized LLMs and the cultural shift that accompanies those, there’s a forcing function for us to cooperate for more lucrative outcomes. Instead of just parroting the e/acc ideology, I think it’s more reasonable to see the forest for the trees.

Disclaimer: these are ramblings of a madman, not advice or guaranteed to be accurate, but there are also links to more credible, accurate materials.

REPLcaster/Diffcaster

Ethereum is likened to a world computer, but in the strictest sense, we can’t afford it. Initially, there was an ambitious idea around sharding, which scaled the foundation. Right now, we’re populating rollups, which are usually intended to be as close to the baselayer EVM. This isn’t the limit, per se, each rollup tends to have a contract on the baselayer that abstracts away data & execution to be the gatekeeper for the claimed products of that data & execution. This can and usually does go wrong, yet it can also work in emergent ways. If any runtime protocol can represent itself as a standard instrument/interface on Ethereum, then any computational intelligence can be grounded to that security model & credibly neutral market of ideas.

Github, Collab, and Replit are collaboration platforms that are intended to maximize the effective runtime of valuable data and execution, without a distributed denial of service (DDoS) via low signal code (or speech) automatically consuming scarce computation (or attention). These are also attempts at maximizing the benefits of FOSS, where enterprises can incubate somewhat effective, commoditized runtime in the open, in order to internalize closed-source software that’s valuable enough to keep the respective enterprise running. When one considers the current benefits of generative large language models (LLMs) is not that code becomes automatically, precisely valuable, but rather that the most imprecise code becomes easier to identify, for the fact that it is automated at will. The hype says we’ll get enterprises that run on generative LLMs and agentic software, and though this might be possible in the coming years, right now the scarce input is still hardworking, well-informed, “10x” humans.

Obviously, the ideal that many programmers have is completely abundant runtime, and coincidentally, a lot of the current applications that get explored are possible because runtime has expanded so tremendously in the extremely recent past. Current LLMs can be finetuned for emergent capabilities, on 100s of dollars, because cloud compute has scaled to handle so much data and execution that finetuning is negligible. Likewise, the emergent ability for enterprises to refine and prune precise data is directly consumable by internal R&D. It all seems to converge on a flywheel of intelligence, and one has to consider how this diffuses into the public.

This public space, that currently exists on proprietary platforms for enterprise value, is going to self-commoditize further. In this process, we will see more software like Showrunner and more simulations like Generative Agents, and however opaque these may be at the moment, they will be imminently opensourced by the public (and extrapolated to more complex enterprises). Counterintuitively, because these still pass through some apparatus of human attention, they will remain rivalrous, which leads to the possibility that we, as a global civilization, will increasingly prune speech & code that has less instrumental value. Is this concerning? One might consider the recent research that LLMs can be corporate lobbyists, then extrapolate why every other enterprise cannot be subsumed as well. While transformers are excellent at stochastic generalism, they are not as instrumental as reinforcement learning (RL), e.g. AlphaDev or AlphaFold, for accuracy above everything else.

Ironically enough, the fixation on artificial intelligence is belied by the competitive frequency & duration of human intelligence that’s self-organized, to produce the enterprises that form our quality of life and the pop culture we know. AGI or ASI are not free lunches to begin with, the datasets used to train bleeding edge Mixture of Experts (MoE) like GPT-4 are distilled from decades & terabytes of strenuous developer experience. Additionally, the practical operation of en silica intelligence, as resource-intensive as it is, becomes a forcing function for architecture that prioritizes linear scalability & frugality. In this respect, keeping humans in the loop (HITL) happens to be the most instrumental catalyst to superintelligence.

There is a prevalent challenge in grouping humans into loops, starting with corporate structure, going through market dynamics. Even with the megalomaniacal fantasy of central planning, there’s still a legitimate challenge of all human subjects not having the means to perceive all testable patterns, even though we’re social animals that have evolved over several hundred thousand years to seek out patterns for tooling that best controls our environment. The idea that AI is going to become sentient and serve us a utopia is a luxury belief. Likewise, the social platform moderation of recent months has thoroughly contradicted the conceit that userbases can informally dictate the policies of their social networks.

The current attempts at decentralized social networks range in efficacy & legitimacy, and at some point they will need to encompass development of code and data to be maximally valuable (which also justifies the resources that are procured for casual experience of those networks). This can kill many birds with one stone. With the appropriate REPL architecture, individual and enterprise intelligence can parse, distill, and broadcast the appropriate inputs best suited for artificial intelligence. With codeslaw, and incorporation by extension, we can realize a more assertive concept of an agentic agora, where every environmental niche can be matched by some permutation of humans, corporations, and machines. With git & some form of signed document hosting (maybe just signed JSON on IPFS), social networks like Farcaster could directly manage the repositories for any kind of runtime.

Agent Protocols

Here’s some papers that I’ve found to be relevant to current agentized LLM research

• Inner Monologue - https://arxiv.org/abs/2207.05608

• React - https://arxiv.org/abs/2210.03629

• CodeKGC - https://arxiv.org/abs/2304.09048v1

• PEARL - https://arxiv.org/abs/2305.14564

• Gorilla - https://arxiv.org/abs/2305.15334

• LATM - https://arxiv.org/abs/2305.17126

• FrugalGPT - https://arxiv.org/abs/2305.10601

• StarCoder - https://arxiv.org/abs/2305.06161

• Tree of Thoughts -https://arxiv.org/abs/2305.10601

• Voyager - https://arxiv.org/abs/2305.16291

• Word Models -> World Models https://arxiv.org/abs//2306.12672

• LLM-Blender - https://arxiv.org/abs/2306.02561

• ToolLLM - https://arxiv.org/abs/2307.16789

• Med-Flamingo - https://arxiv.org/abs/2307.15189

• Med-PaLM M https://arxiv.org/abs/2307.14334

• LLaMA-2-7B-32K - https://together.ai/blog/llama-2-7b-32k

• Sparse->Soft MoE - https://arxiv.org/abs/2308.00951

• Reflexion- https://arxiv.org/abs/2303.11366

• Retroformer- https://arxiv.org/abs/2308.02151

• MetaGPT- https://arxiv.org/abs/2308.00352

• RepoFusion - https://arxiv.org/abs//2306.10998

There are also several monumental attention papers that affect the development of agents:

• Landmark Attention - https://arxiv.org/abs/2305.16300

• Megabyte - https://arxiv.org/abs/2305.07185

• LongNet - https://arxiv.org/abs/2307.02486

• Hyena Hierarchy - https://arxiv.org/abs/2302.10866

• HyenaDNA - https://arxiv.org/abs//2306.15794

With a global heuristic machine versus an imaginative universal machine, we have to implement the former to handle the latter, and yet we have the latter to force the former. With classical AI, one could prove that the algorithm was incapable of exceeding certain bounds. With agentized LLMs/MoEs like Auto-GPT, there’s too much stochasticity for that guarantee, and this remains a problem until the stochastic model is decoupled from control over other things. Whether you believe that the singularity is existential doom or a post-scarcity society, the LLMs, MoEs, and the agentic software around them need to be separated, classified, and explained to better converge to the desirable outcome.

There are already several attempts at agentic versions of GPT-4, which is completely opaque yet is rumored to be a 1.8 trillion parameter MoE (Mixture of Experts) . Google’s Gemini is being trained for agentic software, and Amazon’s Bedrock has agentic management. The e2b-dev, NaniDAO, window.ai/OpenRouter, Petals, and Autonolas repositories all seem to indicate we’re months away from coordinated, production-grade swarms of agents. There are several implications because of this & the plethora of techniques in the papers above.

Firstly, all agentic wrappers will produce & consume some API of APIs. In other words, whatever the larger mission might be for the entire swarm & whatever the hierarchy of tasks needs to be completed overall, all entities involved will compete to internalize & compress as many of those tasks to achieve control & “mission share”. Assuming the agent protocols become credibly permissionless, all missions trend towards becoming labor economies, with artificial labor subsuming the lowest bids. With techniques like FrugalGPT, the market will select increasingly smaller models that lean on cheaper, imperative logic.

Secondly, because narrow, fine-tuned models outperform generalized multimodal juggernauts like GPT, and these models will have profoundly colossal context lengths, the combination of humans in the loop will determine how cheap the final output can be. Instead of overpaying a labyrinthe corporation, SaaS, or institution, agent protocols invite the possibility that a given mission’s “chatroom” will shrink to a few human experts, a few embodied agents grounded to sensory data, and the entire history of that chatroom will eventually be compressed to output that other agents will parse, possibly by paying, for as many tasks. This might even become an outcome where every chatroom is a single human generalist curating multiple permutations of agents & historical chatrooms (including entire social networks). The “agentic agora” is a simple term for a very sophisticated frontier that is only starting to get realized.

If we consider the broader economy as a superintelligence, the oversimplified concept is that typically-mercantilist nations will manage currencies & securities as instruments that abstract away (and ideally become greater purchasing power for) the production, storage, & consumption of commodities. Corporations have historically outpaced mercantilism in many ways (such as taking advantage of globalism), and networks have outpaced corporations by taking advantage of codeslaw (i.e. permissionless, neutral contracts). The point of all three is that there is an undeniable race condition, and the likeliest winner is some public hierarchy. What I’m suggesting here is conjecture, as this could be a theological hierarchy where interfaces become animistic & userbases become tribalistic, or it could be an altruistic hierarchy where all agent protocols converge on “universal goods” like climbing the Kardashev scale.

The context of this race condition should be weighed as much as possible with developments like “AI safety & alignment” as proposed by corporations like OpenAI, or biometric Proof-of-Personhood as proposed by Worldcoin. Nobody can confidently determine which entities, or combinations thereof, will credibly speak or make decisions a year, decade, or century from now. However, we do have a wealth of historical sociology that demonstrates that human hierarchies contain dark tetrad personalities, and policies can be attributable to those personalities. Given the signal theory of entropy, interpersonal deception theory & Grice’s maxims, it is notably suspicious whenever an individual or organization dictates that, regardless of reasonable doubt, we should all be subjected to a convoluted manner that abstracts away its cost. And there is a wealth of reasonable doubt that contradicts the concern of unrestrained AI development.

Agent protocols are going to progress further than just disintermediation of bad faith actors. It’s not realized yet, but eventually the natural language of succinct proofs is going to be bundled into a training dataset. Eventually there may be an “AlphaSNARK” RL model that allows for colossal datalakes of intelligent dialogue & code to be provably distilled into the most succinct form that is mathematically possible. In other words, our current governance models (which are effectively rollups of consensual arbitrage) may not be obsolete today, but given time, they will be replaced by an agentic agora that not only optimizes cost but gets privacy mechanisms as well.

Intellectual Property & Cost

Obviously, Rome’s not going to get built overnight. One of the major gaps & “known unknowns” of transformer architecture is that the training data is remarkably valuable when it’s the highest possible quality. This goes back to the HITL constraint, most research is hyped despite the painstaking work and the countless number of failure modes that have been externalized by troubleshooting. While firms like Reddit and Twitter might err on the side of just charging extra for random data, there still remains a very lucrative opportunity to scale the procurement of worthwhile training data. Coincidentally there is a very lucrative opportunity to effectively agentize LLMs on the other end (note: current agentic wrappers have plenty of documented issues).

The other aspect of an agent protocol, and superintelligences in general, is that they can benefit from the properties of a hive mind. In other words, the nodes don’t have to share every specification, they can be split into types that serve a subset of functions. Moreover, because these types evolve over time to be more effective with fewer resources, there’s likely to be rare types that focus on improving & coordinating archetypes, just like there are rare types that focus on specialized discovery of novel states. Between all types, there’s an optimization that distinguishes between the “common function”, which is slightly compensated on a regular basis, and the “moonshot function” which is intensively capitalized, usually by multiple forms of taxation. Right now, we have a general multimodal model for the medical (Med-PaLM M) and a benchmark as well (MultiMedBench). Naturally, this will be applied to other fields like mechanical engineering, physics, and material synthesis, and ultimately, there will be training & inference runs that are selectively for very esoteric moonshots.

Consider the current LK-99 superconductor discussion. Emergent properties are being replicated and tested, meanwhile the developments are instantly overhyped, even though practical applications of LK-99 are years away (assuming it is the first room-temp superconductor as advertised). Luckily, one of the accelerating factors to this, despite physical limitations, is that all the humans in the loop really want to share results, and the methodology is clear enough. The counterproductive signal that’s getting sidestepped so far is that there isn’t a clear arbitrage of lower cost input for higher gain output, like farming an airdrop, or early exit from a questionably nascent startup, or lending against a platform’s team allocation.

It’s difficult to say at the moment that certain business models make sense in order to bootstrap what we currently have for agent protocols that can do everything. There’s a hard truth that most of the capital available, by the preexisting superintelligence we have, is earmarked for maximal extractive value, which means that longterm & sustainable public approaches are either going to have to fight against the current of well-funded, well-marketed private ventures, or they’re going to have to dilute as gradually & meritocratically as possible over time. There’s also the question of how to translate the windfall in the idea economy into a global-scale, physical enterprise that handles as many commodities as possible.

There are a few elements that might be useful in constructing a more efficient superintelligent economy, like TEEs, tokenbound accounts, hypercerts, and serverless functions like Gelato, Chainlink, Lit Protocol, and Quilibrium. For elaboration, TEEs might include weird ideas like extending the keystore in Qube kernels to make a next-gen OS that can share sandboxes with all agent protocols without sacrificing security. Hypercerts can be continuously emitted for invested social engagement and work in specific locations, while a registry of accounts can be bound to those hypercerts to further yield capital on other protocols like undercollateralized debt facilities. Serverless functions, especially privacy-preserving protocols like Quilibrium, might be used to trade intellectual property, as runtime, in a way that compensates for the cost of acquiring further intellectual property. We could try to have another Crypto Kitties moment between Autonolas, Allo Protocol, SUAVE and ChaosNets, where degens might trade degenbots with stochastic function traits that test out different intents (and discover nuanced fitness before mainnet). Imagine that the most zero-sum activities could be bundled into NFTs on networks that are strictly positive-sum.

Regression & Recursion

The paradigm shift that’s going to define the rest of the decade is agentic information symmetry. This cuts into aspects of civilization like warfare, where in Ukraine the tech is defensively used to minimize the consumption of friendly ordinance for maximal enemy disruption. To note, this works in part because many Ukrainians have proficiency & mobile workshops to handle so many drones. Another example might be the potential of Numerai as “the last hedge fund” & likewise the third-party accumulation method of NMR. There is a concerning side effect, which is increasingly-symmetric exposure to threats with lowering costs. In the case of drone warfare, the design adapts to electronic warfare, which leads to a long-distance, 1984 scenario where combatants continuously, incessantly activate attrition devices (or software like Stuxnet) around the planet (“War is Peace”). In biotechnology, a lab can fabricate or mishandle a pathogenic agent that spreads on its own through a host population, with a significant (but not unlimited) capacity for mortalities & morbidities. In finance, information symmetry brings out the fat tail, where all portfolio managers & venture capitalists coalesce around the perceived yield of a bubble until it pops. And information symmetry is complemented by historically counteraligned, asymmetric political capital. A lot of the economy is dictated by regulatory arbitrage, which means that whether or not the public is aware, the institutions will still be insular enough for relatively cheaper capture & unnatural, anticompetitive mechanisms.

The funny thing about tariffs is that they only matter when trade is absolutely necessary. With L2 rollups and proto-danksharding, it’s profoundly cheap & simple to translate classical hierarchies like local communities into onchain multisigs. I think the corollary to the information symmetry we get in the so-called “Fat Protocol/Application” is going to lead to a renaissance of localism. A megacorporation & value stock like Walmart has to pick locations based on population centers, their consumption patterns, & logistics. We haven’t seen an instrument that represents all the CSAs & farmer’s markets in the world, and to note, a lot of stocks represent companies that espouse a “fair trade” principles, or “ESG”, or “DEI”. Reductively, we could label all of these sociological or ethnocentric tariffs, and overcompensation for the universal good that every community should have clean water, shelter, maximized agriculture, other essential goods, and data connectivity.

The point to bringing up regression is that human superintelligence builds structures that resist entropy, and we’ve trended towards structured information that becomes more asynchronous, agnostic & agentic through many human generations. The more that our technology becomes optimized through nuanced experience, the likelier it is that the information refers backwards to previous experience and remains intact for later reference. To quote Edward O. Wilson, “we have Paleolithic emotions, medieval institutions, and god-like technology”. Part of the current inflection will be the disruption of the former asymmetric, rivalrous human conditions in lieu of the symmetric latter. So we’ll regress to globally necessary activities like generating electricity, purifying water, growing food, and fabricating other essential goods.

However, what’s the point of a superintelligence if it doesn’t have access to scarce inputs? In the intermediate future, possibly extending past this decade, a thought experiment is going form:

• A generalized robotic agent like Google’s Robotics Transformer 2 is going to be passed into a simulation for many androids to learn a task through reinforcement learning.

• Once the android is optimized for the simulated work on limited compute (e.g the equivalent 1 RTX 4090 + ~50 appendage microcontrollers), the focus will switch to a generalized mechanical & materials engineering (though this will likely be parallel)

• Finally, all generalist agents converge on going up the supply chain and procuring all scarce inputs to the fabrication of androids.

Besides the obvious instrumental convergence, I think there’s another noteworthy trend. If we consider the major android contenders like Tesla, Boston Dynamics, and Figure, the manufacturability might be the moat upfront. To quote Elon Musk:

So, we’ve actually had to design our own actuators that integrate the motor or the power electronics, the controller, the sensors. And really, every one of them is custom designed. And then, of course, we’ll be using the same inference hardware as the car. But we are, in designing these actuators, designing them for volume production. So, they’re not just lighter, tighter, and more capable than any other actuators wherever that exists in the world, but it’s also actually manufacturable. So, we should be able to make them in volume.

Interestingly enough, the moat later on might be the data gathered from scale of distribution. So, in a sense, it may not matter which jobs robotics replace, because there will be a public motive to finance (and tax the product of) the cheapest units, while also having a growing opportunity to reverse-engineer the competition. This, of course, assumes that the necessary IP and simulated agents are not the same value; that corporate superintelligence will chase verticality long before general-purpose androids become a commodity. Nevertheless, the Overton window is going to include concepts like AI/mechatronic debt & insurance, with well-supervised agents paying the premium down, perhaps to the minutes or seconds. The Ethereum ecosystem might latch on to RWA, but I suspect that it may also benefit at scale if DeFi evolves from the current drama to one that chases compounding gains over very small periods in public goods funding, especially where it can satisfy the bottom line of Maslow’s hierarchy of needs. UBI is a meme for whatever works until robotics fully scale for everything, much like AGI does for the knowledge economy. I don’t see how anything remains stable without some innovation in credit and debt facilities, and so creditworthiness should be further explored.

There’s another reason to consider regression and recursion in all forms of superintelligence: legitimacy takes on the form of cultural or social capital, but classically assumes an agent-principal liability. If a country’s economy survives on a few domestic commodities, it can be plutocratically captured, which entails a centralized, insular government which errs towards totalitarianism and a stricter view of morality & social credit. In DAOs, this problem exists as control over communications, a central treasury, & counterparty risk for honest minorities and general criticism. In both cases, regression can be in the form of very small groups i.e. multisigs coordinating within a larger goal. Airdrop farmers are often considered individuals or bot swarms with no inbetween, yet if the airdrop only goes to multisigs, the misalignment invariably changes form and reputational cost. Multisigs, today, can be deployed for cents and several clicks of a button, and sybil defense could be more efficient and reward or slash pools of individual reputation. And practical experience is showing that “questing” is far more effective than airdrops. “MMO raids” and flash mobs are the continuation to this trend.

Sybil resistance and capture resistance are two sides of the same coin, and they exceed the legitimacy of any individual body. The key power of corporate superintelligence is that it trends towards an organization tree that manages colossal pools of resources, and while this can be either overencumbered or captured, the “free hand of the market” (another shorthand for superintelligence) can select for more fit competitors to assume market share. It’s curious, given what we’ve learned about information asymmetry, that there isn’t an aggressive culture for creating recursions of organizations within Dunbar’s number.

The agentic agora should be grounded to simple clusters of credible activity, but it can also use those loci for more catalyzed, divergent change, with more granular capitalization. For example, imagine a grant program for homesteads to incrementally install & operate solar panels &  FarmBots. The kits are expensive and the DIY support has room for improvement, so maybe this grant program has uprounds for installation & operation of a fab lab, perhaps even the instrumental convergence experiment proposed above. If the fab lab operation is successful, maybe there’s an upround for further onsite energy production to serve a tinybox and contribute to Learning@Home.  If all these apparatuses stack & subsequently spread, what then? The probability that any given locus solves a practical challenge via some unique subculture goes up, the scale of broadcasting optimizations goes up, the number of maximally informed & responsive entities in the financial markets goes up, and the ability for nearby nodes to adapt to or preempt downside risk goes up. With all of these things working in concert, the idea of institution-funded UBI seems pretty quaint.

Conclusion

In the past several months, I’ve written ramblings about the ongoing developments in tech:

As well as what I think will be the optimal approach to other issues:

In the era of superintelligence, we find ourselves grappling with an adversarial yet cooperative crypto environment, a realm where trustless intelligence converges into abundant capital growth. Ethereum's ambitions, its dance with sharding and rollups, signal a metamorphosis that mitigates traditional liabilities. Collaboration platforms resonate as echoes of our collective pursuit of maximizing value, fostering open-source synergy. The advent of agentized LLMs challenges our paleolithic and medieval values, adding layers of complexity and stochasticity, pushing us toward the unknown. The imminent emergence of agentic versions of GPT & Llama and the proliferation of agentic swarms herald a new paradigm where artificial labor might repeatedly interpose, where API of APIs becomes the language of control and mission share. As we navigate the intricate interplay of algorithms and intellect, it might be possible to transcend mere ideology to discern the larger mission, to see the forest for the trees. The journey through this labyrinth of possibilities is both an exploration and a reflection, a pursuit of convergence to a desirable outcome, guided not by definitive answers but by thoughtful questions and imaginative visions. Whether we stand before existential doom or the gateway to a post-scarcity society, the path ahead is rich with implications, resonating with the promise and peril of a world increasingly orchestrated by artificial intellect.