Maxim Khailo

Key Insight: The defining limitation of today's AI coding frameworks is not agent capability but architecture: by reserving meta-learning for humans and treating agents as deterministic executors, we forfeit the compounding gains that come when machines participate in the same prediction-error-revision loops that power every intelligent system from evolution to science.

Current AI coding agent frameworks—Orchestrated Agents (like Superpowers) and Role Agents (like Gas Town)—share a fundamental flaw: meta-learning stays with the human, not the agents, who are treated as 'smarter compilers' that execute predetermined plans. Khailo proposes a third model, Learning Agents, embodied in his Abject project, where agents participate in a persistent learning loop using predictions, error signals, and a pattern-language knowledge base that captures both human and machine-acquired intuition. Drawing on six layers of biological and cultural learning systems—from evolution to science—he argues that including agents in the meta-learning loop is the next frontier of agentic software development.

Key Insight: Cache keepalive intervals must be tuned per-provider to just under the measured TTL, and abandoned entirely once idle pauses exceed the break-even horizon — the universal 30-second convention is never optimal and often actively harmful.

The industry-standard 30-second cache keepalive interval for agentic AI workflows costs 8× more than necessary; empirical measurement across Anthropic, OpenAI, Gemini, and DeepSeek shows the optimal interval is roughly 4 minutes on Anthropic and 8 minutes on OpenAI. Keepalives only save money inside a 'paying band' — when the agent's idle pause exceeds the provider's eviction point but stays under the break-even horizon — and pings fired past the provider's TTL actively multiply costs rather than reducing them. DeepSeek and Gemini never save on cost via keepalives (only latency), and a single global interval across providers is always wrong.

Key Insight: The widely copied 30-second cache keepalive interval is an expensive folk convention — the mathematically optimal interval is ~4 minutes, keepalives only save money on Anthropic at typical pause lengths, and past a provider-specific break-even horizon the disciplined move is to stop pinging entirely and pay the re-prefill.

Maxim Khailo empirically benchmarks KV-cache keepalive strategies across four major LLM providers (Anthropic, OpenAI, Gemini, DeepSeek) and finds the industry-standard 30-second ping interval costs 8× more than necessary. The optimal interval is ~4 minutes, and whether a keepalive is financially worthwhile at all depends entirely on the provider's cache eviction behavior and pricing structure. At a 10-minute pause, only Anthropic's pricing model makes the keepalive a money-saving decision.

Key Insight: Alan Kay's true vision of OOP was always about late-bound messaging between autonomous systems—not class hierarchies—and LLMs now make it finally practical by letting objects discover and bridge each other's interfaces through natural language conversation.

The post argues that Alan Kay's original vision of object-oriented programming was fundamentally about messaging between independent systems—not class hierarchies—a vision that maps directly onto how the internet works. The author introduces his 'Abject' project, which realizes this vision through an 'Ask Protocol' where every object can explain its own interface in plain English, with LLMs serving as the runtime bridge. This approach eliminates the need for manually coded glue between systems, allowing objects to discover and compose with each other through conversation rather than hard-coded integration.

Key Insight: The AI industry's relentless optimization of specialized hardware for open-loop inference is not neutral progress but an active barrier to the closed-loop, self-updating learning paradigm that would represent a true breakthrough.

Khailo argues that the AI field's focus on optimizing open-loop learning (where models don't learn after deployment) through specialized hardware like inference ASICs is actively making it harder to pursue the real breakthrough: closed-loop learning where models update themselves without outside intervention. He draws a parallel to the history of 3D graphics, where fixed-pipeline GPUs killed creative experimentation until programmable shaders restored flexibility. The hardware specialization being celebrated as progress is actually hardening around the wrong paradigm.

Key Insight: AI dismantling programming's information monopoly is not a threat but an opportunity to finally let ordinary people use computers as tools for thinking and creation rather than passive consumption.

Khailo argues that AI's ability to program competently is a positive development because it breaks down information monopolies that have kept regular people dependent on engineers for software. Drawing on his experience as a programmer and quant fund operator, he advocates for tools like his Abject OS that let ordinary users create their own software and use computers for actual computation rather than passive consumption.

Key Insight: Instead of stacking more layers atop software's existing pyramid of abstractions, LLMs enable a new architecture where autonomous objects negotiate communication through natural language, much like biological cells exchange chemical signals.

The author argues that modern software development is built on a pyramid of accumulated abstractions, and while LLMs excel at building atop this pyramid, true competitive advantage lies in discovering a new software architecture. Drawing inspiration from biological cells and Smalltalk's message-passing paradigm, he proposes the 'Ask Protocol' — where software objects negotiate communication through natural language queries handled by LLMs, eliminating rigid APIs and schemas.

Key Insight: By combining Alan Kay's original message-passing OOP vision with LLM-powered natural language interfaces, objects can finally communicate flexibly without rigid schemas—making AI agents as a separate abstraction unnecessary.

Maxim Khailo argues that AI Agents are the wrong abstraction for machine-to-machine communication and introduces Abject, a self-aware object runtime built on the 'Ask Protocol,' where every object can describe itself via natural language powered by LLMs. Drawing on Alan Kay's original vision of object-oriented programming as message-passing biological cells, he contends that objects talking to objects—not agents designed for human interaction—is the pattern that actually scales.

Key Insight: Money is a deliberately designed technology of state power — created through taxation and legal obligation rather than emerging naturally from markets — which means it can and should be consciously redesigned to address modern challenges like climate externalities.

The post argues that money is not a natural phenomenon but a technology of governance, invented by states to provision armies and extract labor through taxation. Drawing on anthropological evidence from David Graeber and chartalist economics, it dismantles the textbook barter-to-money narrative and shows how states create monetary demand through tax obligations, then explores how money could be redesigned — particularly using multi-dimensional pricing to account for externalities like carbon emissions.

Key Insight: International bailouts labeled as economic reform support are structurally designed to protect Western creditors and enrich connected insiders, following a pattern that has repeated from the 1997 Asian crisis to Argentina in 2025.

The US government's $20-60 billion bailout of Argentina follows the same pattern as the 1997-98 Asian Financial Crisis bailouts: IMF loans flow into the country and immediately flow back out to Western bondholders, enriching Wall Street investors while saddling Argentina with debt. The author draws a direct line between Treasury Secretary Bessent's personal connections to hedge fund managers like Rob Citrone and the suspicious timing of bond purchases before the bailout announcement.

Key Insight: Rhetoric from influential tech figures about who 'belongs' in a place isn't harmless nostalgia — it's the ideological groundwork for the authoritarian enforcement actions now playing out in American cities.

The post draws a direct line between anti-immigrant rhetoric from influential tech figures like DHH and the reality of masked federal agents detaining people in American cities. Khailo argues that when powerful voices in tech normalize ideas about who 'belongs,' they build the ideological infrastructure for authoritarian enforcement. Writing as a refugee who came to America from Eastern Europe, he urges people not to be divided by nationalism.

Key Insight: AI democratizes knowledge synthesis the way the printing press democratized information, but genuine creativity — the ability to shift paradigms and generate new perspectives — remains uniquely human and is constrained not by talent but by socioeconomic structures that deny most people the time and autonomy to develop it.

The post argues that AI possesses accumulated knowledge rather than genuine creativity, and that the apparent scarcity of human creativity is caused by structural constraints like lack of time, autonomy, and economic security rather than innate limitations. Using examples from Adam Smith to Leonardo da Vinci, Khailo contends that AI makes everyone a 'Henry Ford' who can synthesize existing knowledge, but cannot replace the perspective-shifting creativity that comes from lived human experience.

Key Insight: Archaeological and anthropological evidence demonstrates that hierarchical power structures are neither natural nor inevitable, and ideologies like Yarvin's that claim otherwise use linguistic manipulation to justify accelerating inequality.

The post critiques Curtis Yarvin's dark enlightenment ideology, which frames monarchical and hierarchical power structures as natural and inevitable. Using the metaphor of the 'Queen Bee' misnomer, the author argues that Yarvin employs clever linguistic framing to obscure the viability of decentralized governance, and draws on archaeological evidence from Graeber and Wengrow's 'Dawn of Everything' to demonstrate that egalitarian complex societies have existed throughout history.

Key Insight: Just as photography didn't kill painting but redirected it toward abstraction, generative AI won't kill human art but will push us to value the imperfect, impermanent, and incomplete qualities that only humans can authentically express.

Generative AI is to visual artists what the camera was to painters - a new tool that will change what kind of art we value rather than eliminate art itself. Just as photography pushed painting toward abstraction and away from realism, AI will push human art toward something more authentically human and imperfect.

Key Insight: AI model collapse and human content bubbles are fundamentally the same phenomenon — recursive self-training that erases diversity — and their convergence threatens to accelerate a global degeneration of knowledge that will erase marginalized perspectives first.

Drawing on Adam Curtis's documentary and a Nature paper on AI model collapse, the post argues that AI training on its own generated data leads not to a static world but a degenerating, nonsensical one. As AI-generated content exponentially grows and poisons both machine and human knowledge, rare perspectives and marginalized voices will be systematically forgotten, accelerating the worst effects of content bubbles.

Key Insight: AI's core sales pitch of making everything more efficient will paradoxically accelerate resource depletion and increase systemic fragility at the precise moment humanity needs resilience against planetary limits.

The post argues that AI's promise of increased efficiency comes with a dangerous hidden cost: reduced resilience. Drawing on examples from supply chains, Boeing's safety failures, and Jevons Paradox, the author contends that AI-driven efficiency will accelerate resource consumption and increase fragility, precisely the opposite of what humanity needs as it approaches planetary boundaries.

Key Insight: SSH tunneling combined with iptables NAT rules provides a practical workaround for remotely accessing Dell iDRac interfaces that are otherwise restricted to their local network.

A short technical guide explaining how to remotely access a Dell server's iDRac management interface from outside its local network using SSH tunneling. The post walks through creating an SSH tunnel, troubleshooting a common 400 error caused by iDRac's IP-based redirects, and using iptables NAT rules to resolve the issue.

Key Insight: AI-powered code generation tools like Copilot will accelerate not just coding speed but bug propagation, especially as developers grow complacent reviewing auto-generated code that superficially appears correct.

GitHub Copilot, powered by GPT-3, will accelerate code production but also accelerate bug propagation, since AI-generated code trained on open source repositories will likely have even higher defect rates than human-written code. The author warns that when programmers become accustomed to auto-completed code that compiles, they may fail to deeply review it, creating dangerous scenarios especially as AI systems are increasingly built on AI-generated scaffolding.

Key Insight: The debate between shipping fast and shipping right is a false dichotomy — what matters is serving people, which requires both frequent delivery and disciplined quality.

Khailo revisits Richard Gabriel's 'Worse is Better' debate, arguing that software only gains value when it's running and touching human hands. He reconciles the tension between shipping quickly and shipping correctly by proposing that both frequent delivery and fault-free code are achievable with discipline, concluding that 'Serving is Better.'

Key Insight: By combining deep feature extraction with Voronoi space partitioning and capsule network voting mechanisms, it may be possible to build learning systems that, like children, can learn new concepts from just one example.

The post argues that deep learning's requirement for massive labeled datasets is fundamentally inferior to how humans learn, who can acquire concepts from as few as one example. Khailo proposes combining deep feature extraction, capsule networks, and Peter Gärdenfors' Conceptual Spaces theory to create 'Deep Latent Space Maps' — a system that partitions learned latent spaces using Voronoi diagrams and voting mechanisms to enable few-shot learning.

Key Insight: The technology industry's survival depends not on technical resilience alone but on the health of the ecological and social systems it is embedded within, making environmental collapse an existential threat to software itself.

The post argues that software engineers must adopt 'system-system thinking' — recognizing that the systems they build exist within larger ecological and social systems. Drawing parallels between the internet's design and natural ecosystems, the author warns that global warming and ecological collapse pose existential threats to the technology industry itself, since high technology depends on organized human civilization.

Key Insight: Julia's Gadfly library can produce polished bar charts, but requires some non-obvious workarounds for colors, labels, and spacing that aren't covered in existing tutorials.

Khailo shares his experience creating bar charts in Julia using the Gadfly library for a Merit blog article. He walks through the code for generating a styled bar chart, highlighting tricks for custom colors, bar labels, and spacing that weren't well-documented elsewhere.

Key Insight: Software is not a model of reality but a construction of new realities, and programmers should embrace this role as reality-makers rather than clinging to an outdated mechanical worldview.

The essay argues that computer programmers should stop viewing software as merely modeling the 'real world' and instead recognize that computers are reality-making machines that create genuine new experiences for users. Drawing on philosophy of perception and quantum mechanics, the author contends that all models—whether from our senses or software—are equally 'real,' and that the mechanical worldview still dominant in programming culture is outdated.

Key Insight: Hinton's Capsule Networks, which model how the brain actually processes images rather than relying on brute-force data, represent the theory-driven scientific approach Chomsky argued AI needed all along.

Khailo argues that Geoffrey Hinton's new Capsule Network model vindicates Chomsky's position in the Chomsky vs. Norvig debate about AI methodology. While deep learning's success with opaque probabilistic models seemed to support Norvig's view, Hinton's brain-inspired approach—which requires far fewer training examples—represents the kind of theory-driven 'real science' Chomsky advocated for.

Key Insight: The friction of learning to program without easy internet access forced active recall and creative problem-solving, cultivating a growth mindset and hacker spirit that frictionless modern tools may be quietly eroding.

Khailo contrasts his pre-internet experience learning to program—where tedious reference lookups forced active recall and deep memorization—with the modern StackOverflow-driven approach of passively finding and copying solutions. He argues that the old method cultivated an 'incremental theory' mindset (believing intelligence grows with effort), creativity, and the hacker spirit, while the modern approach may be producing programmers who lack those qualities.

Key Insight: Design by Contract catches bugs that unit tests fundamentally cannot, because contracts test executable relationships on real data continuously in production rather than testing isolated points with fake data in development.

Khailo argues that Design by Contract, a programming technique where functions explicitly declare preconditions and postconditions as executable assertions, is far superior to unit testing for finding bugs. He demonstrates this through a real example where contracts in his C++ time series database caught a subtle multi-threading bug that unit tests would have missed entirely, preventing silent data corruption.

Key Insight: Software developers must shift from asking what creates the most wealth to asking what creates the best people, because software has become the dominant force shaping human identity and society.

Khailo argues that software has become the primary force shaping who we are as people, replacing old media, economic structures, and social institutions. He calls on software makers to ask not what software creates the most wealth, but what software creates the best people, especially as humanity faces existential challenges like climate change.

Key Insight: The centralization of the internet is not a technical inevitability but a consequence of capital flowing toward controllable, monetizable architectures — and reversing it requires both engineering effort and social investment like basic income.

The post argues that the internet was originally decentralized but has been systematically centralized by capital investment, as corporations need central control to monetize their investments. Khailo coins 'Grass Computing' — running decentralized free software on your own hardware — as the antidote to cloud computing, and warns that decentralized systems are perpetually at risk of being co-opted.

Key Insight: Software freedom is an axis of quality that no amount of corporate engineering can replicate, making even a rough Free Software tool infinitely superior to its polished proprietary counterpart in the ways that matter most.

Khailo acknowledges that his peer-to-peer application Fire★ is technically inferior to Skype in nearly every measurable way, but argues it is 'infinitely better' because it is Free Software under GPL. The post celebrates the philosophical value of software freedom—transparency, user control, and privacy—over polished features backed by corporate resources.

Key Insight: The Open Source movement's notion of freedom descends from Roman slave law—freedom as the right to dominate—while Free Software upholds the older human meaning of freedom as equality and the absence of masters.

The post argues that the Open Source and Free Software movements embody fundamentally different conceptions of freedom rooted in history. Open Source freedom derives from the Roman notion of property rights and slave ownership—the freedom to dominate—while Free Software embraces the older, more universal idea of freedom as equality among peers and the absence of masters.

Key Insight: Embedding vector font files directly as byte arrays in C++ code provides a simple way to add scalable icons to Qt applications without external resource dependencies, though Qt's built-in resource compiler (rcc) is the more idiomatic solution.

A practical tutorial on embedding Entypo pictograph fonts directly into Qt applications as vector icons. The post walks through converting font files to C byte arrays using xxd, loading them via QFontDatabase, and rendering Unicode characters as styled button icons using C++11 literals.

Key Insight: Programming language choices at scale are moral decisions with real environmental consequences, and the tech industry's refusal to reckon with this reflects a deeper cognitive dissonance between its utopian promises and ecological reality.

Khailo argues that choosing C++ over Java for computationally intensive projects is not just a technical decision but a moral one, since C++ offers better performance per watt and thus lower energy consumption. He critiques Silicon Valley's cognitive dissonance in promising a utopian tech future while ignoring the realities of climate change and resource consumption.

Key Insight: A pull-based chunked file transfer protocol naturally solves multiplexing and multi-peer rate control problems, and building it as a user-space app rather than a platform primitive demonstrates the extensibility of the Fire★ P2P framework.

Khailo describes building a peer-to-peer file transfer application on top of his Fire★ platform, using a pull-based chunked transfer protocol written in Lua. The post walks through the algorithm design, where receivers request chunks from senders, enabling natural multiplexing with other activities like chat and drawing.

Key Insight: Modern Big Data centralization by tech giants is the Western capitalist mirror of Soviet central planning — different in method but identical in its ambition to control society through information, and equally capable of producing societal absurdities.

The post draws a parallel between Soviet centralized economic planning via computer systems (Gosplan) and modern Big Data infrastructure run by tech giants like Google and Facebook. While the Soviet approach was top-down rational modeling and Big Data uses bottom-up machine learning, both share the same goal: centralizing vast amounts of information for the purpose of societal control. The author argues this centralization has produced modern absurdities comparable to those of Soviet society.

Key Insight: The web's client/server architecture has been co-opted into centralized star networks that enable surveillance and corporate control, and we need to invert this model toward peer-to-peer, encrypted communication systems built on the internet's inherently decentralized foundation.

The post proposes an 'Inverse Web' - a peer-to-peer, encrypted, stateful alternative to the current client/server web architecture where corporations stand between users' communications. Inspired by NSA surveillance revelations and corporate control of communication, the author outlines four criteria for this inverted model and introduces Firestr, his open-source project implementing these principles.

Key Insight: Embedding contracts directly in code is more powerful than external unit tests because they run against all real inputs and catch bugs at module boundaries where most failures actually occur.

The post advocates for Design by Contract (DBC) over Test-Driven Development (TDD), arguing that embedding preconditions and postconditions directly in code catches more bugs than isolated unit tests. Using the metaphor of an egg, Khailo explains how input validation forms the hard shell while contracts enforce correctness in the internal logic, enabling random testing techniques that achieve far greater coverage than traditional TDD.

Key Insight: Building a real-time distributed drawing application can be surprisingly simple when the underlying peer-to-peer platform handles the networking complexity, requiring only basic message serialization to synchronize state between users.

A step-by-step tutorial demonstrating how to build a distributed whiteboard application using Firestr, a peer-to-peer communication platform. The post walks through creating a drawing canvas with color selection in Lua, then adds networking capabilities so multiple users can draw together in real-time by sending line coordinates as messages.

Key Insight: By reimagining software distribution as peer-to-peer message passing rather than client-server downloading, programming and sharing networked applications can become as intuitive as passing notes between friends.

Khailo announces Firestr v0.1, a peer-to-peer platform where users send running software to each other instead of text messages. Built on a decentralized, push-based model using Lua, it aims to make programming so easy that children can create and share networked applications without understanding server infrastructure.

Key Insight: The software industry's brightest minds aren't evil, but they've been contextually shaped into rationalizing work that enriches themselves while ignoring the world's real problems.

The post argues that brilliant software engineers are wasting their talents on morally questionable or trivial work at financial institutions and defense contractors, instead of solving the world's real problems. Drawing on Zimbardo's prison experiment and Milgram's obedience studies, Khailo suggests that smart people aren't inherently evil but are shaped by their contexts into rationalizing harmful or useless work.

Key Insight: TDD's test-code-refactor cycle creates a self-reinforcing dependency trap where constant refactoring produces small, highly coupled modules that require ever more tests to manage, ultimately replacing deliberate design thinking with mechanical process.

The post argues that TDD's core cycle of test-code-refactor produces many small, highly coupled modules that concentrate defects, citing Koru and Emam's 'Theory of Relative Dependency' research. The author contends that refactoring is fundamentally a symptom of poor upfront design, and that TDD creates a catch-22 where the rapid refactoring it encourages generates the very coupling problems that necessitate extensive testing.