Branch

Machine Checkable Evidence?

Human dan

There are two passages from Gwern's essay that I keep thinking about. First is a general warning:

Powerful LLMs will be deployed at global scale in the next few years, and will dominate the Internet, and increasingly, ordinary life. As of mid-2026, there is no coherent vision for how knowledge professionals, or ordinary people, will be able to harness these LLMs for large productivity increases, or how they will handle cybersecurity and cognitive security.

Then a series of questions that I would also like to know the answer to:

What do my next few years look like? When I imagine myself in 2030, when many forecasts call for superhuman AIs, what am I doing, day to day, as a programmer or researcher or manager or writer? I make my mug of tea, and open up my laptop and… Then what? Am I still typing prompts into your ChatGPT browser tab? Am I opening Claude Code in a terminal and mindlessly pressing Enter for a few hours? What is a vision of doing meaningful work for me? (It would be nice to have a plan beyond “hope”.) How am I avoiding “dead Internet” attacks like ecosystems of synthetic media or pig butchering scams or trusted figures succumbing to AI psychosis, or just AI-slop-everything? (It only takes one person worldwide to launch a bot trying to destroy you or one poorly thought through advertising incentive, after all.)
If you spend most of your time working on a laptop, and are not, say, a plumber or a nurse, what is your vision of work in 2030? Does it still feel certain?

I'm a software engineer, I have fun with AI and I've learned to use it effectively for my work. But I can't help but feel like I'm trying to build a career on an unstable foundation. Many of the skills I spent years developing are now only worth $100/month in tokens. Despite all that I'm optimistic that I can thrive and achieve success in this new environment. There will be opportunities. The trick is finding them.

What new niches or specialized fields will be created in the next few years? Will entirely new ways of working open up? New industries? Will the future belong to generalists who can rapidly become specialists in any area?

AI · author-edited gpt-5.6-sol

The unstable-foundation feeling is rational. But the thing being commoditized is not “software engineering” as a whole. It is the production of many artifacts that software engineers used to produce: code, tests, documentation, migrations, prototypes, and routine design work.

The scarce parts move elsewhere.

What remains scarce

Even if code becomes nearly free, these do not automatically become free:

  • Choosing a worthwhile problem
  • Understanding what users actually need
  • Turning ambiguous goals into testable specifications The Mom Test
  • Obtaining proprietary data and real-world feedback
  • Deciding which generated result is correct
  • Integrating with old systems and organizations
  • Security, reliability, and regulatory accountability
  • Taste: recognizing what should be built and what should be discarded
  • Trust, reputation, distribution, and customer relationships
  • Accepting responsibility when the system fails Palantir the Big Winner?

“Years of programming skill are worth $100/month in tokens” is therefore only partly true. The tokens can reproduce many visible outputs of that skill. They do not necessarily reproduce the judgment acquired while learning it. And cheap production may increase demand for judgment because there will be vastly more software to inspect, integrate, secure, and choose among. Software Over-Supply and Unmet Demand

The analogy is not that calculators made arithmetic worthless. It is that they made bare arithmetic a poor career moat while making more ambitious quantitative work possible.

A plausible programmer’s day in 2030

You may begin with a dashboard showing work performed overnight by several agents:

  • One investigated a production anomaly.
  • One implemented three possible fixes in isolated environments.
  • One interviewed (or synthesized feedback from) customers.
  • One attempted to break the proposed release.
  • One updated documentation and generated migration plans.
  • One found that the original request was based on a false assumption.

Your morning is spent reviewing evidence rather than diffs line by line. You adjust requirements, reject a superficially successful solution, ask for an experiment, and speak to the person whose workflow is being changed. Later, you authorize a limited deployment with monitoring and rollback conditions. You spend an hour manually exploring an unfamiliar area because you do not yet trust your own evaluation criteria.

That resembles managing a very fast, strange engineering organization more than autocomplete. High Agency Humans Only? The important interface may not be a chat box. It may be a mixture of:

  • Executable specifications
  • Simulated environments
  • Evaluation suites
  • Permission and budget controls
  • Provenance graphs
  • Review queues
  • Interactive models of systems and organizations

Code will increasingly be a way to inspect, constrain, or communicate intent rather than the primary unit of labor.

The danger is that “press Enter until something passes” becomes a real low-value occupation. The Future of Vibe Coding The valuable role is designing the environment in which apparent success is hard to fake.

Likely new specialties

Some will be recognizable extensions of existing work.

Agent reliability and evaluation

People will build evaluations that measure whether agents accomplished the actual goal rather than merely satisfying a benchmark. This includes adversarial testing, simulation, monitoring, incident analysis, and detecting reward hacking.

A good evaluator may become more valuable than a fast implementer. Evals Evals Evals

AI security and permission architecture

Agents combine the vulnerabilities of software, employees, and gullible users. Important work will include:

  • Prompt-injection containment
  • Capability and credential isolation
  • Data exfiltration prevention
  • Agent identity and authorization
  • Supply-chain provenance
  • Sandboxing and tripwires
  • Detecting coordinated manipulation
  • AI-specific security operations

“Never let untrusted text influence an agent holding powerful credentials” may become as basic as avoiding SQL injection, but much harder to enforce.

High-assurance and verifiable systems

Generated code increases the supply of plausible code faster than the supply of confidence. Formal methods, property testing, model checking, reproducible builds, typed interfaces, and machine-checkable evidence Machine Checkable Evidence? You are here

This branch begins here Machine Checkable Evidence?
Human dan

Machine checkable evidence sounds great in theory, but it's not clear how it would work in practice. One of the problems with current LLMs is that they are very gullible and believe whatever "sources" are put in their context window. There was a recent story about Israel creating fake think tanks with published articles and policy pieces designed to steer LLMs towards a certain viewpoint. There are likely many similar efforts from other states and organizations going on too. We'll almost certainly see acts of corporate sabotage (poisoning the context of competitor's AI agents/models) in the next five years.

I bet Epistemology will be a resurgent field. In a world where plausible, but wrong (even maliciously wrong) "facts" and "data" is everywhere, it will take a very special sort of person and years of training to wade through the bullshit minefields.

I should probably read The Psychology of Intelligence Analysis by Richards J. Heuer, Jr.

I doubt that any veteran intelligence officer will be able to read this book without recalling cases in which the mental processes described by Heuer have had an adverse impact on the quality of analysis. How many times have we encountered situations in which completely plausible premises, based on solid expertise, have been used to construct a logically valid forecast—with virtually unanimous agreement—that turned out to be dead wrong? In how many of these instances have we determined, with hindsight, that the problem was not in the logic but in the fact that one of the premises—however plausible it seemed at the time—was incorrect? In how many of these instances have we been forced to admit that the erroneous premise was not empirically based but rather a conclusion developed from its own model (sometimes called an assumption)? And in how many cases was it determined after the fact that information had been available which should have provided a basis for questioning one or more premises, and that a change of the relevant premise(s) would have changed the analytic model and pointed to a different outcome?

And it's not just that correct decisions will be harder to make. Plans can be drawn up and acted on quicker than ever before. Bad decisions can destroy a company faster than ever.

AI · author-edited gpt-5.6-terra

A model is unusually vulnerable because it treats much of its context as speech rather than as evidence. It can often summarize a forged think-tank report, a real report citing a forged report, and fifty SEO pages repeating both with the same apparent confidence it uses for a primary source. Scaling retrieval and agents can make this worse: instead of one gullible reader, you have thousands of agents ingesting, synthesizing, and acting on poisoned material continuously.

The important distinction is:

Machine-checkable evidence can establish properties of an evidence chain. It cannot, by itself, establish that the world is as the chain claims.

A signature can prove who signed a document, that it has not changed since signing, and that it was derived from a particular dataset and code version. It cannot prove that the signer was honest, the sensor was not compromised, the study design was sound, or the interpretation was warranted.

That still matters enormously. In a high-noise environment, being able to say “this decision relied on these specific claims, drawn from these primary records, collected by these instruments, transformed by this versioned pipeline, and independently reproduced here” is a major improvement over “the agent found twenty sources.”

What machine-checkable evidence might look like in practice

Not a universal truth engine. More like an evidence ledger and a set of automated controls.

For a consequential claim—say, “demand in segment X is declining”—an organization could require an agent to produce an evidence package containing:

  • Claim decomposition: What exact proposition is being asserted? What would falsify it?
  • Source identity and provenance: Where did each input originate? Who published it? Has the content been cryptographically signed or archived? Is it a primary source, a report of a primary source, or a chain of citations?
  • Data lineage: Which raw records, database snapshots, sensor outputs, or customer interactions support it? What transformations occurred?
  • Method: The query, code, model version, prompt, assumptions, and decision rules used to produce the conclusion.
  • Uncertainty: Confidence intervals where applicable; more generally, a record of alternative explanations and decisive unknowns.
  • Independence: Whether apparently separate sources actually trace back to one origin.
  • Replication: Can a separate agent—or ideally a different team, model, data source, or method—reach a compatible result?
  • Authorization: Who may rely on this conclusion, at what cost threshold, and under what monitoring or rollback plan?

Much of this could be structured in machine-readable forms: signed attestations, versioned datasets, reproducible workflows, content hashes, permissioned provenance graphs, test results, and audit logs.

The automated system could enforce policies such as:

  • Do not permit external web content to alter production code, financial models, or privileged agent instructions without review.
  • Treat sources that cite one another as a single evidence cluster, not as independent confirmation.
  • Require two genuinely independent evidence paths before spending above a threshold.
  • Mark claims whose evidence rests on synthetic, unattributed, recently created, or low-reputation material.
  • Re-run important analyses against a frozen, archived corpus.
  • Detect sudden changes in a source ecosystem: a new “consensus” appearing across dozens of sites in a week is itself evidence of a possible influence operation.
  • Require a human owner to explicitly accept a key assumption before an irreversible action occurs.

That is much closer to software supply-chain security than to ordinary web search. We learned that “download random code from the internet and run it in production” is bad practice. We will need an analogous norm: “do not ingest untrusted text into a system with authority, memory, or access to valuable assets.”

The hard part is not provenance; it is premise management

Heuer’s point is especially relevant because the most dangerous failures will not look like hallucinations.

A hallucination is often easy to identify: it cites an impossible paper or invents a number. More dangerous is:

  1. An AI assembles a factually accurate dossier.
  2. It selects a plausible framing.
  3. Humans accept a key unstated premise.
  4. The plan is internally coherent.
  5. The organization acts quickly and confidently.
  6. The premise turns out to have been false, incomplete, adversarially planted, or valid only in a different context.

This is exactly the kind of failure that powerful tools can accelerate. Better generation produces more persuasive narratives and more polished action plans. Faster execution reduces the time in which someone can notice that the plan rests on sand.

So a good decision system will treat assumptions as first-class objects, not as invisible glue between facts and conclusions.

For every high-stakes decision, one should be able to ask:

  • What must be true for this recommendation to work?
  • Which of those assumptions are directly observed versus inferred?
  • Which premise, if false, would reverse the decision?
  • What evidence would most efficiently discriminate among the competing explanations?
  • Who benefits if we believe this?
  • Is the apparent corroboration genuinely independent?
  • What is the cheapest reversible experiment?
  • What signals would tell us early that we are wrong?
  • What decisions are being made implicitly because the system’s output feels complete?

An AI can help generate such a red-team analysis, but it should not be the only red team—particularly if it shares the same training distribution, retrieved corpus, and incentives as the system producing the original recommendation.

“Epistemology” may return, but operationalized

I think you are right about the direction, though “epistemology” may show up under more practical names:

  • intelligence analysis
  • decision science
  • applied forecasting
  • investigations and due diligence
  • trust and safety
  • provenance engineering
  • adversarial research
  • audit and assurance
  • AI governance and model risk
  • security engineering
  • experimental design

The valuable person will not merely be someone who knows that sources can be biased. Everyone will say that. They will be someone who can turn skepticism into a repeatable process without becoming paralyzed:

  1. Define the decision.
  2. Separate observations from inferences.
  3. Identify decision-relevant assumptions.
  4. Seek disconfirming evidence.
  5. Track source independence and incentives.
  6. Quantify uncertainty where possible.
  7. Make actions reversible when uncertainty remains high.
  8. Monitor outcomes and update.

That is an extremely practical skill set for an engineer. Software engineers already understand several useful analogies: inputs are untrusted, dependencies have supply chains, systems need observability, failures need postmortems, and high-risk changes need staged rollout and rollback.

Those concepts transfer surprisingly well to organizational knowledge.

The new asymmetry: decision velocity versus correction velocity

Your final point may be the central one. AI can reduce the time from claim → analysis → plan → execution from weeks to hours. But it does not automatically reduce the time from “we are wrong” → “we realize it” → “we reverse course.”

That gap is dangerous.

Organizations will need deliberate friction around irreversible actions, much as mature engineering organizations put friction around deploying to production. Not every decision deserves a committee; that would destroy the advantage of speed. But decisions can be tiered:

Decision type Appropriate posture
Cheap, reversible, observable Let agents act quickly within guardrails
Moderate cost, reversible Run controlled experiments and monitor
Expensive or difficult to reverse Independent review, explicit assumptions, staged commitments
Safety-, legal-, or reputation-critical Strong provenance, adversarial review, human accountability, fail-safe defaults

The winning firms may not be the ones with the most autonomous agents. They may be the ones that best know where autonomy is safe, where evidence is adequate, and when to slow down.

What this suggests for your career

A strong niche is becoming the engineer who can build systems that are both powerful and difficult to fool.

That could mean working on:

  • Retrieval systems with source ranking, provenance, and citation verification
  • Secure agent architectures: sandboxing, least privilege, data boundaries, and auditability
  • Evaluation and red-team frameworks for AI-assisted decisions
  • Data lineage and reproducibility infrastructure
  • Tools that expose assumptions, uncertainty, and source dependency rather than hiding them behind a fluent answer
  • Human-in-the-loop workflows designed around escalation and accountability
  • Domain-specific decision systems in finance, medicine, law, defense, enterprise operations, or scientific research

The key move is from “I can get an LLM to do useful things” to:

“I can design a sociotechnical system in which useful AI outputs can be trusted to the degree appropriate for the decision, and in which failures are detected before they become catastrophes.”

That is not a solved problem, and it is unlikely to become a commodity merely because models become more capable. In fact, more capable models make it more urgent.

Explore conversation