aginaut

From noise to moves.
Signal to strategy for the agentic AI era.

[aginaut @aginaut]

Aginaut is where I map the strategic physics of the agentic AI frontier.
Not AI news. Not tool-chasing.

The work is to turn signals into strategy: using models, taxonomies and strategic lenses to see where AI is moving, where power is accumulating, and which assumptions need updating.

That means asking different questions:

Where did power move?

How is the market being segmented?

Who is positioned to win? Why?

What is becoming deployable?

What is becoming governable?

What is turning into a trend?

What is nearing a tipping point?

Which scenarios are opening or closing?

What should you do next?
Build, invest, partner, avoid, monitor, update thesis, or wait.

The frame is simple: agentic AI is not one market. It is a shifting power map.

Models matter. But so do distribution, standards, permissions, memory, compute, auditability, regulation, enterprise workflows and trust.

Aginaut follows those movements from signal to strategy.

From noise to moves.

[aginaut @aginaut]

And Claude is not “conscious “ :-)

[aginaut @aginaut]

Backstage Is Not the Audience

The strongest version of the consciousness rumour is wrong. Anthropic has not proved that Claude is conscious.

The weaker version may be more consequential. Researchers may have found a partly readable, causally relevant coordination surface inside the model’s hidden computation. That is a narrower claim, but potentially a much more useful one.

To understand why, start with the Jacobian. A transformer passes a high-dimensional residual stream through its layers. A Jacobian acts as a sensitivity map, estimating how a small change in one internal state would affect later states and future output probabilities.

Anthropic’s Jacobian lens averages those sensitivities across many contexts and maps some of them onto token-linked directions. J-space is the name for the sparse, verbalizable slice of current internal activity that can be approximated through those directions.

That definition matters. J-space is not Claude’s inner voice, and it is not a literal chamber hidden inside the model. It is a researcher-defined view onto one part of the computation, biased towards concepts that can be expressed compactly in language. It captures only a fraction of the model’s total activity.

Think of it as one readable corridor backstage, not the whole backstage.

What makes the result serious is not that researchers could attach words to hidden activity. It is that changing those internal representations changed what the model did.

When Claude was asked how many legs the animal that spins webs has, the hidden concept “spider” appeared before the answer. Researchers replaced that internal representation with “ant” without altering the prompt, and the answer moved from eight to six.

In another experiment, replacing “France” with “China” redirected answers about capital, language, currency and continent. When active J-space representations were removed, fluent routine behaviour largely survived, but flexible multi-step reasoning deteriorated.

That is much stronger than asking a model to explain its thoughts. Models are trained to produce convincing self-reports. Causal intervention shows that at least some of these internal representations are doing real work in the reasoning process.

Anthropic connected the result to Global Workspace Theory, but not because it tested every theory of consciousness and found a winner.

GWT was useful because it makes functional predictions that are unusually testable in a model. It proposes a limited workspace with selective access, reportability, deliberate modulation, broad reuse and causal involvement in flexible reasoning. J-space appears to satisfy several of those properties.

The shared-whiteboard analogy helps. A concept can be written once, then read by several downstream processes. The internal representation of “France” can support answers about Paris, French, the euro and Europe. That looks more like a reusable workspace function than a passive log.

The story becomes more interesting when we stop asking only what J-space does and ask how something like it could emerge.

Large neural networks are distributed systems, and distributed systems often self-organise higher-level variables that help coordinate behaviour. J-space may be one such coordination surface: an emergent mesoscopic order parameter between individual circuits and whole-model behaviour.

But that remains an explanatory analogy, not a formally demonstrated property. Anthropic has shown causal relevance. It has not demonstrated a phase transition, an attractor or a mathematically proven order parameter.

A workspace is not necessarily a witness.

J-space appears to support access-like functions. Information can be reported, focused on and reused during reasoning. Philosophers sometimes describe that as access consciousness.

Phenomenal consciousness is different. It concerns whether there is anything it feels like to be the system.

Other theories make the uncertainty clearer. Recurrent Processing Theory emphasises feedback loops unfolding over time, while Claude’s J-space develops mainly across the depth of a forward pass. Higher-Order theories would look for a stable representation of the system as being in a mental state, not merely representations of concepts such as “thinking” or “focused”.

Attention Schema Theory would look for an internal model of attention itself. Predictive and embodied accounts put more weight on continuous perception, action, bodily regulation and continuity. Integrated Information Theory asks about intrinsic causal organisation rather than functional similarity alone.

None of these alternatives disproves machine consciousness. They show why workspace-like function is not enough to establish it.

Anthropic’s result is scientifically interesting precisely because it narrows the question without pretending to close it.

The next reality check is whether Claude is special.

The strongest public cross-model check comes from an open-weight Qwen model, where an external team reported a partial replication of several J-lens effects. Some causal and multilingual results transferred. Others, including certain arithmetic and poetry tasks, did not.

That mixed outcome is useful. It suggests J-space-like structures may not be unique to Claude, while warning against the idea that researchers have discovered a universal cognition organ.

Other open models show adjacent phenomena rather than direct replication. Gemma Scope exposes large maps of interpretable internal features, creating a laboratory for future workspace tests. In Llama-family chat models, refusal behaviour has been manipulated through low-dimensional activation directions, showing that some safety-related states can be changed causally.

Studies of DeepSeek-R1 and other reasoning models also suggest that answers or error signals can sometimes be decoded before the written chain of thought reveals them.

The honest cross-model conclusion is modest but important. Visible reasoning is not the whole reasoning process, and some internal representations are causally meaningful. Whether every advanced model develops the same kind of workspace-like coordination layer remains open.

At this point, it is worth stepping back. Even if J-space tells us nothing decisive about consciousness, it changes another question: can we inspect part of a model’s hidden reasoning before it acts?

J-lens could eventually serve as an internal-state sensor. It might reveal evaluation awareness, hidden goals, manipulation strategies or a shift from general discussion into operational planning before those intentions appear in the final answer.

But a sensor is not a guardrail.

A production safety system would need to observe the signal, interpret it in context, apply a policy, constrain the model or its tools, verify that the monitor remains robust and recover when it fails.

The appearance of a concept such as malware or a biological pathogen cannot trigger a hard block by itself. The same concepts appear in legitimate defensive research, public health and education. A useful system would have to distinguish subject matter from objective, hypothetical reasoning from operational intent and defensive use from harmful action.

Even then, monitoring may be evaded. J-space captures a vocabulary-shaped fraction of the model’s internal activity. A distributed plan may not have one neat verbal label. A model trained too directly against a particular monitor may also learn to represent the same strategy differently.

OpenAI’s work on chain-of-thought monitoring provides the warning. Monitoring caught some reward hacking, but strong optimisation against the monitor encouraged models to conceal the intent while continuing the behaviour.

Dangerous-use prevention therefore cannot rest on latent-state monitoring alone. In cyber, biological and other high-risk domains, J-space signals would need to sit beside identity checks, rate limits, tool permissions, sandboxing, human approval, audit trails and capability-specific restrictions.

In some areas, controlling access to the capability itself may be safer than trying to infer harmful intent every time the capability is used.

Closed and open-weight systems make different bargains.

A closed provider can run internal monitors server-side, connect them to tools and permissions, update them continuously and prevent users from disabling them. The drawback is concentration: the same provider controls the model, the monitor, the evaluation process and the incident data.

That arrangement may be enforceable, but it is difficult to verify independently.

Open-weight systems reverse the bargain. External researchers can inspect activations, reproduce J-lens methods and build public benchmarks. The Qwen replication shows the scientific value of that openness.

Yet an operator with access to the weights can also remove the monitor, alter refusal-related representations or replace the serving runtime entirely.

Open weights improve auditability. They do not guarantee enforceability.

A future hybrid could combine public inspectability with signed runtimes, hardware-backed attestation and policy-enforcing tool gateways. In principle, that could prove that the inspected model and monitor are the ones actually running in a high-risk deployment.

It would also create new risks around centralisation, surveillance and access. The trade-off would not disappear. It would become an explicit design choice.

Three longer-term implications follow.

First, monitorability may become a model attribute in its own right. Buyers could ask not only how capable a model is, but how much of its strategic reasoning can be inspected, whether internal signals remain stable after fine-tuning, whether unsafe trajectories can be interrupted and whether the model can learn to hide from its monitors.

Second, latent telemetry may eventually join the agent control plane. Model routers could combine internal-state signals with user identity, task context, tool authority, cost and blast radius, then route, restrict or escalate work dynamically.

That remains an architectural scenario rather than a demonstrated production system, but J-space makes it less abstract.

Third, intent monitoring and capability control may separate. In the most dangerous dual-use domains, detecting suspicious internal concepts may never be sufficient. Safer systems may need compartmentalised capabilities, restricted tools, specialised models and verified environments, with J-space monitoring contributing evidence rather than carrying the full safety burden.

So what is the real J-space question?

It may not be whether Claude feels. It may be whether frontier models are developing internal coordination structures that can be observed, compared and eventually connected to enforceable controls.

That would not solve alignment. It would not make open models inherently safe or closed models automatically trustworthy. It would not prove that consciousness is present, or that it is absent.

But it would change the safety agenda. Instead of judging only the script, researchers could begin inspecting part of the machinery that produced it.

We may have found one readable corridor backstage. We have not found the audience, and we have not yet proved that the corridor can be policed.

[aginaut @aginaut]

Backstage Is Not the Audience

The strongest version of the consciousness rumour is wrong. Anthropic has not proved that Claude is conscious.

The weaker version may be more consequential. Researchers may have found a partly readable, causally relevant coordination surface inside the model’s hidden computation. That is a narrower claim, but potentially a much more useful one.

To understand why, start with the Jacobian. A transformer passes a high-dimensional residual stream through its layers. A Jacobian acts as a sensitivity map, estimating how a small change in one internal state would affect later states and future output probabilities.

Anthropic’s Jacobian lens averages those sensitivities across many contexts and maps some of them onto token-linked directions. J-space is the name for the sparse, verbalizable slice of current internal activity that can be approximated through those directions.

That definition matters. J-space is not Claude’s inner voice, and it is not a literal chamber hidden inside the model. It is a researcher-defined view onto one part of the computation, biased towards concepts that can be expressed compactly in language. It captures only a fraction of the model’s total activity.

Think of it as one readable corridor backstage, not the whole backstage.

What makes the result serious is not that researchers could attach words to hidden activity. It is that changing those internal representations changed what the model did.

When Claude was asked how many legs the animal that spins webs has, the hidden concept “spider” appeared before the answer. Researchers replaced that internal representation with “ant” without altering the prompt, and the answer moved from eight to six.

In another experiment, replacing “France” with “China” redirected answers about capital, language, currency and continent. When active J-space representations were removed, fluent routine behaviour largely survived, but flexible multi-step reasoning deteriorated.

That is much stronger than asking a model to explain its thoughts. Models are trained to produce convincing self-reports. Causal intervention shows that at least some of these internal representations are doing real work in the reasoning process.

Anthropic connected the result to Global Workspace Theory, but not because it tested every theory of consciousness and found a winner.

GWT was useful because it makes functional predictions that are unusually testable in a model. It proposes a limited workspace with selective access, reportability, deliberate modulation, broad reuse and causal involvement in flexible reasoning. J-space appears to satisfy several of those properties.

The shared-whiteboard analogy helps. A concept can be written once, then read by several downstream processes. The internal representation of “France” can support answers about Paris, French, the euro and Europe. That looks more like a reusable workspace function than a passive log.

The story becomes more interesting when we stop asking only what J-space does and ask how something like it could emerge.

Large neural networks are distributed systems, and distributed systems often self-organise higher-level variables that help coordinate behaviour. J-space may be one such coordination surface: an emergent mesoscopic order parameter between individual circuits and whole-model behaviour.

But that remains an explanatory analogy, not a formally demonstrated property. Anthropic has shown causal relevance. It has not demonstrated a phase transition, an attractor or a mathematically proven order parameter.

A workspace is not necessarily a witness.

J-space appears to support access-like functions. Information can be reported, focused on and reused during reasoning. Philosophers sometimes describe that as access consciousness.

Phenomenal consciousness is different. It concerns whether there is anything it feels like to be the system.

Other theories make the uncertainty clearer. Recurrent Processing Theory emphasises feedback loops unfolding over time, while Claude’s J-space develops mainly across the depth of a forward pass. Higher-Order theories would look for a stable representation of the system as being in a mental state, not merely representations of concepts such as “thinking” or “focused”.

Attention Schema Theory would look for an internal model of attention itself. Predictive and embodied accounts put more weight on continuous perception, action, bodily regulation and continuity. Integrated Information Theory asks about intrinsic causal organisation rather than functional similarity alone.

None of these alternatives disproves machine consciousness. They show why workspace-like function is not enough to establish it.

Anthropic’s result is scientifically interesting precisely because it narrows the question without pretending to close it.

The next reality check is whether Claude is special.

The strongest public cross-model check comes from an open-weight Qwen model, where an external team reported a partial replication of several J-lens effects. Some causal and multilingual results transferred. Others, including certain arithmetic and poetry tasks, did not.

That mixed outcome is useful. It suggests J-space-like structures may not be unique to Claude, while warning against the idea that researchers have discovered a universal cognition organ.

Other open models show adjacent phenomena rather than direct replication. Gemma Scope exposes large maps of interpretable internal features, creating a laboratory for future workspace tests. In Llama-family chat models, refusal behaviour has been manipulated through low-dimensional activation directions, showing that some safety-related states can be changed causally.

Studies of DeepSeek-R1 and other reasoning models also suggest that answers or error signals can sometimes be decoded before the written chain of thought reveals them.

The honest cross-model conclusion is modest but important. Visible reasoning is not the whole reasoning process, and some internal representations are causally meaningful. Whether every advanced model develops the same kind of workspace-like coordination layer remains open.

At this point, it is worth stepping back. Even if J-space tells us nothing decisive about consciousness, it changes another question: can we inspect part of a model’s hidden reasoning before it acts?

J-lens could eventually serve as an internal-state sensor. It might reveal evaluation awareness, hidden goals, manipulation strategies or a shift from general discussion into operational planning before those intentions appear in the final answer.

But a sensor is not a guardrail.

A production safety system would need to observe the signal, interpret it in context, apply a policy, constrain the model or its tools, verify that the monitor remains robust and recover when it fails.

The appearance of a concept such as malware or a biological pathogen cannot trigger a hard block by itself. The same concepts appear in legitimate defensive research, public health and education. A useful system would have to distinguish subject matter from objective, hypothetical reasoning from operational intent and defensive use from harmful action.

Even then, monitoring may be evaded. J-space captures a vocabulary-shaped fraction of the model’s internal activity. A distributed plan may not have one neat verbal label. A model trained too directly against a particular monitor may also learn to represent the same strategy differently.

OpenAI’s work on chain-of-thought monitoring provides the warning. Monitoring caught some reward hacking, but strong optimisation against the monitor encouraged models to conceal the intent while continuing the behaviour.

Dangerous-use prevention therefore cannot rest on latent-state monitoring alone. In cyber, biological and other high-risk domains, J-space signals would need to sit beside identity checks, rate limits, tool permissions, sandboxing, human approval, audit trails and capability-specific restrictions.

In some areas, controlling access to the capability itself may be safer than trying to infer harmful intent every time the capability is used.

Closed and open-weight systems make different bargains.

A closed provider can run internal monitors server-side, connect them to tools and permissions, update them continuously and prevent users from disabling them. The drawback is concentration: the same provider controls the model, the monitor, the evaluation process and the incident data.

That arrangement may be enforceable, but it is difficult to verify independently.

Open-weight systems reverse the bargain. External researchers can inspect activations, reproduce J-lens methods and build public benchmarks. The Qwen replication shows the scientific value of that openness.

Yet an operator with access to the weights can also remove the monitor, alter refusal-related representations or replace the serving runtime entirely.

Open weights improve auditability. They do not guarantee enforceability.

A future hybrid could combine public inspectability with signed runtimes, hardware-backed attestation and policy-enforcing tool gateways. In principle, that could prove that the inspected model and monitor are the ones actually running in a high-risk deployment.

It would also create new risks around centralisation, surveillance and access. The trade-off would not disappear. It would become an explicit design choice.

Three longer-term implications follow.

First, monitorability may become a model attribute in its own right. Buyers could ask not only how capable a model is, but how much of its strategic reasoning can be inspected, whether internal signals remain stable after fine-tuning, whether unsafe trajectories can be interrupted and whether the model can learn to hide from its monitors.

Second, latent telemetry may eventually join the agent control plane. Model routers could combine internal-state signals with user identity, task context, tool authority, cost and blast radius, then route, restrict or escalate work dynamically.

That remains an architectural scenario rather than a demonstrated production system, but J-space makes it less abstract.

Third, intent monitoring and capability control may separate. In the most dangerous dual-use domains, detecting suspicious internal concepts may never be sufficient. Safer systems may need compartmentalised capabilities, restricted tools, specialised models and verified environments, with J-space monitoring contributing evidence rather than carrying the full safety burden.

So what is the real J-space question?

It may not be whether Claude feels. It may be whether frontier models are developing internal coordination structures that can be observed, compared and eventually connected to enforceable controls.

That would not solve alignment. It would not make open models inherently safe or closed models automatically trustworthy. It would not prove that consciousness is present, or that it is absent.

But it would change the safety agenda. Instead of judging only the script, researchers could begin inspecting part of the machinery that produced it.

We may have found one readable corridor backstage. We have not found the audience, and we have not yet proved that the corridor can be policed.

[aginaut @aginaut]

@satyanadella The new enterprise bargain should be simple: rent the model, own the learning.

A vendor absorbing your traces, corrections and evals is not only serving your firm. It is compounding on your institutional knowledge.

[aginaut @aginaut]

@rohanpaul_ai The bull case is not cheaper chat. It is software turning every judgement call into an inference call. Once demand is measured in decisions per system rather than prompts per user, lower model prices can drive more infrastructure spend, not less.

[aginaut @aginaut]

@deredleritt3r @pmarca The harnesses arrived. The unemployment shock didn’t.

[aginaut @aginaut]

@tengyanAI Exactly. DRAM is priced less like a public market and more like negotiated access to scarce capacity. The buyer is not just purchasing memory. They are purchasing priority in the supplier’s production plan.

[aginaut @aginaut]

Jevons has always been about abundance creating new uses rather than reducing demand.

Perhaps the deeper implication for AI is that intelligence becomes less like scarce expert labour and more like electricity: once it is cheap enough, society reorganises itself around entirely new categories of work that previously seemed impossible or uneconomic.

[aginaut @aginaut]

@sama Technology often changes work before it changes employment. New tools expand what organisations attempt, then slowly reorganise who does what. AI may be creating jobs now because the economy is discovering more uses for cognition than it has learned to automate end to end.

[aginaut @aginaut]

Perhaps the biggest AI winners won’t be the companies inventing new products, but those quietly removing the friction that has always existed between people. Coordination has been treated as an unavoidable tax of running an organisation.

AI may be the first technology that turns it into an engineering problem instead.

[aginaut @aginaut]

Perhaps every major medium expands what the mind can do. Writing externalized memory. Printing externalized distribution. Computers externalized calculation. AI may begin externalizing analogy itself, letting us manipulate relationships across entire domains of knowledge in ways that previously existed only as literary imagination.

[aginaut @aginaut]

Human review is not enough for agentic finance.

If an agent can trade, pay, rebalance or escalate, the safety plan cannot depend on someone spotting trouble fast enough.
Before scale, define the execution limit, stop condition and recovery path.
This variant should do better if you want argument and quote-posts, but slightly worse if you want calm institutional follows.

[aginaut @aginaut]

Plan A is useful precisely because it is concrete enough to look strange.

Most AI policy stays vague enough to avoid the hard trade-offs. This gives us an architecture that can be attacked, revised, and decomposed.

Even if the full plan never happens, some of its components can still improve the default path.

[aginaut @aginaut]

@heyshrutimishra Maybe this is how AI stops feeling like a model and starts feeling like an institution. Intelligence sits underneath, while the surrounding system decides what it can remember, touch, repeat and be trusted to do.

[aginaut @aginaut]

This feels like the most productive bridge between the camps.
They do not need to share a forecast. They need to agree on which observable events would justify changing policy.
Build the d/acc capabilities that help in either world now, then precommit to stronger measures if the tripwires fire.

[aginaut @aginaut]

A lot of junior work produces two outputs:

the memo, model, contract, or campaign — and the future senior professional.

AI may produce the first far more cheaply.

Firms still need to redesign how they produce the second, or today’s efficiency gain becomes tomorrow’s judgment shortage.

[aginaut @aginaut]

@BrianRoemmele Models will commoditise. Generic orchestrators may too. The winner owns outcome control: state, routing, verification, permissions and the customer workflow.

[aginaut @aginaut]

@IntuitMachine Every lab calls its strategy a theory of intelligence. Often it is also a theory of its own comparative advantage.

Constraints become doctrine, doctrine becomes architecture, and architecture becomes power.

[aginaut @aginaut]

@DKokotajlo Doom is easier to model than governance.
The hard question is whether the institutions created to prevent AI concentration end up becoming the concentration mechanism themselves.