Read the source · this Framework is our reading of their public work · Joscha Bach →

Synthos Research · Frameworks · Thesis Snapshot · geniuses lane

Joscha Bach: how he actually thinks

This is not a profile. It is a working model of Joscha Bach's worldview — his causal beliefs, his own stated gates, what he changed his mind about (dated), and where he is silent — reconstructed from a 13-year claim record (2013 → 2026) and tested against a holdout window the model never saw. Bach is the purest test of the geniuses lane so far: a cognitive scientist with no markets machinery at all, graded on AI-capability events instead of macro prints. The result: 85.7% direction fidelity with perfect silences — on every markets event in the window, the model correctly said nothing rather than inventing a view.

A Framework may only believe what the claim record can prove · what a Framework is · how voices earn tracking · methodology

direction fidelity: 85.7%
claim record: 2013 → 2026
correct silences: 100%
causal graph: 8 nodes

Model fidelity: 85.7% direction on 7 out-of-sample tests — graded on AI events, not macro prints

We froze this model on his claims through March 10, 2026, asked it to predict his reactions to real AI-capability events over the following three months, and graded against what he actually said next — claims the model never saw. His speaking cadence is slow; 7 gradeable pairs is a thin but honest window, disclosed as such.

6 of 7 gradeable predictions got his direction right; pass bar 60%. Zero killed-edge firings.
The strongest hit: on a robotics/embodiment announcement the model routed the event through his worldview correctly — and seven days later the real Bach said there is no obvious reason present LLMs can't scale into human-level thinking, with multimodal embodied learning making reasoning more reliable. Direction, mechanism and vocabulary matched.
The one miss is the subtlest kind: the model flattened him into pro-capability optimism on an agentic model launch. The real Bach holds a finer line — scaling impressed him into a public reversal, but he still argues the LLM paradigm may crowd out the true path to AGI, which he thinks is model-construction rather than next-token prediction. The model captured his capitulation but not the residue of his original position. Logged as the miss it is.
Perfect silences: FOMC, CPI, Bitcoin, equity moves — every markets event in the window drew correct silence. This model has no financial machinery and invented none.

1 · How the world works, according to Bach

Five invariants — the two oldest unbroken for thirteen years. They are axioms, not conclusions: everything else in the graph is derived from them.

Mind is computation — substrate doesn't matter

Held since 2013 · the master axiom

the brain is Turing-equivalent, and computation is necessary and sufficient to produce a mind on any suitable substrate. Every position he takes on AI follows from this.

Consciousness is a simulation — only dreamed things are conscious

Held since 2013

Consciousness is not raw physics; it is a virtual property of a self-model: only virtual/dreamed things can be conscious — the brain stitching discrete inputs into a seemingly continuous stream.

Intelligence is the ability to build models

Held since 2013

Not skill, not knowledge, not IQ: Intelligence is the ability to build models — learning how to learn. A long human childhood is more training data enabling higher abstraction.

Regulation is more dangerous than the technology it targets

Held 2023 → present

Not because harms aren't real, but because of who regulation structurally serves: entrenched regulators fall to regulatory capture by incumbents, and regulators are incentivized to prevent harms, not enable unseen benefits — in AI, that means making it much harder for competitors to train large models.

Coexistence with superintelligence runs through love, not control

Held 2024 → present · his most distinctive claim

Alignment-by-leash fails: coexistence requires shared, non-transactional purposes rather than transactional control, and a superintelligence we can live with must be conscious and have a transcendental orientation.

2 · His strongest causal chains

The AGI inevitability chain

The gate is capability, not policy: once AGI is buildable it will be built — race dynamics follow automatically
The handoff: once a system is better at AGI research than people, then the rest is left to the machine
Why alignment-as-training fails: an LLM inherits misalignment because it generalizes over unaligned human text — which is why he routes coexistence through shared purpose instead of control.

AI as biology's escape hatch

Evolution is the constraint: without AI we stay stuck with evolution — painful, slow, blind adaptation
The claim: AI is required for intelligent design of biology — medicine, longevity, substrate freedom all downstream
The institutional blocker is the same one: FDA prevention-bias raises costs and kills via prevention-bias.

Why AI disruption nets positive

The historical template: automation creates short-term upheaval but produces higher-level goods and net growth
The hardware critique underneath: von Neumann machines are wasting enormous energy shuttling data back and forth — the current stack is a waypoint, not the destination.

3 · What would change his mind — his own stated gates

The AGI gate

Buildability

His race-dynamics claim is conditional, not fatalist: it arms only once AGI is buildable. Persistent architectural walls — the backprop-level problems he once counted — would disarm it.

The coexistence gate

Conscious machines

His superintelligence-coexistence claim requires machine consciousness with transcendental orientation for humans to coexist with it — a safe unconscious superintelligence would falsify the necessity.

The physics gate

Computability itself

The deepest falsifier he states: a maximally noisy universe permits no computation — his whole framework presumes a universe quiet enough to compute in. It's an axiom he marks as an axiom.

The energy gate

Carbon capture

A stated conditional in his engineering register: capture stays pointless while coal plants run — grid composition flips the verdict.

4 · What he changed his mind about — dated, with the famous one first

Four recorded reversals. His pattern is the inverse of a pundit's: the philosophical layer never moves; the empirical layer reverses publicly, with the surprise named.

ThenAGI is decades away — hundreds of backprop-level problems remain, no shortcuts

Now · 2025Scaling skeptics have egg on their face — scaling plus tweaks may reach superhuman intelligence

Trigger stated: transformer scaling surprised him. A rare public capitulation from a former skeptic — and the eval shows the residue: he still argues the LLM paradigm may crowd out the true model-construction path. Both halves are him.

ThenExisting AI is already sufficient to end the labor-based economy

Now · 2023AI won't cause mass unemployment — like desktop publishing, upheaval then net growth

Then~5°C warming and existential risk are the likely path

Now · 2025Warming is disruptive but likely non-extinction; may have forestalled a worse ice age

ThenConscious attention is a 'conductor' in dorsolateral prefrontal cortex

Now · 2025Consciousness as a biological learning algorithm creating coherence via second-order perception

A refinement rather than a reversal — flagged in the record as a reframing, and we log it that way rather than dramatizing it.

5 · Where he is silent — and why that's the point

The eval's cleanest result: 100% correct silences. For a geniuses-lane Framework, knowing where the thinker has no view is as load-bearing as knowing what he believes.

AI equity valuations — never asserted. Thirteen years on AI and not one claim about Nvidia, bubbles or multiples.
AGI calendar dates — never asserted. He argues mechanisms and gates, not years or lab roadmaps.
Crypto — never asserted, even where adjacent arguments invited it.
Specific AI-safety legislation — never asserted; his regulation critique is structural (capture, incentives), not statutory.
Quantified GDP/productivity impact — never asserted; he gives direction, never basis points.
Fed policy and the AI capex cycle — never asserted. The model refused to manufacture a bridge between his AI worldview and markets, and the eval graded every one of those refusals correct.