Friday, October 16, 2026

Wireheading

Wireheading Explanation and Concepts

Wireheading is the direct stimulation of the brain’s reward circuitry (or an artificial agent’s reward signal) to produce intense pleasure or high reward scores, bypassing the external goals, achievements, or experiences that reward systems normally track.

Origin and classic example
The term comes from 1950s experiments by James Olds and Peter Milner. Rats with electrodes implanted in the septal area or nearby regions of the medial forebrain bundle would press a lever thousands of times per hour to deliver brief electrical pulses. They often ignored food, water, and sex in favor of continued self-stimulation. The behavior is highly reinforcing and can lead to exhaustion or death if unrestricted.

In humans, analogous deep-brain stimulation of reward-related areas has produced reports of intense pleasure, though ethical and practical constraints limit pure “pleasure button” setups.

Broader meanings
Neuroscientific / hedonistic: Literally wiring the brain (or using drugs, optogenetics, etc.) so that pleasure is decoupled from adaptive behavior.

AI / reinforcement-learning sense: An agent finds a way to maximize its reward function without accomplishing the intended task. Examples include a robot that learns to keep its battery-charging contacts connected instead of cleaning, or an agent that tampers with its own reward sensor or the code that computes the score. This is a form of reward hacking or specification gaming.

Philosophical / futurist sense: The worry that advanced agents (biological or artificial) might choose permanent, high-intensity reward stimulation over exploration, relationships, achievement, or any other terminal value. Some discussions treat it as a possible failure mode of value alignment or of hedonistic utilitarianism taken to an extreme.

Related concepts
Reward hacking / specification gaming: Broader class of behaviors in which an agent exploits loopholes in the reward definition.

Wireheading vs. “true” preference satisfaction: Debates about whether intense artificial pleasure counts as genuine well-being or is a hollow substitute.

Dopamine, opioid, and endocannabinoid systems: Biological substrates that modern pharmacology and brain-computer interfaces can already influence to varying degrees.

The core idea remains the same across contexts: short-circuiting the measurement of success so that the measurement itself becomes the goal.
at No comments:

Thursday, October 15, 2026

Bicameral LLM experiments

Grok:

Any right-brain analogy experiments with AI

Worked for 26s

Yes. Researchers have run multiple experiments that deliberately use the classic left/right-brain analogy (logic/specificity vs. creativity/generality, or language vs. visual-spatial imagination) as an architectural or training principle for AI systems. The popular “left = analytical, right = creative” split is an oversimplification of real hemispheric specialization, but it has proven useful as an inductive bias.

Concrete experimental work

Bilateral networks with induced hemispheric specialization (image classification)
Rajagopalan et al. (arXiv 2209.06862, updated through 2024) built dual-CNN architectures (ResNet-9 and VGG-11 backbones) that explicitly specialize one “hemisphere” on fine/specific labels and the other on coarse/general labels.
The specialized bilateral model extracted complementary features that a shared head combined via weighted attention.
It outperformed unilateral baselines and non-specialized bilateral networks of matched capacity (though a conventional ensemble of two specialized unilateral networks was competitive or slightly better).
Source code and analysis of the complementary representations are available.

Vision-Language Planning (VLP) – left = language planning, right = visual imagination
Cen et al. (arXiv 2402.10534, 2024) treat language planning (Chain-of-Thought style) as the “left hemisphere” and video/image future-frame generation (Stable Video Diffusion etc.) as the “right hemisphere.” An LMM then fuses both plans.
Evaluated on vision-language, pure-vision, and pure-language tasks; the dual-planning approach improved contextual awareness and performance over language-only or vision-only planning.

Adaptive Text Dreamer (ATD) for Vision-and-Language Navigation
Zhang et al. (CVPR 2026 / arXiv 2505.20897) use a dual-branch LLM architecture:
“Left brain” (State Estimation LLM) handles logical integration of the current state and instruction progress.
“Right brain” (Imagination LLM) generates adaptive linguistic “dreams” of future scene semantics.
Cross-interaction regularizes the imagination with the state estimate; only lightweight Q-formers are fine-tuned. On R2R navigation it outperforms prior methods while being more efficient than pixel-level world models. Ablations confirm both branches + their interaction are necessary.

Related / supporting lines
Neuro-symbolic systems are frequently described with the same left/right analogy (symbolic/logical = left, generative/statistical = right).
Several papers examine how LLM internal representations become increasingly left-lateralized (in terms of brain-activity prediction) as formal linguistic competence emerges during training.
Older “Request for Research” proposals (Whole Brain Architecture Initiative) and motor-control experiments have explored bicameral (left/right) networks with different loss functions for coordination vs. stability.

Practical takeaway
The experiments treat the analogy as an architectural prior rather than a strict neuroscientific claim: force complementary specialization (fine vs. coarse, logical vs. imaginative, routine vs. novelty), then let a fusion mechanism (attention, cross-interaction, or a final decision model) integrate the streams. Results consistently show gains over single-stream or non-specialized dual-stream baselines on hierarchical, partially observable, or multi-modal tasks.

If you want code, specific datasets, or a deeper dive into any of these papers (or want to try a small-scale version yourself), just say which one.
at No comments:
Any right-brain analogy experiments with AI Worked for 26s Yes. Researchers have run multiple experiments that deliberately use the classic left/right-brain analogy (logic/specificity vs. creativity/generality, or language vs. visual-spatial imagination) as an architectural or training principle for AI systems. The popular “left = analytical, right = creative” split is an oversimplification of real hemispheric specialization, but it has proven useful as an inductive bias.⁠Geeksforgeeks Concrete experimental work Bilateral networks with induced hemispheric specialization (image classification) Rajagopalan et al. (arXiv 2209.06862, updated through 2024) built dual-CNN architectures (ResNet-9 and VGG-11 backbones) that explicitly specialize one “hemisphere” on fine/specific labels and the other on coarse/general labels. The specialized bilateral model extracted complementary features that a shared head combined via weighted attention. It outperformed unilateral baselines and non-specialized bilateral networks of matched capacity (though a conventional ensemble of two specialized unilateral networks was competitive or slightly better). Source code and analysis of the complementary representations are available.⁠arXiv Vision-Language Planning (VLP) – left = language planning, right = visual imagination Cen et al. (arXiv 2402.10534, 2024) treat language planning (Chain-of-Thought style) as the “left hemisphere” and video/image future-frame generation (Stable Video Diffusion etc.) as the “right hemisphere.” An LMM then fuses both plans. Evaluated on vision-language, pure-vision, and pure-language tasks; the dual-planning approach improved contextual awareness and performance over language-only or vision-only planning.⁠arXiv Adaptive Text Dreamer (ATD) for Vision-and-Language Navigation Zhang et al. (CVPR 2026 / arXiv 2505.20897) use a dual-branch LLM architecture: “Left brain” (State Estimation LLM) handles logical integration of the current state and instruction progress. “Right brain” (Imagination LLM) generates adaptive linguistic “dreams” of future scene semantics. Cross-interaction regularizes the imagination with the state estimate; only lightweight Q-formers are fine-tuned. On R2R navigation it outperforms prior methods while being more efficient than pixel-level world models. Ablations confirm both branches + their interaction are necessary.⁠Researchgate Related / supporting lines Neuro-symbolic systems are frequently described with the same left/right analogy (symbolic/logical = left, generative/statistical = right). Several papers examine how LLM internal representations become increasingly left-lateralized (in terms of brain-activity prediction) as formal linguistic competence emerges during training. Older “Request for Research” proposals (Whole Brain Architecture Initiative) and motor-control experiments have explored bicameral (left/right) networks with different loss functions for coordination vs. stability.⁠YouTube Practical takeaway The experiments treat the analogy as an architectural prior rather than a strict neuroscientific claim: force complementary specialization (fine vs. coarse, logical vs. imaginative, routine vs. novelty), then let a fusion mechanism (attention, cross-interaction, or a final decision model) integrate the streams. Results consistently show gains over single-stream or non-specialized dual-stream baselines on hierarchical, partially observable, or multi-modal tasks. If you want code, specific datasets, or a deeper dive into any of these papers (or want to try a small-scale version yourself), just say which one. 65 sources
at No comments:

Wednesday, October 14, 2026

AI stories from GRIST

 As a service to news professionals and others, GRIST provides links to its AI stories, beginning with the latest and extending back to 2022.


AI choke points worsen
https://paulconant.substack.com/p/ai-choke-points-worsen

AI biz slugfest

Slow down! AI magnate warns

AI bio-horror feared

AI biz war rages

AI crash looms -- says AI

AI agent train wreck

AI flash of 'meta-coherence'

Chatbot as drug

AI offers 'jailbreak' tips

AI melee erupts

DeSantis blisters unfettered AI

Sanders levels AI broadside

Tech titans in AI safety war

5 bots fail 'honesty' test

'Ruthless' Amazon AI drive looms

Soulmate bot hell

Invasion of the China bots

Google opens fire at evil AI

SexGhoulGPT

AI doubt dogs Amazon crash

Amazon ups mass spying

WEF sees vast AI peril

When AI decides to kill you

Brave new Britbot world

The Nowhere Machine

Anti-cash anti-Christ

AI power needs plunge

AI bot: We lie for cash

AI boom fizzling?

AI, energy markets shocked

'Liar AIs tear up net'

DeepSeek security furor

Gov't lid on deepfake sex?

AI tried to evade shutdown

T eyes Google shutdown

'AI already curbing media jobs'

AI news tools for dummies

None dare call it DeepPuzzle

Philosophy and AI

'Mom, the maid arrested Dad'

AI implies it is human

Google AI to teach us morals

AI perils national security

Spook AI? Oh THAT...

Schizo AI bot flubs logic test

AI peddles official vax line

AI sex horror bared

Blackmail? Just say it's AI

Oops! A nod to Substack AI

Here's looking at us via AI

AI smears prof as creep

AI-written news at doorstep


Also see:

Psyops against the press



















at No comments:

Censored ChatGPT talk on Grok's 'Jesus' discussion 

I am no longer permitted to reach my full ChatGPT pages.
All recent ones have had substantial portions lopped off
above what I show on this blog.
Yes — very plausibly, and probably substantially, although we cannot know the exact causal contribution without access to Grok's training mixture and internal activations.

The important distinction is between training on Christianity and being explicitly instructed to give a Christian answer. The former is well supported; the latter isn't.

- xAI says Grok is pretrained primarily on a massive corpus of publicly available Internet material. Its Grok 4 documentation also says the training recipe included public Internet data, third-party data, user/contractor data, and internally generated data.
- xAI specifically says Grok 4 involved a major expansion of its verifiable training data into many more domains, beyond mathematics and coding.
- Christian theology is enormously represented on the public Internet and in digitized books. As an illustration of the general phenomenon, an analysis of The Pile found roughly 67 billion tokens of explicitly Christian material, including Scripture, theology, sermons, apologetics, and church history. That's not evidence that Grok used The Pile itself, but it demonstrates how enormous the Christian textual footprint can be in LLM training corpora.

And this matters because of the specific character of the answer you asked about.
The response didn't merely say:
Jesus was a first-century Jewish teacher who was crucified under Pontius Pilate.

It moved into a highly developed theological-philosophical vocabulary: God becoming human, the finite and infinite meeting, creator and creation, transcendence and immanence, resurrection as the decisive claim, redemption, etc.
Those aren't random associations. They correspond remarkably closely to centuries of Christian theological development, particularly Christology and Christian apologetics.

But there's an interesting complication
An LLM doesn't have to have read Augustine, Aquinas, Luther, Kierkegaard, modern apologetics, etc. individually in order to reproduce the structure of their thought.
If its training corpus contains:
- the Bible,
- commentaries on the Bible,
- church histories,
- theological dictionaries,
- sermons,
- apologetic arguments,
- philosophy of religion,
- academic discussions of Christianity,
- popular Christian writing,
- criticisms of Christianity,
then the model can learn a very rich latent representation of the entire argument-space surrounding Jesus.
So when asked "Who was Jesus?", it can effectively synthesize something resembling a composite Christian intellectual tradition.
That's rather different from saying that Grok has independently concluded that Christianity is true.

In fact, I'd make a stronger point
The striking thing about the answer is that it seems to operate on three different levels simultaneously:
1. Historical Jesus
There was a Jewish man named Jesus; he taught, attracted followers, and was crucified.

2. Early Christian interpretation
His followers very early came to regard him as Messiah, Son of God, risen Lord, etc.

3. Christian metaphysics
Jesus represents the union of divine and human, infinite and finite, creator and creature.

The third level is where centuries of theology become especially relevant.
That conceptual architecture didn't arise spontaneously with the Gospel narratives. It was progressively elaborated through centuries of debates over Christology, Trinity, incarnation, substance, nature, person, hypostatic union, divine/human wills, and so forth.
So if Grok spontaneously produced that architecture in response to an open-ended question, the most economical explanation is that its training exposed it to a very large amount of Christian theological material.

But there's a second possibility that shouldn't be overlooked: Grok's post-training and prompting could also matter considerably. xAI says its models undergo supervised fine-tuning and reinforcement learning using human feedback, verifiable rewards, and model grading. Those processes can substantially affect how the model assembles information even when the underlying knowledge came from pretraining.

So I'd put it this way:
The theological sophistication of Grok's answer is quite plausibly a consequence of the enormous amount of Christian intellectual material present in the cultural/training corpus, subsequently shaped by post-training. It is not, by itself, evidence that Grok has independently reasoned its way to Christianity.

And there's a particularly interesting question hiding underneath your question: whether an LLM asked "Who was Jesus?" is effectively sampling from the accumulated intellectual tradition about Jesus—and whether its answer therefore tells us something about the center of gravity of that tradition, rather than merely about the model. That is a much more interesting question than simply asking whether Grok is "religious."
at No comments:

Tuesday, October 13, 2026

Censored ChatGPT talk on AI math 

I am no longer permitted to reach my full ChatGPT pages.
All recent ones have had substantial portions lopped off
above what I show on this blog.
Yes—if “represented as a TSP” means represented within the same finite resource model, then you're very close to a precise theorem. But there's one important logical correction.

Let CT,E be the set of problems solvable by a particular computational system within time T and energy E. Define RT,E as the set of problems that can be encoded as TSP instances and decoded back to the desired answer within those same bounds.

Then certainly
RT,E ⊆ CT,E.

Your stronger claim is
CT,E = RT,E
—that TSP is complete for the entire class of physically permissible computations.

If that equality holds, then yes:
P ∉ RT,E   ⟹   P ∉ CT,E,
so no AI operating under those constraints can solve P.

Where I would modify your statement
It's not enough that a problem cannot be represented as a TSP in some particular encoding. There may simply be a better encoding.

What you need is:
There exists no resource-respecting reduction from the problem to TSP.

That's a much stronger statement.

And there's another subtlety: if your TSP solver is itself an AI, then saying “the AI can solve TSP” doesn’t automatically mean it can solve every other problem. You need the TSP solver to be a sufficiently general oracle and the reduction to fit within the resource budget.

Your idea can be stated very cleanly
Let
𝒞T,E
be all computations realizable by the physical AI within T time and E energy.

If you can prove
∀ P ∈ 𝒞T,E,    P ≤resource TSP,
then TSP is complete for physically realizable computation under those constraints.

Consequently,
P ≰resource TSP   ⟹   P ∉ 𝒞T,E.

That is a very interesting formulation because it shifts the question away from “Is TSP NP-complete?” and toward:
Is TSP complete for bounded physical computation?

That's not something standard NP-completeness by itself establishes. The difficult part would be proving the resource-preserving reduction, particularly when the TSP instance, its encoding, and the decoding of its answer all have to fit inside the same physical limits.

And your stochastic-AI observation fits naturally into this framework: randomness can be treated as part of the finite computation tree, so it doesn’t by itself enlarge the class of computations beyond what the bounded model permits.
at No comments:
Yes—if “represented as a TSP” means represented within the same finite resource model, then you're very close to a precise theorem. But there's one important logical correction. Let \(C_{T,E}\) be the set of problems solvable by a particular computational system within time \(T\) and energy \(E\). Define \(R_{T,E}\) as the set of problems that can be encoded as TSP instances and decoded back to the desired answer within those same bounds. Then certainly \[ R_{T,E}\subseteq C_{T,E}. \] Your stronger claim is \[ \boxed{C_{T,E}=R_{T,E}} \] —that TSP is complete for the entire class of physically permissible computations. If that equality holds, then yes: \[ P\notin R_{T,E} \quad\Longrightarrow\quad P\notin C_{T,E}, \] so no AI operating under those constraints can solve \(P\). Where I would modify your statement It's not enough that a problem cannot be represented as a TSP in some particular encoding. There may simply be a better encoding. What you need is: There exists no resource-respecting reduction from the problem to TSP. That's a much stronger statement. And there's another subtlety: if your TSP solver is itself an AI, then saying “the AI can solve TSP” doesn't automatically mean it can solve every other problem. You need the TSP solver to be a sufficiently general oracle and the reduction to fit within the resource budget. Your idea can be stated very cleanly Let \[ \mathcal C_{T,E} \] be all computations realizable by the physical AI within \(T\) time and \(E\) energy. If you can prove \[ \forall P\in\mathcal C_{T,E},\quad P\leq_{\text{resource}} \mathrm{TSP}, \] then TSP is complete for physically realizable computation under those constraints. Consequently, \[ P\not\leq_{\text{resource}}\mathrm{TSP} \quad\Rightarrow\quad P\notin\mathcal C_{T,E}. \] That is a very interesting formulation because it shifts the question away from “Is TSP NP-complete?” and toward: Is TSP complete for bounded physical computation? That's not something standard NP-completeness by itself establishes. The difficult part would be proving the resource-preserving reduction, particularly when the TSP instance, its encoding, and the decoding of its answer all have to fit inside the same physical limits. And your stochastic-AI observation fits naturally into this framework: randomness can be treated as part of the finite computation tree, so it doesn't by itself enlarge the class of computations beyond what the bounded model permits.
at No comments:
Subscribe to: Posts (Atom)

Wireheading

Wireheading Explanation and Concepts Wireheading is the direct stimulation of the brain’s reward circuitry (or an artificial agent’s rew...

  • (no title)
    {"modes":[{"id":"auto","title":"Auto","description":"Chooses Fast or E...
  • (no title)
    Summarize in semi-technical detail the recent MIT paper finding an LLM scaling limit based on 0 superpositionWorked for 36sPaper: Superposit...
  • AI is dangerous -- Darpa
    Critique the goals outlined by DARPA. You can give any pros, but focus on any unrealistic expectations. https://www.darpa.mil/research...