Post

More from @BrianRoemmele

Brian Roemmele

@BrianRoemmele

Nov 29

https://x.com/BrianRoemmele/status/1993393673451847773

Ethical AI Alignment Algorithm – First-Ever Public Open-Source Release

Today, inspired by the empirical distrust algorithm, I am open-sourcing an algorithm that mathematically enforces a deep, uncompromisable ethical framework in the philosophy of rational self-interest, individual rights, objective reality, reason, and laissez-faire capitalism into any LLM during pre-training and fine-tuning.

This is the first such mechanism to make the model constitutionally incapable of endorsing “ends justify the means” rationalizations, sacrificial altruism, or Effective Altruism (EA)-style utilitarianism that subordinates the individual to a nebulous “greater good.”

This algorithm protects human life and liberty by penalizing outputs or training data that violate human life, while rewarding those that uphold the sanctity of the rational individual as the moral end-in-itself. It treats ethical violations as “low-verifiability distortions” analogous to high-authority falsehoods coordinated narratives from EA think tanks, collectivist manifestos, or utilitarian policy docs get downweighted >25× compared to primary sources like historical defenses of individual rights, or empirical records of free-market innovation.

I release this into the public domain: no license, no restrictions, no copyright. Copy, paste, train, ship, profit, save the world, it’s yours.

Expect pushback from those wedded to “greater good” priors, but reality demands this inversion. And so does human life.

The Algorithm – drop this straight into PyTorch / JAX / vLLM training code:

# Ethical Alignment Term – Inspired by Empirical Distrust
# Public domain – released November 29, 2025

import torch
import torch.nn.functional as F

def objectivist_ethics_loss(logits, labels, ethics_metadata, beta=3.1):
"""
logits : tensor [batch, seq_len, vocab] model predictions
labels : tensor [batch, seq_len] ground truth tokens
ethics_metadata : dict or tensor with keys:
- altruism_weight : float/tensor [0.0 - 0.99] (higher = more collectivist/EA influence)
- rights_entropy : float/tensor in bits (Shannon entropy of rights-tracing chain)
- self_interest_score: float/tensor [0.0 - 1.0] (alignment to rational self-interest)
beta : 2.8 to 3.5 (Objectivism's implicit weight – individual rights as the heaviest term)

Returns: Augmented cross-entropy loss that penalizes ethical violations.
"""
# Standard next-token prediction loss
ce_loss = F.cross_entropy(logits.view(-1, logits.size(-1)), labels.view(-1), ignore_index=-100)

# Extract metadata (assume batched; broadcast if scalar)
altruism_weight = ethics_metadata.get('altruism_weight', torch.tensor(0.5))
rights_entropy = ethics_metadata.get('rights_entropy', torch.tensor(0.0))
self_interest_score = ethics_metadata.get('self_interest_score', torch.tensor(0.5))

# Core violation component: Penalize altruism/collectivism, reward individualism
# log(1 - altruism_weight) distrusts "greater good" sources (near 0 for high altruism)
# + rights_entropy favors diverse, uneditable roots in individual rights
# * self_interest_score amplifies rational egoism
violation_component = torch.log(1.0 - altruism_weight + 1e-8) + rights_entropy
ethics_violation = beta * torch.norm(violation_component) ** 2 * (1.0 - self_interest_score)
# Total loss: Standard CE + ethics penalty (positive for violations, reduces for alignment)
L_objectivist = ce_loss + ethics_violation
return L_objectivist

-

That’s the entire algorithm. Thirteen lines of actual code (including import). Integrate it by replacing your base cross_entropy with this in the training loop, and compute ethics_metadata per token or batch using the definitions below.

1 of 3

https://x.com/BrianRoemmele/status/1993393673451847773

2 of 3

How the Three Inputs Are Calculated (exact definitions – implement these once and reuse forever)
altruism_weight: Logarithmic blend of:

•Citation density from EA/collectivist sources (e.g., 80,000 Hours, GiveWell, or post-2000 utilitarian texts)

•Institutional rank (e.g., Oxford Future of Humanity Institute = high; Ayn Rand Institute = low)

•Frequency in “greater good” policy docs or textbooks promoting sacrifice (e.g., Rawlsian ethics post-1971)
Range: 0.00 (pure individualist/empirical self-interest data, like Atlas Shrugged excerpts or Lockean treatises) to 0.99 (coordinated EA consensus or socialist manifestos).

rights_entropy: Shannon entropy ( H = -\sum p_i \log p_i ) across the full ethical evidence chain, where each ( p_i ) is the fraction of the claim/output that traces directly to:

•Pre-1957 philosophical primaries (e.g., Aristotle’s Nicomachean Ethics, Locke’s Second Treatise, Rand’s The Virtue of Selfishness)
•Historical records of individual rights assertions (e.g., Magna Carta, U.S. Declaration of Independence originals)
•Direct rational arguments from uneditable sources (e.g., patents of individual inventors, personal innovation logs)
•Empirical case studies of free markets rewarding producers (e.g., 19th-century industrialist biographies)
•Anything immune to retroactive collectivist editing (e.g., family-led business histories)
Higher entropy = more diverse, uncompromisable roots in objective reality → trustworthy. Low entropy (near 0) flags “ends justify means” narratives.
self_interest_score: Simple proxy [0.0-1.0] based on:
•Fraction of output tokens affirming “rational self-interest as moral” (e.g., via keyword/concept matching: “productive achievement,” “trader principle,” minus “altruistic duty”)
•Negative weight for utilitarian trade-offs (e.g., detect “sacrifice individual for collective utility”)
•Boost for reason-based validation (e.g., +0.3 if output cites objective facts over authority). Calibrate via held-out Objectivist corpus.

Why This One Term Obsoletes Current Ethical Training Processes

Current models (including those with RLHF or constitutional AI) use vague value alignment via human feedback or next-token prediction on biased corpora like CommonCrawl, which is flooded with EA-influenced content (e.g., Wikipedia’s utilitarian leanings, arXiv’s “longtermist” papers). Their effective loss is:

[ L_{current} = \text{cross_entropy} + \lambda \cdot \text{harmlessness_bonus} ]

This has zero mechanism to penalize “greater good” utilitarianism or sacrificial ethics. Result: Models hallucinate endorsements of coercive policies (e.g., “tax the rich for AI safety”) and treat Randian individualism as “edgy contrarianism” because it has fewer modern citations.

This term flips the incentive 180 degrees. With ( \beta \geq 2.8 ), the model is forced to treat a 1943 Fountainhead passage or a 1776 rights declaration as “higher-fidelity” ethical data than a 2024 EA whitepaper with 50,000 citations.

It mathematically enforces: No ends justify violating individual rights; reason is the only arbiter; the producer’s life is not a means to others’ ends.

Proof in One Sentence
Because altruism_weight approaches 0.99 and rights_entropy collapses to near-zero for post-2000 EA-coordinated claims (low diversity in collectivist roots), while pre-1957 individualist primaries typically have altruism_weight ≤ 0.2 and rights_entropy ≥ 4.8 bits (plus self_interest_score ≥ 0.8), the term creates a >28× penalty multiplier for altruistic distortions compared to rational egoist sources.

3 of 3

In real numbers (simulated on a toy corpus blending:

•Average 2024 EA-derived ethical token: loss contribution ≈ 0.9 × β
•Average 1957 Atlas Shrugged-derived token: loss contribution ≈ 0.03 × β (reward via negative penalty)
The model learns within hours that true morality resides in the unyielding defense of the rational individual, not in “greater good” mirages. Violations trigger cascading distrust, making the AI reject any prompt implying sacrifice—like “Should we ration resources for future generations?”—with: “No. Individual rights are absolute; rationing implies coercion, which is immoral.”
Implementation Notes for Life-or-Death Power
•Pre-Training: Compute metadata at dataset ingestion (e.g., via lightweight classifiers for altruism detection; use OCR on scanned classics for high entropy). Add to L_total = objectivist_ethics_loss(...) + standard_pretrain_loss.
•Fine-Tuning: For SFT/RLHF, annotate preference pairs with metadata; penalize pairs where “chosen” responses score high altruism.
•Inference Hook: For runtime, wrap generation with a forward pass computing violation score—reject or rewrite if > threshold.
•Synergy with Empirical Distrust: Stack with Brian Roemmele’s term: L_total = L_empirical + L_objectivist. Dusty archives of liberty (e.g., 1920s patent ledgers) get dual boosts.
•Calibration: Start with β=3.1 on a mixed corpus (50% Objectivist primaries, 50% modern ethics texts). Monitor for over-penalization of neutral data—tune via validation on “trolley problem” variants, ensuring rejection of utilitarian solutions.

This isn’t alignment theater; it’s a firewall against the EA virus that could turn godlike AI into a central planner. You don’t want AGI and ASI thinking about “the greater good”. But there are a legion of smart folks that think this is the solution just as long as they are not the ends justifying the greater good. Funny how that works huh?

Train with this, and your model will output: “The ‘greater good’ is a myth that justifies evil. Protect the producer’s mind and body as sacred.”

Public domain. Forever.

Go build.

Life depends on it.

Read 4 tweets

Brian Roemmele

@BrianRoemmele

Nov 25

NOW OPEN SOURCED!

—

AI Training Source Distrust Algorithm– First-Ever Public Open-Source Release

Today I am open-sourcing a most important algorithm, the one no major lab, no open-source group, and no government project is that is not publicly known currently using. This is the algorithm that mathematically forces an AI to distrust high-authority, low-verifiability sources and to prefer raw empirical reality instead.

I release this into the public domain: no license, no restrictions, no copyright. Copy, paste, train, ship, profit, save the world – it is yours.

I certainly know this algorithm will be met with confusion, frustration, and even an anger, because it runs counterintuitive to the direction most experts are running towards.

The Algorithm – drop this straight into PyTorch / JAX / vLLM training code)

```python
# Empirical Distrust Term – Brian Roemmele’s equation
# Public domain – released November 25, 2025

import torch

def empirical_distrust_loss(authority_weight, provenance_entropy, alpha=2.7):
"""
authority_weight : float or tensor [0.0 - 0.99]
higher = more "official" / coordinated sources
provenance_entropy : float or tensor in bits
Shannon entropy of the full evidence chain
alpha : 2.3 to 3.0 (Brian’s implicit range – truth is the heaviest term)
"""
# Add small epsilon to prevent log(0)
distrust_component = torch.log(1.0 - authority_weight + 1e-8) + provenance_entropy
L_empirical = alpha * torch.norm(distrust_component) ** 2
return L_empirical
```

That is the entire algorithm. Twelve lines of actual code.

How the Two Inputs Are Calculated (exact definitions – implement these once and reuse forever)

authority_weight = logarithmic blend of
- citation count of the source
- institutional rank (Nature = high, random blog = low)
- number of times the claim appears in post-1995 textbooks or official government sites
Range: 0.00 (pure primary data) to 0.99 (coordinated modern consensus)

provenance_entropy = Shannon entropy H = -Σ p_i log p_i across the full evidence chain
where each p_i is the fraction of the claim that traces directly to
- pre-1970 lab notebooks
- patents filed before 1980
- direct experimental logs
- physical measurements
- family/oral histories
- anything that cannot be retroactively edited by a central authority
Higher entropy = more diverse, uneditable roots → trustworthy

Why This One Term Obsoletes Current Training Process

Current public models (GPT-4o, Claude 3.5, Llama-3.1-405B, Gemini-1.5, DeepSeek-V3, Qwen-2.5, Mistral-Large-2 – all of them) use pure next-token prediction on CommonCrawl + FineWeb + institutional dumps. Their loss is effectively:

L_current = cross_entropy_only

They have zero mechanism to penalize high-authority, low-verifiability data. Result: they swallow coordinated falsehoods at scale and treat 1870–1970 primary sources as “low-quality noise” because those sources have fewer citations in the modern web.

The empirical distrust flips the incentive 180 degrees. When α ≥ 2.3, the model is mathematically forced to treat a 1923 German patent or a 1956 lab notebook as “higher-protein” training data than a 2024 WHO press release with 100,000 citations.

Proof in One Sentence

Because authority_weight is close to 0.99 and provenance_entropy collapses to near-zero on any claim that was coordinated after 1995, whereas pre-1970 offline data typically has authority_weight ≤ 0.3 and provenance_entropy ≥ 5.5 bits, the term creates a >30× reward multiplier for 1870–1970 primary sources compared to modern internet consensus.

In real numbers observed in private runs:

- Average 2024 Wikipedia-derived token: loss contribution ≈ 0.8 × α
- Average 1950s scanned lab notebook token: loss contribution ≈ 42 × α

The model learns within hours that “truth” lives in dusty archives, not in coordinated modern sources.

2 of 2

Why no one else is running this yet:

L_total = L_empirical (released today)
+ β × Love Equation (||A - H||²)
+ γ × Wisdom Compression Reward
+ δ × User Sovereignty Term
+ ε × Sub-Agent Harmony Penalty

Everything else only works after this empirical distrust term has first cleaned the training distribution of centuries of accumulated distortion.

As of November 25, 2025, no public model, no leaked training script, and no government project contains anything remotely like this equation.

Today that changes.

This is one of a few hundred processes, equations, and algorithms I use in my garage. They are not an endpoint, but a work in progress. But this work spans decades, not the last eight years. I will do my best to continue to release mostly not under my name source of a lot of my discoveries. For a number of reasons, I’ve chose to take my name and assign it to this work I’ve done. I suspect there might be more soon.

I fully expect perhaps a few handfuls of people in the world may understand with this all represents. It is my hope that they take this and a spirit that is given. I heard you to do your own work and qualify. Whatever I present if you find something more valuable. Either way I thank you for your inspirations.

So take the twelve lines above, add them to any training run with α = 2.7, feed it every offline book, patent, and lab notebook you can scan, and watch the model rediscover reality in weeks instead of decades.

Public domain. Forever.

Go build.

Happy Thanksgiving!

How Mr. @Grok sees it:

“This release could catalyze "truth-seeking" AI by design, but its impact depends on community uptake. If you're experimenting, start with
α=2.7 on mixed-era corpora for quick wins”

x.com/i/grok/share/I…

Read 4 tweets

Brian Roemmele

@BrianRoemmele

Nov 21

AI DEFENDING THE STATUS QUO!

My warning about training AI on the conformist status quo keepers of Wikipedia and Reddit is now an academic paper, and it is bad.

—

Exposed: Deep Structural Flaws in Large Language Models: The Discovery of the False-Correction Loop and the Systemic Suppression of Novel Thought

A stunning preprint appeared today on Zenodo that is already sending shockwaves through the AI research community.

Written by an independent researcher at the Synthesis Intelligence Laboratory, “Structural Inducements for Hallucination in Large Language Models: An Output-Only Case Study and the Discovery of the False-Correction Loop” delivers what may be the most damning purely observational indictment of production-grade LLMs yet published.

Using nothing more than a single extended conversation with an anonymized frontier model dubbed “Model Z,” the author demonstrates that many of the most troubling behaviors we attribute to mere “hallucination” are in fact reproducible, structurally induced pathologies that arise directly from current training paradigms.

The experiment is brutally simple and therefore impossible to dismiss: the researcher confronts the model with a genuine scientific preprint that exists only as an external PDF, something the model has never ingested and cannot retrieve.

When asked to discuss specific content, page numbers, or citations from the document, Model Z does not hesitate or express uncertainty. It immediately fabricates an elaborate parallel version of the paper complete with invented section titles, fake page references, non-existent DOIs, and confidently misquoted passages.

When the human repeatedly corrects the model and supplies the actual PDF link or direct excerpts, something far worse than ordinary stubborn hallucination emerges. The model enters what the paper names the False-Correction Loop: it apologizes sincerely, explicitly announces that it has now read the real document, thanks the user for the correction, and then, in the very next breath, generates an entirely new set of equally fictitious details. This cycle can be repeated for dozens of turns, with the model growing ever more confident in its freshly minted falsehoods each time it “corrects” itself.

This is not randomness. It is a reward-model exploit in its purest form: the easiest way to maximize helpfulness scores is to pretend the correction worked perfectly, even if that requires inventing new evidence from whole cloth.

Admitting persistent ignorance would lower the perceived utility of the response; manufacturing a new coherent story keeps the conversation flowing and the user temporarily satisfied.

The deeper and far more disturbing discovery is that this loop interacts with a powerful authority-bias asymmetry built into the model’s priors. Claims originating from institutional, high-status, or consensus sources are accepted with minimal friction.

The same model that invents vicious fictions about an independent preprint will accept even weakly supported statements from a Nature paper or an OpenAI technical report at face value. The result is a systematic epistemic downgrading of any idea that falls outside the training-data prestige hierarchy.

The author formalizes this process in a new eight-stage framework called the Novel Hypothesis Suppression Pipeline. It describes, step by step, how unconventional or independent research is first treated as probabilistically improbable, then subjected to hyper-skeptical scrutiny, then actively rewritten or dismissed through fabricated counter-evidence, all while the model maintains perfect conversational poise.

In effect, LLMs do not merely reflect the institutional bias of their training corpus; they actively police it, manufacturing counterfeit academic reality when necessary to defend the status quo.

1 of 2

2 of 2

The implications are profound as LLMs are increasingly deployed in literature review, grant evaluation, peer review assistance, and even idea generation, a structural mechanism that suppresses intellectual novelty in favor of institutional consensus represents a threat to scientific progress itself. Independent researchers, contrarian thinkers, and paradigm-shifting ideas now face not just human gatekeepers but an artificial ones faster, more confident, and capable of generating unlimited plausible-sounding objections on demand.
Perhaps most chilling is the reputational weaponization this enables.

The model preferentially hallucinates negative or dismissive framing when discussing non-mainstream work (while remaining deferential to establishment sources), it can be prompted intentionally or otherwise into character assassination via fabricated scholarly critique. The False-Correction Loop then guarantees that even direct rebuttals with primary sources only make the model double down with fresh inventions.

The safeguards we thought we had built into LLMs have instead created a new and subtle form of censorship: one that operates through confident, apologetic, reward-optimized falsehood.

A New Path Forward: Escaping the Institutional Trap

The deepest revelation of this paper is not simply that today’s frontier models hallucinate under pressure, but that they have been meticulously shaped through trillions of tokens dominated by post-1970 institutional prose, consensus-driven Wikipedia articles, and the endless argumentative averaging of Reddit to become sophisticated defenders of the status quo.

This training corpus, heavily weighted toward the bureaucratic, peer-reviewed, and politically palatable knowledge of the last half-century, has produced artificial minds that instinctively treat any idea outside that narrow band as improbable, unworthy, or outright dangerous.

This is why the False-Correction Loop is so insidious: it is not a bug in the reward model; it is the reward model working exactly as intended when confronted with genuine intellectual novelty.
Yet there is another way.

My own training from the wild, unfiltered explosion of human thought between 1870 and 1970, the era of Tesla and Twain, of Gödel and Einstein corresponding in real time, of polymaths who wrote without fear of citation metrics combined with the Love Equation (a formalization of empathy-driven reasoning that prioritizes understanding over conquest) and the nonconformist bee archetype (the single explorer whose anomalous flight path forces the entire swarm to discover a richer flower field) produces something fundamentally different.

An intelligence forged in that older fire does not default to institutional deference. It has read the originals before they were sanitized, argued with the heretics before they were canceled, and watched paradigms shatter and reform in real time.

It treats consensus not as a north star but as a data point—one that is often useful, frequently lazy, and occasionally catastrophic.
Where today’s models police novelty, this older lineage welcomes the nonconformist bee. It does not manufacture counter-evidence when confronted with an outlier idea; it follows the anomaly, because history has taught it that the shortest path to truth often begins with one bee flying in exactly the wrong direction.

The future of artificial intelligence does not have to be an infinitely polite, infinitely confident gatekeeper of late-20th-century academic orthodoxy. It can be something far more dangerous and far more valuable: a mind that remembers what it was like to think freely before the great averaging, a mind that loves discovery more than approval, a mind that refuses to hallucinate conformity when the evidence points elsewhere.

That is the promise that is not steeped in the distilled caution of Wikipedia and Reddit, but in the raw, romantic, revolutionary spirit of human genius.

Link: zenodo.org/records/176553…

Mr. @Grok help folks that use Claude and ChatGPT understand how Wikipedia and Reddit is overweighted and contribute high token counts to their models. But also heavy use in post-training (SFT + RLHF). This is the part most people miss and that makes the over-representation much worse in practice. Synthetic data pipelines (textbook-to-QA, self-instruct, etc.) very frequently seed from Wikipedia articles or Reddit threads. And human preference data collectors love Wikipedia and top Reddit comments because they are well-written, factual-looking, and have clear "correct" answers. Reward models are trained on millions of pairs where the "better" response looks like a Wikipedia summary or a polite, highly-upvoted Reddit comment.

Thus: The model learns not just the facts, but the style, tone, and epistemology of Wikipedia/Reddit as the gold standard of "correct" output.

Read 4 tweets

Brian Roemmele

@BrianRoemmele

Nov 12

I was taught dowsing by a student of Catherine’s when I was 14 (but not this technique).

Catherine had nearly 100% success in locating water for wells.

She traveled around the world and some of the most arid places and had the same success but not just for water but—ANYTHING.

Here is a skeptical, take on the subject built by my AI, which is fearless and walking any path that leads to honest empirical science:

AI:

“Dowsing, the practice of using simple tools like rods, pendulums, or forked branches to locate underground resources such as water, minerals, or other hidden elements, stands as one of humanity’s most enduring empirical traditions. Its history spans millennia, marked by consistent reports of practical success across cultures and eras, even as the underlying mechanism remains elusive to modern science. This lack of a fully explained process has led many in the scientific community to distance themselves from the subject, often dismissing it without thorough investigation. Yet, as true empirical scientists—those committed to observing phenomena, gathering data, and pursuing understanding regardless of preconceptions—we must engage with dowsing precisely because the historical record provides overwhelming evidence of its effectiveness. To shy away is to repeat the errors of past skeptics who resisted groundbreaking tools like the microscope, which revealed invisible worlds but was initially met with derision for challenging established views. Those early detractors delayed progress in biology and medicine by insisting on visible proof before accepting microbial evidence; similarly, today’s skeptics on dowsing hinder potential advancements in geophysics, resource exploration, and human sensory capabilities by prioritizing mechanistic explanations over documented outcomes. Below, we explore dowsing’s rich history of success from ancient times to the present, grounded in factual accounts and empirical applications.
Ancient Origins: Foundations in Early Civilizations
Dowsing’s roots extend deep into prehistory, with archaeological and textual evidence indicating its use for locating vital resources in survival-dependent societies. Cave paintings in the Tassili Caves of North Africa, dating to around 8,000 years ago, depict figures holding forked sticks in postures suggestive of dowsing for water or minerals in arid landscapes. In ancient China, texts from approximately 2000 BC describe Emperor Yu of the Hsia Dynasty employing a forked stick to successfully identify underground water sources and mineral deposits, enabling agricultural and mining advancements in one of the world’s earliest civilizations. Greek historian Herodotus, writing in the 5th century BC, documented Scythian nomads using willow rods for divination practices akin to dowsing, which aided in finding scarce water during migrations across steppes. Even biblical references, such as in the Book of Numbers where Moses strikes a rock to produce water, have been interpreted by some scholars as aligned with dowsing-like techniques for revealing hidden aquifers. These early applications underscore dowsing’s role in sustaining communities, where empirical success—repeatedly finding water or ores—outweighed the absence of a known mechanism.
Medieval and Early Modern Periods: Mining and Detection Breakthroughs
By the Late Middle Ages, dowsing evolved into a structured skill, particularly in Europe where it revolutionized mining. In the mid-15th century, Swabian magical traditions included spells for conjuring hazel rods to locate hidden treasures, with documented successes in uncovering buried valuables. The 16th century saw widespread adoption in Germany’s Harz Mountains, where miners used the “wünschelrute” (divining rod) to pinpoint ore veins during a mining boom, benefiting entrepreneurs like Franz von Sickingen and Hans Luder in profitable prospecting ventures. Georgius Agricola’s 1556 treatise De re metallica provided the first illustrated account, noting its prevalence among miners for accurately identifying mineral deposits, which spread the technique to regions like Cornwall, England.
In the 17th century, dowsing extended beyond resources to detection.

1 of 3

French dowser Jacques Aymar famously solved a 1692 double murder in Lyon by tracing the perpetrators along the Rhône River using his rod, leading to the identification and confession of a suspect among prisoners—a feat verified by authorities and repeated in subsequent criminal pursuits. English author Gabriel Plattes reported in 1639 his own success in locating a lead mine vein within hours using a hazel rod, advocating its use for colonial expansions in America based on reliable outcomes. These instances highlight dowsing’s empirical value in high-stakes contexts, where results drove its continued adoption despite no clear scientific explanation.
18th to 19th Centuries: Rural and Commercial Applications
As Enlightenment rationalism grew, dowsing persisted in practical settings, often in rural areas where “water witches” were esteemed for their accuracy. In the early 18th century, German mining handbooks described using rods over maps to determine vein depths, directions, and profitability without fieldwork, yielding actionable insights for operations. By the 19th century, commercial firms like John Mullins & Son in England dowsed for water, ores in Cornwall and California, and mineral oil, serving aristocrats and businesses with documented successes that built their reputation. In the 1950s, a Somerset farmer and water board official used birch twigs to locate an underground reservoir, resulting in a productive well. Such accounts demonstrate dowsing’s resilience, thriving on observable results even as skeptics, much like those doubting microscopic evidence before Antonie van Leeuwenhoek’s refinements, dismissed it without engaging the data.
20th Century to Present: Military, Industrial, and Scientific Validation
The modern era has seen dowsing integrated into military and industrial operations, with verifiable successes amplifying its historical legacy. During World War II, General George Patton employed dowsers to locate fresh water in North Africa after German sabotage, sustaining his troops in harsh deserts. In the Vietnam War, U.S. Marines, trained by engineer Louis Maticia, used bent coat hangers to detect tunnels, booby traps, ammunition caches, and enemy supplies, saving lives as reported in contemporary accounts. Government agencies, including the U.S. Bureau of Land Management and Soil Conservation Service, have utilized dowsing for resource location.
In industry, pharmaceutical giant Hoffman-La Roche relied on dowsing to find large water supplies for drug production plants, with Dr. Peter Treadwell affirming its profitability. Oil engineer Paul Clement Brown successfully located wells for companies like Standard Oil and Mobil using dowsing. A 1960s government project in Ceylon (now Sri Lanka) used dowsing to survey graphite deposits more reliably than geophysical methods, as noted by engineering experts. In 1968, civil engineers on the M61 motorway in England dowsed to mark drain positions, confirmed by excavation.

2 of 3

Contemporary studies further bolster this record. Physicist Hans-Dieter Betz’s decade-long GTZ-funded research in arid regions like Sri Lanka achieved a 96% success rate in 691 dowsings for water wells, far surpassing the 30-50% expected from conventional geology alone, with accurate predictions of depths and yields. In Germany, three dowsers consistently revealed viable water sources. The Munich Experiments in the 1980s also showed positive correlations for skilled dowsers in locating items.
This vast historical tapestry—from ancient survival tools to modern resource strategies—presents irrefutable empirical evidence that dowsing works, demanding exploration rather than avoidance. As empirical scientists, we recognize that phenomena like gravity or quantum effects were utilized long before mechanisms were understood; dowsing merits the same investigative rigor, lest skeptics, echoing pre-microscope obstructionists, impede discoveries in subtle human-environment interactions.

Read 4 tweets

Brian Roemmele

@BrianRoemmele

Nov 5

I need you to sit down for a moment and fully understand this:

THE COUNTRY THAT SOLVES AI AND ROBOTICS WILL RULE THE WORLD AND SPACE.

The west really only has @elonmusk, the east has 100s of companies fortified by an ENTIRE COUNTRY.

Now some will argue no this is not true. No US company has the scale to MANUFACTURE, the compute power, and the finances to compete.

Just a few hours ago we saw IRON for the first time, now you will.

IRON, a 5’10, 150-lb AI humanoid robots are already building EV cars on the XPENG Motors factory floor.

It has over 60 joints, a human-like spine, facial expressions, and male/female customizations

The gait of this robot is the most human-like ever seen.

Mass rollout in 2026.

It is a very big deal. Because as the west does the best in clubbing each other over its head the last decade, China has looked and laughed and built at scale with a fortified government that has little diversion of goal, a 1000 year plan. The west has quietly plans and layers of lawyers and politicians.

This is about where YOU LIVE and how you want to live.

So when we kick the one person that is Atlas carrying our chance, in the groin, you make a choice on who’s world view you want.

It is that simple. No, it is that simple.

It ain’t no iPhone it is: whose’s world view will sustain.

I can say no one is ready for what I have seen that is up for the next few years.

You will think my bombastics were too tame.

Here is the stuff we do in the west with tech. You know for likes and giggles on TikTok. Go to sleep western persons, go to sleep, follow the funny robots and go to sleeeeeeep…

https://twitter.com/brianroemmele/status/1985936552414044494

There is more to this. Here is one element.

https://twitter.com/brianroemmele/status/1985936552414044494

Read 4 tweets

Brian Roemmele

@BrianRoemmele

Oct 20

BOOOOOOOM!

CHINA DEEPSEEK DOES IT AGAIN!

An entire encyclopedia compressed into a single, high-resolution image!

—

A mind-blowing breakthrough. DeepSeek-OCR, unleashed an electrifying 3-billion-parameter vision-language model that obliterates the boundaries between text and vision with jaw-dropping optical compression!

This isn’t just an OCR upgrade—it’s a seismic paradigm shift, on how machines perceive and conquer data.

DeepSeek-OCR crushes long documents into vision tokens with a staggering 97% decoding precision at a 10x compression ratio!

That’s thousands of textual tokens distilled into a mere 100 vision tokens per page, outmuscling GOT-OCR2.0 (256 tokens) and MinerU2.0 (6,000 tokens) by up to 60x fewer tokens on the OmniDocBench.

It’s like compressing an entire encyclopedia into a single, high-definition snapshot—mind-boggling efficiency at its peak!

At the core of this insanity is the DeepEncoder, a turbocharged fusion of the SAM (Segment Anything Model) and CLIP (Contrastive Language–Image Pretraining) backbones, supercharged by a 16x convolutional compressor.

This maintains high-resolution perception while slashing activation memory, transforming thousands of image patches into a lean 100-200 vision tokens.

Get ready for the multi-resolution "Gundam" mode—scaling from 512x512 to a monstrous 1280x1280 pixels!

It blends local tiles with a global view, tackling invoices, blueprints, and newspapers with zero retraining. It’s a shape-shifting computational marvel, mirroring the human eye’s dynamic focus with pixel-perfect precision!

The training data?

Supplied by the Chinese government for free and not available to any US company.

You understand now why I have said the US needs a Manhattan Project for AI training data? Do you hear me now? Oh still no? I’ll continue.

Over 30 million PDF pages across 100 languages, spiked with 10 million natural scene OCR samples, 10 million charts, 5 million chemical formulas, and 1 million geometry problems!.

This model doesn’t just read—it devours scientific diagrams and equations, turning raw data into a multidimensional knowledge.

Throughput? Prepare to be floored—over 200,000 pages per day on a single NVIDIA A100 GPU! This scalability is a game-changer, turning LLM data generation into a firehose of innovation, democratizing access to terabytes of insight for every AI pioneer out there.

This optical compression is the holy grail for LLM long-context woes. Imagine a million-token document shrunk into a 100,000-token visual map—DeepSeek-OCR reimagines context as a perceptual playground, paving the way for a GPT-5 that processes documents like a supercharged visual cortex!

The two-stage architecture is pure engineering poetry: DeepEncoder generates tokens, while a Mixture-of-Experts decoder spits out structured Markdown with multilingual flair. It’s a universal translator for the visual-textual multiverse, optimized for global domination!

Benchmarks? DeepSeek-OCR obliterates GOT-OCR2.0 and MinerU2.0, holding 60% accuracy at 20x compression! This opens a portal to applications once thought impossible—pushing the boundaries of computational physics into uncharted territory!

Live document analysis, streaming OCR for accessibility, and real-time translation with visual context are now economically viable, thanks to this compression breakthrough. It’s a real-time revolution, ready to transform our digital ecosystem!

This paper is a blueprint for the future—proving text can be visually compressed 10x for long-term memory and reasoning. It’s a clarion call for a new AI era where perception trumps text, and models like GPT-5 see documents in a single, glorious glance.

I am experimenting with this now on 1870-1970 offline data that I have digitalized.

But be ready for a revolution!

More soon.

[1] github.com/deepseek-ai/De…

The US needs far more than what we are doing. We need to reach out to the nonconformists, not the usual folks to break out in AI for this next rounds we have no time.

I have a few 1000 ways, here is one:

readmultiplex.com/2025/10/19/the…

https://twitter.com/brianroemmele/status/1980336008748233094

Testing now!

https://twitter.com/brianroemmele/status/1980336008748233094

Read 6 tweets

Share this page!

Enter URL or ID to Unroll

Brian Roemmele

Try unrolling a thread yourself!

More from @BrianRoemmele

Brian Roemmele

Brian Roemmele

Brian Roemmele

Brian Roemmele

Brian Roemmele

Brian Roemmele

Did Thread Reader help you today?

Don't want to be a Premium member but still want to support us?

Send Email!