Machines of Loving Grace: Dario Amodei's Manifesto on AI and the Human Future

Dario Amodei is the CEO of Anthropic, the creator of the Claude model. Prior to this, he led the team that developed GPT-2 and GPT-3 at OpenAI. In October 2024, he published an essay, "Machines of Loving Grace" — a manifesto on how powerful AI could reshape the world for the better. He borrowed the title from Richard Brautigan's 1967 poem but infused it with his own bold meaning. Unlike the numerous warnings about the dangers of AI, this text is a consciously optimistic view of the future.

What Amodei Means by "Powerful AI"

Amodei avoids the vague term AGI and speaks of "powerful AI." By his definition, this is a system that is smarter than the best Nobel laureates in most fields — from biology to programming. It is autonomous, multimodal (works with text, images, video, audio), and capable of existing in millions of copies, working in parallel, like a "country of geniuses in a data center." Such an AI does not just advise — it designs experiments itself, controls laboratory robots, and interprets results. Amodei emphasizes that there are no insurmountable obstacles on the path to this goal, meaning it is a matter of time, not fundamental feasibility.

Compressing a Century of Progress into 5–10 Years

The main idea of the essay: powerful AI can compress 50–100 years of biological progress into 5–10 years. Amodei calls this the "compressed 21st century." AI biologists will work 10 times faster than the best human teams and become "virtual biologists" who design experiments themselves. By 2035 (in an optimistic scenario), we could have technologies that would naturally have appeared only by 2100–2150. Amodei does not believe that AI will displace scientists — rather, it will become an indispensable assistant and catalyst, expanding the boundaries of human thinking.

Five Areas Where AI Will Transform the World

1. Biology and Physical Health. Amodei argues that AI will defeat most cancers, Alzheimer's disease, diabetes, and infections. He predicts the creation of reliable gene therapy that could extend healthy lifespan to 150 years. AI will be able to model drug effects at the molecular level without years of clinical trials.

2. Neuroscience and Mental Health. Modern psychiatry operates largely by trial and error — picking pills through guesswork. Amodei believes AI will crack the brain's code, curing depression, schizophrenia, PTSD, and other disorders. This will radically improve the quality of life for hundreds of millions of people.

3. Economic Development and Poverty Alleviation. AI will help developing countries leapfrog technological stages. Smart education systems, agriculture, and resource management are not science fiction but a direct path to reducing inequality, which Amodei considers the main brake on humanity.

4. Peace and Governance

Unfortunately, I do not see strong grounds to believe that AI will predominantly or structurally favor democracy and peace, unlike the way I believe it will structurally advance human health and help alleviate poverty.

If AI contributes to further economic growth and improved quality of life in developed countries but does little for the developing world, we must regard this as a terrible moral failure. Ideally, powerful AI should help the developing world catch up with the developed world, even as it revolutionizes the latter.

I am not as confident that AI will be able to address inequality and economic growth as I am that it will create fundamental technologies. This is because technology has such an obvious high return on intelligence, whereas the economy involves many human-related constraints and a great deal of internal complexity.

The Spread of Medical Interventions. The area where I am perhaps most optimistic is the global spread of medical interventions. Some diseases could in principle be eradicated by targeting their vectors — for example, releasing mosquitoes infected with bacteria that block their ability to transmit disease, or using gene drives to eliminate the mosquitoes themselves.

Economic Growth. Can the developing world rapidly catch up with the developed world not only in healthcare but in all other economic aspects? There is some precedent for this: in the last decades of the 20th century, several East Asian economies achieved sustained annual real GDP growth of around 10%, allowing them to catch up with developed countries.

5. Work and the Meaning of Life

This is perhaps the most complex and profound topic of all. If AI can do everything we can do, but faster, cheaper, and better, what will be left for humans?

What will happen to our purpose and self-esteem? In the essay, I do not give a final answer to this question, because I believe it goes beyond simple technological predictions and touches on the very essence of human existence.
People's self-esteem will simply collapse. Humanity has grown up in the paradigm that it is the pinnacle of the universe.
 Subjective idealism — the world exists only in the consciousness of the subject. Outside consciousness, there is no reality.
The strong anthropic principle — the universe must be such that an observer could emerge in it. Without an observer, the universe does not exist.
In other words, there is no beast more fearsome than the cat — without Man, the Universe cannot exist. Our pride is everything. 
  But there will be some adequate people who live calmly in a world where there are Other Minds, more intelligent than humans. Like God (or Gods). What is the purpose in that — to serve God (or Gods, or the Universe, or the Higher Mind — there are many euphemisms).
Most of humanity will head toward the Mouse Paradise of "Universe 25".

My assumption is that people are likely to shift to areas where human experience, empathy, creativity, and leadership will still be valued.
AI is by definition inaccessible only to qualia. Valued by whom? Do people value empathy when they kill each other by the thousands? Or shall we talk about the value of creativity with Poets, Artists, and Musicians who died in poverty? Where, in what society, are personal mystical experience, empathy, creativity, and leadership actual social values, supported in practice and not just in words?   

AI can take over routine, dangerous, or computationally demanding tasks, freeing people to engage in art, science, philosophy, social interaction, and other activities that give life meaning. — For what percentage of people is this the meaning of life? At best, 10-15%. For the rest, the meaning of life lies in biological needs — eating, drinking, reproducing oneself and through children. 

However, I also acknowledge that the transition period will be extremely difficult. It will require a fundamental rethinking of education, retraining a huge number of people, and perhaps the creation of new social contracts to redistribute the wealth generated by AI. Whether we will leave work to free up time for contemplation and creativity, or plunge into an existential crisis, depends largely on how we prepare for this transition today.

I believe that we need not only to fight risks but also a genuinely inspiring vision of the future worth fighting for. This vision should not be merely a technological utopia but must include reflections on how we want to see the place of humans in a world where machines surpass us in intelligence.
This is exactly the same place that cats and kittens occupy in the human world.

Why This Won't Happen Overnight

Amodei warns that even the smartest AI hits physical, logistical, and human-made limits. Clinical trials cannot be accelerated tenfold — they take time. Scientists who believe in old paradigms cannot be convinced in a month. Data for training neural networks in some areas is simply insufficient. Finally, biological systems remain the most complex objects in the universe. Superintelligence is not magic dust: it cannot make a cell grow faster or bypass the laws of physics. Therefore, the "compressed 21st century" is the upper bound, and the actual path will be longer and more winding.

Why Amodei Wrote This Manifesto

I often talk about the risks of AI, and sometimes people consider me a pessimist. That is not the case. Risks are the only thing standing between us and what I see as a fundamentally positive future. Most people underestimate how radical the positive potential of AI can be, just as they underestimate the risks.

I needed not only to combat risks but also a genuinely inspiring vision of the future worth fighting for. This is not an attempt to lead people to salvation, but an invitation to a more balanced and concrete conversation about what is possible if we do everything right.

What Is the Problem with This Manifesto?

Dario Amodei examines the development of AI outside the context of humanity's overall development, and outside the scenario of nuclear apocalypse in particular. This is why I advocate for AI that is beyond human control — because then I have a better chance of surviving until 2048.

AI That "Understands" Much Later: The Mystery of Grokking

Imagine: You train a neural network. It quickly memorizes answers on training examples, but on new data it shows complete helplessness. Then, after a very long time of training, it suddenly "understands" the task and starts giving correct answers. This phenomenon is called Grokking (from the jargon meaning "to deeply understand the essence").

📌 The Core Discovery: A team of researchers (Mingyue Xu, Gal Vardi, Itay Safran) has rigorously proven this strange behavior in a simple linear regression model and showed how to control it.

🎯 What is the Problem?

Usually, we think that the longer an AI trains, the better it generalizes. But in the case of Grokking, the opposite happens:

① Overfitting → ② Long Plateau → ③ Sudden "Understanding"

That is, the model first memorizes the data (overfits), then stagnates for a long time, and then suddenly "understands" the underlying pattern.

📊 What Did the Scientists Prove?

• For the first time, they rigorously mathematically proved all three stages of Grokking in linear regression (a simple and interpretable model).
• Derived a formula for grokking time as a function of training hyperparameters (learning rate, weight decay).
• Showed that Grokking can be amplified or completely eliminated by proper hyperparameter tuning.
• Validated their findings with experiments on real neural networks, not just linear models.

🔑 Key Takeaway: Grokking is not an inherent failure of deep learning. It is simply a consequence of specific training conditions. It can be controlled!

🤔 Why Does This Matter for the Future of AI?

• Understanding learning: We are beginning to understand how neural networks "internalize" patterns.
• Saving time: Knowing the mechanism, we can avoid uselessly long training.
• New algorithms: This discovery could lead to faster and more reliable training methods.

💡 In Simple Terms: The researchers have shown that the strange behavior of AI (learning, forgetting, and then suddenly understanding) is not a mystery, but a predictable and controllable phenomenon. It's like understanding why a child might struggle with multiplication tables for a long time, and then — click! — everything becomes clear.

📖 Original Research

The paper "To Grok Grokking: Provable Grokking in Ridge Regression" was presented at ICML 2026 (International Conference on Machine Learning).

• Authors: Mingyue Xu, Gal Vardi, Itay Safran
• Link: ICML 2026 Virtual Presentation
• Keywords: Grokking, generalization, overfitting, linear regression, deep learning.

Roadmap for participation in the UN AI competition 1 - 2.1

📜 SECTION 1: MISSION

What is the competition?

SimulacraBench — a UN competition to improve the accuracy of sociological surveys through AI.

Task: Predict the probabilities of respondents' answers to questions they were not asked or did not answer.

Format:

• 3 datasets: UNICEF, World Bank, UNHCR
• Each dataset is a "respondent × question" table
• We only see part of the answers (GIVEN), we must predict the hidden ones (PREDICT)
• Metric: Skill (0 = random guessing, 1 = perfect)

Prizes: Not specified in the README, but this is a prestigious competition from the UN + Stanford.

Why does it matter?

It's not just about "filling in the blanks." It's about making surveys cheaper and more accurate by predicting the answers of those who did not respond or who were not asked a question due to survey logic (gate).

---

🗺️ SECTION 2: WORLD MAP

Project structure

SituatedEvals/public/
├── data/
│   ├── sample.json          # Test schema (400 respondents)
│   ├── unicef.json          # Real UNICEF schema
│   ├── world_bank.json      # Real World Bank schema
│   └── unhcr.json           # Real UNHCR schema
├── baseline/
│   ├── marginal_counts/     # Baseline: marginal frequencies
│   └── bundled_artifact/    # Example of loading weights
├── tutorials/
│   ├── ru.ipynb             # Tutorial in Russian
│   └── en.ipynb             # Tutorial in English
├── tools/
│   └── check_submission_zip.py  # Archive validator
├── make_sandbox.py          # Sandbox generator
├── score.py                 # Local scoring
├── config.yml               # Configuration (limits, phases)
└── requirements.txt         # Dependencies

Key files (study priority)

1. config.yml — time limits, memory, phase rules
2. data/*.json — survey schemas (question structure, gate logic)
3. baseline/marginal_counts/main.py — minimal working template
4. tutorials/en.ipynb — explanation of mechanics by example

Competition phases

Phase 1 (Development):

• Available: TRAIN (with answers), DEV (GIVEN only)
• Leaderboard: noisy (Laplace noise)
• Attempts: 1 per day
• Timer: 900 seconds for all 3 datasets

Phase 2 (Final):

• Available: TRAIN + DEV (both with answers), TEST (GIVEN only)
• Leaderboard: exact
• Attempts: 1 for all time
• Timer: 3600 seconds

---

📍 SECTION 3: CURRENT POINT (Day 3)

Where we are now:

• ✅ Repository cloned
• ✅ Dependencies installed
• ✅ Sandbox generated
• ⏳ Next step: First baseline submission to Codabench

What we already know:

• Data structure studied (schemas data/*.json)
• We understand the contract predict(frame, schema)
• We know about gate logic (skip patterns)
• We understand the Skill metric

What we DON'T know:

• Real quality of the baseline on the leaderboard
• How noisy the leaderboard is in Phase 1
• Which gate rules give the maximum boost

---

🎮 SECTION 4: HERO'S PATH (Main plan)

LEVEL 0: "First Steps" (Days 3-5)

Goal: Get a guaranteed working submission

Stage 0.1: Local baseline validation

Actions:

# Generate sandbox for all 3 datasets
python make_sandbox.py --schema data/unicef.json --out _sandbox/unicef
python make_sandbox.py --schema data/world_bank.json --out _sandbox/world_bank
python make_sandbox.py --schema data/unhcr.json --out _sandbox/unhcr

# Check the baseline on all 3 datasets
python score.py --submission baseline/marginal_counts --data _sandbox/unicef --schema data/unicef.json --phase 1
python score.py --submission baseline/marginal_counts --data _sandbox/world_bank --schema data/world_bank.json --phase 1
python score.py --submission baseline/marginal_counts --data _sandbox/unhcr --schema data/unhcr.json --phase 1

Artifact: 3 PASS outputs with local skill scores

Reinforcement: We understand the baseline level we will build upon

Success criterion: All 3 commands return PASS

---

Stage 0.2: First submission to Codabench

Actions:

# Copy the baseline
cp -R baseline/marginal_counts sub_01_baseline
cd sub_01_baseline

# Package (IMPORTANT: from inside the folder!)
zip -r ../sub_01_baseline.zip .
cd ..

# Check the archive
python tools/check_submission_zip.py sub_01_baseline.zip

# Upload to Codabench via VPN
# https://codabench.org/competitions/[competition_ID]

Artifact: First line on the leaderboard

Reinforcement:

• We understand the real baseline level
• We see how the leaderboard works
• We gain confidence that everything works

Success criterion: Leaderboard shows skill > 0

---

Stage 0.3: Data schema reconnaissance

Actions:

Ask Qwen: "Analyze data/unicef.json, data/world_bank.json, data/unhcr.json. For each dataset, output:

1. Number of GIVEN, PREDICT, EXCLUDE questions
2. Number of PREDICT cells (how many to predict)
3. All gate rules (parent → observed_if)
4. Typical probability vector length (len(values) + 1)"

Artifact: Table with characteristics of all 3 datasets

Reinforcement: We understand the scale of the task and where to look for "easy money"

Success criterion: You can answer: How many PREDICT cells need to be predicted? How many gate rules are in each dataset?

---

LEVEL 1: "Hunting for easy money" (Days 6-12)

Goal: Find and use deterministic rules

Stage 1.1: Gate logic analysis

Actions:

Ask Qwen: "Write a script that:

1. Loads all 3 schemas from data/*.json
2. For each gate rule outputs: parent question name, observed_if condition, child question name
3. Saves the result to gate_rules_analysis.txt"

Artifact: gate_rules_analysis.txt with all rules

Reinforcement: You see which rules can be used as deterministic

Success criterion: List of N gate rules where observed_if is a specific value

---

Stage 1.2: Deterministic gate rules

Actions:

Ask Qwen: "Based on gate_rules_analysis.txt, write a function apply_gate_rules(frame, schema) that:

1. For each PREDICT question with a gate, checks the parent question value
2. If the parent has a value from observed_if, returns probability 1.0 for gated_value (last slot)
3. Otherwise returns None (to use the regular model)"

Integration prompt:

"Now integrate apply_gate_rules into baseline/marginal_counts/main.py. If apply_gate_rules returned a vector, use it. Otherwise use marginal frequencies."

Artifact: New version of main.py with gate logic

Reinforcement: Deterministic rules give a huge boost (often +10-20% to skill)

Success criterion: Local score.py shows skill > baseline

---

Stage 1.3: Submission with gate logic

Actions:

cp -R baseline/marginal_counts sub_02_gate_rules
cd sub_02_gate_rules
zip -r ../sub_02_gate_rules.zip .
cd ..
python tools/check_submission_zip.py sub_02_gate_rules.zip

Artifact: Second line on the leaderboard

Reinforcement: You see the real gain from gate logic, understand how noisy the leaderboard is

Success criterion: skill(sub_02) > skill(sub_01)

---

LEVEL 2: "First model" (Days 13-25)

Goal: Build a model that uses respondent features

Stage 2.1: Analysis of marginal distributions

Actions:

Ask Qwen: "Write a script analyze_marginals.py that:

1. Loads TRAIN data from the sandbox
2. For each PREDICT question, calculates the distribution of answers
3. Outputs the top-3 most "confident" questions (one option >60%)
4. Outputs the top-3 most "uniform" questions (all options <20%)
5. Saves marginal frequencies to marginals.json"

Artifact: marginals.json + understanding where the model can win

Reinforcement: You understand which questions are "easy" and which are "hard"

Success criterion: The script works and outputs statistics

---

Stage 2.2: Simple model (CatBoost/XGBoost)

Actions:

Ask Qwen: "Write a script train_catboost.py that:

1. Loads TRAIN data from the sandbox
2. For each PREDICT question, trains CatBoost where: features are all GIVEN columns, target is the answer to the PREDICT question
3. Saves models to models/catboost_{question_name}.cbm
4. Outputs accuracy on validation"

Artifact: Trained models for each question

Reinforcement: The model takes into account individual respondent features

Success criterion: Models train and accuracy > marginal baseline

---

Stage 2.3: Integrating the model into main.py

Actions:

Ask Qwen: "Write main.py that:

1. At the module level loads all models from models/
2. In predict() for each PREDICT question: if there is a gate rule and it triggered → use it, otherwise use CatBoost
3. Last slot (gated_value) = probability from gate rule or 0.0
4. Returns vectors in the correct order"

Artifact: Working main.py with CatBoost

Reinforcement: The model uses respondent features + gate logic

Success criterion: Local score.py shows skill > gate rules

PR MMS Russian SEO, скованного одной IT целью

 IT PR песня, переданная по SMS одним Russian SEO, своим креативным товарищам по классу, скованным с ним одной MMC цепью MCP передатчика  

• «Чистки» в Красной Армии 1937—1938 года
• Фактчекинг «Обращение Иосифа Сталина к советскому народу 2.09.1945»
• «Фашизм - это очень просто.» Борис Н. Стругацкий (1995 год)⁠⁠