LIVE
US Military Avoids Close Call After AI-Generated Intelligence Hallucination18/09/26|Study finds gender bias in GPT models is not reduced but reshaped across generations17/09/26 · OpenAI|Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|US Military Avoids Close Call After AI-Generated Intelligence Hallucination18/09/26|Study finds gender bias in GPT models is not reduced but reshaped across generations17/09/26 · OpenAI|Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|
Beginner🧠

AGI: when is it really coming? (2026 analysis without hype)

Sam Altman says 'few years', Yann LeCun '20 years min', expert median 2047. We analyzed predictions, benchmarks, 4 major obstacles, 3 scenarios.

16 min readPublished May 7, 2026· Updated September 17, 2026

In one sentence

AGI (Artificial General Intelligence) is the focus of all fears and fantasies in 2026: Sam Altman says "in a few years", Yann LeCun says "20 years minimum", Yoshua Bengio says "5-10 years is plausible". The truth? Nobody knows, and the very definition of AGI is not consensus. But we can look at the facts, benchmarks, and serious predictions to form an informed opinion without falling into hype or denial.

🧠
The analogy that works
AGI is a bit like level 5 autonomous cars (no steering wheel, in all conditions). In 2015, we were promised that for 2020. In 2020, for 2025. In 2026, we're told 2030. Progress is real but the 'last 10%' is 10x harder than the first 90%. Same with AGI: GPT-5 covers 80% of the ground on many tasks, but the remaining 20% (true understanding of the world, authentic creativity, robust causal reasoning) is a technical chasm we don't know how to cross. Not through lack of money or GPUs: through lack of fundamental scientific understanding.

🤖 Before AGI, here's what GPT-5 can already do

Our complete guide to GPT-5: strengths, limitations, pricing.

Read the GPT-5 article

The timeline of AGI predictions: who says what?

AGI predictions from AI leaders in 2026
🔮 When will AGI arrive according to AI leaders? 2026 Now 2030 2035 2040 2050 Expert median 2060 2075+ Cautious Sam Altman OpenAI CEO "few years" Demis Hassabis DeepMind CEO "5-10 years" Yoshua Bengio Turing 2018 "plausible" AI Impacts survey 2,778 researchers median 2047 Yann LeCun Meta Chief AI "20 years min" Note: CEO predictions are biased (need investors). Academic researchers are generally more cautious.
From most optimistic (Sam Altman) to most cautious (Yann LeCun). Opinions range from 4 to 50 years.

But what really is "AGI"?

The problem: there is NO single definition
Here are 5 competing definitions of AGI used in 2026: 1. OpenAI definition (vague, business-focused) > "A highly autonomous system that outperforms humans at most economically valuable work." → Critique: "economically valuable" excludes consciousness, authentic creativity, etc. 2. DeepMind definition (by levels) - Level 1: Emerging AGI (equal to non-specialized human), partially reached in 2024 - Level 2: Competent AGI (top 50% professionals), partially reached in 2026 - Level 3: Expert AGI (top 10% professionals), not yet - Level 4: Virtuoso AGI (top 1%), no - Level 5: Superhuman (above all), no 3. François Chollet definition (ARC-AGI) > "AGI = ability to efficiently acquire new skills when facing novel tasks." → His benchmark ARC-AGI-2 is still far from solved (5% by the best LLMs in 2026). 4. Yann LeCun definition (physical capabilities) > "System that understands the physical world at the level of a domestic cat." → Nobody is there. According to him, this requires changing paradigms (beyond LLMs). 5. Philosophical definition (consciousness) > "System that has a subjective conscious experience of the world." → Off-topic for most: we don't even know how to define consciousness. TL;DR: depending on which definition you choose, we're between "already there" and "never". That's why the debate is so confused.

Benchmarks: where are we in 2026?

AGI benchmarks: human vs LLM 2026 performance
📊 2026 Benchmarks : Human (grey) vs Best LLM (cyan) 0% 25% 50% 75% 100% Image recog. (ImageNet) 94% 99% Speech recog. (LibriSpeech) 93% 98% Code (SWE-Bench) 42% (avg dev) 78% (Claude Opus) Math olympiad (AIME) 85% (top 5%) 90% (o3) Knowledge (MMLU) 90% (PhD) 89% Abstract reasoning (ARC-AGI-2) 80% (avg human) 5% (massive fail) Physical world (CommonsenseQA) 95% 73% (gap) Continual learning 100% ~0% (major flaw) Human Best LLM 2026
On some tasks, LLMs surpass humans. On others (ARC-AGI), it's still a disaster.

The 4 MAJOR obstacles to AGI

📚Why it's so hard to go from GPT-5 to AGI

1. 🌍 No model of the physical world

Ask GPT-5: "I'm holding a glass of water. If I turn it upside down, what happens?" → it answers well (text). Ask it to mentally simulate the trajectory of the water and predict where it will land → fails.

GPT-5 learned from text about the world, not from the world itself. Yann LeCun calls this the inverted "Moravec paradox": what's easy for a 3-year-old child (physical intuition) remains inaccessible to the best LLMs.

Potential solution: world models (DeepMind Genie 3, World Labs Marble) that learn from video/simulation. Still experimental.

2. 🔄 No continual learning

You ask GPT-5 a complex question. It makes a mistake. You explain the error. It learns it for this conversation (in-context learning), but FORGETS EVERYTHING in the next conversation.

A human who makes a mistake learns permanently. LLMs, no. Everything must go through costly re-training (months, millions of dollars).

Potential solution: neuromorphic computing, online learning. No major breakthrough in 2026.

3. 🧩 Causal reasoning flaw

GPT-5 is very good at correlations (X often occurs with Y). But mediocre at causation (X causes Y).

Example: "There's a correlation between ice cream sales and drownings. Conclusion?" → GPT-5 can say "common cause (summer)". But on new real cases, it invents causal links that are plausible but false.

This is what Judea Pearl (Turing 2011) calls the "ladder of causation": seeing > intervening > imagining. LLMs remain stuck at the first level.

4. 🎯 ARC-AGI-2: the test no LLM passes

Invented by François Chollet (creator of Keras), ARC-AGI measures the ability to solve NEW visual puzzles with very few examples, exactly what a 5-year-old human does easily.

| Model | ARC-AGI-1 | ARC-AGI-2 | |--------|-----------|-----------| | Average human | ~80% | ~80% | | Expert human | ~98% | ~85% | | GPT-4 (2023) | 9% | <1% | | GPT-5 (2025) | 75% | 5% | | o3-preview | 87% | 12% |

→ On ARC-AGI-2, the best LLMs are still under 15%. This is the wall that nobody knows how to cross.

Why it matters: if a system can't solve puzzles that children do, how can it claim to be AGI?

The 3 credible scenarios 2026-2050

3 scenarios for AGI arrival

 🔮Scenario📊Probability (estimate)
🚀 FAST SCENARIO (2030-2035)AI breakthrough enables crossing the wall of causal reasoning and continual learning. Combination of LLM + world models + RL. 'Competent' level AGI achieved.10-20%
🚶 MODERATE SCENARIO (2040-2050)Incremental progress. No single breakthrough but accumulation. LLM + agents + multimodal + memory become robust enough for 'Competent AGI'. Not truly 'general' but hard to distinguish.40-50%
🐢 SLOW SCENARIO (2060+)LLM plateau reached. Need a new paradigm (neuromorphic, quantum, other). Several decades of fundamental research needed.20-30%
❌ 'NEVER' SCENARIOAGI is impossible with current architecture, and no new paradigm works. Minority hypothesis but seriously defended (Yann LeCun, Gary Marcus).10-20%

My analysis of the analysis:

  • The moderate scenario (2040-2050) is most probable according to the majority of academic experts
  • CEOs (Altman, Hassabis) lean towards fast (commercial interest in hype)
  • Nobody serious says "never before 2100"
  • Nobody serious says "in 2 years" without marketing quotes

What should YOU do in 2026?

🎯
No-regret strategy facing AGI uncertainty
Whatever scenario materializes, these actions are winning: 1. Learn to use AI in your job If AGI arrives in 2035, you'll have 9 years of pre-AGI experience = massive advantage. If AGI arrives in 2060, you'll have gained productivity for 30 years = career transformed. 2. Invest in "moat" skills Protected professions until AGI: - Soft skills (negotiation, leadership) - High-end creativity (artists, brands) - Skilled physical trades (surgeon, plumber) - Legal responsibility (senior lawyer, doctor) 3. Build economic assets Own assets (real estate, stocks, business) rather than depending 100% on your salary. If AGI massively disrupts employment, capital owners are better protected. 4. Train continuously 3-5 hours/week on AI and new trends. Newsletter, courses, experimentation. 5. Keep critical thinking When a CEO says "AGI in 2 years", ask yourself who benefits. Same when a detractor says "it will never happen". The truth is usually between the two.

The question of danger: should you be afraid?

The 3 serious risks (and the 3 that are fantasy)
🚨 Serious risks (short-medium term): 1. Massive employment disruption (already underway, 8% jobs at risk) 2. Industrial-scale disinformation (deepfakes, automated propaganda) 3. Extreme concentration of power (5-10 companies control global AI) 🎬 Movie fantasies (unlikely before decades): 1. Skynet / conscious AI that rebels: we don't even know what consciousness is, much less how to create it involuntarily 2. AGI that turns matter into paperclips ("paperclip maximizer"): assumes AGI AND recursive self-improvement AND absence of safeguards, very low probability scenario 3. Singularity before 2030: predicted since 1965, always "in 10 years". Take it like fusion: "30 years" forever. Reasonable position: - Take short-term risks seriously (regulation, security, redistribution) - Don't neglect the possibility of AGI but don't panic either - Prepare for both extreme scenarios like taking insurance (low probability, big impact)

The metaphor that sums it all up

🏔️
AGI is Everest when you don't know where it is
Imagine you want to climb Everest, but: - 🗺️ Nobody knows exactly where it is (no consensus definition of AGI) - 🌫️ A permanent fog masks the summit (we've never seen AGI) - ⛰️ You're making good progress on some slopes (math, code, language), you doubt on others (causal reasoning, physics) - 👥 Some climbers tell you "we'll get there tomorrow" (they're selling the expedition), others "it's inaccessible" (they stayed at base camp) - 🧗 The final ascent (the remaining 20%) is infinitely harder than the 80% already covered The truth? We're climbing. Fast. But we don't know how much remains. Nobody. Not even Sam Altman. What we know for certain: continuing to climb changes the world. Whether you believe it or not, prepare. The summit is just a stage, the journey is what transforms societies.

Absolutely remember

  • AGI = no consensus definition, so debate sometimes confused
  • GPT-5 / Claude / Gemini beat humans on certain specific tasks (chess, code, math), not on everything
  • 4 major obstacles: physical world, continual learning, causal reasoning, generalization (ARC-AGI-2)
  • Academic expert median: AGI around 2040-2047 (50% chance)
  • Sam Altman says "a few years" but it's marketing for investors
  • Yann LeCun says "20 years min" because LLMs can't reach AGI according to him
  • No-regret strategy: learn AI + level up + build your "moats" + invest (real estate/stocks)
  • Serious risks: employment, disinformation, concentration of power, not Skynet

Nobody knows when AGI will arrive, or even if it will arrive. What's certain: current progress is already enough to transform the world. Focus on how to adapt today, not on tomorrow's fantasies.

🧠 Quiz
Question 1 of 3

What is the median AGI prediction according to the AI Impacts survey of 2,778 AI researchers?

To go further

Tags
AGIAvenir IAPrédictionsIA & Société

Read next