The Road to AGI

AlphaZero co-creator David Silver launched London-based Ineffable Intelligence and raised $1.1 billion at a $5.1 billion valuation to build what he calls a 'superlearner' — an AI system that acquires capability through reinforcement learning rather than human-generated training data. The architectural bet is a deliberate move away from the LLM paradigm of pre-training on internet text, extending Silver's own DeepMind research thesis (from AlphaZero through MuZero) that self-play and reward-based learning can reach superhuman performance without imitating human examples. If the approach scales beyond narrow domains, it would reframe a central assumption of the road to AGI — that capability is bounded by the breadth and quality of available human data. The round, led by Sequoia and Lightspeed with Google, Nvidia and the UK Sovereign AI fund participating, makes Ineffable one of the largest single bets ever placed on a non-LLM path to general intelligence.TechCrunch ↗

DeepSeek published V4-Flash and V4-Pro in preview on 24 April 2026, its first major model drop since R1. V4-Flash is a 284-billion-parameter mixture-of-experts model with 13 billion active parameters, a 1-million-token context window, and open weights on Hugging Face. It scores 88.4% on MMLU and 91.6% on LiveCodeBench — within two points of the larger V4-Pro — while the API charges $0.14 per million input tokens and $0.28 per million output tokens, cheaper than GPT-5.4 Nano. The release continues DeepSeek's pattern of shipping frontier-adjacent capability with sharply lower pricing than closed US competitors.Simon Willison review ↗

In an essay for The Ideas Letter, Bright Simons argues that the intelligence of large language models is not a property of architecture or compute alone but a compressed reflection of the social complexity of the civilisation that produced their training data. As an illustration, he contrasts a hypothetical model trained on 3000 BC Egypt — which would lack syllogistic reasoning — with one trained on 300 BC Athens, which would gain logical inference, and so on through history. He links this to recent work on AI-assisted writing showing that individuals produce more creative outputs but populations converge on similar ones, to Shumailov et al.'s 2024 Nature paper on model collapse from recursively generated data, and to Andrew Peterson's research on knowledge collapse. The implication for the road to AGI is structural: as firms — IBM, Klarna, Duolingo, Atlassian, Block among them — replace humans with AI to cut headcount, they thin out the very social reasoning that future training corpora depend on. Simons predicts the most successful organisations of the coming decade will counterintuitively use AI to generate more, not less, human interaction; he notes IBM has already begun reversing its earlier redundancy stance.The Ideas Letter ↗

OpenAI released GPT-5.4 with a 1-million token context window, unifying previously separate reasoning, coding, and general capabilities into a single model. The system demonstrated 33% fewer factual errors and improved agentic workflow completion.

METR, the research group that tracks how long an AI can stay productive on a task without supervision, added Claude Opus 4.6 to its time-horizons benchmark with a 50%-reliability estimate of ~14.5 hours (95% CI 6h–98h). The series since 2022 reads: GPT-3.5 ~30 seconds, GPT-4 ~4 minutes (2023), o1 ~40 minutes (2024), GPT-5.2 ~6 hours (2025), Opus 4.6 ~14.5 hours (Feb 2026). METR forecaster Ajeya Cotra has said it isn't unreasonable to expect ~100 hours by end of 2026, though METR cautions that measurements above ~16h are unreliable on its current saturated task suite. The metric matters because delegation requires a unit of independent work that matches a human researcher's session length. Once an AI can hold a 12+ hour task without re-anchoring, the dominant pattern shifts from assistance to autonomy — directly relevant to Jack Clark's argument that the engineering side of AI R&D is now mostly delegable.METR ↗

Anthropic released Claude Opus 4 and Sonnet 4, models capable of sustained, multi-hour agentic work — coding, research, and analysis across complex, multi-step tasks.

Chinese lab DeepSeek released R1, an open-source reasoning model that matched proprietary frontier models at a fraction of the training cost. The model demonstrated strong chain-of-thought reasoning and was released with full weights.

Anthropic released Claude 3.5 Sonnet with the ability to see, understand, and control a computer screen. For the first time, an AI could autonomously navigate software, fill forms, write documents, and execute multi-step workflows across applications.

OpenAI released o1, a model that spends more time "thinking" before answering. Using reinforcement learning to develop internal reasoning chains, o1 achieved state-of-the-art results on mathematics, coding, and scientific reasoning benchmarks.

India surpassed the US and China in the number of AI and machine learning engineers, with its IIT system and tech ecosystem producing more AI practitioners than any other country.

OpenAI released GPT-4, a multimodal model that could process text and images. It passed the bar exam in the 90th percentile, scored in the top percentiles on AP exams, and demonstrated reasoning abilities that led some researchers to publish papers about "sparks of AGI."

Three former Google DeepMind and Meta researchers founded Mistral AI in Paris. Within months, they released Mixtral 8x7B, an open-source mixture-of-experts model that rivalled GPT-3.5.

OpenAI released ChatGPT, a conversational interface to GPT-3.5. It reached 1 million users in 5 days and 100 million in 2 months — the fastest-growing application in history.

DeepMind researchers published findings showing that most large language models were significantly undertrained. Their "Chinchilla" model, with 70 billion parameters trained on 1.4 trillion tokens, outperformed the 280-billion parameter Gopher.

AI21 Labs released Jurassic-1, a 178-billion parameter model rivalling GPT-3, built in Tel Aviv. Israel — with more AI startups per capita than any other nation — demonstrated disproportionate impact in AI development.

DeepMind's AlphaFold 2 solved the 50-year-old protein folding problem, predicting 3D structures of proteins to near-experimental accuracy. It later predicted the structure of nearly every known protein — over 200 million structures.

OpenAI released GPT-3, a 175-billion parameter model that could perform tasks it was never explicitly trained for — translation, code generation, arithmetic — simply from a few examples in its prompt.

The United Arab Emirates established the Technology Innovation Institute (TII) and began building what would become the Falcon series of large language models.

OpenAI trained GPT-2, a 1.5-billion parameter language model that generated remarkably coherent text. They initially withheld the full model, citing concerns about misuse — the first major public debate about whether AI capabilities should be freely shared.

In 2018, Google released BERT (bidirectional pretraining) and OpenAI released GPT-1 (autoregressive pretraining), two competing approaches to making language models understand context.

Eight researchers at Google published the transformer architecture, replacing recurrence entirely with self-attention mechanisms. The paper's deceptively simple title belied its revolutionary impact.

China's State Council released a national strategy aiming to make China the world leader in AI by 2030, with a domestic AI industry worth $150 billion. The plan committed massive government funding and integrated AI into education at all levels.

DeepMind's AlphaZero mastered chess, Go, and shogi (Japanese chess) in hours, starting from only the rules — no human games, no human knowledge, no opening books.

DeepMind's AlphaGo defeated world Go champion Lee Sedol 4-1 in Seoul. Go has more possible positions than atoms in the universe, making brute-force search impossible. AlphaGo combined deep neural networks with Monte Carlo tree search to develop intuitive, human-like play.

Google acquired London-based DeepMind Technologies for approximately $500 million. Founded by Demis Hassabis, Shane Legg, and Mustafa Suleiman, DeepMind's explicit mission was to "solve intelligence."

Ian Goodfellow at the University of Montreal invented GANs — a framework where two neural networks compete, one generating data and one judging it. The idea reportedly came to him during a conversation at a bar.

Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton at the University of Toronto trained AlexNet, a deep convolutional neural network, on GPUs. It won the ImageNet competition with a top-5 error rate of 15.3% — crushing the runner-up at 26.2%.

Jeff Dean and Andrew Ng launched Google Brain, a deep learning research project within Google. Using 16,000 CPU cores across 1,000 machines, the team trained a neural network that learned to detect cats in YouTube videos without being told what a cat was.

Stanford professor Fei-Fei Li and her team published ImageNet, a dataset of 14 million hand-labelled images across 20,000+ categories. The associated annual competition (ILSVRC) became the benchmark that drove computer vision forward.

Geoffrey Hinton at the University of Toronto published a breakthrough method for training deep neural networks using layer-by-layer unsupervised pretraining followed by fine-tuning. For the first time, networks with many layers could be trained effectively.

The Canadian Institute for Advanced Research (CIFAR) launched its Neural Computation and Adaptive Perception programme, providing sustained funding to Geoffrey Hinton, Yoshua Bengio, and Yann LeCun when neural network research was deeply unfashionable.

IBM's Deep Blue defeated world chess champion Garry Kasparov in a six-game match. The system used brute-force search evaluating 200 million positions per second, combined with hand-tuned evaluation functions and an opening book crafted by grandmasters.

Sepp Hochreiter and Jürgen Schmidhuber published Long Short-Term Memory (LSTM), a recurrent neural network architecture with gated memory cells that could learn to store, retrieve, and forget information over long sequences.

Sepp Hochreiter's diploma thesis formally identified the vanishing gradient problem — the mathematical reason deep neural networks were failing to learn. Gradients shrank exponentially through layers, making training beyond a few layers impractical.

Yann LeCun at Bell Labs demonstrated that convolutional neural networks (CNNs) trained with backpropagation could recognise handwritten digits with high accuracy. The system was deployed commercially to read ZIP codes on US mail.

Roboticist Hans Moravec at Carnegie Mellon articulated what became known as Moravec's Paradox: high-level reasoning (chess, logic) is computationally cheap for machines, but low-level sensorimotor skills (walking, catching a ball) are enormously hard.

David Rumelhart, Geoffrey Hinton, and Ronald Williams published a clear, practical method for training multi-layer neural networks using backpropagation of errors. Though the algorithm had been discovered earlier, this paper demonstrated it could learn useful internal representations.

Japan's Ministry of International Trade and Industry (MITI) launched the Fifth Generation Computer Systems project, a $400 million national initiative to build intelligent computers using logic programming and parallel processing. It aimed to achieve conversational AI and expert reasoning within a decade.

Philosopher John Searle at UC Berkeley proposed the Chinese Room thought experiment, arguing that a computer manipulating symbols according to rules does not truly "understand" anything. The argument challenged whether symbol-processing AI could ever achieve genuine intelligence.

British mathematician James Lighthill published a devastating government-commissioned report concluding that AI had failed to deliver on its promises. The UK government cut nearly all AI funding, and the report influenced funders worldwide.

Marvin Minsky and Seymour Papert published "Perceptrons," a mathematical analysis proving that single-layer perceptrons could not learn certain functions (like XOR). The book was widely interpreted as a death blow to neural network research.

Joseph Weizenbaum at MIT created ELIZA, a simple pattern-matching program that simulated a psychotherapist. Despite using no real understanding, many users became emotionally attached to it, revealing humanity's readiness to attribute intelligence to machines.

John McCarthy at MIT created LISP (List Processing), a programming language built on lambda calculus with features like recursion, dynamic typing, and garbage collection. It became the standard language for AI research for the next three decades.

At Cornell, Frank Rosenblatt built the Mark I Perceptron, the first machine that could learn from examples. Funded by the US Navy, it was a hardware neural network that learned to classify simple visual patterns using adaptive weights.

John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon organised the Dartmouth Summer Research Project on Artificial Intelligence. The workshop coined the term "artificial intelligence" and established AI as a distinct academic discipline. McCarthy later founded the Stanford AI Laboratory.

Turing published "Computing Machinery and Intelligence," posing the question "Can machines think?" and proposing a practical test (now called the Turing Test) to evaluate machine intelligence. He predicted that by the year 2000 machines would fool human judges 30% of the time.

British mathematician Alan Turing published "On Computable Numbers," introducing the concept of a universal machine that could simulate any computation. This theoretical framework laid the mathematical bedrock for every computer and every AI system that would follow.

Despite extraordinary progress, researchers continue to debate what's still needed for AGI. Open challenges include: genuine causal reasoning (not just pattern matching), persistent memory and learning from experience, embodied intelligence and physical world understanding, robust common sense, goal-setting and autonomous motivation.