The Bible of AI

Computing with light: what photonics could change for AI

Computing with light: what photonics could change for AI

Light can carry information, but it can also perform selected calculations inside a purpose-built circuit. Photonic computing is moving from isolated optical operations towards experimental systems that run neural networks. This review examines what recent demonstrations actually establish, from programmable tensor processors to tiny task-specific optical structures and computing-centre trials. Its central question is what these advances could change for artificial intelligence: useful computing capacity, response time, energy demand and cost. The answer depends on the whole system, including memory, conversion, precision, calibration and data movement. Published classification results are distinguished from general AI capability; projected power budgets are separated from measured consumption. The article explains why a fast optical operation does not guarantee a faster application, and why silicon can remain part of a photonic future. It identifies credible opportunities for hybrid and specialised accelerators while setting out the evidence still needed before broad efficiency or economic claims can be accepted.

Computing with light: what photonics could change for AI Leer más »

Living Neurons and Silicon: Where Does the Learning Happen?

Living Neurons and Silicon: Where Does the Learning Happen?

Biohybrid computing combines living neural tissue with electronic interfaces and digital algorithms. Its central challenge is to determine which components account for improved performance. This article examines experiments using neuronal cultures and brain organoids, distinguishing biological adaptation from the training of digital readouts and the selection of stimulation patterns. It compares feedback-dependent performance, reservoir computing and adaptive control, including findings that limit claims about reproducibility and retention. Published criticism of sentience claims is considered alongside the researchers’ response. Human cellular origin is distinguished from human cognition, and laboratory demonstrations from evidence of practical computational superiority. The analysis identifies four priorities for stronger evidence: separating component contributions, testing retention and transfer, reporting independent biological preparations, and comparing complete systems under matched conditions. Biohybrid computing offers an experimental approach to neural information processing, provided that claims remain proportionate to the evidence and that the biological material is studied under appropriate ethical oversight.

Living Neurons and Silicon: Where Does the Learning Happen? Leer más »

Subbarao Kambhampati seated beside robots, in a photograph supplied by Arizona State University.

Subbarao Kambhampati: planning, explanation and the discipline of verification

Subbarao Kambhampati studies what makes a plan workable and an explanation useful. His career connects doctoral research on adapting plans with temporal planning, collaboration between people and robots, and the evaluation of language models. With collaborators, he developed Sapa, helped formulate explanation as reconciliation between different models of a situation, and contributed to PlanBench. His LLM-Modulo proposal gives language models constructive roles alongside external systems that check their outputs. A professor at Arizona State University and former AAAI president, he received the association’s joint educator award with EAAI in 2025. This biography follows the research behind his criticism of unsupported claims about AI, including studies of self-verification and intermediate model outputs. It also considers the progress his group acknowledges, the assumptions behind formal guarantees and the continuing distinction between obtaining a correct answer and understanding why a system produced it.

Subbarao Kambhampati: planning, explanation and the discipline of verification Leer más »

AI Logic and Reasoning: Past, Present and Future

AI Logic and Reasoning: Past, Present and Future

What would justify saying that an AI system reasons? This article follows the question from explicit symbolic procedures to learned representations, using Othello as a controlled setting in which a model’s predictions can be checked against a known world. It distinguishes successful behaviour, recoverable internal information and evidence of causal influence. The discussion examines why some apparent jumps in capability depend on the scoring method, what interventions can reveal about a model, and why a written chain of thought is not automatically a faithful explanation. Studies that support structured representations are read alongside work that tests their consistency and exposes dependence on training data. The final section considers human–AI collaboration as an empirical question: combining a person and a model does not guarantee a better result. Drawing on primary scientific studies, the article identifies useful standards for future evaluation without treating current findings as proof of general understanding or consciousness.

AI Logic and Reasoning: Past, Present and Future Leer más »

Gary Marcus photographed in Washington Square Park, New York, in 2011.

Gary Marcus: rules, reasoning and the search for trustworthy AI

Gary Marcus has spent his career asking how a learner can go beyond the examples it has encountered. His research on language development and infant rule learning led to a sustained interest in the structures that support generalisation. In The Algebraic Mind, he explored how neural systems could implement symbolic operations; later, with Ernest Davis and Doug Lenat, he developed proposals connecting learning, knowledge and reasoning. His career also includes founding roles at Geometric Intelligence and Robust.AI, books on human cognition and participation in debates about AI governance. A Professor Emeritus at New York University, Marcus continues to argue for hybrid approaches, including a 2026 paper with Vaishak Belle. This biography follows the scientific origins of that programme, the evidence and objections surrounding it, and the distinction between identifying the limitations of existing systems and demonstrating that an alternative can reliably overcome them.

Gary Marcus: rules, reasoning and the search for trustworthy AI Leer más »

Portrait of Danijar Hafner.

Danijar Hafner: learning to act through world models

Danijar Hafner studies how artificial agents can learn models of their surroundings and use them to decide what to do. His work connects visual experience, prediction and action through a sequence of collaborative projects, including PlaNet, Dreamer and DayDreamer. These systems investigate planning and behaviour learning inside representations acquired from experience, with experiments ranging from simulated control and Minecraft to physical robots. Hafner completed doctoral research at the University of Toronto with Jimmy Ba, following studies at UCL and the Gatsby Unit and research experience at Berkeley and Google Brain. A former Staff Research Scientist at Google DeepMind, he was named among MIT Technology Review’s Innovators Under 35 in 2026. His trajectory offers a perspective on artificial intelligence centred on acting in an environment. The distinctions between learning a model, producing plausible predictions and achieving reliable behaviour remain essential to understanding both the significance and the limits of that work.

Danijar Hafner: learning to act through world models Leer más »

Pierre-Yves Oudeyer during a filmed conversation at librairie mollat.

Pierre-Yves Oudeyer: learning through curiosity

Pierre-Yves Oudeyer studies how a learner chooses what to explore. A research director at Inria and head of the Flowers team, he has connected developmental robotics, computational models of language and intrinsic motivation. His early work investigated how shared speech patterns could emerge among interacting agents. Later studies, developed with collaborators including Frédéric Kaplan and Adrien Baranes, examined curiosity, learning progress and the selection of goals. This research has led to experiments in which robots organise aspects of their own practice, as well as investigations of exploratory methods in simulated scientific environments. Oudeyer’s trajectory offers a perspective on artificial intelligence centred on development and experience. It also includes work on grounding language models through interaction. Across these projects, the central question remains concrete: how can a machine discover which experiences will help it learn, within the structures and constraints supplied by its designers?

Pierre-Yves Oudeyer: learning through curiosity Leer más »

Fei-Fei Li speaking at AI for Good in 2017.

Fei-Fei Li

Fei-Fei Li helped make the visual world a shared resource for artificial intelligence. Her path began with physics at Princeton and doctoral research at Caltech, before leading through university laboratories, ImageNet and the development of spatial models. The images and categories assembled by her collaborators gave researchers a common basis for learning and comparison. They also raised questions about representation and the limits of what a label can tell us about a person. Alongside this scientific work, Li helped establish AI4ALL and Stanford’s institute for human-centred AI, connecting research with education and public responsibility. At World Labs, she and her co-founders are exploring environments that computers can generate and reconstruct. A proposed acquisition by AMD marks another possible transition. Across these settings, her career reveals the people, judgements and institutions behind machine vision, and the continuing challenge of deciding what it means for a computer to understand what it sees.

Fei-Fei Li Leer más »

Olmo Hybrid 7B: architecture, evidence and reproducibility

Olmo Hybrid 7B: architecture, evidence and reproducibility

Olmo Hybrid 7B is a research language model from Ai2 that combines recurrent layers with attention. Its relevance extends beyond the availability of its weights: the release connects an architectural hypothesis with published experiments, training materials and a scientific study accepted at COLM 2026. This profile examines what those materials make possible, how the reported efficiency gains should be interpreted, and where reproducibility remains uncertain. It distinguishes the base model from instruction-following and reasoning-oriented variants, separates training infrastructure from practical deployment requirements, and assesses licensing without treating every component as legally interchangeable. The review also considers results that qualify the headline gains and an unresolved discrepancy in the documentation of long-context training. The resulting assessment supports Olmo Hybrid as a valuable object of scientific investigation, while reserving stronger conclusions about general superiority, exact reproduction and reasoning transparency. The Bible of AI (BAI) has not executed the model or independently reproduced its published measurements.

Olmo Hybrid 7B: architecture, evidence and reproducibility Leer más »

Alexandr Wang: from the data behind AI to Meta’s frontier models

Alexandr Wang: from the data behind AI to Meta’s frontier models

Alexandr Wang’s career connects two often separate stories about artificial intelligence: the human work that makes data useful, and the organisation of laboratories that build frontier models. Raised in Los Alamos by physicist parents, he showed an early aptitude for science before leaving MIT and co-founding Scale AI with Lucy Guo in 2016. Scale turned data preparation and evaluation into a major business. Meta’s 2025 minority investment then brought Wang into its own AI leadership. This profile follows that transition through his upbringing, the company’s formation, its defence-planning work and his responsibilities at Meta. It also examines what the headline transaction actually meant, why labelled data is important without being the sole basis of AI, and how the contribution of researchers and annotation workers fits into the story. A dated chronology, a biographical record and linked sources place the individual within the wider institutions that shaped his influence.

Alexandr Wang: from the data behind AI to Meta’s frontier models Leer más »