Back to Adabra

Adabra World Model

Estimated 5 minutes · Sound starts when you press Play

Build your clone with Adabra
Research and sources

Research behind the digital self

  1. Su et al. · ICML · 2026

    NVIDIA research

    An agent can coordinate models and tools to do more: ToolOrchestra improved tested reasoning and tool-use performance while reducing cost versus its GPT-5 baseline

    Paper ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration
    Status Peer-reviewed, 2026
    Results HLE 37.1 vs 35.1; FRAMES 76.3 vs 74.0; τ²-Bench 80.2 vs 77.7, against the tested GPT-5 configuration
    Scope Benchmarks; not better than every baseline

  2. Park et al. · arXiv preprint 2411.10109 (v3, June 2026)

    Google DeepMind and Stanford University, with co-authors from Northwestern, the University of Washington and Sciences Po

    Personal context improves simulations of individual responses: agents grounded in interviews and surveys reached 86% of participants’ own two-week retest consistency on held-out General Social Survey questions

    Paper LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals
    Status Preprint, 2026
    Sample 1,052 U.S. adults
    Scope Normalized accuracy on held-out survey answers; not 86% of a whole person

  3. Zhang et al. · Findings of ACL · 2026

    Amazon-affiliated researchers, with university co-authors

    Personal memory makes AI responses more individual: PersonaAgent combines remembered experiences and a changing persona to improve personalized predictions

    Paper PersonaAgent: Bridging Memory and Action for Personalized LLM Agents
    Status Peer-reviewed, 2026
    Scope Offline benchmark tasks with Wikipedia and memory retrieval; not real-world automated actions

  4. Anthropic · How we built our multi-agent research system · 13 June 2025

    90.2% higher performance on Anthropic’s internal research evaluation: a Claude Opus 4 lead agent with Claude Sonnet 4 subagents, versus Claude Opus 4 working alone

    Status Company-reported engineering evaluation, not peer reviewed
    Scope Not accuracy or time saved; no sample size or uncertainty published; the multi-agent system uses more tokens

  5. Yang, Zyskowski, Yonack & Ma · arXiv working paper · June 2026

    Perplexity researchers, with a Harvard Business School–affiliated co-author

    87% less estimated task time: 269 minutes with search and manual execution → 36 minutes with an agent and human oversight

    Paper How AI Agents Reshape Knowledge Work: Autonomy, Efficiency, and Scope
    Status Working paper (not peer reviewed), 2026
    Sample 10,000 matched task pairs
    Method Estimated, not timed: tool-based counterfactual estimates; the 36 minutes include an assumed 10 minutes of human oversight per task
    Summary Harvard Business School AI Institute article

  6. Adank & Wang · Journal of the Acoustical Society of America · 2026

    UCL and University of Roehampton

    Clearer words in background noise: listeners identified 67.5% of words in cloned voices, versus 54.1% in the original voices

    Paper Voice clones are easier to understand in noise than their human originals: The voice cloning intelligibility benefit
    Status Peer-reviewed, April 2026
    Sample 80 British-English listeners aged 18–35; ten voices, cloned with one system (ElevenLabs)
    Scope A 13.4 percentage-point difference in speech-shaped noise; not a hearing-loss trial
    Open access UCL Discovery

Technical foundations

  1. Attention Is All You Need

    Google Brain, Google Research and the University of Toronto

    Introduced the Transformer, a neural architecture that processes relationships in language through attention.

    Authors Ashish Vaswani, Noam Shazeer, Niki Parmar et al.
    Published NIPS 2017 · Peer-reviewed
    PDF proceedings.neurips.cc

  2. Training Compute-Optimal Large Language Models

    DeepMind

    Examines how to balance model size and training data within a fixed computing budget.

    Authors Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch et al.
    Published NeurIPS 2022 · Peer-reviewed (proceedings title: “An empirical analysis of compute-optimal large language model training”)
    PDF proceedings.neurips.cc

  3. Training language models to follow instructions with human feedback

    OpenAI

    Uses human demonstrations and ranked feedback to train language models to follow instructions.

    Authors Long Ouyang, Jeffrey Wu, Xu Jiang et al.
    Published NeurIPS 2022 · Peer-reviewed
    PDF proceedings.neurips.cc

  4. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

    Facebook AI Research (Meta), University College London and New York University

    Combines language generation with retrieved documents as an external source of knowledge.

    Authors Patrick Lewis, Ethan Perez, Aleksandra Piktus et al.
    Published NeurIPS 2020 · Peer-reviewed

  5. ReAct: Synergizing Reasoning and Acting in Language Models

    Princeton University and Google Research

    Interleaves reasoning steps with actions that gather information from tools and environments.

    Authors Shunyu Yao, Jeffrey Zhao, Dian Yu et al.
    Published ICLR 2023 · Peer-reviewed
    PDF openreview.net

  6. Scaling and evaluating sparse autoencoders

    OpenAI

    Develops methods for extracting and evaluating interpretable features inside language models.

    Authors Leo Gao, Tom Dupré la Tour, Henk Tillman et al.
    Published ICLR 2025 · Peer-reviewed
    PDF proceedings.iclr.cc

Human learning

  1. A Neural Substrate of Prediction and Reward

    University of Fribourg, Massachusetts Institute of Technology and Baylor College of Medicine

    Connects dopamine reward-prediction signals with computational models of learning.

    Authors Wolfram Schultz, Peter Dayan & P. Read Montague
    Published Science 1997 · Peer-reviewed review

  2. Perception in chess

    Carnegie Mellon University

    Examines how chess expertise relates to recognizing and remembering meaningful patterns.

    Authors William G. Chase & Herbert A. Simon
    Published Cognitive Psychology 1973 · Peer-reviewed

  3. Intersubject synchronization of cortical activity during natural vision

    Weizmann Institute of Science

    Shows that shared film viewing can align patterns of cortical activity across viewers.

    Authors Uri Hasson, Yuval Nir, Ifat Levy, Galit Fuhrmann & Rafael Malach
    Published Science 2004 · Peer-reviewed

  4. Experimental Study of Inequality and Unpredictability in an Artificial Cultural Market

    Columbia University and the Santa Fe Institute

    Demonstrates how seeing other people’s choices can reshape cultural popularity.

    Authors Matthew J. Salganik, Peter Sheridan Dodds & Duncan J. Watts
    Published Science 2006 · Peer-reviewed

  5. Identification of the Social and Cognitive Processes Underlying Human Cumulative Culture

    University of St Andrews, Durham University, University of Texas MD Anderson Cancer Center, CNRS and Université de Strasbourg

    Examines how teaching, imitation and cooperation relate to building on shared knowledge.

    Authors L. G. Dean, R. L. Kendal, S. J. Schapiro, B. Thierry & K. N. Laland
    Published Science 2012 · Peer-reviewed

Digital clones: frontier research

Memory and continual learning

Voice: cloning and full-duplex conversation

Face and body: avatars

Persona and mind

Psychology: attachment, trust and disclosure

Evaluation

Safety, consent and provenance

Frontier papers

Lab updates and news

In the film

Figures shown in the film, by reference, with their sources.

S1

  • About a billion bits a second in; about 10 bits a second out · Zheng & Meister, Neuron 2025Order-of-magnitude estimates from a review; not a measurement of any one person.
  • Next-hour place: up to 93% predictable in principle · Song et al., Science 2010An upper bound for mobility in one mobile-phone dataset, not achieved forecasting accuracy.
  • Life outcomes stay hard to forecast (also in S6) · Salganik et al., PNAS 2020 · 2021 correction recordFor the outcomes and data studied.
  • 58% next-hour prediction; 11.5 hours backScene model, no external source: 7 of 12 fictional Fridays; a week model with an assumed 80% clone share. Not an observed result.

S2

  • 8.3 billion people (also in S3 and S5) · UN DESA, World Population Prospects 2024Medium-variant projection for mid-2026, not a live count.
  • 117,000 points, one per million lives born by 2022; alive in 2022: 6.8% · PRB, Kaneda & Haub, 2022Estimates that depend heavily on prehistoric assumptions.
  • Morning mood peak: 84 countries, 509 million messages · Golder & Macy, Science 2011Language-based mood among about 2.4 million sampled Twitter users; the moving band is an illustration.
  • City lights, population rings, sun, moon and clockIllustrative construction with approximate city weights; no external source.

S3

S4

S5

S6

S7

  • Applause slows to about half speed and synchronizes · Néda et al., Nature 2000The scene’s population and coupling are simulation settings.
  • One shared clock can lock people who never hear each other · Childs & Strogatz, 2008A model result for idealized oscillators.
  • People watching the same film show synchronized brain activity · Hasson et al., Science 2004Correlated activity, not identical thoughts.
  • A feed runs on one clock; her clone on each person’s · YouTube on personalizationThe thread count is a scene simulation.

Illustrative calculator

The finale’s calculator is an illustration with chosen inputs, not a measured result or a forecast.

  • Inputs 5,000 paying subscribers; $300 per subscriber per month; one thread per subscriber; 5 conversations per subscriber per month; 10 messages per conversation (both sides); 5 minutes to answer a conversation manually; 100% of conversations handled (an illustrative input, not measured performance).
  • Gross revenue 5,000 × $300 = $1,500,000 per month, before fees, costs and tax; not profit or guaranteed income.
  • Threads 5,000 × 1 = 5,000.
  • Conversations 5,000 × 5 = 25,000 per month.
  • Messages 25,000 × 10 = 250,000 per month.
  • Hours 25,000 × 5 minutes ÷ 60 = 2,083.3, shown as 2,083+: the manual-answering equivalent combined across all subscribers, not one person’s hours back.

Additional research

AI helping people work and learn

Peer-reviewed research on AI-assisted work and learning.

  1. Brynjolfsson, Li & Raymond · Quarterly Journal of Economics · 2025

    An AI assistant that learned from top agents’ conversations helped less-experienced agents most

    Task Customer-support chats at one software firm
    Compared with Agents before, or without, the assistant
    Sample 5,172 agents
    Scope The AI suggested; people decided. About 15% more issues resolved per hour on average; top agents gained little.

  2. Noy & Zhang · Science · 2023

    40% less task time and 18% higher quality ratings, on average, on professional writing tasks

    Task Bounded, occupation-specific writing tasks
    Compared with Random assignment to ChatGPT access or not (about 80% of those given access used it, as did 10–20% of controls)
    Sample 453 college-educated professionals overall; the headline estimates exclude a fixed-time incentive arm
    Scope Short writing tasks, not a whole working week. Quality means ratings by blinded professional graders, not factual accuracy; time was self-reported, with activity checks.

  3. Dell’Acqua et al. · Organization Science · 2026

    25.1% less time on tasks within GPT-4’s capabilities; on another task, GPT-4 access reduced accuracy

    Task Realistic consulting tasks, inside and outside GPT-4’s capabilities
    Compared with Random assignment to GPT-4 access (with or without a prompting overview) or no AI
    Sample 385 consultants in the inside-capability comparison and 373 in the outside one; 758 at one firm in total
    Scope Time was measured over the first 17 questions of a fixed-length session. On the outside task, correct answers fell by about 19 percentage points.

  4. Kestin et al. · Scientific Reports · 2025

    Structured AI tutoring improved immediate physics learning compared with active-learning classes

    Task Two physics lessons in one university course, tested straight after
    Compared with The same lessons taught as in-class active learning (crossover design)
    Sample 194 eligible students; the post-test counts (142 with AI, 174 in class) are not fully reconciled with that figure in the published text
    Scope Immediate learning, not retention. AI time was logged (median 49 minutes; 70% under 60); class time was assumed to be 60 minutes, not measured.

  5. Létourneau et al. · npj Science of Learning · 2025

    AI tutoring systems were generally positive against teacher-led instruction, and weaker against other digital tutors

    Task K-12 school learning with intelligent tutoring systems
    Compared with Teacher-led instruction, or non-intelligent digital tutors
    Sample 28 studies, 4,597 learners
    Scope Mostly pre-LLM systems; no meta-analysis. Not personal clones.

  6. Bastani et al. · PNAS · 2025

    A GPT-4 practice assistant lowered later unaided exam scores; a teacher-guided tutor showed no statistically detectable difference from control

    Task High-school maths practice sessions, then exams without AI
    Compared with Practice with books and course notes, randomised by class
    Sample About 1,000 students in one school; the main analysis excludes honours classes
    Scope “GPT Base” resembled ChatGPT; its exam scores were about 17% below control. The guided tutor’s null result does not show equivalence or absence of harm.

Memory, tools and human control

Research on personalization, tool use and human control in AI systems.

  1. Cai et al. · The Web Conference, 2025

    PUMA used remembered behaviour to improve personalized tool use in shopping tasks.

    Paper Large Language Models Empowered Personalized Web Agents
    Evidence Capability benchmark
    Scope Simulated profiles and web functions; no owner-time study.

  2. Hu et al. · AAAI, 2026

    Orion combined preference tracking, clarification and tool actions, improving its benchmark results.

    Paper Orion: Steering Personalized Web Agents via Global-Micro Profiling and Adaptive Intent Tracking
    Evidence Capability benchmark
    Scope Simulated shopping interactions; life benefits were not measured.

  3. Zhou et al. · CHI, 2026

    In a 48-person study, intermediate checks cut simulated agent-task completion time by 13.5%.

    Paper When Should Users Check? Modeling Confirmation Frequency in Multi-Step Agentic AI Tasks
    Evidence Controlled human study of task time
    Scope Compared with checking only at the end, not doing the work manually.

  4. Peng et al. · Science Advances, 2026

    Across 19 studies, survey-based digital twins showed weak agreement with the individuals they represented.

    Paper Digital twins are funhouse mirrors: Five systematic distortions
    Evidence Boundary: how faithfully a model represents a person
    Scope Survey-response simulations; expert-service performance was not tested.

Also referenced