ODEI
Create yours

SCIENCE

WHY YOUR AI NEEDS A MODEL OF YOUR WORLD.

What the research says about working with AI, where it goes wrong, and the bet ODEI makes. Every claim links to its source, including the ones that argue against us.

The argument

THE ARGUMENT

YOUR WORLD IS WHAT YOUR AI IS MISSING.

Everyone has access to the same models. Only you have your world.

THE EVIDENCE

SEVEN FINDINGS. ONE BET.

  1. 01

    THE WORK MOVED TO YOU.

    AI helps, and the checking, joining and re-explaining land on you.

  2. 02

    YOUR AI FORGETS YOU.

    Assistants lose about 30% accuracy when they must remember across sessions.

  3. 03

    STEERING TAKES A MODEL.

    The simplest system that steers something well is a model of it.

  4. 04

    DIRECTION CHANGES WHAT PEOPLE DO.

    Specific goals and concrete next steps measurably raise results.

  5. 05

    PARTNERSHIP CAN BE MEASURED.

    On average, you and AI together do not beat the better of the two. Sometimes you do.

  6. 06

    INITIATIVE NEEDS LIMITS.

    Who proposes and who decides are separate choices, and approval alone is not enough.

  7. 07

    YOUR MODEL SHOULD STAY YOURS.

    What holds your life should outlive any company, including ours.

FINDING 01

THE WORK MOVED TO YOU.

AI REALLY DOES HELP.

In a controlled experiment with 453 professionals, ChatGPT cut the time on short writing tasks by 40% and raised quality by 18%.2 Across 5,172 support agents, issues resolved per hour rose 15%.3

THE SAVINGS SHRINK AT WORK.

Across about 25,000 workers in Denmark, users report saving 2.8% of their hours. Of the new tasks AI created, 59% are implementation and oversight.4

YOU BECOME THE CHECKER.

A study of 319 knowledge workers found that GenAI “shifts the nature of critical thinking toward information verification, response integration, and task stewardship.”5

YOU SUPPLY THE CONTEXT.

One survey of 6,000 workers put it at 2.3 hours a week spent giving AI the context it needs. It is self-reported, and the company behind it sells a context product.6

EVERY SWITCH HAS A PRICE.

Researchers who shadowed 24 information workers saw 57% of their threads of work interrupted. Work resumed the same day came back about 25 minutes later, after two other activities.7 In the lab, a 2.8 second interruption doubled sequence errors.8

YOUR HEAD HAS A LIMIT.

Short-term memory holds about four chunks.9 You cannot keep your whole life in mind while you shuttle between tools.

A RANDOMIZED TRIAL

MEASURED

19% SLOWER

16 experienced developers on 246 real tasks took 19% longer when AI was allowed.1

BELIEVED

20% FASTER

Afterwards the same developers still believed AI had sped them up by 20%.1

On the 44 tasks with usable recordings, about 9% of their time went to reviewing and cleaning AI output. One reason the paper names: “AI doesn’t utilize important tacit knowledge or context.” Coding only, early-2025 tools, and the authors warn against generalizing.

FINDING 02

YOUR AI FORGETS YOU.

ABOUT 30% ACCURACY, GONE.

A benchmark of 500 questions found that commercial assistants and long-context models lose about 30% accuracy when they must remember across sustained interactions.10

A BIGGER WINDOW IS NOT THE FIX.

“Current language models do not robustly make use of information in long input contexts.” What sits in the middle gets missed. Pasting your life into a chat is not memory.11

MEMORY HAS TO LIVE OUTSIDE THE MODEL.

One line of work manages memory the way an operating system does, outside the model’s window.12 Another showed that memory, reflection and planning each contribute to how believably agents behave.13

STRUCTURE HELPS.

The same benchmark improved results by organizing memory: indexing facts, tracking time, then retrieving what matters.10 A model of your world is structured memory about you.

FINDING 03

STEERING TAKES A MODEL.

“If the organism carries a ‘small-scale model’ of external reality and of its own possible actions within its head, it is able to try out various alternatives, conclude which is the best of them, react to future situations before they arise…”Kenneth Craik, 194314

MODEL-MAKING IS COMPULSORY.

Control theory’s good regulator theorem: a regulator that is both maximally successful and simple must be a model of the system it regulates. In its authors’ words, this changes “the status of model-making from optional to compulsory.”15

AI RESEARCH ARRIVED AT THE SAME PLACE.

Agents that learn a model of their environment can rehearse inside it.16 In Nature in 2025, one configuration with a world model beat specialized methods across more than 150 tasks.17

ROBOTS NEED PHYSICS. PEOPLE NEED DIRECTION.

Those models are of games and physics. ODEI’s bet is a model of a person: your goals, your people, your projects and where you are going.

A LOOP, NOT A CHAT.

Goal-directed systems run on feedback: act, measure against the goal, correct. That idea goes back to cybernetics.18 Observe, Decide, Act, Verify and Evolve follow that tradition.

FINDING 04

DIRECTION CHANGES WHAT PEOPLE DO.

“DO YOUR BEST” DOES NOT WORK.

Specific, difficult goals beat “do your best”, with effects from d = .42 to .80. “In short, when people are asked to do their best, they do not do so.” Goals plus feedback beat goals alone.19

THE NEXT STEP IS WHERE PEOPLE FAIL.

Goals often do not turn into action. Plans that say what you will do and when had a medium-to-large effect on reaching goals across 94 independent tests, d = .65.20

WHY THIS IS THE BET.

An AI that holds your goals and brings you the next step works exactly where the research says people get stuck.

WHAT THIS DOES NOT SHOW.

These studies are about people, not AI. Whether goals held by an AI, or plans proposed by one, work as well is still untested.

FINDING 05

PARTNERSHIP CAN BE MEASURED.

S = P(you + AI) − max( P(you), P(AI) )

Below zero, it is assistance. Above zero, it is partnership.

What 106 experiments foundHuman and AI together, effect size (Hedges’ g) with 95% confidence interval21
Better than you alone +0.64+0.53 to +0.74
Better than the better of the two −0.23−0.39 to −0.07
When you beat the AI alone +0.46+0.28 to +0.66
When the AI beats you alone −0.54−0.71 to −0.37

Synergy, the same comparison as S. On average it is below zero: the combination loses to whichever of you is stronger.

Our reading: AI usually makes you better than you were. On average it does not make the pair better than the stronger of the two. The gains appear when the person is the stronger side, and we read that as the person bringing something the AI lacks. Your goals, your people and your direction are exactly what it lacks. Most of these studies are short lab tasks from 2020 to 2023, none tests a personal model, and creation tasks only trended positive (g = +0.19, not significant).

FINDING 06

INITIATIVE NEEDS LIMITS.

DO NOTHING, ASK OR ACT.

Mixed-initiative design chooses between the three by weighing its uncertainty about your goals.22 Knowing your goals is what makes asking well possible.

AUTOMATION IS NOT ONE DIAL.

Gathering, analysing, deciding and acting can each be automated to a different level.23 Who holds the initiative is its own choice, separate from who does the work.24

A VERY OLD PROBLEM.

Anyone who acts on your behalf can drift from your interest. Economics prices it as agency cost: monitoring, bonding and residual loss.25 Applying this to AI is an analogy, and a useful one.

APPROVAL ALONE IS NOT ENOUGH.

Automation bias hits experts too and “cannot be prevented by training or instructions.”26 A yes button does not make you safe, which is why how much reaches you and how it is shown matters. ODEI’s desktop app, in development, brings one next step at a time, and no agent can approve it.

FINDING 07

YOUR MODEL SHOULD STAY YOURS.

COMPANIES SHUT DOWN.

“If a service shuts down, the software stops functioning, and data created with that software is lost.” Local-first software is the answer to that.27 In ODEI’s desktop app, your World Model is a local file on your computer.

IT BECOMES PART OF HOW YOU THINK.

One philosophical argument holds that a resource you trust and can always reach can be part of your mind.28 It is contested. If it is even half right, it should not belong to someone else.

WHAT IS NOT PROVEN

WHERE THE EVIDENCE STOPS.

  • No study here tests a personal world model. Nothing here shows that one improves a life.
  • Partnership is not the norm. On average, people and AI together fall below the stronger of the two.
  • Many of these numbers come from small samples, short tasks, or models older than today’s.
  • Approving every step does not remove bias.
  • ODEI has not published a measured S.

THE TEST

THE SAME TASK. THREE RUNS.

You alone. The AI alone. You and the AI, working from your World Model. If the third does not beat the better of the first two, that is assistance, however good it feels. The METR developers felt faster too. This is the bar ODEI holds itself to.

SOURCES

CHECK US.

  1. Becker, J., Rush, N., Barnes, E. & Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR. arXiv:2507.09089
  2. Noy, S. & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192. doi:10.1126/science.adh2586
  3. Brynjolfsson, E., Li, D. & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889-942. doi:10.1093/qje/qjae044
  4. Humlum, A. & Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. NBER Working Paper 33777. nber.org/papers/w33777
  5. Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R. & Wilson, N. (2025). The Impact of Generative AI on Critical Thinking. CHI ’25. doi:10.1145/3706598.3713778
  6. Glean (2026). Work AI Index. Vendor survey of 6,000 workers, self-reported. glean.com
  7. Mark, G., Gonzalez, V. M. & Harris, J. (2005). No Task Left Behind? Examining the Nature of Fragmented Work. CHI 2005, 321-330. doi:10.1145/1054972.1055017
  8. Altmann, E. M., Trafton, J. G. & Hambrick, D. Z. (2014). Momentary Interruptions Can Derail the Train of Thought. Journal of Experimental Psychology: General, 143(1), 215-226. doi:10.1037/a0030986
  9. Cowan, N. (2001). The magical number 4 in short-term memory. Behavioral and Brain Sciences, 24(1), 87-114. doi:10.1017/S0140525X01003922
  10. Wu, D., Wang, H., Yu, W., Zhang, Y., Chang, K.-W. & Yu, D. (2025). LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. ICLR 2025. arXiv:2410.10813
  11. Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F. & Liang, P. (2024). Lost in the Middle: How Language Models Use Long Contexts. TACL, 12, 157-173. doi:10.1162/tacl_a_00638
  12. Packer, C., Wooders, S., Lin, K., Fang, V., Patil, S. G., Stoica, I. & Gonzalez, J. E. (2023). MemGPT: Towards LLMs as Operating Systems. arXiv:2310.08560
  13. Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P. & Bernstein, M. S. (2023). Generative Agents: Interactive Simulacra of Human Behavior. UIST ’23. doi:10.1145/3586183.3606763
  14. Craik, K. J. W. (1943). The Nature of Explanation, chapter 5. Cambridge University Press.
  15. Conant, R. C. & Ashby, W. R. (1970). Every good regulator of a system must be a model of that system. International Journal of Systems Science, 1(2), 89-97. doi:10.1080/00207727008920220
  16. Ha, D. & Schmidhuber, J. (2018). World Models. arXiv:1803.10122
  17. Hafner, D., Pasukonis, J., Ba, J. & Lillicrap, T. (2025). Mastering diverse control tasks through world models. Nature, 640, 647-653. doi:10.1038/s41586-025-08744-2
  18. Wiener, N. (1948). Cybernetics: Or Control and Communication in the Animal and the Machine. The Technology Press and John Wiley & Sons. Reissued by MIT Press, 2019. doi:10.7551/mitpress/11810.001.0001
  19. Locke, E. A. & Latham, G. P. (2002). Building a practically useful theory of goal setting and task motivation. American Psychologist, 57(9), 705-717. doi:10.1037/0003-066X.57.9.705
  20. Gollwitzer, P. M. & Sheeran, P. (2006). Implementation intentions and goal achievement: A meta-analysis of effects and processes. Advances in Experimental Social Psychology, 38, 69-119. doi:10.1016/S0065-2601(06)38002-1
  21. Vaccaro, M., Almaatouq, A. & Malone, T. W. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8(12), 2293-2303. doi:10.1038/s41562-024-02024-1
  22. Horvitz, E. (1999). Principles of mixed-initiative user interfaces. CHI ’99, 159-166. doi:10.1145/302979.303030
  23. Parasuraman, R., Sheridan, T. B. & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics, Part A, 30(3), 286-297. doi:10.1109/3468.844354
  24. Beer, J. M., Fisk, A. D. & Rogers, W. A. (2014). Toward a framework for levels of robot autonomy in human-robot interaction. Journal of Human-Robot Interaction, 3(2), 74-99. doi:10.5898/JHRI.3.2.Beer
  25. Jensen, M. C. & Meckling, W. H. (1976). Theory of the firm: Managerial behavior, agency costs and ownership structure. Journal of Financial Economics, 3(4), 305-360. doi:10.1016/0304-405X(76)90026-X
  26. Parasuraman, R. & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381-410. doi:10.1177/0018720810376055
  27. Kleppmann, M., Wiggins, A., van Hardenberg, P. & McGranaghan, M. (2019). Local-First Software: You Own Your Data, in spite of the Cloud. Onward! ’19. doi:10.1145/3359591.3359737
  28. Clark, A. & Chalmers, D. (1998). The Extended Mind. Analysis, 58(1), 7-19. doi:10.1093/analys/58.1.7

CREATE YOUR WORLD MODEL.

See how it works