|
ResearchAudio · Paper of the week
Stop asking how smart AI is. Ask how much smarter it makes you.
7-minute read · 9 diagrams · Paper: arXiv 2611.05376
|
|
Think about one normal day.
In the morning, you have a good idea. It is gone before you write it down. In a meeting, someone asks about a number you reported three years ago. You cannot remember it. In the evening, you must make a hard choice. You made a similar choice before, but you cannot remember how you decided.
None of these is a problem of intelligence. You had the knowledge. It just did not reach you at the right moment.
And somewhere, an AI that could help was waiting in a chat box. It did not know who you were, where you were, or what you needed.
A new paper from Oxford says this gap is the next big problem in AI. It also says the answer brings real dangers, such as an AI that whispers answers to you in an exam. Part 5 shows the author’s plan for that.
|
|
Note: This is a perspective paper. It gives a new way to think about AI, with names, definitions, and a research plan. It does not build a system or run experiments. The author says so directly.
|
|
|
In this issue
1. The other half of the AI race
2. The “cognitive gap” you pay every day
3. Six things a real AI partner must be
4. Why the best answer is often silence
5. The dangers, and Exam Day
6. How we would know it works
|
|
|
Part 1 of 6
The other half of the AI race
|
|
For ten years, we scored AI on one question: how much can the machine do alone? That race leads to AGI (artificial general intelligence).
The author, Hao Fei, says there is a second race. He calls it AHI: AI-augmented human intelligence. Its question is different: how much more can one particular person do with the AI?
Both races use the same models. The difference is what you measure. One measures the machine. The other measures you.
|
| Diagram 1 · Two halves of the endgame |
|
AGI
“How smart can the machine be?”
Autonomous agent runs long tasks alone |
| ↑ |
Tool-using agent does a task with tools |
| ↑ |
Chatbot answers one message |
Score: what the AI completes
|
AHI
“How much smarter can it make you?”
Cognitive symbiont knows your life; lives on you |
| ↑ |
Personal agent knows your preferences; in the cloud |
| ↑ |
Reactive assistant knows this message; in a chat box |
Score: what the person can now do
|
| Shared base: the same models, agents, memory, and hardware. Every gain in the base helps both sides. |
|
|
There is a physical side too. Today, people imagine one kind of smart being in the real world: the robot. The author says there will be a second kind of equal importance: a person with an AI that lives on their body, through glasses, earbuds, or a pendant.
Why now? The author argues that AI keeps getting stronger, but the connection between an AI and you has barely changed. How much does it know about you? How fast can it help? How much does it interrupt? How much do you trust it?
|
| Diagram 2 · The bottleneck has moved (a concept sketch, not data) |
|
What AI can do on its own
How well AI connects to you
The space between the two is help you could get but do not. Three signs you feel it: you explain your background in every chat; help comes too late or too loud; and you do not trust the AI with anything important.
|
|
|
Part 2 of 6
The “cognitive gap” you pay every day
|
|
The central idea of the paper has a name: the cognitive gap.
On one side is everything you could bring to a moment: your past work, your notes, your old decisions, what you saw last week. On the other side is the small part your working memory can hold right now.
The gap between them costs you every day. You forget. You do not notice. You cannot find the right words. The author is clear that this is not about how intelligent you are. Everybody pays it.
|
| Diagram 3 · The cognitive gap |
Everything you could bring
past work · old decisions · notes · people you met · what you saw · what you said you wanted
|
≠ |
What is in your head right now a few items
|
| The symbiont’s job: bring back the right slice, at the right moment |
|
No mind reading needed. Just your own context, returned when it matters.
|
|
|
The system that closes this gap is a cognitive symbiont. (In biology, symbiosis means two living things that help each other.)
A symbiont sees what you see through small devices. It keeps your context for years. It understands your situation. It takes part in your thinking at the right time. It does tasks for you. And it keeps learning about you as you change. One rule is fixed: final judgement and goals stay with you.
|
|
Part 3 of 6
Six things a real AI partner must be
|
|
The author uses a famous example: Jarvis, the AI helper in the Iron Man films. Jarvis has known its owner for years. It is always there. It speaks without being asked. And it never says, “Please stop and explain.”
This gives the “Jarvis test”: six properties. Each one is necessary. Here is how today’s personal agents score.
|
| Diagram 4 · The Jarvis test |
| Property |
Without it… |
Jarvis |
Agents today |
Lifelong one memory over years | every chat starts from zero | ✓ | PARTLY |
Embedded present in real life | help only comes afterwards | ✓ | ✗ |
Real-time helps while it still matters | the right words, too late | ✓ | ✗ |
Private stays between you two | you cannot confide in it | ✓ | PARTLY |
Low-friction no stopping to explain | using it costs more than it gives | ✓ | ✗ |
Proactive decides when to speak | you never ask about what you do not know you missed | ✓ | PARTLY |
| ✓ has it PARTLY has some of it ✗ does not have it |
|
|
The properties also fight each other. Proactive pulls against low-friction: an AI that speaks more also interrupts more. Lifelong pulls against private: an AI that remembers more also holds more of your secrets. The author says these tensions are the real research problems.
And the goal is to go beyond Jarvis. Jarvis needs a lab its owner built; a symbiont must work with a few low-cost wearables. Jarvis mostly runs systems and warns; a symbiont must understand your goals, feelings, and blind spots. Jarvis is for one rich man; a symbiont must be for everyone.
|
|
Keep reading: Next, the strangest design rule in the paper. To be truly helpful, the AI must say nothing most of the time. After that: what happens when you walk into an exam with an AI in your ear.
|
|
|
Part 4 of 6
Why the best answer is often silence
|
|
Today’s AI tools wait for a request, then answer. A symbiont is always there. So it has a new problem: when should it speak at all?
The paper gives it five choices at every moment. The first choice is silence. And silence is not a failure. It is a formal output, and usually the correct one.
|
| Diagram 5 · Five ways to take part, from quiet to active |
Silence the default |
|
Hint a word or two |
|
Suggest a full idea |
|
Ask a question back |
|
Act do it for you |
Each time, it must decide three things: what to help with, when, and how. The answer to “how” always includes silence, or a quiet note saved for later.
|
|
|
To make this work, the paper describes two loops that run at the same time. A slow loop searches your memory and checks facts. A fast loop decides the exact moment to show something. The author calls this “cognitive full-duplex”: like a phone call where both sides can talk at once. Here is the paper’s own example:
|
| Diagram 6 · One moment in a team meeting |
| 1 |
A colleague asks about a number you reported last quarter. |
| 2 |
SLOW LOOP finds the number in your record and checks it. |
| 3 |
FAST LOOP puts one line in your glasses, within about a second. |
| 4 |
You answer, in your own words. |
| 5 |
The line disappears. The AI goes back to silence. |
No phone. No typing. No “let me check.” Nobody else in the room sees it.
|
|
|
This is one stage of a bigger cycle. The paper calls it the symbiosis loop. Each stage feeds the next:
|
| Diagram 7 · The symbiosis loop (trust surrounds all of it) |
1. Sense sees and hears what you do |
→ |
2. Keep one record of your life |
→ |
3. Understand your situation and mind |
| ↑ |
|
YOU + symbiont |
|
↓ |
6. Grow learns you better over time |
← |
5. Act & think with errands, decisions, people |
← |
4. Live alongside silence, hint, act, at the right time |
| Underneath: the personal runtime · devices, permissions, scheduling, and the cost of being always on |
|
|
|
|
Part 5 of 6
The dangers, and Exam Day
|
|
Here is the uncomfortable part. A chatbot that writes code does not change what you can do in front of other people. A symbiont does.
The paper gives real cases. In 2025, a product called Cluely was sold as a way to cheat. Chatbot companies turned off features during China’s college entrance exam. And the same privacy that protects you also hides the help from the examiner, the interviewer, or the friend across the table.
The author maps harm in three directions, and treats it as part of the design, not an afterthought:
|
| Diagram 8 · Three directions of harm |
↓ To you
Lost skills from relying on it. Manipulation. Flattery (telling you what you want to hear). Wrong cues you repeat out loud. False memories: in one study, a chatbot caused over 3× more false memories in witness interviews than a control.
|
|
→ Through you, to others
Secret help in exams. Secret scripts in conversations. A hidden advantage in a negotiation. Fraud with live cues.
|
|
↑ From institutions
An employer, school, insurer, platform, or state using the system for its own goals.
|
|
|
Added together: harm to society, including pressure on everyone to use one, just to keep up.
|
|
|
|
The author takes skill loss seriously. People who rely on GPS tend to have weaker spatial memory. Knowledge workers who trust AI more report less critical thinking. And a language model with a few facts about its opponent was more persuasive than humans in debates. A symbiont would know far more than a few facts.
So what is the plan? The key insight: unlike a drug or a brain implant, a symbiont is outside your body, removable, and can be inspected. So it can follow the rules of each room you enter, and prove that it did.
|
| Diagram 9 · Exam Day (the paper’s own scenario) |
🏫 Room announces: “No assistance” + a list of allowed aids |
→ |
🤖 Symbiont switches to unassisted mode |
→ |
🔒 Proof examiner can check the switch, without seeing your data |
→ |
✅ Exam ends full help returns |
| Fair access: a classmate with dyslexia keeps text-to-speech. It is on the list of allowed aids. |
|
Honest learning: all term, the symbiont tutored with hints. Its record shows it never gave answers outright. |
|
|
|
This leads to a bigger idea about tests. Credentials should measure what the job requires. If a doctor must act alone in an emergency, test them without the symbiont. If a job is always done with tools, test with the tools.
The author also expects exams to change shape: more oral exams, more observed process, and more tasks where the value is in reasoning that cannot be passed on in one line.
|
|
Part 6 of 6
How we would know it works
|
|
The author calls this the most important move in the paper: change what we measure. Not how smart the AI is, but how much more capable the person became. He proposes four numbers, and says reporting only one is not enough:
|
Augmentation Do you do better with it than without it? |
|
Synergy Does the pair beat the better of the two alone? Often, human + AI teams do not. |
|
Retention After you stop using it, are you stronger than before, or weaker? |
|
Personal gain How much comes from knowing you, compared with the same AI without your context? |
|
|
The last one is clever. If you compare “with symbiont” against “no AI,” you give the symbiont credit for what the base model already does. The fair comparison is the same model, with and without your personal context.
Because each symbiont serves one person, the author recommends N-of-1 trials (borrowed from medicine: the same person, tested with and without) and long studies over weeks and months. These studies must also report the people who quit and the cases that went badly.
The paper ends with 23 open problems. Four are marked as most urgent: modelling a whole life’s record at scale, a speed budget from sensor to ear, the cost of an AI that is always on, and better wearable hardware.
|
What to keep in mind
|
| It is a map, not a machine. There is no new algorithm and no experiment. The value is in the names, the framework, and the standard of proof it sets. The author says this himself. |
|
| The diagrams are concepts. The “bottleneck has moved” graph is a sketch. The paper labels it as not measured. |
|
| The hard parts are open. The safety plan depends on proofs that a symbiont stayed quiet, and those still need to be built. The paper lists them as open problems. |
|
|
The takeaway
For ten years, we asked: how smart is the AI?
This paper asks a more personal question: how much stronger did it make you? And then a harder one: are you still strong when it is switched off?
In the author’s view, the next decade of AI should be measured not by the machine, but by the person wearing it.
|
|
| Quick facts |
| Paper | From Autonomous Agents to Cognitive Symbionts: Toward AI-Augmented Human Intelligence |
| Author | Hao Fei (University of Oxford) |
| Posted | 3 November 2026 |
| Type | Perspective paper (no experiments) |
| Project page | cognitive-symbiont.github.io |
|
|
|
|
One question for you: would you wear an AI that remembers your whole life, if you could switch it off and prove it?
Hit reply and tell me. I read every answer.
|
ResearchAudio · Hard research papers, in simple English.
Forward this to one friend who thinks about the future of AI.
|