The rarest state

Ask a working scientist how often they are at their best, fully absorbed, learning quickly and seeing connections that were invisible the day before, and the answer is rarely “every day”. Such states tend to be episodic: a few days in a season, separated by long stretches of ordinary work.

For most of history that rhythm was acceptable. It may not remain so. Generative AI now drafts, summarizes and solves faster than any person, and employers already rank analytical thinking, AI literacy, creative thinking and lifelong learning among the fastest-rising skills to 2030 (World Economic Forum, 2025). What is left for humans is the hardest part: framing problems, integrating across fields, judging fluent machine output and learning disciplines that are still being invented. The rare state would need to become the ordinary one.

Why now, and why caution

The question is old, but the answers are new. Randomized trials of AI tutors now exist (Kestin et al., 2025; Bastani et al., 2025; Wang et al., 2024a); non-invasive stimulation can reach the hippocampus without surgery (Violante et al., 2023; Jung et al., 2026); and implanted interfaces restore speech in real time (Card et al., 2024; Wairagkar et al., 2025; Littlejohn et al., 2025).

History counsels humility. Working-memory training was expected to raise intelligence and did not (Melby-Lervåg et al., 2016). “Smart drugs” increase effort while lowering the quality of the solutions people reach (Bowman et al., 2023). The celebrated effect of exercise on cognition in healthy adults nearly vanishes once publication bias is corrected (Ciria et al., 2023). And a drug class that looked protective against Alzheimer’s disease in millions of health records failed in randomized trials (Cummings et al., 2026).

One test for every method

So we put one question to every method: what remains when the tool is taken away? We call it the remainder test. A tutor that raises practice scores but lowers unassisted exam scores fails it (Bastani et al., 2025); a stimulation protocol whose benefit persists a day after the current is switched off passes, at least provisionally (Miyatsu et al., 2024).

Seven borders, five levels

The review is organized around seven borders of the learning mind. Three lie along the path knowledge travels: intake, forgetting and transfer. One governs the learner’s state: attention. Two are set by biology: plasticity and time. The last is the skull, which AI partners and brain–computer interfaces try to cross from outside.

For each border we climb a ladder of five levels of invasiveness: L0 behavioral and environmental methods; L1 digital and AI tools, where nothing enters the body; L2 non-invasive neuromodulation by electric, magnetic, ultrasonic, sensory or nerve stimulation; L3 chemical and biological interventions; and L4 surgical, cellular and genetic interventions.

Peak is timing, not intensity

A popular intuition holds that a better mind is simply a more active one. The evidence suggests otherwise. Waves of activity sweep the cortex every few seconds, and sensory encoding and memory retrieval are each enhanced at different phases of the wave: the brain alternates between facing outward and inward rather than doing both at once (Yang et al., 2026a). Flow, too, depends on a balance of challenge, skill and autonomy rather than maximal effort (Durcan et al., 2024). Peak cognition is the right state at the right time.

How to use this page

• Listen. Press ▶ Play narration at the top of any chapter. The page scrolls with the voice and highlights each block.
• Follow the evidence. Every citation is a link to the original article.
• See the originals. Figures from open-access papers are shown unchanged, with their license. Click any figure to enlarge it.
• Replay. Animated diagrams play when they scroll into view; press ↻ to watch again.

L0 · The oldest methods still set the bar

Two principles have survived decades of testing: spacing study over time and retrieving information from memory rather than re-reading it. Their paradox is that both feel harder and less productive than massed re-reading, so learners systematically underrate them (Carpenter et al., 2022).

Where they have been tested at scale, the benefits are large. Across 13 studies and 21,415 learners in medical education, spaced repetition outperformed standard study with a standardized mean difference of 0.78 (Maye & Hurley, 2026). The effects are domain-dependent, though: spacing helps less in mathematics (Murray et al., 2025), and interleaving helps with paintings but hurts with word lists (Brunmair & Richter, 2019).

The most encouraging finding is that the environment can do the persuading. In two randomized trials, awarding points for each day of practice rather than each question answered spread students’ practice and raised exam scores, most of all for students with the lowest prior grades (YeckehZaare & Resnick, 2025).

L1 · The crutch and the scaffold

Generative AI can deliver spacing, retrieval and feedback to every learner at scale; it can also remove the need to learn at all. When nearly 1,000 high-school students practiced mathematics with GPT-4, practice grades rose, but once the AI was removed, students who had used the standard chat interface scored 17% lower than students who never had it. A version designed to guide rather than tell largely avoided the harm (Bastani et al., 2025).

The same technology, built as a scaffold, accelerates learning. In a randomized trial in university physics, an AI tutor designed on pedagogical principles produced more than twice the learning gains of an in-class active-learning session (Kestin et al., 2025). An AI “co-pilot” that suggested expert moves to 900 human tutors raised their 1,800 students’ mastery by 4 percentage points, and by 9 points for students of the weakest tutors, for about USD 20 per tutor per year (Wang et al., 2024a).

L1 · The brain as an input to teaching

EEG alpha power tracked moment-to-moment cognitive load in 108 adults watching lectures of varying difficulty (Pei et al., 2024); alpha neurofeedback from a single-channel headset improved mathematics learning in 80 students compared with sham feedback (Hou et al., 2024). A tutor that adapted to EEG-measured engagement raised engagement but not short-term learning (Baradari et al., 2025), a reminder that an engaged learner is not necessarily a learning one.

L2 · Learning under stimulation

Stimulation is now applied during real learning, and the most interesting results concern who benefits. In 72 adults learning arithmetic, random-noise stimulation helped mainly those with strong baseline frontoparietal connectivity and a particular excitation–inhibition balance (Zacharopoulos et al., 2025). If neuromodulation has a future in education, it is a personalized one.

Vagus nerve stimulation, reachable through the skin of the neck or ear, gave some of the most practical results: in career linguists at a military language school, cervical stimulation during a five-day vocabulary course accelerated recall, and the gains survived a day without stimulation (Miyatsu et al., 2024).

L2 · Reaching deep without surgery

Temporal interference crosses two high-frequency electric fields so that their slow difference frequency acts only where they meet. Its first human validation steered the focus to the hippocampus and improved episodic-memory accuracy (Violante et al., 2023). Focused ultrasound offers millimeter precision at depth: theta-burst ultrasound to the anterior temporal lobe, a hub for word and concept meaning, improved semantic memory in 23 adults and even increased local gray-matter volume (Jung et al., 2026). Oscillatory electrical stimulation is less consistent: gamma tACS improved recognition a week later in one study (Honma & Nomura, 2026), while theta–gamma tACS added nothing to cognitive training in older adults (Diedrich et al., 2025).

L4 · Writing memory directly

Using electrodes implanted in epilepsy patients for clinical reasons, researchers have played a model of the hippocampus’s own code back into it, improving short-term memory by 37% (Hampson et al., 2018). Content-specific patterns for categories of images changed memory in 22.4% of patient–category combinations, with improvements outnumbering impairments about two to one (Roeder et al., 2024). Closed-loop systems that stimulate only when encoding is failing improved recall (Ezzyat et al., 2018).

L0 · Protecting the state before boosting it

Sleep loss impairs every core executive function, including working memory, inhibition and cognitive flexibility, with medium-to-large effects across 79 publications (Cao et al., 2025). The benefits of the enhancers reviewed below are small by comparison, which makes protecting sleep one of the most effective “cognitive enhancers” available.

Mindfulness is the best-evidenced way to strengthen attention deliberately: across 111 randomized trials and 9,538 participants it improved executive attention, working memory and inhibition, with effects of g = 0.19 to 0.39 against active controls (Zainal & Newman, 2023). Flow correlates moderately with performance (Harris et al., 2021).

L1 · Listening to one’s own brain

A meta-analysis of 15 trials found a small benefit of neurofeedback for attention overall, but none in the eight trials that used sham feedback (Kimura et al., 2024): the benefit seems to come from expectation and engagement rather than the feedback itself. Consumer headsets remain less reliable than research systems for tracking attention in daily life (Ronca et al., 2026).

L2 · A replication reckoning, and one exception

A 2023 meta-analysis of 102 tACS studies reported modest-to-moderate gains (Grover et al., 2023); a 2024 meta-analysis restricted to working memory in 1,386 healthy adults found g = 0.076, falling to zero after bias correction, although stimulation designed to disrupt synchrony reliably impaired memory, so the current does act on the brain (Chuderski & Chinta, 2024). The four most common rTMS protocols showed neither reliable group effects nor reliable individual responses against sham (Passera et al., 2026).

Against this background, transcutaneous auricular vagus nerve stimulation stands out: 53 studies and more than 1,500 participants show a moderate benefit, largest for executive function and cognitive flexibility (Liu & Li, 2026). Temporal interference aimed at frontoparietal cortex improved high-load working memory in a double-blind crossover trial (Zheng et al., 2025).

L3 · More effort, not better thought

In healthy, rested adults, modafinil and methylphenidate produce small effects and d-amphetamine none (Roberts et al., 2020). On complex knapsack problems, all three drugs made people work longer, yet the value of their solutions fell; strong performers often dropped below average as their strategies became erratic (Bowman et al., 2023). Caffeine reliably sharpens attention a little (Kløve & Petersen, 2025).

Sleep is a finely timed program

During non-REM sleep, the hippocampus replays recent experience, and three rhythms must align for that replay to strengthen memory: slow cortical oscillations, spindles nested in their up-states, and hippocampal ripples locked to the spindles. Recording inside the brain and on the scalp, researchers found that spindle-locked ripples mark successful memory reactivation in humans (Schreiner et al., 2024), and that cueing helped some memories more than others, in two distinct waves of reactivation (Duan et al., 2025). In mice, recent and older memories are even replayed in different sub-states of the same sleep (Chang et al., 2025).

Every success works with the timing

Targeted memory reactivation, replaying during sleep a sound or smell that accompanied learning, can strengthen declarative, procedural and emotional memories (Carbone & Diekelmann, 2024). After learning, frontal-midline theta tACS improved recall at two hours, one day and one week (Shtoots et al., 2024).

The most striking demonstration comes from the top of the ladder: deep-brain stimulation in human prefrontal cortex, delivered in synchrony with slow waves in the medial temporal lobe, improved recognition memory, while identical stimulation without precise timing did not help and sometimes hurt (Geva-Sagiv et al., 2023). Timing is not a detail; it is the mechanism. The same lesson may explain a disappointment: closed-loop sound reliably enlarged slow oscillations at home, but produced no measurable memory benefit in 34 participants (Baldassarri et al., 2025).

Reopening the windows of youth

In mice, a shared property of psychedelics was the ability to reopen a critical period for social-reward learning, for a duration proportional to each drug’s acute effects in humans (Nardou et al., 2023). LSD and psilocin bind directly to TrkB, the receptor for brain-derived neurotrophic factor, and their plasticity-promoting effects depend on TrkB rather than on the serotonin receptor that drives hallucinations (Moliner et al., 2023).

Dramatic brain effects, no cognitive gains yet

A single dose of psilocybin disrupted functional connectivity more than three times as much as methylphenidate, with some changes lasting weeks (Siegel et al., 2024); whether this is a window for learning is unknown. Psychoplastogens do not raise blood BDNF (Calder et al., 2024), and controlled microdosing trials show no cognitive benefit over placebo (Murphy et al., 2024; Prochazkova et al., 2026).

The furthest frontier is genetic. In mice, CRISPR-based editing of the epigenetic state of a single gene promoter, Arc, in the neurons that stored a memory was enough to switch that memory’s expression on or off, reversibly (Coda et al., 2025). No comparable human work exists, and even heritable editing for disease lies decades away (Visscher et al., 2025).

Keeping the learning brain for a lifetime

Exercise improves general cognition, memory and executive function across ages, according to an umbrella review of 2,724 trials (Singh et al., 2025). Yet a critical re-analysis restricted to healthy adults found that the effect shrank to a non-significant 0.05 after correcting for publication bias (Ciria et al., 2023). Mediterranean-style diets are associated with lower dementia risk (Fekete et al., 2025), and the Lancet Commission attributes up to 45% of dementia cases to 14 modifiable risks, including less education: learning builds the reserve that protects learning (Livingston et al., 2024).

Biological maintenance, mostly in animals

A single low dose of the longevity protein klotho improved memory in aged monkeys (Castner et al., 2023); the platelet factor PF4, abundant in young blood, restored cognition in aged mice (Schroer et al., 2023); a cheap prebiotic improved cognitive scores in a twin trial of adults over 60 (Ni Lochlainn et al., 2024). Forty-hertz light and sound cleared amyloid in mice (Murdock et al., 2024), but did not entrain gamma rhythms or reduce amyloid in independent work (Soula et al., 2023).

A cautionary tale, and a clinical reality

Meanwhile, replacing lost neurons has become a clinical reality, for restoration: dopaminergic cells grown from embryonic or induced pluripotent stem cells were transplanted into people with Parkinson’s disease without serious adverse events, and several improved (Tabar et al., 2025; Sawamoto et al., 2025).

L1 · AI as a partner in thought

Clark (2025) argues that it is our nature to build hybrid thinking systems that fluidly incorporate non-biological resources. AI increasingly acts as a “System 0” that pre-processes information before human intuition and deliberation begin (Chiriatti et al., 2024). The risk is that extension becomes replacement.

In an EEG study, 54 people wrote essays with an LLM, a search engine or no tools; LLM users showed the weakest brain connectivity and struggled to quote their own essays (Kosmyna et al., 2025). The study is a small preprint, but its message matches survey evidence: frequent AI use was associated with lower critical thinking through cognitive offloading, most strongly in the young (Gerlich, 2025), and knowledge workers who trusted AI more reported thinking less critically (Lee et al., 2025).

L4 · Opening the skull

In a man with ALS, 256 electrodes in speech motor cortex decoded attempted speech with 97.5% accuracy over a 125,000-word vocabulary, in a synthetic version of his pre-illness voice (Card et al., 2024). Newer systems stream speech in 80-millisecond steps (Littlejohn et al., 2025) and reproduce intonation well enough to sing short melodies (Wairagkar et al., 2025). Hardware is becoming less invasive: a 1,024-channel thin-film array can be placed without a craniotomy (Hettick et al., 2025), and an electrode delivered through a blood vessel let paralyzed patients control computers with no serious device-related adverse events over a year (Mitchell et al., 2023). Still, the literature documents only about 80 implanted-BCI users worldwide (Dohle et al., 2025). Every system restores; none enhances.

Thinking tissue

“Organoid intelligence” proposes to use stem-cell-grown brain organoids, interfaced with electrode arrays, as computing substrates, with ethics embedded from the start (Smirnova et al., 2023); because consciousness in such systems cannot yet be defined, its proponents recommend measuring intelligence rather than consciousness (Caraballo-Bosch et al., 2025).

The border that matters most

Preparing people to learn sciences that do not yet exist is a problem of far transfer: whether gains carry into domains that were never trained. Across 87 publications, working-memory training produced gains on similar tasks but no reliable improvement in reasoning, reading or arithmetic compared with treated controls (Melby-Lervåg et al., 2016). A synthesis of cognitive-training and bilingual-advantage meta-analyses found both indistinguishable from zero after correction for publication bias (Paap et al., 2024).

Three leads that keep the door ajar

Bilingualism may improve cognition not by transferring skills but by making attention more efficient (Bialystok, 2024). Music training improves executive function in children, mostly inhibitory control (Cai et al., 2025). And action video games appear to improve attentional control and, through it, the rate at which new tasks are learned, an ability to learn how to learn (Bavelier & Green, 2024), with small but consistent associations with cognitive ability across 133 studies (Zhao et al., 2026). What transfers may not be knowledge or skill, but the efficiency of attention and the speed of new learning.

Where optimism needs discipline

Effects shrink as designs tighten, and observational promise can vanish in trials. No stimulation, drug or implant in this review has demonstrated far transfer to a genuinely new domain.

The ledger

The largest and most precise effects are behavioral or digital; among non-invasive stimulation methods only auricular vagus nerve stimulation shows a moderate, precise effect; and drug effects in healthy adults are small.

Readiness in five dimensions

Read column by column, the table tells a coherent story. Benefit does not rise with invasiveness: the best-evidenced, safest and cheapest gains sit at the bottom of the ladder. The transfer column is the weakest in every row. Restoration leads enhancement by a decade or more (Card et al., 2024; Tabar et al., 2025). And the governance column reminds us that readiness is social as well as technical.

Crossing well: for teachers and institutions

Crossing well: neurotechnology, drugs and rights

Approach neuro-products with informed caution: neurofeedback benefits largely vanish against sham (Kimura et al., 2024), consumer EEG is not yet a reliable attention monitor (Ronca et al., 2026), and stimulants trade quality of thought for effort (Bowman et al., 2023). Critical perspectives on “neuropedagogy” deserve a hearing (Williamson et al., 2025). Whether enhancement counts as cheating depends on what education is for (Huttunen, 2025), and a freedom to enhance can become a pressure to enhance (Istace & Van Assche, 2026).

Neural data may reveal mental states, and rights to mental privacy and integrity are being given legal form (Ligthart et al., 2023; Szoszkiewicz & Yuste, 2025): Chile’s Supreme Court ruled on brain data in 2023 (Cornejo-Plaza et al., 2024), Colorado and California treat neural data as sensitive (Colorado General Assembly, 2024; California Legislature, 2024), and UNESCO adopted the first global Recommendation on the Ethics of Neurotechnology in 2025 (UNESCO, 2025; OECD, 2024).

What comes next

2026–2030: convergence. Learning science, recovery and scaffolding AI, with brain-state-aware systems entering research and professional training (Pei et al., 2024; Baradari et al., 2025). The defining skill is learning with AI without outsourcing understanding.

2030–2035: personalized neuromodulation, if replicated. If independent, preregistered trials confirm vagus nerve stimulation, temporal interference or focused ultrasound for learning (Liu & Li, 2026; Zheng et al., 2025; Jung et al., 2026), and markers can guide who benefits (Zacharopoulos et al., 2025), neuromodulation may become an adjunct to intensive training (Tyler et al., 2025). The null results make this far from certain (Passera et al., 2026; Diedrich et al., 2025).

Beyond 2035: restoration first. Implants, psychoplastogens and epigenetic editing will expand as restorative medicine long before any use in healthy learners. No evidence supports a timetable for “knowledge upload” or genetic enhancement of intelligence.

The experiment the field needs

Participants learn an unfamiliar, formally specified discipline with an active-learning baseline, AI assistance, or an intervention combined with training. After a delay and with all tools withdrawn, they are assessed on retention, transfer to novel problems, calibrated uncertainty and reasoning under fatigue, alongside well-being. Studies should report comparators, blinding and adverse effects, following checklists such as RATES (Nejati et al., 2025).

The mind we can build

We began with the rare peak state and asked whether it could become ordinary. The answer is a qualified yes, but not in the way popular imagination expects. Effort beats ease: the strongest gains come from methods that make the learner work, and from AI that scaffolds that work. Timing beats force: consolidation, stimulation and peak cognition itself depend on acting in step with the brain’s rhythms. And the remainder is what counts.

The learner of the coming decade is not a cyborg but a disciplined, well-rested person working with a tutor that asks rather than tells, perhaps with a small device at the ear if replication allows. Implants and genes will transform the lives of patients long before they change the classroom. And one border remains almost untouched: transfer into the unknown, the very capacity an AI-shaped world will demand most.