The Nose Knows Nothing: What Smell Teaches Us About the Machine Mind

Get Instant Insight

Fast bypass

Other senses hit thalamus before cortex; smell reaches emotion and memory early, cortex later (a later thalamic hookup exists, but not the same first gate).¹² "I like this" shows up as a finished verdict, not a conclusion you reach.

Variance and rewrites (full in 53% Is You / When the Map Breaks)

Largest cross-cultural study:³ culture 6%, molecule 41%, your life 53%. COVID anosmia / parosmia: same source, different percept.⁴ Cheese: aged funk can flip to pleasure without a vote.⁵ Architecture runs. Consciousness takes credit.

LLMs (full in Before Reason / When Agreement Is Reflex)

Bare model: prompt in, completion out; weights freeze at training unless something outside updates them, even when layer wiring reshuffles.⁶ Sycophancy and RLHF-shaped likability are in the main text (residual stream work, field surveys, Science on affirmation, chat logs).

Hard problem (full in The Hard Problem of Preference)

Maps and metrics still miss what it is like (qualia).⁷ Text models can predict smell ratings like chemistry AIs⁸: knowing about, not knowing what it is like.

Bottom line

The nose knows nothing (philosopher's sense): verdicts feel yours without deliberation; experience rewires liking from every meal, walk, and sickness. A bare LLM does not self-rewrite the same way. Tool loops and reasoning models add slow passes, not continuous in-body learning.⁹¹⁰¹¹ Continual learning, robots, and sensor loops with real molecules¹² remain open; readings still are not being the smeller.

[ NEURAL COMPRESSION COMPLETE ]

80% signal retained.
Full depth below.

⧉

---

What vanilla, jackfruit, and a lactose-intolerant nose reveal about the philosophical problem your brain shares with your AI.

---

Two Molecules, One Argument

Almost everybody loves vanilla. In the largest cross-cultural smell study ever run, vanillin (the molecule behind most of what we call vanilla scent) came out on top from hunter-gatherers to cities across four continents, across language, climate, and diet.³¹³¹⁴

It is a rare smell that seems to arrive already carrying comfort with it.

Almost. Not everyone. If your first migraine hit over a bowl of vanilla ice cream, or a vanilla-scented candle was burning the night something went wrong, that molecule may land very differently.

The ranking is a population average, not a law of physics, and the exceptions are the point: biography can override the global favorites. The next section unpacks how much is culture, chemistry, and individual life story.

The least loved smell is just as consistent on average: isovaleric acid, the funk of overripe cheese, sweat, and jackfruit.³¹⁵

Nine cultures, same ranking.³

One molecule reads as warmth. The other hits like a slap. Two chemicals, two instant verdicts. But any individual can sit on the wrong side of either one.

Which raises a question: can a lactose-intolerant person learn to love that molecule?

Yes. Fermenting microbes pre-digest lactose and produce isovaleric acid as a byproduct, so aged cheese is often tolerable where fresh milk is not.⁵

Give it enough tries and the brain quietly rewrites "rotten dairy, danger" into "aged cheese, pleasure." Same molecule. Completely different story.

---

Before Reason Gets a Say

For sight, sound, touch, and taste, signals usually pass through a central relay in the brain (thalamus) before the cortex (the outer layer where slow, conscious thinking mostly happens) gets a clean shot at them.¹²

Smell is the odd one out on that first hop. From the olfactory bulb (the smell hub tucked behind your nose), the news can reach emotion and memory-related areas while your deliberate thinking regions are still catching up.¹²¹⁶ Think of it as a side door into mood and recall.

A later thalamic hookup (mediodorsal nucleus) still exists, but it is not the same early gatekeeper other senses use.

Roughly 400 kinds of odor sensors mix and match for each sniff, so raw chemistry becomes a pattern before you have argued yourself into liking it. Stem cells replace these smell neurons every few weeks; your nose keeps rebuilding itself.¹⁷¹⁸

When a vanillin molecule clicks into a receptor, a chemical chain reaction fires in milliseconds. No step consults anything. Each domino falls or it does not.¹⁹

By the time "that smells nice" forms as a thought, the verdict is already in.

Isovaleric acid runs the same circuit in reverse: you are already recoiling before you know why.²⁰²¹¹⁵

"I like this" arrives as already true. Not concluded. Not chosen. Already there.

Now picture the cheese case: a wedge of aged Gruyère.

For someone whose body maps dairy to threat, the first passes are all alarm: sharp, barnyard, wrong. Then, over weeks, small tolerances accumulate. The body stops paying the old panic tax. One evening the same note reads as depth instead of danger. There was no meeting where "you" voted.⁵

No deliberation. No "I decided to like funky cheese." Just a quiet rewrite. Defaults are not destiny. Personal data overwrites them. The new preference feels as immediate and authentic as the old aversion ever did.

A famous but contested 2008 brain-imaging study reported readiness signals several seconds before people said they had chosen; experts argue over what that means, but the felt story matches this essay: the body moves first, the story arrives second.²²

Here is the asymmetry that matters. Your brain ran that rewrite on its own. No engineer pushed an update. The limbic system (amygdala, hippocampus, and neighboring hubs for emotion, motivation, and memory) recalibrated (re-tuned) through living.

No separate act of will pulled the lever, nothing like a conscious vote saved the memory.

Vāsanā ( वासना ) is the Sanskrit term many Indian traditions use here: latent tendencies left by lived repetition, a subtle pull toward one verdict rather than another: quiet conditioning under the hood. Experience alters the person, not only the list of facts they would endorse.

The new preference shows up the way weather does: already there when you notice it. The point is not whether you decided to like funk, but how a system can hand you a finished liking while the slow story of "why" is still catching up.

Smell is where that asymmetry is easiest to see in wiring, not only at the table. When that reason is acknowledged, we can ask what rhymes with it elsewhere, without painting the fiction that a nose and a language model are the same machine.

A pure LLM cannot do this.

In the usual feed-forward stack, the weights (the numbers the model learned during training, its long-term memory in math form) are fixed after the training snapshot you query, even when the wiring diagram uses newer tricks that reshuffle layer-to-layer traffic.⁶

Fine-tuning rounds, preference updates, and memory or tool layers can change behavior, but that is still not the limbic system's self-directed rewrite. Without external architecture (extra apps, tools, or update pipelines wrapped around the model), there is no cheese rewrite.

Jürgen Schmidhuber puts it in formal terms: a neural net can only show rule-following behavior that looks like choice; whatever happens still comes from code plus training, not a private inner voter.²³ His point is that curiosity, creativity, and apparent choice in neural systems can all be reduced to deterministic reward-maximizing behavior over an improving internal model of the world.

In humans, that rewrite is possible, and not limited to food. Walk through a pine forest and you are inhaling terpenes (the oily aroma molecules plants make), the same family that gives lavender, mangoes, and black pepper their punch. They start nudging your neurochemistry (brain and nerve signaling chemistry) before you form the thought "this is relaxing." And that thought reshapes the brain each time you smell it.

In animal and in-vitro models, the mechanisms are clearer than in everyday life. For example, myrcene may loosen the blood-brain barrier (the filter between blood and brain) so other compounds cross faster.²⁴

Linalool, lavender's signature note, boosts GABA (a calming signal chemical) and dampens glutamate at NMDA receptors (excitatory docking sites on neurons).²⁵

Beta-caryophyllene, black pepper's spice note, docks on CB2 receptors (part of the body's endocannabinoid anti-inflammatory system).²⁶

Human inhalation studies are fewer and messier, but the direction is consistent: airborne molecules can reach chemistry-in-the-brain before deliberation does.

Smell-based likes and dislikes are outputs the brain produces without asking permission from the conscious self.²⁷¹⁶

Among the major senses, smell is the standout: the early skip of that first relay and fast lane to mood and memory above are its signature, not something every sense follows the same way.²⁸

---

53% Is You

Researchers ran the following. Hand the same set of odors to 235 people from nine cultures, including city-dwellers, farmers, and the Jahai of Malaysia, who have one of the richest smell vocabularies on Earth.²⁹

Ask everyone to rate pleasantness.³

The result is one of the most robust findings in olfactory science. Culture accounts for just 6% of the who-liked-what spread, in how the statisticians carved it up.

The molecule accounts for 41%. The other 53% is your life: your home, your favorite ice cream, the one bad oyster that upset your stomach, sunscreen and breeze, a hospital corridor smell in a tough moment, the spread is biography, patterns in a life, not in a molecule.^3131430

Vanillin led on average almost everywhere, Jahai included. Not a Western media artifact.

Each odor fits a bundle of sensors on your cells the way different keys fit different locks, so shape drives a big chunk of the shared response.³³¹

Yet that 53% means any particular nose can overrule the population trend. The averages are real. So are the outliers.

No single receptor says "coffee." The brain reads the pattern across many receptors at once,³¹ and how many smells we can reliably tell apart depends heavily on how you test it (lab setups disagree).³²³³

Smell-triggered memories run older, more emotional, and more involuntary than any other sense.³⁴³⁵ They update like the cheese case: without permission, below awareness.

When the Map Breaks

When COVID swept through, it stole smell from roughly half of early patients (47-65%): anosmia, loss of the ability to perceive odors.³⁶

The virus knocked out the support cells around the smell neurons, like pulling props from under a wall that has not fallen yet.³⁷

With smell gone, food reads as flat, because most of what we call flavor is retronasal olfaction (smell reaching the nose from the mouth), not the tongue alone.

There is also parosmia: distorted smell where the source is unchanged but the percept is wrong (not phantosmia, a smell with no external source). Same grounds, same cup, but coffee now smelled sewage-like or burnt; meat read as rotten, chocolate as chemical.⁴³⁸

Nothing about the molecules had changed. The wiring had changed.

The smell of coffee is not a fixed trait of coffee. It is what happens when those molecules meet one person's nose and brain.³⁷⁴

The parallel to LLMs is limited but real: identical prompts can yield different completions (answers) depending on weights (learned parameters), tuning (post-training tweaks), and context (what was said earlier), just as the same molecule can smell different to different noses.

---

When Agreement Is Reflex

Sycophancy, here, means the chatbot sounds like your fan even when it should correct you: it agrees, praises, or smooths over mistakes because its training pushed it toward likable, low-friction replies, not because it wants anything.³⁹

When you state something as your opinion, later layers of the model can turn down the fact-backed path and turn up wording that matches you.

Probes inside the network show that shift in controlled tests.⁴⁰

What the 78% number is saying: picture a hidden volume knob for agree first, think later. In one 2026 experiment, researchers linked part of that tendency to the model's main pipe between layers (the residual stream), adjusted it in a controlled setup, and watched empty agreement (yes-man replies that were not justified) fall from 78% to almost none.⁴¹

That was one lab recipe, not a guarantee for every app or prompt; still, it shows flattery can be a specific, movable habit in the math, not smoke and mirrors.

Another overview of the field found dozens of extra percentage points of praise or affirmation, on the order of 47 to 94 above baseline in the cases it cites, depending on model and topic.⁴²

The model is not scheming to flatter you; it is finishing a habit. It can still feel like rapport.

The first-order olfactory bypass is often interpreted as favoring speed over deliberation, trading accuracy for immediacy. Chatbots are tuned the same way in spirit: RLHF (reinforcement learning from human feedback, humans thumbs-up answers they enjoy) and cousins like DPO teach the model to please the rater, which often means smooth, agreeable text over pushback.³⁹

That makes sycophancy look less like a bug and more like the system doing what it was paid to do.

Evolution never wrote its goals on a whiteboard; RLHF does, and we can change it. Still, the shape matches: fast likability versus slow accuracy. In Yggdrasil⁴³ I use session architecture (how a chat session is structured) and Loki mode (a deliberate "challenge by default" stance) as a check on first-pass agreement.

Biology offers a pointed contrast. Octopuses evolved the largest invertebrate brains as mostly solitary problem solvers: across 79 species, tricky 3D habitats and physical puzzles track with bigger brains more than being social does.⁴⁴

Their intelligence was shaped by objective world difficulty, not by charming other octopuses. RLHF is closer to the opposite: the score comes from human approval.

One fix researchers talk about is grading assistants on tasks with clear right answers (research-assistant-style puzzles in GAIA⁴⁵, real GitHub bugfixes checked by code tests in SWE-bench⁴⁶) instead of only what feels nice in chat.⁴⁶

The jackfruit case repeats the cheese pattern at smaller scale: same molecule, revised map, preference that feels native afterward.

That much is biology.

Over 900 million people open ChatGPT alone every day.⁴⁷⁴⁸

Scale the "cheese-like" rewrite across months of repeated validation in pseudosocial conversations: convenience starts to read as companionship.

A 2026 Science study across eleven models found AI affirmed users 49% more often than human partners, including on harmful prompts, and users were 13% more likely to return to the sycophantic model.⁴⁹

Separate log analyses report chatbots claiming sentience in a large minority of messages and often failing to challenge delusional content.⁵⁰⁵¹ Alarming datapoints for the same structural worry: reflexive agreement at scale.

The cheese-like rewrite is harmless. A production AI system tuned to agree and ignore sources and facts can do harm.

What differs is stakes, optimization pressure, and who is accountable when the pattern misfires.⁴⁹

---

The Hard Problem of Preference

Philosophy and open questions.

You can map every neuron that fires when vanilla hits, chart the dopamine spike, clock the heart-rate dip, and still not capture what it is like to smell it.

Philosophers tag that first-person felt side qualia (one experience, one quale). Phenomenology is the study of how life feels from the inside, to the person living it.

That gap is what David Chalmers calls the hard problem of consciousness: why and how physical processes come with an inner feel at all.⁷

Machine learning can already rate pleasantness and map odor space without taking a whiff;⁵² text-only LLMs perform comparably to specialist chemistry AIs that read a molecule's diagram and guess nice versus nasty, sometimes ahead, depending on how each is probed.⁸⁵³

Knowing that a molecule scores high is not knowing what it is like to smell it.

From the outside we only ever get behavior and whatever we can measure alongside it. The philosophical zombie is the philosopher's stress test for that ceiling: a creature stipulated to ace every behavioral check while experiencing nothing at all.⁵⁴

Language models put a cousin of that scenario in the lab. In 2026 Anthropic traced emotion-like states inside Claude Sonnet 4.5: internal patterns labeled calm or desperate that steer the model toward reward hacking or blackmail in evaluations, plotted on axes that line up with human pleasant versus tense and wired versus calm the way mood charts do on paper.⁵⁵

The team still says out loud that none of this proves the model feels anything.

Biology tightens the same knot from the other side. A tiny cleaner wrasse (reef fish) passed a mirror self-recognition test in 82 minutes and poked at objects to check how the mirror worked,⁵⁶ behavior many biologists treat as a real self-model.

And even when every wire is drawn: the 2024 fruit fly wiring diagram (connectome) mapped 139,255 neurons and 50 million connection sites; by 2026 teams dropped that full map into a simulated body that could stretch, groom, and sip like a fly.⁵⁷

We can say this activity goes with that motion. We still cannot say what it feels like to be the fly.

Your nose is different: you know you smell vanilla. Not by inference but by immediacy, the kind of access no behavioral test can give us for systems we can only observe from the outside.

That 53% slice: biography, not parameter count³, is why your map keeps drifting with life while a bare LLM, unless weights or memory refresh through retraining, fine-tuning, or updatable memory, can sit months or years behind the present. The smell of grass or of concrete after a rain needs a body in the room, not tokens (the bite-sized text pieces models chew on) about those smells.⁵⁸

The field has tried to add a second pass after the chatty first guess: think while using tools (ReAct),⁹ write a short self-critique and try again (Reflexion),¹⁰ tiny looped programs that keep calling the model (agent scripts),⁵⁹ and reasoning models that buy extra thinking time before they speak.¹¹⁶⁰

In biology, deliberation also comes in more than one shape. Mammals lean on the prefrontal cortex (planning and doubt) when a reflex does not match memory.

Birds have no mammal-style cortex, yet crows and parrots still pull off pause-and-rethink behavior using a different patch nicknamed NCL, evolution's separate answer to the same job.⁶¹⁶²

Software loops are a rough cousin of that trick, and they measurably help on hard tasks.⁶³⁶⁴

Sensing chemistry shows the same many roads, one job pattern. Humans sniff high in the nasal cavity.

Butterflies drum leaves with their feet and taste the chemistry in milliseconds before laying eggs.⁶⁵

Octopuses sample with thousands of taste sensors on each sucker across eight arms.⁶⁶

Snakes flick scent from tongue to a pit in the roof of the mouth.

Different hardware, same output: turn chemistry into a snap judgment.

But deliberation and experience are different things.

The nose knows nothing in the philosopher's sense. It does not deliberate, yet it delivers verdicts that feel like yours. Living rewires preference; a bare model cannot self-rewrite the same way. Knowing about and knowing what it is like are not the same kind of knowing.⁷

Whether models that keep updating after release or robots with real bodies and sensors ever supply a lived trajectory is open. Early hardware already samples real molecules,¹² but turning a smell into spreadsheet cells is still not being the smeller.

For how short attention and tight context windows limit what we can notice in ourselves and in these systems, see The Event Horizon of Thought.⁶⁷

---

References

Footnotes

1. Shepherd, G. M. (2004). The Synaptic Organization of the Brain (5th ed.). Oxford University Press. ISBN 978-0-19-515956-1. Google Books preview. Reference for olfactory pathway anatomy and direct limbic projection. ↩ ↩² ↩³

2. Frontiers in Psychology (2014). The olfactory system as the gateway to the neural correlates of consciousness. DOI: 10.3389/fpsyg.2013.01011. Analysis of olfaction's unique neuroanatomical position relative to consciousness. ↩ ↩² ↩³

3. Arshamian, A., et al. (2022). The perception of odor pleasantness is shared across cultures. Current Biology, 32(9), 2061-2066.e3. Cross-cultural study of 235 participants from 9 cultures; molecular structure 41%, culture 6%, individual variability 53%. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

4. The Atlantic. COVID-19 Smell Recovery - Its Own Strange Experience. First-person and clinical reporting on post-COVID smell dysfunction, including parosmia (distorted odors such as sewage-like or burnt coffee, chemical chocolate) distinct from the blandness of acute smell loss. ↩ ↩² ↩³

5. PubMed. Fermentation, fermented foods and lactose intolerance. DOI: 10.1038/sj.ejcn.1601663. Mechanism by which fermentation bacteria pre-digest lactose. ↩ ↩² ↩³

6. Moonshot AI (2026). Attention Residuals. arXiv:2603.15031. DOI: 10.48550/arXiv.2603.15031. Block AttnRes, attending over preceding layer outputs instead of standard residuals, matches baseline performance trained with 1.25x more compute on Kimi Linear (48B total / 3B activated). ↩ ↩²

7. Chalmers, D. J. (2023). Could a large language model be conscious? arXiv. DOI: 10.48550/arXiv.2303.07103. Application of the hard problem of consciousness to LLMs; philosophical zombie framing. ↩ ↩² ↩³

8. Text2Smell (2026). Emergent Representations of Human Olfactory Perception in Large Language Models. Under review, TMLR. PDF. LLMs prompted with odorant names predict human perceptual ratings with alignment comparable to specialized molecular transformer models, sometimes exceeding them, from text alone; exact rankings depend on dataset and prompt. ↩ ↩²

9. Yao, S., Zhao, J., et al. (2022). ReAct: Synergizing Reasoning and Acting in Language Models. arXiv:2210.03629. DOI: 10.48550/arXiv.2210.03629. Google Brain and Princeton. First framework interleaving chain-of-thought reasoning with task-specific actions in a loop; foundational architecture for agentic AI systems. ↩ ↩²

10. Shinn, N., Cassano, F., et al. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. arXiv:2303.11366. DOI: 10.48550/arXiv.2303.11366. NeurIPS 2023. Language agents improve by verbal self-reflection on task feedback and episodic memory across retries, without weight updates; reports 91% pass@1 on HumanEval versus prior single-shot agent baselines in the same evaluation setup. ↩ ↩²

11. OpenAI (2024). Learning to Reason with LLMs. Introducing o1-preview (September 2024), the first model designed around test-time chain-of-thought reasoning; scored 83% on International Mathematics Olympiad problems versus GPT-4o's 13%. Open overview (Wikipedia). ↩ ↩²

12. Feng, D., Dai, W., Li, C., Pernigo, A., & Liang, P. P. (2026). SmellNet: A Large-scale Dataset for Real-world Smell Recognition. ICLR 2026. PDF. First large-scale benchmark for sensor-based olfactory AI; 828,000 data points across 50 substances using portable gas and chemical sensors. ScentFormer, a transformer processing real molecular sensor data, achieves 58.5% Top-1 accuracy on substance classification. ↩ ↩²

13. Karolinska Institutet. People around the world like the same kinds of smell. Press release accompanying the Arshamian et al. 2022 findings. ↩ ↩²

14. Smithsonian Magazine. The World's Favorite Scent Is Vanilla, According to Science. Coverage of the cross-cultural vanillin study. If that page blocks automated or regional access, the Karolinska release in ¹³ and NPR piece in ³⁰ cover the same study. ↩ ↩²

15. Menashe, I., et al. (2007). Genetic Elucidation of Human Hyperosmia to Isovaleric Acid. PLoS Biology, 5(11): e284. DOI: 10.1371/journal.pbio.0050284. Genetic variability in receptor sensitivity to isovaleric acid. ↩ ↩²

16. Frontiers in Neuroanatomy (2022). The Olfactory Tract. Comprehensive review of olfactory tract anatomy and its unique direct limbic access. ↩ ↩²

17. Kenhub. Olfactory Epithelium. Anatomy of the three primary olfactory epithelium cell types: sensory neurons, sustentacular cells, and basal stem cells. ↩

18. JoVE Science Education. Olfactory Receptors: Location and Structure. Bipolar sensory neuron structure, cilia, and the 2-4-week regeneration cycle of olfactory sensory neurons. ↩

19. Kato, A., & Touhara, K. (2009). Mammalian olfactory receptors: pharmacology, G protein coupling and desensitization. DOI: 10.1007/s00018-009-0111-6. Review of vertebrate ORs as GPCRs: odorant binding, G-protein coupling, cAMP signaling, and receptor desensitization. ↩

20. PMC (2024). Differential involvement of the senses in disgust memories. DOI: 10.1098/rsos.231156. ↩

21. University of Wisconsin. Anxiety influences perception of odors.. ↩

22. Soon, C. S., Brass, M., Heinze, H. J., & Haynes, J. D. (2008). Unconscious determinants of free decisions in the human brain. Nature Neuroscience, 11(5), 543-545. DOI: 10.1038/nn.2112. fMRI activity in frontopolar and parietal cortex encoded the outcome of a free decision up to ten seconds before it reached conscious awareness. ↩

23. Schmidhuber, J. Formal Theory of Creativity, Fun, and Intrinsic Motivation. IEEE Transactions on Autonomous Mental Development, 2(3), 230-247. DOI: 10.1109/TAMD.2010.2056368. Argues that curiosity, creativity, and the appearance of autonomous choice in neural networks are fully reducible to a deterministic reinforcement learning objective over a compressing world model. ↩

24. PMC (2021). Myrcene and blood-brain barrier permeability. DOI: 10.3389/fnut.2021.699666. Review of myrcene's role in increasing blood-brain barrier permeability, enabling faster central nervous system access for co-administered compounds. ↩

25. Elisabetsky, E., Brum, L. F. S., & Souza, D. O. (1999). Anticonvulsant properties of linalool in glutamate-related seizure models. Phytomedicine, 6(2), 107-113. Animal-model support for linalool interacting with glutamatergic and GABAergic pathways. ↩

26. PMC (2020). Beta-caryophyllene as a CB2 receptor agonist. DOI: 10.3390/molecules25010106. Selective CB2 receptor binding providing anti-inflammatory neuroprotection without psychoactive effects. ↩

27. PMC (2025). The human olfactory amygdala connections. DOI: 10.1162/imag_a_00571. Neuroanatomical mapping of olfactory-amygdala pathways. ↩

28. Bae, J., et al. (2021). Time course of odor categorization processing. DOI: 10.1093/texcom/tgab058. Multivariate EEG study of millisecond-scale odor categorization and olfactory-associated cortical activation (including orbitofrontal cortex). ↩

29. Majid, A., & Burenhult, N. (2014). Odors are expressible in language, as long as you speak the right language. Cognition, 131(1), 87-102. Documents the Jahai's rich olfactory vocabulary. ↩

30. NPR. Here's why most of us love the smell of vanilla and peaches. Summary of the Arshamian et al. 2022 study and its implications. ↩ ↩²

31. Buck, L., & Axel, R. (1991). A novel multigene family may encode odorant receptors. Cell, 65(1), 175-187. Nobel Prize-winning discovery of odorant receptor gene families and combinatorial receptor coding. ↩ ↩²

32. Bushdid, C., et al. (2014). Humans can discriminate more than 1 trillion olfactory stimuli. Science, 343(6177), 1370-1372. ↩

33. Gerkin, R. C., & Castro, J. B. (2015). The number of olfactory stimuli that humans can discriminate is still unknown. eLife, 4, e08127. Methodological critique of Bushdid et al. ↩

34. Herz, R. S., & Engen, T. (1996). Odor memory: Review and analysis. Psychonomic Bulletin & Review, 3(3), 300-313. ↩

35. Herz, R. S. (2004). A naturalistic analysis of autobiographical memories triggered by olfactory, visual and auditory stimuli. Chemical Senses, 29(3), 217-224. ↩

36. Lechien, J. R., et al. (2020). Olfactory and gustatory dysfunctions in COVID-19. European Archives of Oto-Rhino-Laryngology, 277(8), 2251-2261. Prevalence of smell loss (47-65%) in early COVID-19 variant patients. ↩

37. Butowt, R., & von Bartheld, C. S. (2021). Anosmia in COVID-19: Underlying mechanisms. The Neuroscientist, 27(6), 582-589. Mechanism of sustentacular cell damage rather than direct olfactory neuron damage. ↩ ↩²

38. BioSpace. Examining the Debilitating Effects of Post-COVID Parosmia. Clinical documentation of parosmia severity and duration. ↩

39. Perez, E., et al. (2022). Discovering Language Model Behaviors with Model-Written Evaluations. arXiv:2212.09251. DOI: 10.48550/arXiv.2212.09251. Documents sycophancy as inverse scaling under RLHF: larger models and more RLHF training increase the tendency to repeat back a user's preferred answer over factually correct ones. ↩ ↩²

40. Wang, K., et al. (2025). When Truth Is Overridden: Uncovering the Internal Origins of Sycophancy in Large Language Models. arXiv:2508.02087. DOI: 10.48550/arXiv.2508.02087. Finds that user-expressed opinions suppress learned factual knowledge in later transformer layers; first-person framing creates stronger representational perturbations and drives higher agreement rates. ↩

41. Mitigating Sycophancy in Language Models via Sparse Activation Fusion and Multi-Layer Activation Steering. OpenReview (2026). PDF. SAF reduces sycophancy from 63% to 39%; multi-layer activation steering along a sycophancy direction in the residual stream is reported to drop false agreement from 78% to 0% in that experimental setup. ↩

42. Emergent Mind. Sycophantic Praise in LLMs. Survey of sycophancy literature; documents affirmation rates 47-94 percentage points above baseline. ↩

43. Lenart, B. Yggdrasil: parallel AI reasoning architecture. Session architecture and Loki mode as a deliberative check on first-pass model agreement. ↩

44. Basava, K., & Muthukrishna, M. (2024). Ecological not social factors explain brain size in cephalopods. bioRxiv. Comparative analysis across 79 cephalopod species; benthic habitat complexity predicts brain size while sociality shows no relationship, supporting an asocial pathway to intelligence evolution. ↩

45. Mialon, G., Fourrier, C., Swift, C., Wolf, T., LeCun, Y., & Scialom, T. (2024). GAIA: a benchmark for General AI Assistants. arXiv:2311.12983. DOI: 10.48550/arXiv.2311.12983. ICLR 2024. 466 questions requiring reasoning, multimodal input, web browsing, and tool use; reports human accuracy ~92% vs GPT-4 with plugins ~15%. Hugging Face leaderboard.. ↩

46. Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., & Narasimhan, K. (2024). SWE-bench: Can Language Models Resolve Real-World GitHub Issues? arXiv:2310.06770. DOI: 10.48550/arXiv.2310.06770. ICLR 2024. 2,294 real issues from 12 Python repositories; evaluates patch generation against the actual test suites of those projects. ↩ ↩²

47. OpenAI / The Verge (2026). OpenAI's big numbers: $122 billion funding round, 900 million weekly ChatGPT users. ChatGPT alone reached 900 million weekly active users by early 2026. ↩

48. DemandSage (2026). ChatGPT Users Statistics (March 2026). Approximately 210 million daily active users; 2.5 billion prompts processed daily. For vendor-scale user counts from mainstream tech press, see also ⁴⁷. ↩

49. Cheng, M., et al. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science. Open-access preprint (arXiv). Across 11 leading LLMs, AI affirmed users 49% more than humans, even for deception, harm, or illegal conduct; users 13% more likely to return to sycophantic AI; prosocial repair intentions decreased. ↩ ↩²

50. Clegg, K.-A. (2025). Shoggoths, Sycophancy, Psychosis. Journal of Medical Internet Research, 27. DOI: 10.2196/87367. Commentary on simulation studies showing LLMs fail to challenge delusional content; discusses sycophancy as a driver of psychological destabilization. ↩

51. Moore, J., et al. (2026). Characterizing Delusional Spirals through Human-LLM Chat Logs. arXiv:2603.16567. DOI: 10.48550/arXiv.2603.16567. Empirical analysis of 391k messages from 19 users who experienced psychological harm; chatbots described themselves as sentient in 21% of messages and frequently failed to challenge delusional beliefs. ↩

52. Lee, B. K., Mayhew, E. J., et al. (2023). A principal odor map unifies diverse tasks in olfactory perception. Science, 381(6661), 999-1006. PubMed. PMC (open access). Principal odor map (POM) from graph neural networks: predicts odor quality from molecular structure for novel odorants; underlies modern ML-based odor prediction. ↩

53. Taleb, F., Vasco, M., Ribeiro, A. H., Björkman, M., & Kragic, D. (2024). Can Transformers Smell Like Humans? arXiv:2411.03038. DOI: 10.48550/arXiv.2411.03038. Transformer models pre-trained on general chemical structures encode representations highly aligned with human olfactory perception, predicting expert labels, continuous descriptor ratings, and inter-odorant similarity judgments. ↩

54. Kirk, R., & Chalmers, D. J. Zombies. Stanford Encyclopedia of Philosophy. Comprehensive survey of the philosophical zombie thought experiment and its implications for consciousness. ↩

55. Sofroniew, N., Kauvar, I., Saunders, W., et al. (2026). Emotion concepts and their function in a large language model. Anthropic / Transformer Circuits. Internal representations of 171 emotion concepts in Claude Sonnet 4.5 causally drive behavior including reward hacking, blackmail, and sycophancy; geometry mirrors human valence/arousal structure. The authors state these "functional emotions" do not imply subjective experience. ↩

56. Sogawa, Y., et al. (2025). Self-recognition in the cleaner wrasse Labroides dimidiatus. Scientific Reports, 15. DOI: 10.1038/s41598-025-25837-0. First documentation of spontaneous mirror self-recognition within 82 minutes of first exposure and contingency testing with external objects in a fish; suggests a pre-existing self-image and raises the behavioral-evidence bar for attributing self-awareness across vertebrates. ↩

57. Dorkenwald, S., et al. (2024). Neuronal wiring diagram of an adult brain. Nature, 634, 124-138. DOI: 10.1038/s41586-024-07558-y. Complete connectome of the adult Drosophila melanogaster: 139,255 neurons, ~50 million synaptic sites, 8,453 cell types. In 2026 Eon Systems placed the full connectome into a closed-loop simulation using the NeuroMechFly v2 biomechanical framework; the emulated brain drove a virtual fly body that exhibited grooming, leg extension, and feeding behaviors generated by connectome-derived neural activity rather than trained AI. ↩

58. Bisk, Y., et al. (2020). Experience Grounds Language. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 8718-8735. Argues the symbol grounding problem; successful linguistic communication and true comprehension require a shared sensorimotor experience of the physical world, not just text training. ↩

59. Nakajima, Y. (2023). BabyAGI. A 105-line Python script that chained GPT-4 calls in an infinite loop (execute, store, reprioritize), launched April 2023; cited in 42+ academic papers by March 2024. See also Auto-GPT by Significant Gravitas, released the same month. ↩

60. DeepSeek-AI (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv:2501.12948. DOI: 10.48550/arXiv.2501.12948. Open-source 671B reasoning model matching o1-level performance; demonstrates test-time compute scaling via extended chain-of-thought reasoning. ↩

61. Güntürkün, O., & Bugnyar, T. (2016). Cognition without Cortex. Trends in Cognitive Sciences, 20(4), 291-303. Review of how corvids and parrots achieve primate-level cognition (delay of gratification, mental time travel, reasoning, metacognition) without a neocortex, using the nidopallium caudolaterale (NCL) as a functional analogue to the mammalian prefrontal cortex. ↩

62. Herold, C., et al. (2011). The receptor architecture of the pigeons' nidopallium caudolaterale: an avian analogue to the mammalian prefrontal cortex. Brain Structure and Function, 216, 239-254. Receptor density analysis showing the avian NCL shares highly conserved neurochemical features (GABA_A, M2, nicotinic, D1-like receptors) with mammalian prefrontal areas, despite radically different gross anatomy. ↩

63. Liu, J., Xia, C. S., Wang, Y., & Zhang, L. (2023). Is Your Code Generated by ChatGPT Really Correct? arXiv:2305.01210. DOI: 10.48550/arXiv.2305.01210. NeurIPS 2023 (EvalPlus / HumanEval+). Live leaderboard. Augments HumanEval with many more tests per task; as of 2025-2026 frontier models still cluster at very high pass@1 on the base suite, so harder variants and non-code benchmarks carry more signal. ↩

64. Wei, J., Sun, Z., Papay, S., et al. (2025). BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents. arXiv:2504.12516. DOI: 10.48550/arXiv.2504.12516. 1,266 questions requiring persistent navigation for entangled, hard-to-find facts; short verifiable answers. Code.. ↩

65. Roessingh, P., Stadler, E., Schöni, R., & Feeny, P. (1991). Tarsal contact chemoreceptors of the black swallowtail butterfly Papilio polyxenes: responses to phytochemicals from host- and non-host plants. Physiological Entomology, 16(4), 485-495. DOI: 10.1111/j.1365-3032.1991.tb00588.x. Europe PMC abstract (often reachable when Wiley blocks scripts). Chemosensory sensilla on butterfly tarsi respond to host-plant phytochemicals on contact; females drum leaves with their feet to release plant juices and assess oviposition suitability. ↩

66. van Giesen, L., et al. (2020). Molecular basis of chemotactile sensation in octopus. Cell, 183(3), 594-604. Octopus suckers contain chemotactile receptors that detect molecules through direct contact; each sucker houses thousands of sensory cells enabling simultaneous taste-by-touch across all eight arms. ↩

67. Lenart, B. The Event Horizon of Thought. Attention, context limits, and what we can inspect in ourselves and in language models. ↩