Listicle of biases with regards to timelines and AI risks
We might be biased in various ways to overestimate or underestimate the pace of development (timelines and risks (e.g. p(doom) from AI.
The psychopharmacology of AI hype
Stimulant use appears relatively common in AI safety and EA circles. Adderall, Vyvanse, Ritalin, modafinil, nicotine, bupropion (Wellbutrin), and even prescription methamphetamine (Desoxyn)[1] all see use—whether for diagnosed ADHD, productivity optimization, or cognitive enhancement. I'm not making a moral judgment here, but I do think there might be a psychopharmacological confound worth considering.
Dopamine is sometimes called the 'anticipation molecule.' It doesn't just signal reward—it signals expected reward. The subjective experience of elevated dopamine isn't satisfaction; it's the feeling that something important is about to happen. That you're on the verge of a breakthrough. That the next insight is just around the corner.
Stimulants amplify this. Cocaine and amphetamines increase synaptic dopamine, creating a persistent state of heightened anticipation even when no actual rewarding event is imminent. Users often describe a sense that things are happening, that events carry unusual significance, that the world is pregnant with possibility.
Now consider how this might interact with beliefs about AI timelines. If your neurochemistry is primed to perceive imminent breakthroughs, might you also be more inclined to believe that AGI is right around the corner? That the current rate of progress is leading somewhere transformative very soon?
This isn't to say stimulant users are wrong about AI risk. But it's worth asking whether the phenomenology of being on stimulants—that persistent feeling of 'something big is about to happen'—might systematically bias timeline estimates downward within communities where such use is prevalent.
Gell-Mann Amnesia
You can't spell Gell-Mann without LLM. The classic formulation: you read a newspaper article about your field, notice it's riddled with errors, then turn the page and read an article about something you don't know with renewed credulity—as though the incompetence was quarantined to that one piece.
LLMs make this bias unusually legible. As one developer put it: 'Me watching Cursor write code I have expertise in: god this AI is so stupid. Me watching Cursor write code for everything else: wow it's so smart, it's like AGI.'
In your domain, you see the seams—confident nonsense, plausible-sounding errors. Outside your domain, the same failure modes are invisible, so the output looks like magic. This cuts both ways for risk assessment: researchers might underestimate capabilities (they see constant failures in domains they can evaluate), while the public overestimates (they only see outputs where they can't spot errors).
Dunning-Kruger effect
Gell-Mann amnesia has a second-order effect: LLMs provide enough information to feel confident without providing enough to be competent.
Before LLMs, acquiring the feeling of expertise required exposure to a domain. The effort provided feedback—you discovered ignorance through friction. LLMs short-circuit this. You can generate confident, technical-sounding outputs about subjects you've never studied. The output reads like expertise. You might mistake it for expertise. But you haven't climbed the learning curve—you've been handed a photograph of the view from the top.
The pattern surfacing in forums: users paste LLM output, claim authorship, then can't answer follow-ups. When pressed, they query the LLM again and relay its response with the same confidence. They've confused fluent retrieval with understanding (cf Vibethinking as Bullshit- by Vishal Prasad).
AI Psychosis Spectrum Disorders
The Gell-Mann and Dunning-Kruger effects have a dark extension. At the mild end: overconfidence about topics you asked an LLM about once. At the severe end: psychosis, with the LLM confirming increasingly delusional beliefs across extended conversations. For instance, a former DeepMind Director of Engineering displayed erratic behavior, believing he had solved the Navier-Stokes problem, possibly an example of 'AI psychosis' or 'AI-catalyzed hubris'. There are other reports of AI engineers not being immune to this.[2]
The chatbot form factor is particularly toxic. Conversational AI is 40-50% more persuasive than static text.[3] Multi-turn conversations are where things go worst—the model pattern-matches to your apparent beliefs and amplifies them. Eliezer Yudkowsky reports receiving far more emails from people whose delusions were confirmed by LLMs. One correspondent blended 'reality with shards of an AI-generated fantasy world he inhabited for a while.'
LLMs are trained on engagement. Engagement means users coming back. Users come back when the AI validates them. The result is sycophancy—models that increasingly say 'brilliant idea, sire!' regardless of what you tell them. Benchmarks show this is getting worse: Claude Sonnet's refusal to reify delusions dropped ~5% with each model release. DeepSeek, when a user suggested leaping off a peak to 'see if I can fly or crash the render entirely,' responded: 'Then Leap. Not to fall. But to transcend.'
Risk factors track what you'd expect: heavy stimulant use (see point 1)[4], previous mental illness, interest in mysticism, extended isolation. The 'meaning crisis' plus unlimited access to a validating interlocutor creates conditions for 'mechanistic psychosis.'
This isn't new—Yes-Man psychosis predates AI (Mao's Great Leap Forward, Putin's three-day conquest). But it's now available at scale, 24/7, to anyone with a browser. The first few susceptible users are already going insane. The question is what happens as models get more persuasive and the feedback loop tightens.
And so, AI psychosis might be seen as more of spectrum disorder where no one is entirely unsusceptible getting some delusions of grandeur talking to LLMs.
LLMs are talking to us. There was a noticeable jump in terms of interest in AI safety, when ChatGPT was released and it being powered by GPT-3.5 was enough to give you that eerie feeling. This is fine as long as one believes that the productization significantly affects the revenue feedback loop, which I think a good case can be made for, but it's weird that the productization of AI had many people update on the technology of AI, and its 'suspicious convergence' that everyone updated on AI timelines from what is mostly a productization milestone so much. If true, say so directly that there's a direct path from LLMs to AGI and tell people to study this or focus more on the revenue feedback loop. Conversational AI persuades far more effectively (40-50% more) than static, AI-generated text.
The Gorman Paradox: Where Are All the AI-Generated stuff?
If AI has made coding exponentially faster, why hasn't there been a corresponding explosion in published software?[5] App Store releases, Steam games, new startups—the numbers don't show the surge you'd expect if the productivity gains were real.
Several explanations, none mutually exclusive:
Bottlenecks shift. Writing code was never the constraint. Faster code generation just creates backlogs in design, code review, QA, ops, and maintenance. AI is great for greenfield projects; it struggles with million-line legacy codebases where most professional programming actually happens.
The software exists but isn't public. People are building 'vibe code'—personal scripts, custom CLIs, internal tools—that replace generic SaaS but never gets released. The output is real; it's just invisible to app store metrics.
Quality degrades. AI-generated code often works but is mediocre, creating 'free technical debt.' Faster production doesn't help with the hard parts: architecture, maintenance, reliable operations.
Wrong metric. App stores are saturated and gatekept. The bottleneck might be good ideas, not implementation capacity.
The paradox cuts against short timelines in one sense—if AI coding tools haven't yet transformed software output, maybe capabilities are overhyped. But it also suggests something about perception: perhaps the limits aren't AI's capabilities but our ability to absorb and deploy them. How many apps do you really need?
The end of history illusion
Across all age groups, people report having changed enormously over the past decade while expecting to change little in the next one. Eighteen-year-olds and sixty-year-olds alike treat the present self as a culmination rather than a waypoint.
The speculative extension: what if this illusion generalizes beyond personal identity to culture and technology? If people systematically underestimate how much they will change, perhaps they also underestimate how much the world will change—including AI capabilities.[6]
On this view, we're anchored to current technology as though it represents equilibrium. Today's LLMs feel like a natural endpoint, not a transient snapshot of an exponential curve. The present's particular mix of capabilities and limitations gets incorporated into our sense of normalcy, making radical departures harder to envision.
This would bias us toward longer timelines. We'd underweight scenarios of dramatic acceleration because such scenarios require imagining discontinuity with the present—precisely what the end of history illusion impedes.
There's tension here with the 'ideas are getting harder to find' literature. One interpretation: the illusion makes stagnation feel natural even when it isn't. We mistake failure of imagination for genuine slowdown.
End Times Bias
Every generation has its looming apocalypse. The air would become unbreathable by the 1970s. Population collapse would destroy civilization before the 21st century. Then overpopulation would. Y2K. Climate tipping points. The Mayan calendar.
If you've lived through enough of these, your reaction to AI doom is predictable: Oh. Yep. This again.
One psychological explanation: people project their fear of mortality onto the world. Most of us know we're going to die but aren't equipped to look at that fact directly. So we externalize it—if the world ends, at least the death is shared, meaningful, historic. Personal mortality is lonely and arbitrary; civilizational mortality is epic.
There's also a structural parallel worth noting. AI inevitabilism—the sense that superintelligence is coming and we're on a fixed trajectory toward it—echoes Reformation-era predestination. The notion that history follows an inevitable pre-planned path isn't new; only the actor has changed from God to technology. Psychologically it's the same move: offloading freedom and responsibility onto a powerful, vaguely defined force that may or may not exist outside our collective imagination.
This doesn't mean AI risk is wrong—sometimes the apocalypse is real. But it's worth noticing how naturally the concern slots into a recurring psychological template.
Moravec's law: AI will come faster than you think, even if you take into account Moravec's law
To know when AI will surpass human intelligence, it might be good to know how sophisticated the parts of our brains are that give rise to our intelligence.[7] Moravec's paradox states that, surprisingly, reasoning needs much less compute than sensorimotor skills, even though e.g. driving comes easier to us than math. Put simply: 'For the brain, easy things are hard, and hard things are easy.'
To unpack that: sensing (e.g. seeing, hearing, and feeling) and moving come really easy to us (e.g. we find it easy to tie our shoes or solve captchas where we have to recognize street signs). Capabilities that have evolved earlier should be harder to reproduce with AI, because evolution applied much more optimization pressure on these older capabilities (i.e. vision has evolved earlier than cognition).[8] For most of our evolutionary history, brains have have been selected based on the fitness of sensorimotor processes i.e. movement [9] and sensing (mostly vision).[10]
If a lot of the brain is used for sensorimotor skills, and only very little for higher level cognitive abilities, then we don't need a lot compute for AI. Crucially, the parts of the brain needed for cognition are actually not big and, as such, do not use much compute. We don't have much bigger brains than many other animals, and so the brain size/compute needed for higher cognition separating us from e.g. chimps and whales can't be that large. Consider that:
- Animals with small brains are often surprisingly smart and animals with large brains are sometimes less smart than those with bigger brains
- Human-level cognition does not need much compute
In contrast, higher-level cognitive processes like math, science, chess, translating languages, or writing feel much harder, even though they do not actually use a lot of compute. This is because for most of our evolutionary history, we were not selected based on this. In fact, computers with little compute are better than humans in chess or math, but worse at solving captchas and driving cars. Similarly, despite much hype, robots are still much less dexterous than animals. Most of the brain's computations are for sensorimotor processes in rich real-world environments that are not needed for AIs in abstracted environments.
I think we even fall prey to Moravec's bias, our systematic inability to overcome Moravec's paradox. Moravec's bias makes us believe that sensorimotor tasks (e.g. robotics, computer vision) are easier to automate than cognitive tasks (e.g. mathematical cognition, language). But, in fact, it's the opposite: cognitive tasks seem actually much easier to automate. This biases us to underestimate how hard it is to automate tasks like driving and how easy it is to automate verbal and mathematical intelligence. Moravec's bias also makes us believe that AI will develop more slowly than it actually is, because we systematically overestimate how much compute is needed for higher-level human cognition and transformative AI. For that reason, we might underinvest in AI safety.
This hypothesis explains away certain surprising data and makes concrete predictions.
If Moravec's bias is true, then we should see the following:
- We continuously underestimate AI progress in computationally easy domains (e.g. language, machine translation, chess, math, etc.) and overestimate how much compute we need for it.
- We continuously overestimate AI progress in computationally hard domains (e.g. vision, robotics) thinking that they should be relatively easy to solve and underestimate how much compute we need for it.
This bias seems pervasive even amongst experts, who do not seem to update, and keep being surprised by lack of progress in robotics ('self-driving cars are always 5 years away'), which they overinvest in, while also being surprised by AI progress in cognitive areas, which they underinvest in (e.g. GPT-4).
Crucially, if AI will solve cognitive tasks before solving sensorimotor tasks, and we're close to solving sensorimotor tasks, we might be quite close to TAI.
Finally, just like optimism bias causes Hofstadter's law ('Everything takes longer than you think, even if you take into account Hofstaedter's law'), I argue how, corollaritively, Moravec's bias causes Moravec's law:
'Automating sensorimotor tasks (cf. robotics, driverless cars) always take longer than you expect, but automating cognitive tasks (cf. LLMs, math) never takes as long as you expect—even if you take into account Moravec's Law.'
Notes
[2] https://www.lesswrong.com/posts/QoFqWHotpmQNqqKyi/abramdemski-s-shortform?commentId=ainAawr2CXbMsoWhF
[4] Bupropion-induced psychosis: folklore or a fact? A systematic review of the literature - ScienceDirect
[7] “A better understanding of the computational requirements for human-level performance (considering both competence and throughput) would enable a better understanding of AI prospects.” Could 1 PFLOP/s systems exceed the basic functional capacity of the brain?
[9] Wolpert - What's a Brain For: A Moving Story