Science fiction has always been the research agenda nobody funded. Writers proposed the universal translator in 1966, the conversational computer in 1968, the household robot in 1927, and for most of a century the honest scientific answer to all of them was the same: not with any technique we know. Then, in roughly one decade, the answer changed.
This article holds itself to a rule most AI writing does not. Every capability below had to clear two bars: a named science-fiction origin that predates it by decades, and hard evidence today, meaning a product you can use or a result published in a peer-reviewed venue such as Nature, Science or the New England Journal of Medicine. Concept videos, staged demos and vendor claims did not count. Twenty-one references are listed at the end.
The prediction gap: year imagined vs year realized
Each line runs from first fiction appearance to first shipped or peer-reviewed realization
Hollow dot: first fiction appearance (Metropolis 1927; Turing 1950; Asimov 1956; Star Trek 1966; Firefox and Knight Rider 1982-83; TNG holodeck 1987; Iron Man 2008). Filled dot: first shipped product or peer-reviewed result. Median wait: 49 years. Nine of the ten filled dots land after 2015.
The list, with evidence
The universal translator
SHIPPEDFICTION: Star Trek (1966) universal translator · The Hitchhiker's Guide to the Galaxy (1979) Babel fish
Machine translation was a Cold War project that spent fifty years producing phrasebook-grade output. The obstacle was believed fundamental: translation requires understanding, and computers do not understand. Then attention-based neural models made the question empirical. Meta's SeamlessM4T, published in Nature in 2025, is a single model that translates speech and text across roughly 100 languages, speech-to-speech into 36 target languages, about 23% more accurately than the cascaded systems before it, and robust to background noise[3]. Consumer earbuds and phones now run live interpretation in ordinary conversations. The Babel fish took 44 years from joke to product spec.
A machine you cannot tell from a human
PEER-REVIEWEDFICTION: Alan Turing's imitation game (1950) · 2001: A Space Odyssey (1968) · Her (2013)
Turing predicted machines would eventually play his “imitation game” well enough to fool interrogators; for 75 years every claimed pass was a gimmick with loopholes. In 2025, researchers at UC San Diego ran the classic three-party test properly: interrogators held simultaneous five-minute conversations with a human and a model, then judged which was which. GPT-4.5, given a persona prompt, was judged to be the human 73% of the time, significantly more often than the actual humans it was tested alongside[2]. That is not “passing” by a technicality. The machine was more convincingly human than the humans.
The holodeck
EMERGINGFICTION: Star Trek: The Next Generation (1987) · The Matrix (1999) simulated worlds
Generated worlds were considered a rendering problem: you would need artists to build every object in advance. World models dissolved that assumption. OpenAI's Sora showed in 2024 that video generation learns an implicit physics of scenes; Google's Veo 3 added synchronized audio in 2025. The decisive step is DeepMind's Genie 3, announced in August 2025: type a description and it generates an interactive world you can walk through in real time, at 720p and 24 frames per second, holding the scene consistent for minutes and accepting “world events” you inject by prompt[20]. No game engine, no assets, no artists. A first-generation holodeck program, minus the hard-light projectors.
The medicine oracle
NOBEL-VERIFIEDFICTION: Star Trek's tricorder and computed cures (1966) · countless instant-diagnosis machines
Predicting how a protein folds from its sequence was biology's grand challenge for 50 years; a 1969 back-of-envelope argument (Levinthal's paradox) suggested brute force would take longer than the age of the universe. AlphaFold 2 solved the assessment in 2020 with near-experimental accuracy, and its database now holds predicted structures for essentially every catalogued protein, around 200 million, where six decades of global lab work had produced about 214,000[4]. AlphaFold 3 extended prediction to DNA, RNA and drug-molecule interactions[5]. The 2024 Nobel Prize in Chemistry went to its creators[6]. More than three million researchers have used the database; AI-designed drug candidates are now in human clinical trials.
Known protein structures, log scale
Reading thoughts
CLINICAL TRIALSFICTION: telepathy machines everywhere from Firefox (1982) to Minority Report (2002)
Thought-controlled machines were the purest fantasy on this list, because thoughts had no measurable structure. Deep learning found the structure. In 2023, two Nature papers decoded attempted speech from implanted electrodes in paralysed patients at 62 and 78 words per minute[7,8]; a 2024 NEJM study reached 97.5% accuracy on a 125,000-word vocabulary[9]; and 2025 systems stream synthesized voice, in the patient's own reconstructed voice, with well under a second of delay[10]. Non-invasively, fMRI decoding recovers the gist of stories a person is merely imagining[11]. More than twenty people now use Neuralink implants day-to-day in trials. Natural speech runs about 160 words per minute; the machines are closing in.
Decoded speech rate from brain signals, words per minute
Sources: refs [7]-[10]. The 2025 bar streams continuous synthesized voice rather than discrete words; its effective ceiling is conversational cadence.
Self-driving cars
SHIPPEDFICTION: Knight Rider's KITT (1982) · Total Recall's Johnny Cab (1990) · Minority Report (2002)
For decades the field consensus was that open-road driving required human-grade general perception. The evidence now says otherwise, at scale. Waymo's fleet has logged over 220 million fully driverless miles, carrying paying passengers with no human behind the wheel across Phoenix, San Francisco, Los Angeles, Austin and Atlanta. Against insurance-grade human benchmarks on the same roads, the recorded rates are 94% fewer serious-injury crashes, 82% fewer injury-causing crashes and 93% fewer pedestrian injury collisions, results consistent with the peer-reviewed analysis published at the 56.7-million-mile mark[12]. Johnny Cab arrived; it simply skipped the creepy mannequin.
Waymo cumulative rider-only miles, millions
Source: Waymo safety hub and peer-reviewed benchmark studies, ref [12].
Humanoid robots
EMERGINGFICTION: Metropolis (1927) · C-3PO, Star Wars (1977) · I, Robot (2004)
Humanoids existed for decades as scripted stage machines; the missing organ was a brain that could handle the unscripted. That brain arrived as the vision-language-action model: the same architecture behind chatbots, trained to output motor commands instead of words[17]. The result is robots that pick up objects they have never seen because they understand what the objects are. Electric Atlas, Figure's Helix-driven units and Tesla's Optimus are in factory pilots with BMW and others; Chinese manufacturers now sell capable humanoids at motorcycle prices. Ninety-eight years separate Metropolis from a humanoid doing paid warehouse work, the longest wait on this list.
The unbeatable game master
PEER-REVIEWEDFICTION: WarGames' WOPR (1983) · Star Trek's chess-playing computers
After chess fell in 1997, the field agreed on what computers still could not do: intuition (Go), hidden information and deception (poker), and natural-language negotiation (Diplomacy). All three fell in under a decade, each documented in Nature or Science. AlphaGo beat Lee Sedol in 2016 using move evaluation no human taught it[13]. Pluribus beat elite professionals at six-player no-limit poker, bluffing included, for about $1,000 an hour[14]. Cicero reached the top 10% of human Diplomacy players by negotiating alliances, in English, with humans who mostly never suspected they were talking to a machine[15]. WOPR's lesson was that the machine learns by playing itself. That is literally how AlphaGo trained.
The AI scientist
PEER-REVIEWEDFICTION: Asimov's Multivac (1956), the machine that answers science's questions
Machines were supposed to calculate, never discover. In November 2023, DeepMind's GNoME model predicted 2.2 million previously unknown crystal structures, about 381,000 of them stable, an expansion of known stable materials that materials scientists compared to 800 years of conventional progress[16]; a robotic lab at Berkeley then synthesized 41 of 58 attempted new compounds in 17 days, unattended[17]. In mathematics, DeepMind systems scored 28/42 at the 2024 International Mathematical Olympiad, one point short of gold; in 2025 Gemini Deep Think scored 35/42, an official gold medal, solving five of six problems in natural language within the human time limit[18]. Multivac was supposed to arrive around the year 2061. It is early.
DeepMind systems at the International Mathematical Olympiad, points of 42
2024: AlphaProof + AlphaGeometry 2, graded by Fields medallists. 2025: Gemini Deep Think, graded by IMO coordinators. Ref [18].
Jarvis-class agents
EMERGINGFICTION: Iron Man's JARVIS (2008), the assistant that does the work, not just the talking
The newest dream on the list is closing the fastest. In 2023 the best models resolved under 5% of real GitHub engineering issues on the SWE-bench benchmark; by late 2025 frontier agents exceeded 70%, writing, testing and shipping the fix end-to-end. METR's measurement study puts a clock on it: the length of task an AI agent can complete autonomously at 50% reliability has doubled roughly every seven months for six years, from seconds in 2019 to multi-hour engineering tasks today[19]. Extrapolation is not destiny, but the trend line, if it merely holds, reaches week-long projects before 2030. Tony Stark's butler was fiction in 2008. A meaningful fraction of the world's code is now written by its descendants.
The pattern behind all ten
These are ten different problems, translation, biology, driving, mathematics, yet they fell to variations of one method: large neural networks trained on large data with large compute. Training compute for frontier models has grown four to five times per year for over a decade, and the transformer architecture of 2017 turned that compute into capability with unprecedented efficiency. The practical consequence is that “impossible” stopped being a property of problems and became a property of scale. When one method eats problems this different, the sensible question stops being which prediction comes true next and becomes who gets to work on it.
What is still fiction
Honesty demands the other column. Faster-than-light travel, time machines and consciousness uploading are physics problems, and physics has not moved. Artificial general intelligence, a machine matching humans across all domains, remains unproven, whatever the marketing says. Household robots that fold your laundry reliably are still emerging, not arrived. The ten entries above earned their place with receipts; these have none yet. That distinction, evidence versus vibes, is exactly the skill this article is trying to model.
The people building this are hired in India, now
Every entry above runs on the same stack: Python, machine learning, deep learning and LLM engineering. Our Advanced Data Science & AI Engineering program teaches exactly that, live, with real projects, GenAI and model deployment included, and placement support until you hold an offer letter. Recent placements range ₹22 to 29 LPA, with top performers up to ₹45 LPA.
Book a free demoQuestions people actually ask
Is this AI hype or verified science?
Every capability here is backed by a shipped commercial product or a peer-reviewed publication in venues like Nature, Science and NEJM, listed in the references below. Demos, concept videos and vendor claims were excluded by design, and the “What is still fiction” section lists what did not make the cut.
Which of these can I build a career on in India?
All ten sit on one skill stack: Python, machine learning, deep learning, and increasingly LLM and agent engineering. India's GCCs, product companies and startups hire for these roles across Bengaluru, Hyderabad, Pune and Mumbai, and the roles pay among the highest packages in Indian IT.
Do I need a PhD to work in AI?
No. Research scientist roles often want one, but the fast-growing majority of AI hiring is engineering: building, fine-tuning, evaluating and deploying models and agents. Those roles hire graduates who can show working projects, which is what a portfolio-based course is for.
What is still genuinely impossible?
Faster-than-light travel, time travel and mind uploading are constrained by physics, not compute. General intelligence is unproven. This article deliberately separates what has evidence from what has enthusiasm.
References
[1] Turing, A.M. (1950). Computing Machinery and Intelligence. Mind 59, 433–460.
[2] Jones, C.R. & Bergen, B.K. (2025). Large Language Models Pass the Turing Test. arXiv:2503.23674.
[3] Seamless Communication Team, Meta AI (2025). Joint speech and text machine translation for up to 100 languages. Nature 637, 587–593.
[4] Jumper, J. et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589.
[5] Abramson, J. et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500.
[6] The Nobel Prize in Chemistry 2024: Baker, Hassabis, Jumper. nobelprize.org.
[7] Willett, F.R. et al. (2023). A high-performance speech neuroprosthesis. Nature 620, 1031–1036.
[8] Metzger, S.L. et al. (2023). A high-performance neuroprosthesis for speech decoding and avatar control. Nature 620, 1037–1046.
[9] Card, N.S. et al. (2024). An accurate and rapidly calibrating speech neuroprosthesis. NEJM 391, 609–618.
[10] Littlejohn, K.T. et al. (2025). A streaming brain-to-voice neuroprosthesis to restore naturalistic communication. Nature Neuroscience 28; Wairagkar, M. et al. (2025). An instantaneous voice-synthesis neuroprosthesis. Nature.
[11] Tang, J. et al. (2023). Semantic reconstruction of continuous language from non-invasive brain recordings. Nature Neuroscience 26, 858–866.
[12] Kusano, K.D. et al. (2025). Comparison of Waymo rider-only crash rates by crash type to human benchmarks at 56.7 million miles. Traffic Injury Prevention; Waymo Safety Impact hub, data through March 2026: waymo.com/safety/impact.
[13] Silver, D. et al. (2016). Mastering the game of Go with deep neural networks and tree search. Nature 529, 484–489.
[14] Brown, N. & Sandholm, T. (2019). Superhuman AI for multiplayer poker. Science 365, 885–890.
[15] Meta FAIR Diplomacy Team (2022). Human-level play in the game of Diplomacy by combining language models with strategic reasoning. Science 378, 1067–1074.
[16] Merchant, A. et al. (2023). Scaling deep learning for materials discovery. Nature 624, 80–85.
[17] Szymanski, N.J. et al. (2023). An autonomous laboratory for the accelerated synthesis of novel materials. Nature 624, 86–91; Brohan, A. et al. (2023). RT-2: Vision-Language-Action models. arXiv:2307.15818.
[18] Google DeepMind (2024, 2025). AI achieves silver-medal standard at IMO 2024; Gemini Deep Think achieves official gold-medal standard at IMO 2025. deepmind.google.
[19] METR (2025). Measuring AI Ability to Complete Long Tasks. arXiv:2503.14499; SWE-bench leaderboards, 2023–2025.
[20] Google DeepMind (2025). Genie 3: a new frontier for world models. deepmind.google; OpenAI (2024) Sora; Google (2025) Veo 3.
[21] Neuralink (2026). Two Years of Telepathy, clinical trial updates. neuralink.com/updates.
All figures accessed August 2026. Benchmark scores, mileage counts and trial numbers are point-in-time and will change.