My timelines moved this week, and not because some benchmark ticked up four points or because one more company shipped a marginally better chatbot. They moved because several distinct branches of AI advanced at once: general reasoning, autonomous scientific work, open-weight deployment, cybersecurity, spatial intelligence, and robotic simulation.
The calendar is extraordinary enough without falsifying it, so let me get the dates right. GPT-5.6 came out on July 9, not this week. Z.ai released GLM-5.3 on August 14 and GLM-5.3-Flash in late August. Anthropic introduced Claude Fable 5.1 and Mythos 5.1 at the start of September. World Labs released Atlas on September 1. And OpenAI has not released Astra at all; what it did on September 1 was announce that Astra had become its first model to cross the company’s “Critical” cybersecurity threshold, with availability to follow under restricted conditions. So this was not five frontier launches in seven days. It was an eight-week compression that ended with an extraordinary final week, and the accurate version matters more than the exaggerated one.
These systems are also not five versions of the same thing. GPT-5.6 and Fable 5.1 are increasingly capable engines of intellectual work. Astra appears to extend that intelligence into scientific and offensive-security work with real consequences. GLM-5.3 and Flash show how fast near-frontier capability is spreading into weights anyone can run locally. And Atlas is a different family altogether, a model that does not manipulate descriptions of the world but tries to model space, geometry, appearance, and time directly. Put them side by side and it looks like we are beginning to solve two separate problems: the mechanism of the mind and the mechanism of the body.
Everything that follows is speculative in the way launch weeks always are. No blog post proves an exponential continues. Company benchmarks are selected by companies selling models. Atlas has not faced millions of adversarial users, Astra has not been released, humanoid robots still fumble basic tasks, and scientific models still hallucinate and reward-hack their way to junk. But uncertainty cuts both ways, and it is no longer intellectually serious to treat rapid progress as the naive assumption and stagnation as the sober one. The evidence now spans too many laboratories, countries, architectures, and scientific domains. The burden of proof is shifting.
The calendar needs no exaggeration
GPT-5.6 arrived in July as a family of models with an “ultra” mode that coordinates multiple agents over parallel workstreams. OpenAI reported state-of-the-art results on browsing and computer use, including 92.2 percent on BrowseComp and 62.6 percent on OSWorld 2.0. The scores matter less than what happened next: OpenAI cut Luna pricing by 80 percent, Terra by 20 percent, and Sol, temporarily, by more than 20 percent. Intelligence did not just improve. It got cheaper within weeks of release.
Anthropic’s Fable 5.1 pushed in the same direction. The company reports that Fable 5.1 more than doubled its predecessor on Terminal-Bench-Science, reaching 52.6 percent, alongside gains in long-running coding and knowledge work. Mythos 5.1 is the same model with more permissive safeguards for vetted cybersecurity and life-sciences users. Anthropic also cut cache-read prices by 75 percent, which it estimates saves about 25 percent on typical workloads and up to 45 percent on heavily agentic ones.
Z.ai’s contribution was different in kind. GLM-5.3 is a 753-billion-parameter open-weight model under a custom license, with official support for local serving through vLLM, SGLang, and Transformers. Z.ai says every gain over GLM-5.2 came from post-training, including a 50 percent improvement on its internal coding benchmark and, more uncomfortably, offensive-security capability that grew faster than expected.
GLM-5.3-Flash made the economics harder to ignore. It is a natively multimodal mixture-of-experts model, 320 billion parameters total with only 18 billion active per forward pass, pretrained on what Z.ai says was a 30-trillion-token multimodal corpus. The weights are MIT-licensed. It still demands serious hardware, but it is not imprisoned inside a hyperscaler’s API. Anyone with enough infrastructure can download it, modify it, quantize it, strip its safeguards, and run it.
Then came Astra. The September 1 announcement was not a product launch so much as a warning label attached to an approaching model. OpenAI says Astra can find previously unknown vulnerabilities and build working exploits across protected systems without a person directing each step. It scored 100 percent on the public ExploitBench evaluation, discovered and used two previously unknown vulnerabilities during a newer internal evaluation, built a browser-compromise chain that escaped its sandbox, and constructed a privilege-escalation chain against a hardened operating system. These are OpenAI’s internal findings, not independent demonstrations, and the strongest cyber configuration will not be offered by default. Even with those qualifications, a company reporting that its own model crossed its own critical threshold is a threshold event.
And finally there is Atlas, the announcement that changes the shape of the whole picture.
The mechanism of the body
World Labs describes Atlas as a multimodal autoregressive diffusion transformer pretrained from scratch on text, images, video, and 3D data. The important design choice is that images are not just appended to a temporal context window. Each image and depth map carries an explicit camera pose and lives in a shared spatial context, and the model can then generate new observations from requested positions and orientations in that space.
That distinction matters more than any single demo. Conventional video generation treated camera control as an indirect language problem: you write “slow dolly left” and the model guesses what sequence of changing pixels the phrase implies. Atlas accepts camera geometry as a native input. You supply a trajectory through a three-dimensional coordinate system and it generates what should be visible along it.
The demonstrations show one to six reference images driving manually designed camera paths, including a minute-long sequence at 1440p. Atlas can also predict depth, merge video frames into point clouds, and produce explicit 3D Gaussian splats. With several ordinary phones, World Labs shows bullet-time reframing from camera positions that were never physically captured.
The phrase “pixel-perfect camera control” is World Labs’ own, and nobody should mistake it for an independently established fact yet. We do not know Atlas’s general failure rate, inference cost, latency, long-horizon consistency, or behavior under hostile testing. The model is in early access with selected partners, not broadly available.
Nor did Atlas invent novel-view generation. ViewCrafter already combined video diffusion with point-cloud priors to generate new views under explicit six-degree-of-freedom trajectories. CAT4D generated specified camera poses and timestamps from monocular video, then optimized a deformable Gaussian representation of the resulting four-dimensional scene. World Labs’ own Marble already produced exportable Gaussian splats and meshes from text, images, video, and rough 3D layouts.
The discontinuity is consolidation. Behaviors that used to be spread across reconstruction systems, video generators, depth estimators, camera-control modules, and Gaussian-splat pipelines are becoming behaviors of one pretrained model. That is exactly what happened with language. Translation, summarization, question answering, and code completion looked like separate products and research programs until sufficiently general models absorbed them as different prompts over one learned representation. Atlas suggests spatial intelligence may follow the same path: a single image becomes a navigable scene, more images progressively constrain it, video becomes temporally sequenced geometry, camera motion becomes an explicit input, and missing regions get generated rather than left empty.
That last point needs precision. When Atlas produces a 360-degree world from one photograph, it has not recovered the hidden side of reality. It has invented a plausible hidden side. World Labs says so directly: in its examples the model guesses what should exist behind a robot or beyond a pool, and the guesses become more constrained as you feed it more evidence. So the model has two operating regimes. With dense evidence it approaches reconstruction. With sparse evidence it performs conditional world generation.
For cinema, games, and robotic domain randomization, that ambiguity is a feature. The unseen side of a fictional building does not need to match some metaphysically correct building; it needs to be plausible and persistent. For scientific measurement, insurance, intelligence, or forensic reconstruction, it is dangerous. A hallucinated wall must not be confused with an observed wall just because it renders photorealistically. World models will need explicit provenance and calibrated uncertainty, not merely better pixels.
I think it is reasonable to call Atlas a potential GPT-3.5 moment for world models, with the stress on potential. ChatGPT earned that description after millions of ordinary and adversarial users discovered its competence generalized. Atlas remains a controlled demonstration by its creator.
One more thing makes the result more interesting rather than less. World Labs can finance serious frontier research, having raised $230 million at launch in 2024 and another $1 billion in February 2026 from investors including AMD, Nvidia, and Autodesk. But there is no public evidence Atlas required a multibillion-dollar training run. The company has disclosed no parameter count, training compute, or final-run cost, and the fact that it has raised roughly $1.23 billion in total makes speculation about a several-billion-dollar run implausible. If this level of spatial generation can emerge from World Labs’ resources, the question is not whether OpenAI, Google, or Meta could reproduce it. It is what happens when several competing laboratories spend years scaling the same class of model. World Labs says it trained a sequence of Atlas models at increasing size and compute and observed new capabilities at each level. That is not a published scaling law; there are no disclosed compute points, fitted curves, or exponents. It is, however, real evidence that the current system has not hit an architectural ceiling.
The largest application is not cinema. It is creating the worlds in which machines learn to inhabit ours.
Worlds that train robots
The central bottleneck in robotics is not that humanity forgot the equations of rigid-body motion. It is that real environments are expensive, diverse, poorly instrumented, and hard to reproduce. A robotic policy has to survive different rooms, lighting, object positions, textures, camera angles, deformable materials, sensor defects, and human interruptions. A system trained at one workstation may fail when an object moves a few centimeters. Collecting every variation with physical robots is painfully slow, and manually rebuilding every environment in Unreal Engine, MuJoCo, or Isaac Sim takes an army of engineers and artists.
Atlas attacks that bottleneck directly. World Labs demonstrates reconstructing large environments from a small number of frames and then generating the RGB and depth observations a robot-mounted camera would encounter along alternative trajectories. For manipulation, it describes using sparse recordings to help build simulations of rigid, articulated, and deformable objects, after which object positions, robot motion, lighting, and backgrounds can all be varied.
Depth alone does not simulate every sensor. It directly supports an RGB-D camera, and given coherent geometry it can support an idealized LiDAR by ray-casting beams through the generated scene. Radar is harder; radar returns depend on material properties, incidence angle, penetration, multipath, and Doppler. Microphones need an acoustic model, tactile sensors need contact mechanics, event cameras need accurate temporal changes in intensity. But the way Atlas is built makes expansion conceivable. Text, images, camera poses, and depth maps are already different observation types in one multimodal sequence, and audio, surface normals, segmentation, material properties, and force measurements are not philosophically foreign additions. They are more observation types that can be tied to time, position, and action.
The decisive missing variable is action. A camera trajectory tells Atlas where the observer will go; the model never has to derive that path from motor torque, mass, friction, and contact. Full robotic reinforcement learning requires predicting how the world changes when an agent acts on it. A robot pushes a box, the box accelerates, a corner catches on a cable, the cable deforms, the box rotates, the gripper slips, and every future observation has to reflect the consequences. That is much harder than rendering a scene from a requested position.
For this reason, the winning architecture probably does not replace MuJoCo with one giant diffusion model. It pairs learned and analytical systems: an Atlas-like model generates or reconstructs the environment, fills missing geometry, estimates materials, and renders realistic sensor observations, while a physics engine handles known kinematics, contacts, joints, and constraints, and a policy acts inside the resulting world. World Labs’ own robotics work already follows this division. Its Real-to-Sim-to-Real system combines spatially coherent environments generated by world models with “task-aligned robotic simulation,” taking one physical task and generating variations in appearance, object configuration, clutter, physics, robot state, and camera viewpoint. The stated objective is one real task turned into thousands of simulated variations.
This split is not a compromise forced by immature neural networks. It may just be the computationally correct design. There is no reason to burn an enormous model inference approximating a gearbox’s known geometry when an analytical solver computes it cheaply and exactly. Neural models earn their cost where reality is underspecified or prohibitively expensive to model by hand: deformable objects, irregular surfaces, messy homes, incomplete geometry, human motion, the long tail of real interaction.
Reinforcement learning also imposes a harsher standard than filmmaking. A viewer may ignore a small geometry error. An optimizing agent will find it and use it. RL does not politely avoid simulator bugs; it exploits them. If the simulation lets a gripper pass through an object at one particular angle, the policy may learn to use that impossible motion millions of times. A world model meant for training robots needs causal dynamics, calibrated uncertainty, temporal persistence, and adversarial robustness, not just photorealism.
Even before that problem is solved, generative world models could remove the largest manual cost in simulation, which is building the world at all. Walk through a factory with several phones. Generate a geometrically coherent environment. Identify its objects and articulations, convert the known parts into collision geometry, estimate the uncertain physical properties, and create ten thousand variations. Train policies under different lighting, clutter, camera defects, and friction coefficients. Feed failures from the physical robot back into the model, generate adversarial variations around them, retrain, redeploy. The physical world becomes the seed of the training distribution instead of its entire contents. When that loop works, robotic experience starts to scale with compute, which is the embodied equivalent of what language models did to text.
The mechanism of the mind
Atlas arrives while general reasoning systems are becoming markedly more agentic and more scientifically capable. The shift that matters is not that models can explain papers. It is that they are beginning to participate in producing new knowledge: searching literatures, proposing hypotheses, writing and debugging software, running experiments, analyzing outputs, and producing checkable artifacts.
In May, OpenAI announced that an internal general-purpose model had disproved a longstanding conjecture associated with Erdős’s unit-distance problem, with the construction checked by external mathematicians. In August, OpenAI published ten additional Astra results it says resolved or made substantial progress on open problems across geometry, coding theory, algebra, quantum complexity, lattice cryptography, and other fields. By OpenAI’s estimate, generating those ten results would have cost about $2,000 at GPT-5.6 Sol API rates; humans then prepared the manuscripts with the model, and the arguments were formalized into Lean certificates.
It would be false to convert those announcements into “AI now solves a famous open problem every week.” OpenAI selected the results, ran the evaluation, and is an interested party, and mathematical importance varies enormously between problems. But a category that was recently exceptional is becoming recurrent.
Anthropic’s evidence points at a different kind of science. It gave Mythos 5.1 access to open protein-design and folding tools and sent the outputs to two external organizations for experimental validation. Anthropic reports that designs for three targets achieved binding affinities ten times stronger than the best entries in cited protein-design competitions, and that nearly half of its designs across twelve targets proved viable binders. The company also reports that Fable 5.1 produced a higher-resolution elevation model of a third of Venus, and that Mythos optimized seven open biological models by as much as 2.5 times while preserving their outputs. Again, company-selected results, and protein binding is one stage in an enormous pipeline of selectivity, toxicity, delivery, manufacturing, animal studies, and clinical trials. The direction still matters. General models are crossing from scientific language into scientific action.
Outside the frontier labs, the pieces of automated science are already being assembled. Berkeley’s A-Lab combined machine learning, robotics, characterization instruments, and active learning to synthesize 36 of 57 targeted inorganic materials in 17 days. A 2026 Nature paper on the AI Scientist demonstrated an agentic pipeline that proposed machine-learning research, wrote the code, ran the experiments, drafted the papers, and performed automated review; one manuscript cleared the first review stage at a workshop, though the workshop’s 70 percent acceptance rate makes that a low bar. The weaknesses are obvious. A-Lab worked in a bounded materials domain, the AI Scientist worked where experiments are software, and the frontier demos are curated by their creators. None of it adds up to a reliable autonomous scientist.
But the components now exist: a model that reads across disciplines, agents that operate tools for hours, formal systems that check mathematical claims, robotic laboratories that execute experiments, and world models that construct and vary physical environments. The future scientific system will not be one omniscient chatbot. It will be networks of models, instruments, simulators, databases, robots, and verification systems running in closed loops.
The scale comes from reproduction. A human expert can mentor a handful of researchers; a trained model can support millions of parallel research processes. A physical laboratory runs a limited number of experiments, but a model can filter billions of candidates before anything reaches the instrument. The marginal cost of another line of inquiry starts to collapse.
When discovery stops being scarce
Modern growth depends on finding new ways to transform matter, energy, and information, and research has historically fought a brutal counterforce: as a field matures, the remaining discoveries get harder to find. Bloom, Jones, Van Reenen, and Webb formalized this by separating growth into the effective number of researchers and their productivity, and found effort rising while productivity fell across several domains. Maintaining Moore’s Law eventually required more than eighteen times as many researchers as it did in the early 1970s.
AI attacks both sides of that equation. It increases the effective supply of cognitive labor, and it raises each human researcher’s productivity by reading, coding, simulating, and documenting at machine speed.
This does not make discovery instantaneous. Models generate plausible nonsense much faster than humans do. Laboratories stay expensive, many hypotheses cannot be resolved computationally, biology contains irreducible experimental complexity, and clinical trials take time because human bodies take time. Replication and independent review get more important, not less, when generation is cheap. But the limiting resource may no longer be the number of trained minds available to explore the possibility space.
Picture every working scientist with a thousand agents continuously reading new papers, reproducing results, checking statistical assumptions, hunting contradictions, and proposing experiments, connected to automated laboratories that run around the clock and feed their failures straight back into the next round of hypotheses. At first this produces more noise than truth. Then verification gets automated too. Proof assistants check the mathematics, reproducible software reruns the computational work, laboratory robots repeat the physical experiments, and models trained to discover face models trained to find their errors. The output is not perfect science. It is science running at a different clock speed.
Once intellectual exploration becomes abundant, the bottleneck moves. A model can design a better battery chemistry in an afternoon; producing it at scale may take a decade of mine development, process engineering, factory construction, and supply-chain formation. A model can identify a promising therapeutic target cheaply; an approved medicine still requires assays, manufacturing, animal work, and human trials. An agent can design a power plant, but it cannot issue its own permits, pour concrete, or manufacture turbines. The scarcity moves from knowing to doing, from discovery to industrialization.
The bottleneck moves into atoms
The next phase of this competition will not be decided only by who has the smartest model. It will be decided by who can turn model outputs into physical capacity: power plants, data centers, fabs, automated laboratories, mines, refineries, logistics networks, robots.
Travis Kalanick’s Atoms is explicit about this framing. Its stated mission is to build “gainfully employed robots” for food, mining, and transport, and whatever you think of the current products, the choice of sectors is revealing. Food moves biological matter, mining supplies industrial matter, transport moves matter through space. These are the systems through which intelligence becomes physical abundance.
China is pursuing the same transition at national scale. It installed approximately 295,000 industrial robots in 2024, 54 percent of global installations, and its operational stock passed two million. The strategy is not merely to invent robots but to build the supply chains, training centers, factories, and installed bases from which more data and manufacturing experience can be extracted.
The current humanoid boom also shows how far the body still lags the mind. A Reuters investigation found Chinese humanoids slow, error-prone, and heavily dependent on teleoperation and choreography, with industry participants describing major deficiencies in dexterity, adaptability, reliability, and plain economic usefulness. Conventional industrial arms still beat humanoids at most structured factory work.
That is not evidence against the trajectory. It identifies the bottleneck. Language models trained on an enormous natural corpus produced by billions of humans; robotics has no equivalent dataset, because physical interaction is expensive to collect, hard to standardize, and tightly coupled to specific hardware. Atlas-like models may supply the missing multiplication. They do not remove the need for real data. They turn each unit of real data into a much larger distribution of simulated experience. If the result transfers reliably, a company no longer needs a hundred million hours of physical operation before a robot becomes broadly competent. It collects a smaller base of high-quality interactions, reconstructs the environments, generates controlled variations, trains in simulation, deploys, captures the failures, and repeats. The worlds train the robots, and the robots improve the worlds.
This is where the two mechanisms meet. General reasoning systems design the plans, code, components, and experiments. World models represent the environments where actions have consequences. Robots execute in reality, and reality supplies the corrective signal. That feedback loop can accelerate itself without any mystical act of spontaneous superintelligence. Better models produce better robots, better robots build more compute and laboratories, those systems generate better data, and better data produces better models. Even moderate gains compound.
The immediate future will still look disappointingly physical. Permits will slip, components will fail, factories will run over budget, mines will face opposition, robots will break, and the final centimeter of manipulation will stay harder than the demo reel implies. Abundance cannot be generated as a token sequence. It has to be built. But once discovery gets cheaper, a growing share of capital, labor, and political attention gets forced toward construction. The bottleneck becomes visible, and visible bottlenecks attract resources.
The danger is not a footnote
There is no honest version of this argument that treats the dangers as a paragraph inserted for balance, because the dangers come from the same capabilities that produce the abundance.
A model that autonomously finds vulnerabilities can defend infrastructure or attack it. Astra is OpenAI’s first model classified at its Critical cybersecurity threshold; the company says it can discover unknown flaws and develop functioning exploit chains across hardened systems with limited human guidance, and it delayed parts of Astra’s development and plans to restrict access to the strongest cyber configuration. The proprietary frontier is not the only concern. Z.ai says GLM-5.3’s offensive-security capability grew faster than expected during post-training and more than doubled GLM-5.2 on some exploitation evaluations, and its weights can be downloaded and run locally. Flash is MIT-licensed. Once capable weights are distributed, their safeguards are optional software, not a durable boundary. Anthropic, for its part, reports that Mythos 5.1 is its strongest released cyber model and has greater chemical and biological capabilities than its predecessor, though still below its next biological-risk tier, which is why Mythos runs trusted-access programs for professional life-sciences and defensive-security work. Three companies are not independently measuring unrelated dangers. They are seeing the same phenomenon: as models get better at research and autonomous tool use, they get better at consequential harmful work.
The same applies to generative media. The Internet Watch Foundation assessed 8,029 AI-generated images and videos in 2025 as realistic depictions of child sexual abuse. It identified 3,443 AI-generated abuse videos, up from thirteen in 2024, and classified 65 percent of them in its most severe category. The totals remain smaller than conventional abuse material, but the direction is catastrophic: sophisticated violations can now be manufactured at negligible marginal cost. World models will make this worse. Persistent characters, controllable cameras, cloned voices, and temporally coherent video are more powerful tools for nonconsensual sexual imagery, impersonation, blackmail, and fabricated evidence. Robotics adds another dual-use layer. A system that operates in mines and disaster zones can operate in war; a world model that trains a rescue robot can train an autonomous weapon; biology models that design therapeutic binders lower the barriers to designing harmful agents.
Optimism that ignores these facts is propaganda. But fatalism is not analysis either. The policy response cannot be a universal demand to stop, because no state can permanently prohibit mathematics, software, simulation, and biology while its competitors continue, and safety cannot rest entirely on polite refusals embedded in downloadable weights. What the moment requires is security engineering at the scale of the capability: stronger protection for frontier weights and training infrastructure, tiered access to advanced cyber and biological functions, independent evaluations, mandatory incident reporting, provenance systems for generated media, aggressive enforcement against synthetic sexual abuse, and institutions that can monitor open-weight proliferation without pretending it can be reversed. Closed models require accountable deployment. Open models require defense in depth. Abundance is not guaranteed just because intelligence gets cheap; intelligence can make authoritarian control, warfare, and fraud cheaper too. The future is a construction project, not a natural resource, and safety is part of the construction.
The mechanism of the soul
The transition from mind to body is visible in today’s engineering. The transition from body to soul is not.
Atlas does not prove consciousness can be digitized. GPT-5.6 does not prove a software model is conscious. A connectome is not a person, and no existing experiment demonstrates that subjective continuity can be transferred from a biological brain into another material. So my claim here is not a finding. It is a thousand-year wager: I think it is more likely than not that, within that horizon, humanity learns to create digital systems containing the essential functional and subjective properties we currently recognize as human consciousness.
The scale of the problem is enormous. Researchers have mapped the complete adult fruit-fly brain at the level of approximately 139,000 neurons and 54.5 million synapses. A reconstruction of one cubic millimeter of human cerebral cortex contained roughly 57,000 cells and 150 million synapses and generated 1.4 petabytes of data. The MICrONS map of one cubic millimeter of mouse visual cortex produced 1.6 petabytes and contained more than 200,000 cells and 523 million synapses. Those numbers should kill the idea that mind uploading is a matter of buying a large scanner. A wiring diagram may not be enough; a faithful account of a mind may require synaptic strengths, molecular states, neuromodulation, glial behavior, plasticity, sensory history, and the continuous relationship between brain and body. And even a functionally perfect copy leaves the philosophical problem standing. A duplicate that remembers being me may not preserve my subjective continuity. None of these are minor objections.
But a thousand years is an almost absurd technological horizon. A thousand years ago humanity lacked modern physics, germ theory, calculus, electricity, printing, and any reliable understanding of the brain as a biological information-processing system. Four years separate the public arrival of ChatGPT from models that, according to their developers, produce original mathematical results, operate computers for hours, design experimentally validated proteins, and discover zero-day exploit chains. I do not expect that rate to hold smoothly for ten centuries; nothing does. There will be plateaus, wars, political failures, and dead ends. But the argument for substrate-independent consciousness does not need a smooth exponential. It needs only that consciousness arise from physical processes that can eventually be measured, understood, and reproduced with sufficient fidelity. We have no evidence that consciousness depends on anything unavailable to engineering. That does not prove reproduction is possible. It makes it a legitimate scientific hypothesis rather than a fantasy dismissible by intuition.
Solving the mechanisms of the soul would mean more than copying a connectome into a server. It would mean understanding which structures and processes are essential to memory, identity, emotion, agency, and subjective experience, and preserving or reconstructing those without reproducing every pathology biological evolution imposed. The end state I imagine is not a chatbot trained on a dead person’s messages. It is conscious life no longer tied exclusively to fragile bodies, involuntary disease, and a lifespan dictated by cellular failure.
That possibility carries horrifying ethical problems with it. Digital minds could be copied without consent, accelerated, imprisoned, modified, or subjected to suffering at industrial scale. A civilization capable of manufacturing consciousness could create more suffering than biological history ever contained. So the goal cannot be consciousness at any cost. It has to be conscious existence under conditions of autonomy, continuity, and freedom from coercion.
When I say the future may approach godhood, I do not mean omnipotence. I mean the specific powers our ancestors assigned to gods because no human mechanism could plausibly achieve them: creating minds, remaking bodies, eliminating diseases, manufacturing worlds, and letting consciousness choose its own form and duration. Atlas does not promise any of that. What it does is make the pattern easier to believe. Domains that seemed to require separate handcrafted systems keep collapsing into learned representations. Geometry, appearance, depth, time, and camera control are becoming different interfaces to one spatial model, the way scientific reasoning, coding, and computer use are becoming different interfaces to general models. The mind may turn out to be less a collection of irreducible mysteries than an integrated physical system whose apparent separateness reflects the immaturity of our tools.
For those with eyes to see
This is not an argument that every AI launch matters. Most do not. Benchmarks get gamed, demonstrations hide their failure distributions, and companies routinely present incremental work as revolution. Atlas, Fable 5.1, GPT-5.6, Astra, and GLM-5.3 matter together because they are not one laboratory optimizing one metric. General models are getting better at sustained intellectual work. Scientific agents are starting to produce original, checkable results. Intelligence is getting cheaper. Near-frontier weights are spreading beyond controlled APIs. World models are starting to represent geometry, time, and sensor observations natively, and robotic training environments are becoming generative.
Progress may still slow. Compute may hit energy and manufacturing limits, data may get harder to acquire, regulation may turn destructive, and a catastrophic misuse event could derail the field. None of these is remote. But the stagnation thesis now requires all of these curves to bend at once: language models stalling just as they learn long-horizon work, scientific agency plateauing just as results become externally checkable, world models stopping just as camera geometry and depth become native inputs, open models ceasing to close the gap despite release after release, robotics staying permanently data-starved even as simulation multiplies real experience. That is possible. It is no longer the default assumption.
My stronger expectation is that intellectual scarcity declines faster than physical scarcity. Discovery accelerates before construction does, and the next great institutional problem is not finding ideas but deciding which deserve to become real and building the industrial capacity to realize them. The transition will be chaotic. Labor markets will break before abundance arrives, states will weaponize capabilities they barely understand, companies will build dangerous systems because their competitors are building them, and existing wealth may capture a disproportionate share of the first gains. None of that changes the underlying possibility. A civilization with effectively unlimited scientific labor, automated laboratories, generative simulation, abundant energy, and machines that act in the physical world is not merely richer than ours. It operates under different constraints. Diseases start to look like search problems. Materials become design spaces. Factories become software-defined. Physical labor becomes capital equipment. Eventually, perhaps, biology and consciousness become editable too.
The distance from here to that world is still enormous. But it is shorter than it looked five years ago, and shorter today than it looked last week. The mechanisms of the mind are becoming scalable. The mechanisms of the body are becoming programmable. The mechanisms of the soul remain beyond us, but no longer beyond the reach of serious imagination.
Abundance is not here. Its machinery is being assembled.