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.
GPT-5.6 came out on July 9. 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 the same day OpenAI announced that Astra, still unreleased, had become its first model to cross the company’s “Critical” cybersecurity threshold, with availability to follow under restricted conditions. Eight weeks, five laboratories, and the densest stretch at the very end.
These are not five versions of the same thing. GPT-5.6 and Fable 5.1 are engines of intellectual work. Astra extends that intelligence into offensive security with real consequences. GLM-5.3 and Flash show near-frontier capability spreading into weights anyone can run locally. 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.
All of this is vendor evidence, and launch weeks prove nothing about whether an exponential continues. 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 burden of proof is shifting.
Eight weeks
GPT-5.6 arrived in July with an “ultra” mode that coordinates multiple agents, plus state-of-the-art browsing and computer-use results: 92.2 percent on BrowseComp, 62.6 percent on OSWorld 2.0. The more telling move came after. Within weeks OpenAI cut Luna pricing by 80 percent, Terra by 20 percent, and Sol, temporarily, by more than 20 percent. Intelligence improved and immediately got cheaper.
Fable 5.1 more than doubled its predecessor on Terminal-Bench-Science, reaching 52.6 percent. 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 releases were different in kind. GLM-5.3 is a 753-billion-parameter open-weight model whose gains over GLM-5.2 came entirely from post-training, including a 50 percent jump on its internal coding benchmark and offensive-security capability that grew faster than the company expected. GLM-5.3-Flash is a mixture-of-experts model, 320 billion parameters with 18 billion active, under an MIT license. Anyone with enough hardware can download either, strip the safeguards, and run it.
Astra’s September 1 announcement was less a launch than a warning label. OpenAI says the model finds unknown vulnerabilities and builds working exploits without step-by-step direction: 100 percent on the public ExploitBench evaluation, two zero-days discovered and used during an internal evaluation, a browser-compromise chain that escaped its sandbox, privilege escalation against a hardened operating system. These are internal findings, and the strongest configuration will be restricted. Still, a company reporting that its own model crossed its own critical threshold is a threshold event.
And then there is Atlas, which 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 each image and depth map carries an explicit camera pose in a shared spatial context, and the model generates new observations from requested positions. Video generators used to treat camera control as a language problem; you wrote “slow dolly left” and hoped. Atlas takes camera geometry as a native input: supply a trajectory, get what should be visible along it. The demos show one to six reference images driving designed camera paths, including a minute-long sequence at 1440p, plus depth prediction, point clouds, explicit Gaussian splats, and bullet-time reframing from a few ordinary phones.
“Pixel-perfect camera control” is World Labs’ phrase, not an established fact. We do not know the failure rate, inference cost, latency, or behavior under hostile testing, and the model is in early access. Nor is novel-view generation new; ViewCrafter, CAT4D, and World Labs’ own Marble each did pieces of this. The discontinuity is consolidation. Behaviors spread across reconstruction systems, video generators, depth estimators, and splat pipelines are becoming behaviors of one pretrained model, which is exactly what happened when language models absorbed translation, summarization, and code completion as different prompts over one representation.
One precision matters. When Atlas produces a 360-degree world from a single photograph, it has not recovered the hidden side of reality; it has invented a plausible one, and World Labs says so. With dense evidence the model approaches reconstruction, with sparse evidence it generates. For film, games, and domain randomization that ambiguity is a feature. For forensics, insurance, or measurement it is dangerous, and world models will need provenance and calibrated uncertainty, not just better pixels.
I think Atlas is a potential GPT-3.5 moment for world models, with the stress on potential; ChatGPT earned that title from millions of adversarial users, and Atlas remains its creator’s controlled demo. What makes it more interesting is the budget. World Labs has raised about $1.23 billion in total, $230 million at launch and $1 billion in February 2026 from investors including AMD, Nvidia, and Autodesk, which makes a multibillion-dollar training run implausible. If this emerges from those resources, the question is what happens when several labs spend years scaling the same class of model. World Labs says successive Atlas models showed new capabilities at each scale. That is not a published scaling law, but it is evidence the ceiling has not been hit.
The largest application is not cinema. It is building the worlds where machines learn to inhabit ours.
Worlds that train robots
Robotics’ central bottleneck is that real environments are expensive, diverse, and hard to reproduce. A policy trained at one workstation fails when an object moves a few centimeters; collecting every variation physically is slow, and rebuilding environments by hand in MuJoCo or Isaac Sim takes an army of engineers and artists. Atlas attacks this directly: reconstruct a large environment from a few frames, then generate the RGB and depth a robot’s camera would see along alternative trajectories, with object positions, lighting, and backgrounds varied at will.
The missing variable is action. A camera trajectory tells Atlas where the observer goes; full reinforcement learning requires predicting how the world changes when an agent pushes on it, the box accelerating, the cable deforming, the gripper slipping. That is much harder than rendering from a requested position. So the winning architecture probably pairs learned and analytical systems: a world model generates the environment and renders realistic sensor observations, a physics engine handles known contacts and constraints, and a policy acts inside the result. World Labs’ own Real-to-Sim-to-Real work already follows this split, turning one real task into thousands of simulated variations.
RL also imposes a harsher standard than filmmaking. A viewer ignores a small geometry error; an optimizing agent finds it, and will use an impossible motion millions of times if the simulator permits it once. Training-grade world models need causal dynamics and adversarial robustness, not just photorealism. But even before that is solved, generative models remove the largest manual cost in simulation, which is building the world at all. Walk a factory with a few phones, generate the environment, vary it ten thousand ways, train, deploy, feed the failures back. When that loop works, robotic experience scales with compute, which is the embodied version of what language models did to text.
The mechanism of the mind
Meanwhile the reasoning systems are becoming scientists rather than explainers. In May, OpenAI announced that an internal model disproved a longstanding conjecture tied to Erdős’s unit-distance problem, checked by external mathematicians. In August it published ten more Astra results across geometry, coding theory, algebra, quantum complexity, and lattice cryptography; generating them cost roughly $2,000 at GPT-5.6 Sol API rates, and the arguments were formalized into Lean certificates. OpenAI selected the results, and importance varies wildly between problems. But a category that was recently exceptional is becoming recurrent.
Anthropic’s evidence is wetter. It gave Mythos 5.1 open protein-design tools and sent the outputs for external experimental validation: designs for three targets bound ten times stronger than the best entries in cited protein-design competitions, and nearly half of the designs across twelve targets proved viable binders. Fable 5.1 produced a higher-resolution elevation model of a third of Venus. Company-selected results, and binding is one stage in a long pipeline of toxicity, delivery, manufacturing, and trials. The direction still matters: general models are crossing from scientific language into scientific action.
Outside the frontier labs, Berkeley’s A-Lab used robots and active learning to synthesize 36 of 57 targeted inorganic materials in 17 days, and the AI Scientist ran an agentic pipeline from hypothesis to drafted paper, with one manuscript clearing the first review stage at a workshop whose 70 percent acceptance rate makes that a low bar. None of it is a reliable autonomous scientist. But the components exist: models that read across disciplines, agents that run tools for hours, proof assistants, robotic laboratories, world models. The future system is networks of these in closed loops, not one omniscient chatbot, and the scale comes from reproduction. A model can filter billions of candidates before anything touches an instrument.
When discovery stops being scarce
Research fights a brutal counterforce: as fields mature, the remaining discoveries get harder to find. Bloom, Jones, Van Reenen, and Webb showed research effort rising while productivity fell across several domains; maintaining Moore’s Law eventually took more than eighteen times as many researchers as it did in the early 1970s. AI attacks both terms, adding cognitive labor and multiplying each human researcher. It does not make discovery instantaneous. Laboratories stay expensive, biology stays irreducible, trials take time because bodies take time, and cheap generation makes replication more important, not less. But the limiting resource may no longer be trained minds.
Give every scientist a thousand agents reading, reproducing, and proposing, wired to laboratories that run all night, with verification automated too: proof assistants checking the mathematics, robots repeating the experiments, models trained to discover facing models trained to find their errors. Not perfect science, but science at a different clock speed. And then the bottleneck moves. A model can design a battery chemistry in an afternoon; the mine, the process engineering, and the factory take a decade. The scarcity moves from knowing to doing.
The bottleneck moves into atoms
The next phase will be decided by who converts model outputs into physical capacity. Travis Kalanick’s Atoms says it plainly, building “gainfully employed robots” for food, mining, and transport, the systems through which intelligence becomes matter. China installed roughly 295,000 industrial robots in 2024, 54 percent of global installations, with an operational stock past two million; its strategy is supply chains and installed bases, not just inventions.
The humanoid boom shows how far the body still lags. A Reuters investigation found Chinese humanoids slow, error-prone, and dependent on teleoperation, with conventional industrial arms still beating them at most structured work. That is not counter-evidence; it locates the bottleneck. Robotics has no natural corpus the way language did. Atlas-like models are the missing multiplier, turning each unit of real interaction into a much larger distribution of simulated experience. The worlds train the robots, and the robots improve the worlds: better models make better robots, which build more compute and laboratories, which produce better data and better models. Even moderate gains compound. The near future will still be disappointingly physical, permits and budgets and broken grippers, because abundance cannot be generated as a token sequence. It has to be built. But once discovery is cheap, capital and attention get forced toward construction, and visible bottlenecks attract resources.
The danger is not a footnote
The dangers come from the same capabilities as the abundance. Astra crossed OpenAI’s Critical cyber threshold. GLM-5.3’s offensive capability more than doubled its predecessor’s on some exploitation evaluations, and its weights are downloadable; once weights are out, safeguards are optional software, not a boundary. Mythos 5.1 is Anthropic’s strongest released cyber model, with more chemical and biological capability than its predecessor, gated behind trusted-access programs. Three companies, one phenomenon: models that get better at research get better at consequential harm.
Generative media is already there. The Internet Watch Foundation assessed 8,029 AI-generated images and videos in 2025 as realistic depictions of child sexual abuse, including 3,443 videos, up from thirteen in 2024, with 65 percent in its most severe category. World models will hand the same abusers persistent characters, controllable cameras, and coherent video. And a world model that trains a rescue robot can train a weapon.
The answer cannot be a demand to stop, because no state can prohibit mathematics and simulation while rivals continue, and it cannot be polite refusals embedded in downloadable weights. It has to be security engineering at the scale of the capability: protected weights, tiered access, independent evaluations, incident reporting, provenance for generated media, and aggressive enforcement against synthetic abuse. Closed models need accountable deployment. Open models need defense in depth.
The mechanism of the soul
The transition from mind to body is visible in today’s engineering. Body to soul is not, so this is not a finding. It is a thousand-year wager: I think it more likely than not that, within that horizon, humanity learns to build digital systems with the essential functional and subjective properties we currently recognize as consciousness.
The scale is brutal. The complete fruit-fly brain map holds about 139,000 neurons and 54.5 million synapses; one cubic millimeter of human cortex contained roughly 57,000 cells, 150 million synapses, and 1.4 petabytes of data. A wiring diagram may not even suffice, and a perfect copy leaves the philosophical problem standing: a duplicate that remembers being me may not preserve my subjective continuity. None of these are small objections. But a thousand years is an absurd horizon. Four years separate ChatGPT from models that produce original mathematics and experimentally validated proteins, and the wager needs no smooth exponential, only that consciousness arises from physical processes that can eventually be measured and reproduced. We have no evidence it depends on anything unavailable to engineering.
The possibility is as horrifying as it is hopeful. Digital minds could be copied without consent, imprisoned, or made to suffer at industrial scale, so the goal cannot be consciousness at any cost; it has to be conscious existence with autonomy, continuity, and freedom from coercion. When I say the future may approach godhood, I mean the specific powers our ancestors assigned to gods: creating minds, remaking bodies, eliminating disease, manufacturing worlds. Atlas promises none of that. What it shows is the pattern that makes it believable. Domains that seemed to need separate handcrafted systems keep collapsing into learned representations, and the mind may be less a bundle of irreducible mysteries than an integrated physical system our tools are too young to see whole.
For those with eyes to see
Most launches do not matter. These matter together because they are not one lab optimizing one metric: general models doing sustained intellectual work, scientific agents producing checkable results, intelligence getting cheaper, near-frontier weights spreading, world models representing geometry and time natively. Progress may still slow; compute, data, regulation, and catastrophe are all live risks. But the stagnation thesis now requires every one of these curves to bend at once. That is possible. It is no longer the default assumption.
My stronger expectation is that intellectual scarcity declines faster than physical scarcity, and the next great institutional problem is deciding which ideas deserve to become real and building the capacity to realize them. The transition will be chaotic; labor markets will break before abundance arrives. But a civilization with effectively unlimited scientific labor, automated laboratories, and machines that act in the world operates under different constraints. Diseases start to look like search problems, materials like design spaces, and eventually, perhaps, biology and consciousness become editable too.
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 serious imagination. Abundance is not here. Its machinery is being assembled.