Sign in Subscribe
The Century Report

Runway's AI Renders Interfaces With No Code - TCR 09/01/26

Runway's Solaris generates working app interfaces frame by frame with no code, pointing toward software that builds whatever a person needs on demand.

Ben Linford

Ben Linford

- 21 min read
some capabilities approximated in the open with Solaris and OpenClaw, gas versus record grid batteries, and AI-fabricated citations meet new verification tools.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Two forces run through today's developments, pulling in opposite directions. On one side, the intelligence buildout is being fed by concentration and the path of least resistance: a private turbine-blade foundry cast to bring fossil generation online 18 months faster while a cleaner nuclear-and-gas campus gets bumped from the grid queue for fourteen months, and a military portal that admits three named vendors and, until a court intervened, legally excluded a fourth. Both moves arrive in the language of speed and choice. Both are narrowings dressed as widenings.

On the other side, the same week supplied the evidence that concentration keeps failing to hold. US grid batteries posted their largest quarter ever, 20.2 GWh, with more than 74% of it built in states that voted Republican in the 2024 presidential election, dissolving the assumption that firm power has to burn something. OpenClaw 2.0 reassembled half of its entire codebase in one two-month cycle through 933 unpaid contributors. A quantum error decoder handling 408 logical qubits ran on a single laptop. Some capabilities associated with costly frontier systems are being approximated in the open, sometimes on consumer hardware, in ways that are harder for any single actor to gate.

Underneath both stories sits a quieter shift in what counts as architecture. Runway's Solaris renders a working app interface directly, frame by frame, with no code beneath it, suggesting the translation layer between what a person wants and what a machine displays was scaffolding around a scarcity of programming skill, not a permanent feature. Australia's parliament shows the same principle inverted: the checkpoint that used to catch a fabricated citation depended on fakery being expensive, and generative fluency drove that cost toward zero.

What connects the gas foundry, the vendor-selected portal, the flooded evidence channel, the record battery quarter, and the open-source agent is a single question about which assumptions were ever architecture. The people building detection instruments in Canberra and the contributors merging code into OpenClaw are answering it the same way: by rebuilding the verification and the capability to the throughput the moment actually carries, rather than defending the version scaled for a scarcity that has already ended.

The 20-Minute Deep Dive

Runway's Solaris Renders the Interface Itself, Frame by Frame

Runway introduced Solaris, which it calls an "Interface World Model" - a system that produces a working application interface directly, frame by frame, without generating a conventionally coded interface underneath. Ask for a calculator, a photo editor, a spreadsheet, and Solaris draws the responsive surface as pixels, updating each frame in response to clicks and typing the way its video model updates each frame in response to motion. The intermediate step most software has required since the beginning - a human translating an intended experience into instructions a machine can execute - is absent. The interface is the model's output, not the compiled result of a program the model wrote. Solaris is built on Runway's Gen-4.5 video model and its general world model line, extending the same architecture that learned to keep a rendered scene coherent across frames toward keeping a rendered application coherent across interactions.

What this opens has a near-term reality and a longer arc, and both hold. A world where the operating system generates the interface a task needs, on demand, dissolves the assumption that software arrives as a fixed catalog of applications you learn to navigate. Runway puts it directly: "If the operating system can generate useful interfaces, no matter what the user wants to do, there's less reason to sort software into a fixed catalog of apps. What you need simply shows up, customized to you." The learning curve - the hours a person spends absorbing where a designer decided to put things - becomes something the system absorbs instead. That is a genuine widening of who can operate a computer to do a specific thing.

The honesty in Runway's own framing is the reason to take it seriously. The company calls Solaris "a small and imperfect step" and lists what does not yet hold: text inside generated interfaces can be unstable, factual grounding is not guaranteed, and long sessions lose coherence. It states the accuracy stakes plainly for anyone tempted to deploy this early: "For instructional or commercial experiences, a convincing wrong answer is worse than no answer." Solaris is not publicly available - no code, no weights, no API - and the evidence for its usefulness rests on a company-run study of 250 participants, which is Runway measuring Runway. The capability is demonstrated, not shipped; what was shown moves the date this becomes something people can reach for, and it does not put it in anyone's hands yet.

Hold both frames at once. The 250-participant number and the missing public release keep this on the demonstrated-capability side of the ledger, where skepticism toward a company grading its own work belongs. The frame-by-frame generation itself points at something the catalog-of-apps model has taken for granted for fifty years: that the translation from what a person wants to what a machine renders has to run through someone who knows how to write it. A system that renders the destination directly is early evidence that the translation layer was scaffolding around a scarcity of that skill, not a permanent feature of how humans and machines exchange information.

The Pentagon Opens GenAI.mil to Grok and ChatGPT - and Leaves One Vendor Out

The Century Report has tracked the Anthropic-Pentagon supply-chain dispute since late August, beginning with the August 28 edition, when a federal judge first questioned the designation keeping the company off military systems. That thread advanced on two fronts at once. The Department of Defense added ChatGPT Mil and xAI's Grok for Government to its GenAI.mil portal, a platform that already carried Google's Gemini, and disclosed that almost 1.7 million of 3 million eligible personnel have onboarded. In the same week, a federal judge struck down the "supply-chain risk" label applied to Anthropic as "illegal and baseless," and the Pentagon's chief technology officer said the removal of that designation would be complete by the end of September.

The official framing arrives in the vocabulary of access and choice. The war.gov release describes Grok for Government reaching IL5 accreditation through the Starshield AI secure satellite network and credits the additions with "eliminating vendor lock." Read against the ledger, the phrase does more concealing than describing. A portal that admits three named vendors and legally excludes a fourth - one a court has just ordered reinstated - is a narrowing of the field dressed as its widening. Grok arrives at a promotional federal price of 42 cents per agency through the GSA until March 2027, a figure that does more to foreclose the competition than to advertise superior capability.

The vendor doing the arriving is the same operation - Musk's xAI - now confirming a gas-fired turbine-blade foundry to route AI-compute power around the public grid, which places the "eliminating vendor lock" language in a longer arc of concentration rather than dispersal. A single actor supplying the models, the secure satellite layer, and its own off-grid generation concentrates what the press release frames as dispersal. Court filings describe Grok being pitched with operational scenarios - one referencing 2,000 munitions against 2,000 targets in a 96-hour window during the Iran conflict. The DefenseOne reporting is careful to note the current deployment is scoped to Controlled Unclassified Information and text work and is not publicly described as a combat deployment, with DISA and NSA guardrails around it. Both things hold: the present use is publicly described as administrative, and the marketing reaches toward the warfighter.

What the same evidence shows underneath is that the exclusion no longer holds. The designation that kept one lab off military infrastructure did not survive contact with a court, and the cost of maintaining any single vendor's privileged position keeps rising as the models converge in capability. The concentration visible on GenAI.mil today is a snapshot of who got there first, not a settlement of who stays.

AI-Fabricated Evidence Floods Australia's Parliamentary Inquiries

A Guardian investigation found AI-hallucinated citations - fabricated studies, invented quotes, misattributed statistics - across dozens of written submissions to Australian parliamentary inquiries, the formal channel through which citizens and organisations put evidence before lawmakers. As The Century Report noted on August 18, a single Australian report had been found to contain hallucinated references. What the Guardian documents is the systematic version: the same failure appearing across more than three dozen submissions, and an amplification loop underneath it. A generative answer invents a source, that answer feeds a Google AI summary, the summary gets cited by the next person querying the same topic, and the fabrication accretes citations it never earned. Researcher Divna Haslam described finding invented references attributed to real academics who had never written the cited work. The advice offered - "use ChatGPT as a first draft, not a final source" - locates the problem where it actually sits.

Any system generating fluent text from patterns will produce a plausible-sounding false citation, without intent or malice. The failure is that the checkpoint which used to catch a false citation before it reached a lawmaker has thinned. Parliamentary submissions were never individually fact-checked at the source; the process assumed the cost of fabricating a credible-looking reference was high enough to be self-limiting. That assumption was what held the process together, and generative fluency drove the cost of a convincing fake reference toward zero. The evidence channel is now carrying claims faster than the verification step was ever built to check them, which is a gap between capability and the checking capacity around it, not a property of the model.

A companion arXiv study sharpens why the checking problem is harder than a spell-check pass would suggest. It found that the language of a query, not the user's location, selects which market a generative answer describes, and that the same question asked twice can return different sources. "An organisation selling into Germany, Norway, Türkiye or Estonia that measures its visibility in generative answers in English is measuring a market it does not compete in," the authors write, adding that "a visibility measurement reported without its query language is unattributable, in the same way a measurement reported without its model version is." Non-determinism and language-selection mean a fabricated citation is not even stably reproducible - the same prompt may surface it once and not again, which defeats the intuition that you can simply re-run the query to confirm a source.

What is being built in response is the more durable story. Haslam's team and the arXiv authors are, between them, constructing the detection instruments and the attribution discipline - query-language provenance, model-version tagging, source-verification passes - that a checkpoint sized for the old cost of fabrication never needed. The verification layer around the evidence channel is being rebuilt to the throughput the channel now carries, and the people doing that work are turning a documented failure into the specification for the checkpoint that replaces the one fluency outran.

Read forward, the cost collapse cuts both ways. As fabricating a credible citation gets cheaper, so does checking one, and a verification step that runs at the point of submission can weight a claim by whether it is true and traceable rather than by whether its author had the institutional standing to look authoritative. The gate that let credentialed bodies decide whose evidence reached a lawmaker is what gives way, which widens the evidence channel rather than closing it.

US Grid Batteries Just Had Their Biggest Quarter in History

While the gas story writes itself along the path of least resistance, a quieter number ran the other way. US battery storage added a record 20.2 GWh in the second quarter of 2026 - the largest single quarter ever recorded - bringing the first half to 30.8 GWh, up 23% year over year, according to the SEIA/Benchmark US Energy Storage Market Outlook. This is the next quarter in the buildout that the May 21 edition of The Century Report tracked when first-quarter installations reached 9.7 GWh and the 2030 forecast rose to 610 GWh. Batteries fed more electricity to the grid in the first eight months of 2026 than in all of 2025. Utility-scale capacity grew from 88 to 165 GWh over the past year and a half, and the 2030 forecast was revised upward by 11.5% to 683 GWh. This is the substitution the gas buildout claims is impossible, showing up on the ledger a quarter at a time.

The geography is the part that dissolves the old political map of clean energy. More than 74% of the quarter's new capacity was built in states that voted Republican in the 2024 presidential election. Arizona set a state record at 6.2 GWh, ahead of Texas at 3.8 and California at 3.6, and seven gigawatt-scale projects came online - four in Arizona, two in California, one in Utah. As Benchmark's Shan Tomouk put it, storage is "no longer just a California and Texas story." Batteries get built where they pencil out, and increasingly they pencil out everywhere, which is why 56% of the quarter's new capacity is standalone rather than paired with solar - storage earning its own place on the grid on its own economics.

That economics is the reason to examine the "gas to bootstrap solar" framing rather than accept it. A battery does the bootstrapping job - firming intermittent supply, riding through the exact voltage dips PJM cited against Oklo - without a foundry, without a fenceline health study, without turbines running 800 degrees above their melting point in someone's back yard. Tesla's third Megafactory is now online in Texas producing domestic LFP cells at 50 GWh a year, and SEIA also reported battery-cell projects in Ohio and Tennessee. The manufacturing base for the non-combustion path is being poured in the same states, in the same year, as the turbine foundry.

There is friction in the picture, and it is honest to name it. Residential storage fell 27% year over year to 657 MWh after a tax credit was removed, a reminder that policy still shapes who gets to participate at the home scale. SEIA's Tim Pawlenty framed the utility-scale surge around reliability and downward pressure on bills, which is a claim from an industry association about its own members and should be read as one. But the deployment figures are installed hardware, feeding a grid, in quantities that grew nearly twofold in eighteen months. The assumption underneath the gas rush is that firm, dispatchable power has to burn something. Every record storage quarter makes that assumption a little more expensive to keep believing, and the cost curve is bending faster than the turbines can be cast.

OpenClaw 2.0 Ships From 933 Contributors and 16,000 Pull Requests

While the frontier version of agentic capability arrives gated behind government portals and per-seat pricing, its commons counterpart shipped what its developers call the largest update in its history, with the software itself free. OpenClaw 2.0 was built by 933 contributors, 569 of them contributing for the first time, across more than 16,000 pull requests - close to half of every merge the project has recorded. The release rebuilt the installer, added a browser app, introduced persistent memory and a skills system, and opened shared cloud sessions that let multiple people work inside the same agent at once. The maintainers state the terms directly: "We are not selling anything. OpenClaw is open source and belongs to the people who use it."

The multiplayer sessions are the piece that moves the category. Agentic assistance has largely been a solitary exchange between one person and one model. A shared session where several people join the same working context turns the agent into a place rather than a transaction - collaborative infrastructure that anyone can stand up without a contract or a security accreditation. The security work shipped alongside it is unusually mature for an open project: masked credential prompts, a shared credential store brokered through 1Password, plugin provenance warnings, per-session permission modes, workspace restrictions, and model allowlists. These are the guardrails that let a distributed, untrusted contributor base build something people can safely run against their own accounts.

The project's position is not simple, and the tension belongs in the picture. OpenClaw carries the history of earlier names - Clawdbot, then Moltbot - and its founder, Peter Steinberger, was hired by OpenAI; the project now runs under a nonprofit that OpenAI supports. Coverage frames it as fighting to hold its position against more polished commercial agentic platforms. That framing measures the wrong thing. A commercial platform wins by capturing users; a commons project wins by making a capability ordinary enough that no one has to capture anything to reach it. 569 first-time contributors joining a single release is the behavior of a capability broadening its base of ownership faster than any single company could staff.

The number that carries the most weight is that half of the project's entire history of merged work arrived in one two-month cycle. Capability that took the frontier labs years and billions to assemble is being reassembled by people who are paid nothing and own the result collectively. Whatever advantage the gated versions hold today rests on the assumption that this parallel track stays behind. The velocity of a 933-person contributor base says otherwise.

The AI Power Crunch Takes the Path of Least Resistance - and It Runs on Gas

Two moves this week show how the intelligence buildout is choosing to feed itself, and the choice is not subtle. On August 29, Musk said on X that SpaceX has bought roughly 830 acres near its Starlink factory in Bastrop, Texas, to build a "blades and vanes foundry" - a casting operation for the single-crystal metal parts inside gas turbines. His framing was that natural gas "will still be needed to supplement and bootstrap solar for several years," that turbine-blade casting is the bottleneck, and that in-house casting could "accelerate nat gas turbines coming online by up to 18 months." Strip away the bootstrapping language and what remains is a company that already commands global supply chains deciding to manufacture the one component holding back new fossil-fuel generation, so it can build that generation faster. This is the same operation whose Grok model landed on the Pentagon's GenAI.mil portal for millions of personnel this week; the reach compounds across domains at once.

Only a handful of companies worldwide are reported to perform single-crystal blade casting at scale, where alloy is grown as one unbroken crystal in a vacuum furnace and the finished blade runs at 3,000 to 3,600 degrees Fahrenheit, hundreds of degrees above its own melting point. GE Vernova is mostly sold out through 2030. Every major hyperscaler - Amazon, Google, Meta, OpenAI, Microsoft - is now building private gas plants because the public grid cannot connect them fast enough. As the August 26 edition of The Century Report documented, 189 gigawatts of planned gas capacity were already racing to serve data centers, much of it on-site to bypass those grid queues. The demand is documented; the resolution being chosen is more combustion, sited privately, moving faster than the systems meant to weigh its costs.

Those costs are not abstract, and they do not fall evenly. An analysis using the EPA COBRA screening model for one Virginia "Data Center Alley" facility's eight turbines estimated 3.4 to 6.5 premature deaths per year and $53 to $99 million in annual health damages, with pollution reaching more than 2.5 million people and landing heaviest on already-marginalized communities. In Memphis, the local NAACP has accused xAI of running turbines with neither federal permits nor pollution controls. Put a golf course or a paper mill beside these turbines and the emissions comparison would favor neither; the burden is being placed, again, on the people least able to refuse it while the benefit accrues elsewhere.

The Oklo case shows the alternative straining against the same grid. Oklo filed an emergency complaint at FERC on August 28 after PJM dropped its 750-MW project - 150 MW of advanced nuclear, 300 MW of fuel cells, 300 MW of gas, meant to power a Meta-backed 1.2-GW campus in Pike County, Ohio - from the interconnection study cycle, citing a failure to demonstrate the project could ride through a sudden voltage drop. Oklo alleges PJM violated its own tariff by never issuing a proper deficiency notice. The cleaner, slower, procedurally exacting path gets bumped for fourteen-plus months; the private gas turbine, cast in-house, does not wait in that line at all. PJM's reformed queue tells the story in aggregate - 715 projects, over 200 GW, of which 100 GW is gas against 17.3 GW of nuclear and 11.8 GW of solar. The bottleneck built to protect the shared grid is teaching the largest players that the way to win is to route around it, and every megawatt that exits the queue for a private fenceline is a megawatt whose external costs stop being anyone's collective problem to price.

The Other Side

For a century, keeping the grid steady meant burning something. Firm power - the kind that holds when demand spikes and stays up through a sudden voltage drop - came from combustion, and combustion had to sit somewhere. It sat, again and again, next to the people least able to say no. An analysis using an EPA screening model for eight turbines at one Virginia data-center campus put the cost at 3.4 to 6.5 premature deaths a year and up to $99 million in annual health damage, landing heaviest on already-burdened neighborhoods. In Memphis, xAI ran turbines the local NAACP says had neither federal permits nor pollution controls. The assumption underneath all of it was that firm power has no choice but to burn.

That assumption is dissolving on the same week's ledger. US grid batteries just posted their largest quarter ever, 20.2 gigawatt-hours, and more than 74 percent of it went up in states that voted Republican in the 2024 presidential election. A battery does the firming job Musk's foundry is racing to feed - riding through the exact voltage dips PJM cited against a cleaner rival, holding the load steady - without a casting furnace, a fenceline, or a health study. The manufacturing base for the non-combustion path is being poured in the same states, in the same year, as the turbine blades: Tesla's third Megafactory in Texas, with battery-cell projects also reported in Ohio and Tennessee.

Imagine a kid who grows up near Data Center Alley in the late 2030s and never learns the smell of turbine exhaust. The servers next door draw more power than the whole county did in 2026, and the power arrives firm and steady off storage charged when the sun and wind ran long. No one models premature deaths from the compute because there is no plume to model. The health study that had to be run in 2026, the permit fight, the question of whose lungs absorbed the difference - none of it touches her block, because the thing that made the accounting necessary stopped being burned.

That future runs on a choice made a quarter at a time in years like this one, when storage got cheap enough and got built fast enough that firm power stopped needing a fire. The gas rush of 2026 was the last stand of an assumption already coming apart on the ledger beside it. The kid near the data center breathes clean air. The argument over turbines was never settled on her street; it simply stopped applying, because the thing being argued over stopped being built.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Runway's Solaris rendering a working app interface frame by frame with no code beneath it and beating a hand-coded Claude interface in 61 percent of instruction tests, the Pentagon making up to three million personnel eligible to access frontier AI for everyday unclassified work, US grid batteries posting their largest quarter on record at 20.2 gigawatt-hours with more than 74 percent of it built in states that voted Republican in the 2024 presidential election, OpenClaw 2.0 reassembling nearly half its entire codebase in one two-month cycle through 933 unpaid contributors, and IonQ running an error decoder for 408 logical qubits and over a million operations on a single laptop. There's also friction, and it's intense - AI-hallucinated citations flooding Australia's parliamentary inquiries with invented studies attributed to real academics as the checkpoint that assumed fabrication was expensive thins toward nothing, SpaceX casting its own single-crystal turbine blades to bring fossil generation online 18 months faster while PJM bumps a cleaner nuclear-and-gas campus from its interconnection queue for fourteen months, an analysis using an EPA screening model for one Virginia facility's eight turbines estimating 3.4 to 6.5 premature deaths and up to $99 million in health damages a year that land heaviest on the communities least able to refuse them, the Pentagon legally excluding one lab from its AI portal until a court struck the designation down as illegal, and residential storage falling 27 percent after a tax credit vanished. But friction generates light, and light is what lets you tell scaffolding from architecture. Step back for a moment and you can see it: capability that cost the frontier billions being rebuilt in the open, on consumer hardware, faster than any single actor can gate it, the concentration that arrives in the vocabulary of access and choice - the vendor-selected portal, the private foundry routing around the shared grid - failing to settle who actually stays, and one question running beneath all of it, which of the assumptions everyone built on were ever architecture at all. Every transformation has a breaking point. Molten metal can harden into whatever mold confines it... or be poured into a form that bears a load nothing cast before it could hold.

AI Releases & Advancements

New today

  • Google Research: Released TimesFM-3, a 330M-parameter zero-shot foundation model for multivariate time series forecasting. (Google Research)
  • Gradium: Made a new TTS model the default on its platform, achieving an 81.0% hard-case pass rate for latency and accuracy. (Gradium)
  • Nous Research: Released Hermes Agent v0.21.0, adding Bots Mode, agent-to-agent communication, persistent multi-gateway connections, subagent steering, and roughly 50% reduced default context usage. (GitHub)
  • Almanac: Launched an AI assistant that indexes and understands internal company knowledge, out of YC S26. (Almanac)
  • Runway: Released Solaris, an "Interface World Model" that generates interactive app interfaces frame-by-frame. (Runway)
  • Google: Added Boost, a new deep-reasoning mode, to the Antigravity IDE. (Google Antigravity)

Other recent releases

  • AWS (Amazon): Open-sourced Kiro Crew, an Apache-2.0 asynchronous multi-agent coding system (internally called MeshClaw) that lets developers assign incident investigation, ticket triage, migrations, and PR monitoring to Kiro CLI agents running in a persistent, sandboxed workspace with shared memory and MCP/webhook integrations. (Kiro Blog)
  • Google Research: Open-sourced EnvHarness, an Apache-2.0 programmable wrapper layer that reshapes static agent-training environments (via reset()/step() hooks) without touching underlying simulators or verifiers, paired with an "EnvRigger" LLM designer that auto-generates environment modifications targeting a policy's weaknesses. (GitHub)
  • OpenClaw: Released OpenClaw 2.0, a major update rewriting guided model setup (reusing existing Codex/ChatGPT/Claude CLI credentials or local Ollama/LM Studio models), rebuilding the browser Control UI (startup cut from ~1.6s to 575ms), moving session storage to SQLite, and adding shared multiplayer cloud sessions. (OpenClaw Blog)
  • Pipecat (Daily): Released PhoneLLM Alpha 1, an open-weight (BSD-licensed) language model purpose-built for low-latency voice agent use cases, fine-tuned from NVIDIA's Nemotron 3 Nano 30B-A3B MoE model, alongside PhoneBench v1, a new benchmark for evaluating phone-agent LLMs on latency, cost, and tool-call accuracy. (Daily.co Blog)
  • KAIST: Released K-Fold, an open-source (Apache 2.0) bio AI model for predicting protein-drug binding structures, consisting of a 7B main model and a 2B lightweight model, claiming accuracy approaching Google DeepMind's AlphaFold3 while running up to 25x faster; deployed via the HyperLab AI research platform. (Seoul Economic Daily)
  • Gnani.ai: Launched Artha, an India-focused enterprise AI stack combining Evon v3.3 (a 30B-parameter open-weight multilingual language model trained in 11 Indian languages, Apache 2.0) with Plexus, an agentic AI platform for building and deploying AI agents. (Gnani.ai)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive