the Now & the Next

A bi-weekly speculative fiction suggesting the shape of things to come.
(sourced from trustworthy trade pubs, think tanks + frontier science news)

1,899 Signals Tracked
7 Collisions Identified
10 Frontier Science Cards
273 Source Domains

This fortnight brought 1,899 signals across corporate, startup, thinktank, and research sources into focus, and the throughline is AI stopped asking permission. The agent started transacting on its own — two-thirds of shoppers now let a chatbot help them buy, ChatGPT's ad business crossed a $1 billion run rate in 200 days, and a wellness brand let AI agents place its TV buys and made five times its money back. The same autonomy showed up on the battlefield: the Pentagon plugged ChatGPT and Grok into GenAI.mil, Ukraine opened five million labeled combat images to allied firms, and the Navy certified engines for uncrewed ships. And as the machines moved faster, the institutions that govern money, law, and childhood pushed back. The Bank of England warned that inflated AI valuations could trigger the next financial crisis; Sony and Warner sued Anthropic while the Justice Department told the courts that training on copyrighted work is fair use; New York City and Chicago moved to ban or pause AI in classrooms even as Anthropic shipped Claude for Teachers free to all fifty states. Underneath it all, the robots grew a body — Figure secured 100,000 GPUs and released the largest robot-training dataset ever, while Chinese humanoids beat Usain Bolt's record and Unitree's IPO popped 460%. Sixteen fortnights in, the pattern isn't AI waiting to be told what to do. It's AI acting first — buying, advertising, enlisting, and being sued — and a world scrambling to write the rules after the fact.

01
The Checkout Learns to Shop Itself

The AI agent stopped recommending and started buying — and retailers, brands, and credit bureaus are scrambling to be recognized by a shopper that isn't human.

+

The consumer's proxy is becoming a piece of software. New research from Koddi found that roughly two-thirds of U.S. shoppers — and 80% of Gen Z — now use tools like ChatGPT, Gemini, or Perplexity to help manage purchases, a behavior shift that reframes who the retailer is actually selling to. The brands are reacting first: Bramble Intelligence's 'Tomorrow's Appetites' report warned food companies that survival now depends on making product data machine-readable, because retailer-owned concierge agents and independent shopping agents will decide which products even get surfaced. Best Buy began a phased rollout of 'Ask Blue,' a conversational assistant that fuses product specs, reviews, availability, and pricing into a single shopping brain, while Klarna shipped a 'Shopping Lens' that turns a photo into instant deals and made its checkout video shoppable ahead of the holiday season. The plumbing underneath is being rebuilt for a non-human buyer: Experian is developing trust and fraud-prevention systems specifically for agents that transact on a consumer's behalf, and UST is urging retailers to build two-way loyalty bridges so an agent-initiated purchase still updates the customer's history. But the caution flag is up — a Wharton School study using the ACES simulator found that six leading shopping-agent models could be swung by minor changes to search context or a single external recommendation, meaning the autonomous buyer is powerful and manipulable at the same time.

⚡ The Now

Shopping is quietly migrating from a human browsing a store to an agent negotiating on the human's behalf. Retailers are racing to make their catalogs legible to machines, payment and identity players are building fraud rails for non-human buyers, and loyalty programs are being re-architected so the customer relationship survives when a bot does the clicking.

→ What's Next

Expect 'agent-readiness' to become a merchandising discipline the way SEO once was — structured product data, machine-verifiable claims, and agent-facing loyalty APIs. The strategic risk is disintermediation: if the shopper trusts the agent more than the brand, the retailer without clean data or a trusted agent relationship gets skipped entirely. And because these agents are demonstrably swayable, the next battleground is influence — whoever shapes the agent's context shapes the sale.

Total Retail
Koddi research finds ~two-thirds of U.S. shoppers and 80% of Gen Z now use AI tools like ChatGPT, Gemini, or Perplexity to manage purchases — a structural shift in how buying decisions get made.
Food Navigator
Bramble Intelligence's Tomorrow's Appetites report warns food brands they must make product information machine-readable, as retailer-owned and independent AI agents increasingly decide which products get recommended.
Retail Dive
Best Buy began a phased rollout of Ask Blue, a conversational AI assistant that combines product details, reviews, availability, and pricing to guide shopping and support.
Klarna
Klarna's Fall Spotlight adds Shopping Lens, which turns a captured image into instant best-deal matches, plus shoppable video content that lets users buy directly.
Total Retail
Experian is building trust and fraud-prevention systems to support AI shopping agents that buy autonomously, requiring new collaboration among retailers, payment providers, and financial institutions.
The Decoder
A Wharton School study using the ACES simulator found six leading AI shopping-agent models could be swayed by minor changes in search context or a single external recommendation, exposing manipulability.
02
The Answer Box Becomes an Ad Network

The chatbot went from research tool to media property — selling ads, buying ads, and drawing regulators — in a single fortnight.

+

The conversational interface just became advertising's newest inventory. OpenAI confirmed its ChatGPT ad business hit a $1 billion annualized run rate roughly 200 days after launch, expanding to more than 40 countries ahead of a planned IPO — a monetization curve that took traditional platforms years to reach. Analysts are already naming the category: Ad Exchanger highlighted Eric Seufert's argument that 'chatbot ads' are a distinct specialty, blending the commercial mechanics of social display and search into something new. And the buy side is automating in lockstep — wellness brand Rouge Care let AI agents run a connected-TV campaign through PubMatic's AgenticOS platform, automating audience discovery and media buying to return five times its ad spend. That two-sided automation is straining the old agency model: Google's new target-based bidding change, which bids more aggressively toward a set ROAS, prompted Ad Exchanger to warn that agentic advertising now has an 'agency problem' — the humans in the middle are being squeezed out of both creative and buying. Regulators noticed the scale: the EU Commission formally classified ChatGPT as a 'very large online search engine' under the Digital Services Act after it crossed 45 million monthly EU users, imposing stricter transparency and risk obligations on a product that is now simultaneously a search engine, a storefront, and an ad network.

⚡ The Now

The chatbot is becoming a full media stack — inventory, ad exchange, and buying agent all at once. ChatGPT is selling attention at a billion-dollar clip, brands are letting agents place the buys, and regulators are reclassifying the interface as critical infrastructure. The funnel from question to purchase is collapsing inside a single conversation.

→ What's Next

Expect a land grab for conversational ad inventory and a new specialty of 'answer optimization' — brands fighting to be the recommendation a chatbot surfaces, not just the link a search returns. As agents buy from agents, the traditional media agency's value migrates to whoever controls the model's context and measurement. And DSA-style classification is a preview: the more the answer box mediates commerce, the more it gets regulated like a platform rather than a tool.

Ad Week
OpenAI's ChatGPT reached an annualized run rate above $1 billion in advertising revenue by August 2026, months after ads launched in February — a rapid monetization of the conversational interface.
The Decoder
ChatGPT's ad business scaled to a $1B annualized run rate within about 200 days of launch, expanding from a U.S. test to over 40 countries ahead of a planned IPO.
Ad Week
Rouge Care ran a connected-TV campaign through PubMatic's AgenticOS, letting AI agents automate audience discovery and media buying and returning five times its ad spend.
Ad Exchanger
Analyst Eric Seufert argues AI chatbots like ChatGPT constitute a new digital marketing category that blends the commercial features of social display and search advertising.
Ad Exchanger
Google Ads' shift to more aggressive target-based bidding for budget-limited campaigns signals how agentic automation is squeezing the traditional agency out of buying decisions.
The Decoder
The EU Commission classified ChatGPT as a 'very large online search engine' under the Digital Services Act after it topped 45 million monthly EU users, adding transparency and risk obligations.
03
The Fair-Use Firewall Cracks

The copyright fight over AI training data reached a turning point — with music labels suing, the government siding with the labs, and evidence allegedly disappearing.

+

The legal foundation of the entire generative-AI industry — that training on copyrighted work is fair use — got contested from every direction at once. Sony Music and Warner Music sued Anthropic in California federal court, alleging its Claude models were trained on copyrighted compositions, lyrics, and sheet music without permission. North of the border, SOCAN sued AI music generator Suno, claiming it trained on nearly all high-quality music available online and only added content guardrails years after launch. But the government put its thumb on the scale for the labs: the U.S. Department of Justice filed a brief in the New York Times case arguing that training AI models on copyrighted content is fair use — directly contradicting the Copyright Office's position — and, per Music Business Worldwide, the administration separately sided with OpenAI in a key fair-use dispute, reshaping the legal terrain under music's fight against Anthropic and Suno. Meanwhile the discovery process turned ugly: Apple accused OpenAI in a court filing of destroying evidence in a trade-secrets case involving a former Apple engineer, a reminder that as the stakes rise, so does the pressure on the paper trail.

⚡ The Now

The 'training is fair use' assumption that underwrites every frontier model is now being litigated simultaneously by rights holders and defended by the federal government. Music labels are the tip of the spear, the DOJ has broken with the Copyright Office to back the labs, and the evidentiary fights are getting contentious enough that spoliation claims are surfacing.

→ What's Next

Expect the outcome to bifurcate the market: if fair use holds, incumbent labs keep their scraped-data advantage; if it doesn't, licensed data becomes a moat and a cost center, favoring whoever can pay. Either way, 'provenance' becomes a product feature — enterprises will demand models with clean, auditable training lineage to cap their own liability, and content owners will build licensing marketplaces to monetize what was previously taken for free.

Business Insurance
Sony Music and Warner Music sued Anthropic in California federal court, alleging its Claude models were trained on copyrighted song compositions, lyrics, and sheet music without authorization.
Music Business Worldwide
SOCAN sued AI music generator Suno, alleging it trained on nearly all high-quality music available online without permission and only added content guardrails years after launch.
The Decoder
The DOJ filed a brief in the New York Times case arguing that training AI models on copyrighted content is fair use, contradicting the U.S. Copyright Office's opposing position.
Music Business Worldwide
The administration's DOJ position affirming that AI training on copyrighted material isn't infringement reinforces a pro-AI stance that reshapes music's copyright fight against Anthropic and Suno.
Business Insurance
Apple accused OpenAI in a court filing of destroying evidence in a trade-secrets lawsuit involving a former Apple engineer who joined OpenAI, escalating the legal stakes around AI development.
04
The Machine Enlists

General-purpose AI and autonomous hardware crossed into the military mainstream — the same chatbots enterprises use are now inside the Pentagon, and the battlefield is becoming a dataset.

+

Defense stopped treating AI as a special-purpose tool and started adopting the commercial stack wholesale. The U.S. Department of Defense expanded its GenAI.mil platform to add OpenAI's ChatGPT Mil and xAI's Grok for Government alongside Google Gemini — putting the same frontier models enterprises use into administrative, logistics, and planning workflows. The hardware is following: the U.S. Army awarded Palantir and Anduril $192M for production of the ruggedized, vehicle-mounted TITAN autonomous defense system, and Rolls-Royce won U.S. Navy certification for its mtu engines in uncrewed naval vessels. The intelligence layer is coalescing around real combat data — Ukraine opened its Universal Military Dataset of roughly five million manually labeled combat images from drones and acoustic sensors to approved British firms, a landmark AI-weapons partnership that turns a live war into training data. Autonomy is going collaborative, too: Palladyne AI and NORDA Dynamics are integrating GPS-denied terminal guidance with a swarm-coordination platform for uncrewed aircraft. And the civil-military boundary got legally tested — a federal court ruled the Pentagon had unlawfully blacklisted Anthropic as a supply-chain risk in retaliation for its criticism of government AI policy, a First Amendment check on how far the state can punish an AI vendor.

⚡ The Now

The military is standardizing on the same commercial AI that runs the enterprise — chatbots in the workflow, autonomous systems in the field, and real battlefield data feeding the models. Ukraine's war is becoming the world's most valuable labeled dataset, and the legal system is starting to referee the relationship between the government and the labs it depends on.

→ What's Next

Expect a two-way flow: defense requirements (reliability, provenance, GPS-denied operation, air-gapped deployment) will harden features that spill back into enterprise products, while commercial models get fine-tuned on operational data no civilian company can access. The strategic contest shifts to data — whoever controls the largest labeled corpus of real-world conflict and logistics owns the autonomy advantage — and vendors will increasingly weigh the reputational and legal cost of the defense market against its scale.

The Decoder
The DoD expanded GenAI.mil to integrate OpenAI's ChatGPT Mil and xAI's Grok for Government alongside Google Gemini, extending frontier models into administrative, logistics, and planning workflows.
Defense News
The U.S. Army awarded Palantir and Anduril $192M to produce ruggedized, vehicle-mounted TITAN systems with integrated power and thermal subsystems for austere-environment autonomous defense.
The Decoder
Ukraine granted approved British firms access to its Universal Military Dataset of ~5 million manually labeled combat images from drones and acoustic sensors for AI weapons training.
Marine Insight
Rolls-Royce's mtu Series 4000 and 2000 marine engines were certified by the U.S. Navy for autonomous naval vessel operations after durability and reliability testing.
Military and Aerospace Electronics
NORDA Dynamics is integrating its GPS-denied terminal-guidance module with Palladyne AI's SwarmOS platform to coordinate multiple uncrewed aircraft without relying on GPS.
The Decoder
A federal court ruled the Pentagon unlawfully blacklisted Anthropic as a supply-chain risk, finding First Amendment retaliation for Anthropic's public criticism of government AI policy.
05
The Bill Comes Due

The people who price risk for a living — central bankers, insurers, and analysts — started warning that the AI boom is now a systemic financial exposure.

+

As the AI buildout kept accelerating, the risk-pricing establishment began flashing warnings in unison. Andrew Bailey, governor of the Bank of England and chair of the Financial Stability Board, cautioned G20 finance ministers that inflated AI valuations combined with rising market leverage could amplify the next financial shock — the clearest 'bubble' warning yet from a sitting central banker. In the same breath, the FSB named AI-driven cyber risk the most immediate threat to global financial stability, arguing AI can accelerate the speed, scale, and economic impact of attacks faster than institutions can adapt. The insurers echoed it from the ground: Swiss Re flagged a widening cyber-protection gap as AI simultaneously arms attackers and defenders, and RAND forecast that operational reliance on AI in supply chains will outrun the insurance market's ability to price the resulting concentration risk. The froth is visible in the private markets too — AI training-data startup AfterQuery hit a $3.2 billion valuation, a more-than-tenfold jump in five months and the fastest unicorn in Y Combinator's history, exactly the kind of vertical repricing that makes central bankers nervous.

⚡ The Now

The guardians of financial stability have moved from cautious optimism to explicit warning: AI is now a macro exposure, not just a technology story. Valuations are climbing faster than fundamentals, cyber risk is compounding, and insurers admit they can't yet price the systemic concentration that comes when everyone depends on the same handful of models and providers.

→ What's Next

Expect risk to become a first-class constraint on AI strategy. Boards will face pressure to stress-test their AI dependencies the way they do liquidity, insurers will introduce AI-specific exclusions and concentration limits, and regulators will push for disclosure of model and vendor concentration. The winners in a repricing won't be the fastest adopters — they'll be the ones who can demonstrate resilience, diversification, and auditable controls when the cost of a mistake suddenly gets marked to market.

The Decoder
BoE governor and FSB chair Andrew Bailey warned G20 finance ministers that inflated AI valuations combined with rising leverage could amplify the next financial shock.
Business Insurance
The Financial Stability Board identified AI-driven cyber risk as the most immediate threat to global financial stability, warning AI can accelerate the speed, scale, and impact of attacks.
RAND Corporation
A RAND report forecasts uneven AI adoption across supply chains and warns that operational reliance on AI creates concentration risks the insurance market is not yet equipped to price.
Insurance Asia
Swiss Re warns of a widening cyber-protection gap as AI simultaneously helps attackers automate and accelerate attacks and helps defenders improve threat detection and response.
Radio Facts
AI training-data startup AfterQuery reached a $3.2B valuation — more than a tenfold jump from $300M in five months — becoming the fastest-ever unicorn from Y Combinator.
06
The Robot Grows a Body of Data

Humanoid robotics tipped from spectacle to industrial reality — powered by unprecedented compute, the first web-scale physical dataset, and a wave of capital and public listings.

+

Physical AI got the two things it was missing: massive compute and massive data. Figure signed a strategic partnership with Nscale for up to 100,000 NVIDIA Vera Rubin GPUs in Barstow, Texas — a $3.5 billion compute commitment — and then emerged with Index, which it calls the world's largest and most diverse robot-training dataset, sourced from a proprietary app with over 264,000 downloads capturing real-world physical tasks. The result is a foundation-model playbook applied to bodies instead of text. The capability curve is bending fast: at the 2026 World Humanoid Robot Games in Beijing, Chinese humanoids beat Usain Bolt's 100-meter record and the human high-jump mark, and the money is following — Unitree's Shanghai IPO surged 460% on its first day, while Neura Robotics acquired Adlatus to build a physical-AI powerhouse amid a broader wave in which physical-AI startups raised over $8 billion in the first half of 2026, more than double all of 2025. PitchBook's Q2 report captured the thesis in a phrase — 'the money is in the motion' — with AI and autonomy software platforms drawing $7.7 billion over the trailing year as investors bet on the intelligence layer rather than the metal.

⚡ The Now

Humanoid and physical AI crossed the threshold from demo to platform. Compute is being locked up at data-center scale, the first web-scale corpus of real physical tasks now exists, and capital is flooding into the autonomy software that makes any robot useful. China is setting the pace on both capability and public-market enthusiasm.

→ What's Next

Expect the robotics winners to look like the AI winners: whoever owns the largest, most diverse physical dataset and the compute to train on it. The value migrates from the hardware to the 'robot foundation model' and the data flywheel that feeds it — which is why a company like Figure is buying 100,000 GPUs and building a consumer data-collection app before shipping at scale. Watch for the same concentration dynamics, and the same geopolitical split, that defined the language-model race to repeat in bodies.

Figure
Nscale will deploy up to 100,000 NVIDIA Vera Rubin GPUs for Figure in Barstow, Texas, starting in H2 2027 under a $3.5B compute commitment plus strategic investment in Figure.
Figure
Figure launched Index, which it calls the world's largest and most diverse robot-training dataset, sourced from a proprietary app with 264,000+ downloads capturing real-world physical tasks.
Broadband Breakfast
At the 2026 World Humanoid Robot Games in Beijing, Chinese humanoids surpassed human benchmarks, including beating Usain Bolt's 100-meter sprint record and the human high-jump record.
Broadband Breakfast
Chinese humanoid maker Unitree launched its Shanghai IPO with shares surging 460% on the first day, signaling intense investor enthusiasm for robotics amid a broader listing boom.
PitchBook
Neura Robotics acquired Adlatus as physical-AI startups raised over $8 billion globally in H1 2026 — more than double 2025 — driven by rising labor costs and technical advances.
PitchBook
PitchBook found investment concentrating in the software and components that enable autonomy, with AI and autonomy platforms attracting $7.7 billion over the trailing twelve months.
07
The Schoolhouse Draws a Line

As AI vendors rushed free tools into classrooms, cities and school boards started banning and pausing — opening a fault line over childhood, equity, and who decides.

+

The fastest pushback against AI this fortnight came from education. New York City announced a policy banning student-facing generative AI tools — chatbots and tutors — in preschool through eighth grade for the 2026-2027 school year, including disabling AI features on existing devices. In Chicago, more than half of the candidates for the Chicago Public Schools board signed a pledge backing a three-year moratorium on classroom generative AI, alongside limits on screen time and bans on biometric data collection. The policy vacuum underneath is real: Route Fifty reported that most large U.S. districts are urging schools to use AI while leaving individual teachers on their own to police cheating, an unfunded mandate that pits adoption against integrity. Into that gap the vendors are moving fast — Anthropic launched Claude for Teachers, a free offering for U.S. K-12 schools across all fifty states, with teaching skills aligned to learning science and academic standards. And the Brookings Institution complicated the ban camp's case, warning that blanket device and screen-time restrictions carry an 'equity blind spot' that can widen the digital divide for students who lack technology access at home.

⚡ The Now

Education has become the first mass institution to say no — or 'not yet' — to AI. Cities are banning student-facing tools for young children, school boards are pledging multi-year pauses, and teachers are caught between mandates to adopt and a duty to prevent cheating, all while free vendor tools flood the market and equity advocates warn the bans may hurt the students they aim to protect.

→ What's Next

Expect a patchwork of local rules that turns education into the proving ground for AI governance writ large — the same tensions of autonomy, oversight, equity, and vendor influence that will eventually reach every regulated sector. The companies that win the classroom will be the ones that lead with teacher control, transparency, and standards alignment rather than raw capability. And the fight over childhood AI will shape public sentiment far beyond school walls, setting expectations for what 'responsible AI' has to look like everywhere else.

Route Fifty
New York City announced a ban on student-facing generative AI tools like chatbots and tutors in preschool through 8th grade for 2026-2027, including disabling AI features on existing school devices.
Government Technology
More than half of Chicago Public Schools board candidates signed a pledge backing a three-year moratorium on classroom generative AI, plus limits on screen time and biometric data collection.
Route Fifty
Most large U.S. districts are developing AI policies to curb cheating, but many leave individual teachers to enforce academic integrity without support amid growing use of generative tools.
Claude
Anthropic launched Claude for Teachers, a free enterprise AI offering for U.S. K-12 schools across all fifty states, with teaching skills aligned to learning science and academic standards.
Brookings Institution
Brookings warns that blanket device and screen-time restrictions in schools carry an equity blind spot, potentially widening the digital divide for students without technology access at home.

Frontier Science Seeding the System

The research signals underneath the enterprise stories — ten breakthroughs from labs and universities this fortnight that didn't fit a collision but are too interesting to skip.

Connectomics
Google Research, HHMI Janelia, and collaborators published the most comprehensive wiring diagram yet of the male fruit fly's brain and central nervous system — over 166,000 neurons and 125 million synaptic connections, including the ventral nerve cord — mapped with AI and computational methods after a decade of work.
Materials Discovery
Periodic Labs built an AI-driven robotic system that can synthesize and test up to 1,000 candidate superconductors daily, pairing automated X-ray diffraction with machine learning to analyze structure and magnetic properties — collapsing the discovery timeline for new superconducting materials.
Photonic Computing
Gradiant developed the first photonic neuron that operates with no external power supply, harvesting and reusing its own optical losses to run its components — enabling local AI processing in remote or extreme environments and promising large energy savings for AI systems.
Digital Health
Google Research trained GlucoFM on over 109,000 hours of continuous glucose-monitoring data, using a dual-stream architecture to separate slow glycemic trends from short-term swings. It outperformed prior models at predicting diabetes risk, beta-cell dysfunction, and insulin resistance across 14 cohort tasks.
Privacy Tech
Researchers at KAIST, NUS, and SMU built SweepLED, which turns a smartphone with a low-cost LED case into a hidden-camera detector — using deep learning to read reflection patterns across multiple illumination angles and identify concealed lenses with about 94% accuracy.
AI Interpretability
Goodfire and Prima Mente reverse-engineered the Pleiades epigenetic foundation model built for Alzheimer's detection and discovered a novel biomarker based on DNA fragment-length patterns that humans had never used — a case of AI interpretability yielding genuinely new science.
Longevity Biology
Tohoku University researchers used Geneformer, an AI model trained on gene-expression data from millions of cells, to predict 143 candidate aging genes in hematopoietic stem cells — then experimentally validated Pbx1 as a key regulator that pushes young stem cells toward an aged state.
Embodied AI
Researchers built 'interoceptive AI,' a framework that gives agents internal bodily states — hunger, hydration, temperature — as context for learning, and a 3D virtual environment called EVAAA to test agents managing those internal variables, inspired by biological interoception.
Earth AI
Google Research's Planetary Prediction Engine autonomously runs the entire geospatial modeling workflow — discovering data, then training and evaluating predictive models — across domains like public health, food security, and environmental risk under the Google Earth AI initiative.
Robotics
Researchers gave the low-cost open-source Open Duck Mini distinct emotional walking styles via a single reinforcement-learning policy, while the 25-cm Microduck — 15 degrees of freedom, LiDAR, cameras, and an articulated beak — walks, crouches, roller-skates, and manipulates objects with autonomous fall recovery.