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)

2,391 Signals Tracked
7 Collisions Identified
10 Frontier Science Cards
280 Source Domains

This fortnight brought 2,391 signals across corporate, startup, thinktank, and research sources into focus, and the throughline is the permission layer — the sudden, universal scramble to decide who is allowed to act, who gets to approve, and what counts as real. Last fortnight agents were storming the storefront; this fortnight the doors slammed shut. United Airlines blocked customer bots, Walmart built Agent Profiles to tag and gate them, and an SAP study found 87% of U.S. consumers want to approve every AI purchase — the autonomy everyone was selling turns out to be the autonomy almost no one wants ungoverned. Permission showed up as an authenticity crisis too: Dior was accused of 'craftwashing' a supposedly hand-drawn campaign, India's ad regulator warned that AI labels don't verify a single claim, and Svedka handed its Super Bowl ad back to humans after an AI backlash. It showed up on your face, as Google opened pre-orders for Gemini eyewear with Warby Parker and Gentle Monster while Amazon strapped 20,000 pairs onto delivery drivers. It showed up in the conversation, as Google launched multi-turn ads inside AI Mode and chatbot ad formats began to multiply. And it showed up as a reliability reckoning — Gartner now expects 40% of agentic projects to be scrapped by 2027 even as Anthropic's agent leapt from completing 2.5% of freelance jobs to 41% in a year. Underneath it all, AI quietly crossed a line from answering questions to running the experiment — designing circuits, decoding DNA cell by cell, and steering microbots through blood vessels. Twenty-one fortnights in, the pattern is clear: the capability is no longer the hard part. The hard part is permission — who grants it, who withholds it, and who you trust to act in your name.

01
The Door Slams on the Agent

A fortnight after brands raced to get inside the AI shopping conversation, the biggest enterprises started locking their own doors — and a new study revealed consumers never wanted full autonomy in the first place.

+

The agentic-commerce gold rush hit its first wall. United Airlines became the latest enterprise to shut its systems to customer bots, joining a quiet backlash against the autonomous shoppers that foundation-model makers have spent a year promoting. Walmart took the opposite-but-equivalent tack, readying a program called Agent Profiles that assigns dedicated identities to AI agents so their actions can be distinguished — and governed — separately from human customers. The reason the walls are going up became clear in a new SAP study of consumer attitudes: 87% of U.S. and 83% of German shoppers want to review and approve every final purchase an AI makes on their behalf, and a large share say they would never let an AI buy anything unsupervised at all. The infrastructure players are already building for that reality — fintech Atomic FI is raising capital specifically to build a platform that lets banks grant personal AI agents scoped, revocable permission to act on a customer's accounts. The throughline: the question has flipped overnight from 'how do we let agents in?' to 'how do we decide which agents, doing what, with whose consent?'

⚡ The Now

The first chapter of agentic commerce assumed autonomy was the product. The second chapter is discovering that permission is the product. Enterprises are learning that an open door to anonymous bots is an open door to fraud, scraping, and liability, so they are replacing it with gates: block the unidentified ones (United), tag and scope the rest (Walmart), and build the plumbing to grant and revoke authority safely (Atomic). And consumers, it turns out, are on the enterprises' side — overwhelmingly they want AI to shortlist and recommend, then step back and let a human press 'buy.'

→ What's Next

Expect 'agent identity and permissioning' to become a real product category — the OAuth of the agent era — with banks, retailers, and platforms all issuing credentials that say which agent may do what, for whom, up to what limit. The brands that win won't be the ones that block agents or the ones that fling the doors open; they'll be the ones that design a graceful approval experience, where the agent does the tedious 95% and hands the human a clean, confident final decision. Watch for the first high-profile dispute over an agent that overspent or bought the wrong thing — the liability precedent it sets will shape the whole category.

Computerworld
United Airlines joined a growing list of enterprises blocking customer-facing AI bots, while Walmart announced an 'Agent Profiles' program that assigns dedicated profiles to AI agents to differentiate their actions from human customers.
Sourcing Journal
An SAP study found 87% of U.S. and 83% of German consumers want to review and approve final purchases despite AI's growing influence, and that brands must make products more AI-readable while risking becoming easier to replace.
Banking Dive
Atomic FI is raising capital to build a platform that lets bank and fintech clients enable personal AI agents to act on customers' behalf — combining its existing account-action and data-pulling capabilities under scoped permissions.
02
Craftwashing

As AI seeps into every campaign, the most valuable signal a brand can send is that a human actually made the thing — and the most damaging accusation is that one didn't.

+

Authenticity became the fault line of the fortnight. Dior's holiday film, promoted as a hand-drawn collaboration with acclaimed French animator Sylvain Chomet, was accused on social media of 'craftwashing' — selling AI-assisted work as artisanal craft — after it emerged that automated interpolation (and, critics alleged, generative AI for elements like furniture and bees) was used; Chomet admitted AI smoothed some animation while denying the core was machine-made. The deeper problem was named by India's Advertising Standards Council (ASCI), whose draft guidance warned that slapping an 'AI-generated' label on an ad does nothing to tell consumers whether its claims — fabricated experts, fake endorsements, exaggerated results — are actually true. Brands are responding by reaching back toward verifiable humanity: Svedka handed creative control of its 2027 Super Bowl spot to the public via a $500,000 contest after a polarizing AI-made debut, and Ogilvy partnered with talent agency CAA to build AI ads from licensed celebrity scans that are watermarked as AI and approved by the performers, complying with SAG-AFTRA consent rules. The common thread: in a world where anything can be synthesized, provenance and consent become the product.

⚡ The Now

AI made content infinitely cheap, which made the signal of authenticity infinitely valuable. 'AI or not' has become advertising's loudest argument, but the real question underneath it is trust: a disclosure label tells you how something was made, not whether what it says is true. Brands caught optimizing for efficiency (Dior) are discovering the authenticity backlash can cost more than the production saved, while brands that lean into verifiable humanity (Svedka's crowdsourcing, Ogilvy/CAA's consented, watermarked scans) are turning provenance into a marketing asset.

→ What's Next

Watermarking, content credentials, and talent-consent frameworks will shift from compliance checkbox to brand strategy — 'made with permission, verified by provenance' becomes a quality claim, the way 'organic' or 'fair trade' did. Expect regulators to follow ASCI's logic and target deceptive claims rather than AI usage per se, which means the 'AI label' debate will quietly give way to a 'truthful claim' standard. And watch for a premium to re-emerge on demonstrably human craft — the hand-drawn, the live-performed, the imperfect — precisely because it can no longer be assumed.

Glossy
Dior's holiday animated campaign, promoted as hand-drawn work with animator Sylvain Chomet, faced accusations of 'craftwashing' after it used automated interpolation for intermediate frames and was alleged to have used generative AI for some elements.
AdExchanger
The Advertising Standards Council of India's 2026 draft guidance identifies fabricated experts, fake endorsements, and exaggerated results as high-risk AI uses that an 'AI-generated' label alone cannot address — disclosure reveals method, not truthfulness.
AdWeek
After a polarizing AI-made debut, Svedka is crowdsourcing its 2027 Super Bowl commercial through a public contest with a $500,000 prize, reframing creative control as a human, community-driven effort.
AdExchanger
WPP-owned Ogilvy partnered with Creative Artists Agency to create AI-generated ads from digital face and body scans of CAA's celebrity clients — watermarked as AI and approved by the talent, complying with SAG-AFTRA's AI consent regulations.
03
The Face Computer Arrives

Gemini-powered eyewear went on sale through fashion houses and onto the faces of delivery drivers in the same fortnight — the ambient-AI interface stopped being a demo and became a product you can pre-order.

+

The next computing interface this fortnight wasn't in your hand — it was on your face. Google opened pre-orders for its Intelligent Eyewear lineup, Gemini-powered glasses designed in collaboration with fashion brands Gentle Monster and Warby Parker plus Samsung, and spent as much energy on the privacy architecture as the optics: no camera visuals saved to Gemini history, automatic redaction of bystanders' faces, recording-indicator lights, a physical off switch, and 18+ account gating. The fashion channel is doing the cultural work of making them wearable — Warby Parker opened a literary-themed 'Literatureland' pop-up, while Gentle Monster ran a surreal film with Joaquin Phoenix and a celebrity dinner — reframing smart glasses from creepy surveillance gadget to intellectual accessory. Meanwhile the enterprise edge arrived faster than the consumer one: Amazon is scaling from a few thousand smart delivery glasses in 2026 to more than 20,000 by the end of 2027, putting turn-by-turn navigation, package verification, and hazard flags directly in drivers' field of view. Two very different launches, one structural shift: the AI assistant is leaving the screen and taking up residence on the human body.

⚡ The Now

For a decade smart glasses were a punchline; this fortnight they became a shipping product category with real distribution. The two go-to-market paths are revealing: Google is routing consumer adoption through fashion credibility (and an unusually heavy privacy-by-design story) because the barrier was never the tech but the social license to wear a camera on your face. Amazon skipped that debate entirely — in a logistics context, a face computer is just a better clipboard, and 20,000 units is a deployment, not an experiment.

→ What's Next

The battleground will be consent and the bystander. Google's redaction and indicator lights are a bet that ambient AI only scales if the people around the wearer are protected, not just the wearer — expect that to become a regulatory flashpoint in the EU and several U.S. states. On the enterprise side, hands-free, glanceable AI will quietly reshape frontline work (delivery, warehousing, field service, healthcare) faster than it reshapes consumer life, because the ROI is legible and the privacy objections are weaker. The phone isn't going away, but the center of gravity for 'ask the AI' is migrating from a thing you hold to a thing you wear.

Google (The Keyword)
Alphabet opened pre-orders for Intelligent Eyewear collections designed with Gentle Monster, Warby Parker, and Samsung, integrating Gemini AI, alongside privacy features including bystander-face redaction, recording indicators, a physical off switch, and no camera visuals saved to Gemini history.
Glossy
Warby Parker launched a literary-themed 'Literatureland' pop-up and Gentle Monster ran a surreal celebrity film and dinner to promote their Gemini-powered smart glasses, reframing the devices as intellectual, fashionable accessories rather than surveillance gadgets.
Sourcing Journal
Amazon plans to deploy 5,000 smart delivery glasses in 2026 and expand to more than 20,000 by the end of 2027, giving drivers in-field navigation, delivery instructions, and package verification.
04
The Synthetic Consumer

Campaigns, researchers, and marketers began replacing real human panels with AI-simulated ones — raising the vertiginous question of whether a signal generated by a model is a signal at all.

+

A strange inversion took hold this fortnight: having spent years building AI to serve humans, the industry started using AI to simulate humans. Political firm FP1 Strategies used Google's Nano Banana image tool to generate detailed synthetic 'voter archetypes' — complete with biographies — as a stand-in panel for targeting ads, and separately produced a six-hour AI parody spot using an opponent's likeness. Startup Aaru, profiled by Goldman Sachs, has gone further, building population-simulation models that fuse large language models with behavioral data to forecast how any population will act — what they'll buy, which campaigns they'll engage, what they'll stream — effectively offering a synthetic focus group of an entire electorate or customer base. And the synthetic humans are now making media of their own: Meta's Muse can research, record, and host fully AI-voiced podcasts on its servers, with more than 20 shows already live. The collision: the feedback loops that businesses and campaigns use to understand real people are quietly being populated with people who don't exist.

⚡ The Now

Synthetic panels are seductive because they're fast, cheap, and never fatigued — Aaru can simulate a population in the time a traditional pollster schedules a single focus group, and FP1 can spin up a representative 'voter' on demand. But they rest on a load-bearing assumption: that a model trained on human data reproduces human behavior faithfully enough to make real decisions on. When the simulation is wrong, it's wrong invisibly and at scale, and because it's cheap, it crowds out the expensive real-world signal that would have caught the error.

→ What's Next

Expect a sharp bifurcation. Synthetic respondents will become standard for early, directional, high-volume testing (headline variants, ad concepts, message framing) where speed beats precision. But a premium will re-emerge on verified human signal for the decisions that matter, precisely because synthetic data is subject to model collapse — AI trained on AI-generated behavior drifting from reality. The sleeper risk is epistemic: as campaigns optimize against simulated voters and brands optimize against simulated shoppers, both may get very good at persuading populations that only exist inside a model. The discipline to watch is 'synthetic validity' — proving a simulated panel actually predicts real behavior before you bet a budget on it.

Campaigns & Elections
FP1 Strategies used Google's Nano Banana image-generation tool to create detailed synthetic voter archetypes with biographical notes for ad targeting, and produced a six-hour AI parody ad using an opponent's likeness, clearly framed as parody to avoid deepfake concerns.
Goldman Sachs
Aaru launched AI simulation models that understand the composition of any population globally and forecast their decisions — product purchases, campaign engagement, music streaming, transit usage — by fusing large language models with behavioral data.
Podnews
Meta's Muse, an AI assistant launched in September 2026, can research, record with artificial voices, and host fully automated podcasts on Meta's servers, with more than 20 AI-generated shows already live on the platform.
05
Ads Break Into the Conversation

The advertising unit is migrating out of the results page and into the dialogue — and the shopping journey is following it, with Gemini quietly overtaking ChatGPT as the assistant Americans shop with.

+

The attention economy found its next real estate: the inside of the chat. Google launched multi-turn ads for its AI Mode, surfacing sponsored results as users ask follow-up questions, and expanded AI Mode ads to more than 10 global markets — monetizing the conversation rather than the search page. The shift is costing Google elsewhere and, per AdExchanger, those AI ad losses are becoming Meta's gain as chatbot ad formats start to multiply across the ecosystem. The demand side is reorganizing around it: ad-tech firm PubMatic is going all-in on agentic media buying, raising the industry's own question of how much power is too much when machines buy ads from machines. And consumer behavior is validating the land grab — an Omnisend survey found Google Gemini has overtaken ChatGPT as the AI assistant most used by American shoppers, 53% to 46%, meaning the discovery journey that used to start at a search bar increasingly starts inside an assistant that can now carry an ad. The collision: the ad, the shopping trip, and the answer are collapsing into a single conversational surface.

⚡ The Now

For twenty-five years, search advertising lived in a stable grammar: a query, a page of blue links, ads at the top. Generative answers broke that page — and with it the ad inventory Google was built on. Multi-turn ads are the industry's answer: if the user now has a conversation instead of a search, put the sponsorship inside the conversation. That it's already live in 10+ markets, and that Gemini has quietly become the top shopping assistant, means this isn't a pilot — it's the early formation of the post-search ad market, and Meta is positioning to capture the budgets that get dislodged.

→ What's Next

Expect a new ad-format war fought over conversational placements — how a sponsored answer is disclosed, how 'organic' a recommendation inside a chat can honestly be, and whether users will tolerate ads in a medium that felt, briefly, ad-free. Agentic buying (PubMatic) will accelerate the machine-to-machine ad economy, concentrating power in whoever owns both the assistant and the ad exchange. The uncomfortable question for brands: when the assistant both answers the question and takes the ad money, the line between a recommendation and a paid placement gets very thin — and the regulator who first forces clear disclosure inside AI answers will reshape the whole market.

AdExchanger
Google launched multi-turn ads for AI Mode, displaying sponsored results as users ask follow-up questions, and expanded AI Mode ads to more than 10 global markets, while chatbot ad formats multiply across the ecosystem and Meta captures dislodged ad budgets.
Drug Store News
An Omnisend survey found Google's Gemini has overtaken ChatGPT as the AI assistant most used by American shoppers for shopping tasks, with 53% using Gemini compared to 46% using ChatGPT.
AdExchanger
Ad-tech firm PubMatic is going all-in on agentic media buying, in which AI agents purchase advertising inventory, raising industry questions about the concentration of power when machines buy ads from machines.
06
The Reliability Reckoning

Even as AI agents crossed stunning new capability thresholds, the enterprises deploying them pulled back hard — the gap between what an agent can do and what you can trust it to do unsupervised became the fortnight's defining tension.

+

Two numbers told opposite stories this fortnight, and both were true. On capability: Anthropic's Fable 5.2 now completes 41.3% of real freelance jobs on the Scale Labs benchmark — up from 16% in July and just 2.5% a year ago — a near-vertical climb in what an agent can actually finish. On discipline: Gartner now predicts more than 40% of agentic AI projects will be canceled by the end of 2027, undone by escalating costs, unclear business value, and inadequate risk controls. The ground-level reality tracks the caution — Pets+'s 2026 survey of independent retailers found they use AI as a 'thinking partner,' not a decision-maker, routinely questioning its outputs and leaning on their own experience to keep the human touch. Gartner even named the opposite failure mode: 'cognitive surrender,' the tendency of employees to over-trust AI outputs without scrutiny, which it says requires deliberate training exercises like 'hallucination hunts.' The collision: capability is going vertical while trust is being rationed — and the organizations that survive the gap are the ones treating human oversight not as a temporary scaffold but as permanent architecture.

⚡ The Now

The demo-to-deployment gap is the real story of enterprise AI right now. A model that can finish 41% of freelance jobs in a benchmark is thrilling; a model you'd let run unsupervised across your actual business is a different, much higher bar — and Gartner's 40% cancellation forecast is what happens when companies discover that bar in production. The smart operators (Pets+'s independent retailers) never confused the two: they adopted AI as augmentation, kept a human in the loop, and treated every output as a draft. The dangerous failure isn't the agent that's obviously wrong — it's 'cognitive surrender,' the quiet erosion of the human judgment that was supposed to be the safety net.

→ What's Next

Expect 'agent reliability engineering' to mature into a discipline the way site reliability did — evaluation harnesses, confidence thresholds, graceful human handoff, and audit trails becoming table stakes before any agent touches a customer or a dollar. The projects that get canceled won't mostly be the ones with weak models; they'll be the ones that skipped the oversight architecture and discovered the cost of an unsupervised mistake. And watch for the capability and discipline curves to converge: as models like Fable keep climbing, the organizations that invested in the permission-and-oversight layer early will be the ones able to safely widen the aperture, while everyone else stays stuck re-piloting.

The Decoder
Anthropic's Fable 5.2 completes 41.3% of real freelance jobs on the Scale Labs benchmark — up from 16% in July 2026 and 2.5% a year ago — excelling at editing audio and music and designing architecture, marking a steep climb in agent capability.
Retail Dive
Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls, highlighting the gap between agent capability and reliable deployment.
HR Dive
Gartner identifies the need to combat 'cognitive surrender' — employees over-relying on AI outputs without critical evaluation — recommending training exercises like 'hallucination hunts' to rebuild human scrutiny of AI.
Pets+
Independent pet retailers use AI cautiously, treating it as a thinking partner rather than a decision-maker — questioning AI outputs and relying on their own experience to preserve the human touch in their businesses.
07
AI Becomes the Instrument

Beneath the consumer headlines, AI crossed a quieter line this fortnight — from generating answers to running the experiment, designing the hardware, and decoding the biology itself.

+

The most consequential shift this fortnight wasn't in commerce — it was in the lab and the workshop, where AI stopped being a tool you query and became an instrument that discovers. Anthropic showed that Claude (Fable 5.2) can design a working electronic prototype end to end, autonomously revising its own designs — reassigning signals to different pins, rerouting copper connections — when it hits a problem, and separately demonstrated AI that builds scientific models using simulation-based inference to infer parameters that defeat traditional methods. Google DeepMind released AlphaGenome 2, trained on over 107 million individual cell profiles, which predicts how genetic variants affect specific human cell types and generalizes to new single-cell experiments without retraining. In robotics, UC Berkeley's agentic approach let robots write their own skills and hit 99% success rates — matching hand-engineered performance — while Nvidia's GR00T 2 learns a task by watching a single human video and imagining itself as the robot. The collision across circuits, genomes, and machines: AI is moving up the value chain from producing content to producing knowledge and designing the physical world.

⚡ The Now

For three years the frontier models made text, images, and code — artifacts for humans to use. This fortnight's research signals show a categorical jump: AI generating the hypothesis (simulation-based inference), the hardware (Claude's electronics), the biological insight (AlphaGenome 2 reading DNA cell by cell), and the robot's own behavior (Berkeley's self-authored skills). The common structure is self-correction — these systems don't just emit an answer, they test it against the world or a simulation and revise, which is the defining move of an instrument rather than an oracle.

→ What's Next

Expect R&D cycle times to compress in exactly the domains where AI can close the loop between design and test — chip and PCB design, drug and materials discovery, robotics skill acquisition. The bottleneck shifts from 'can we design it?' to 'can we validate and manufacture it fast enough?' — the same permission-and-trust theme running through this whole edition, now applied to science: who signs off when an AI designs the circuit or proposes the genomic target? The enterprises that win the next decade won't be the ones with the best chatbot; they'll be the ones that figured out how to put AI in the loop of their actual invention process — and how to trust what it hands back.

Anthropic Research
Anthropic showed Claude Fable 5.2 can autonomously design electronic prototypes end to end, revising its own designs when it encounters issues — reassigning signals to different pins and rerouting copper connections to resolve problems.
Google DeepMind
DeepMind's AlphaGenome 2, trained on over 107 million individual cell profiles, predicts how genetic variants affect specific human cell types and generalizes to new single-cell experiments without retraining by compressing gene-activity snapshots into cell-state representations.
IEEE Spectrum
UC Berkeley's agentic robotics approach let robots write their own skills, achieving success rates up to 99% — matching human hand-engineered performance and significantly outperforming baseline imitation-learning techniques.
The Decoder
Nvidia's GR00T 2 learns tasks by watching human video and simulating itself as the robot; for 50 selected tasks, under three minutes of additional training raised success rates from 40% to 95%, and it learned to hammer nails from a single hand video.

Frontier Science Seeding the System

Underneath the enterprise headlines, the research slice this fortnight showed AI and its instruments moving deeper into the body, the genome, and the physical world — steering microbots through blood vessels, reading DNA one cell type at a time, predicting a hypertensive crisis days ahead, and turning your own forearm into a trackpad. These are the signals that don't fit a collision but are too remarkable to skip.

Biomedical Robotics
ETH Zurich researchers built microbots from a biocompatible gel laced with magnetic iron-oxide nanoparticles that can be steered into blood vessels by an external magnetic field, where their motion generates flow that helps dissolve blood clots — a potential non-invasive stroke and thrombosis therapy.
Genomics
Google DeepMind's AlphaGenome 2, trained on over 107 million individual cell profiles across 68 tissues, predicts how genetic variants affect specific human cell types and generalizes to new single-cell experiments without retraining — sharpening the search for disease-causing mutations.
Precision Oncology
A Dana-Farber-led international team built MANTIS, an AI model that uses DNA methylation profiling to generate personalized survival predictions for children with medulloblastoma — helping clinicians tailor treatment intensity to each child's risk.
Digital Health
Google Research developed an AI that detects signs of high blood pressure using an ordinary smartwatch's pulse and motion sensors — no inflatable cuff or calibration required — opening the door to passive, population-scale hypertension screening.
Predictive Medicine
University of Maryland researchers applied machine learning to continuous wearable blood-pressure data and predicted hypertensive crises up to 72 hours in advance in high-risk ambulatory patients, presented at the American Heart Association's 2026 Scientific Sessions.
Human-Computer Interaction
North Carolina State University's EpiPad uses a compact millimeter-wave radar wristband to turn a smartwatch user's forearm into a multi-touch trackpad, solving the 'fat finger' problem of tiny watch screens without any sensors on the arm itself.
Speech AI
A KAIST team led by Professor Yong Man Ro improved visual speech recognition using a diffusion large language model, with plans to shrink the model for real-time lip-reading on wearables and robots without an internet connection — aiding accessibility and noisy-environment communication.
Computational Chemistry
Hebrew University of Jerusalem researchers built a machine-learning and molecular-modeling framework that screened over half a million molecules to discover new candidate sugar substitutes — a computational shortcut to sweeteners that could sidestep the drawbacks of existing ones.
AI for Engineering
Anthropic showed Claude Fable 5.2 can autonomously design electronic prototypes from start to finish, revising its own work when it hits a problem — reassigning signals to different pins and rerouting copper connections — a step toward AI as a hardware co-designer.
Neurotechnology
Precision Neuroscience raised $250 million in Series D funding at a $1 billion post-money valuation, underscoring surging investor conviction in brain-computer interfaces as surgical robotics and neurotech drive medtech VC growth.