World Changing Technology in 2026

ELI5/TLDR

Four Stanford scholars sit down to map the technologies actually bending the world in 2026 — AI, synthetic biology, the cyber arms race, and the algorithmic shaping of what we believe. The takeaways: biology is quietly turning into a manufacturing technology as general-purpose as compute; the US-China AI contest is really seven contests, not one; Anthropic’s “Mythos” model just spooked the world by being good enough at cyber to find vulnerabilities in almost any software; and most of what we think we know about the world is filtered through digital feeds engineered to keep us engaged, not informed. The mood is sober but not apocalyptic — closer to: the tools are powerful, the guardrails are missing, and the adults need to show up.

The Full Story

A new word: technopolitics

Colin Kahl, the new FSI director, opens with a borrowed phrase from Ian Bremmer — technopolitics. The idea has two parts. First, we now live in a world where technologies will reshape national security, prosperity, and daily life. Second, those technologies are coming almost entirely from private companies whose market caps are larger than the GDPs of G20 countries. Trillion-dollar firms with near-sovereign control over the digital spaces where we actually live. They are political actors in their own right, whether or not anyone asked them to be.

That sets the frame for the panel. The four guests handle different corners of the same animal: Drew Endy on synthetic biology, Andy Grotto on the AI race and cyber, Jeff Hancock on how platforms shape minds, and Jennifer Pan on digital authoritarianism in China.

Biology is becoming a general-purpose technology

Drew Endy makes the case that biotech is on track to become what AI is — a general-purpose technology, but for manufacturing instead of computing. His numbers: biotech is already a $4 trillion global industry, with about $1 trillion of it counted as the US genetically-engineered domestic product. McKinsey thinks that grows to $30 trillion over 25 years. If that happens, most of the physical inputs to the economy will be biomanufactured — which, oddly, is what they were a thousand years ago, before petroleum elbowed everything else out.

The newest layer is synthetic biology, which is what happens when you stop just editing organisms and start composing them from scratch. The enabling tech is synthetic DNA chemistry — machines that print DNA from a database. Endy compares its importance to silicon wafer manufacturing for computing, except the stuff being printed encodes life itself.

But a DNA printer is only as useful as the sequences you can feed it, and writing in DNA is hard. Enter AI. A Stanford colleague, Brian He, trained a large language model called EVO (and now EVO-2) on natural DNA sequences. Prompt it, and it emits new sequences. Pair that with AlphaFold-style structure prediction — built on about $100 billion of public investment in protein structures — and you have a flywheel. The bottleneck now is data: there is plenty of sequence and structure data, but very little systematic data on how living systems actually behave. Endy wants “large language laboratories” that generate that data at scale.

He adds a third leg: energy. Civilization needs both primary energy (electricity, fuel) and embodied chemical energy (the stuff inside materials). The latter is the part most people forget. By 2050, the world will need maybe 8 more terawatts of embodied chemical energy, and it is unclear where it comes from. His pet idea is electro-biosynthesis, or e-bio — use surplus solar electricity to pull carbon from air, convert it to formate, and feed it to microbes instead of glucose. Photovoltaics replacing leaves. Pacific Northwest National Labs and ADM are early movers. Endy ranks the potential alongside the invention of synthetic fertilizer.

The AI race is actually a heptathlon

Andy Grotto doesn’t love the word “race.” He prefers decathlon, but settles on seven contests:

Frontier — who has the most advanced models. Closest thing to the “country of geniuses in a data center” Dario Amodei talks about.
Applications — who builds the killer apps customers actually want.
Diffusion — who reaps the productivity gains as AI spreads through the economy, while managing the labor-market disruption.
Warfighting — who adopts AI militarily. Sometimes civilian and military adoption move at different speeds.
Manufacturing — for AI outputs that aren’t software (medicines, materials), who can actually make them.
Market share — whose AI stack captures the world. Important because technologies carry the values of their makers.
Governance — whose rules and values shape how AI is used and regulated globally.

The US is six to nine months ahead at the frontier. China is competitive or ahead almost everywhere else. The US strategy is roughly: let the market flood the zone. The Chinese strategy is all-of-the-above, including subsidising adoption in markets the US private sector won’t bother with — large parts of the Global South where the unit economics don’t work for Google or Microsoft.

The deeper challenge: countries everywhere want AI sovereignty. They want their laws to apply, their languages supported, their problems addressed. The US needs to develop what Grotto calls “strategic empathy” for what this looks like from a middle power’s chair.

The advertising trap, and what AI inherits

Jeff Hancock pivots the conversation to what platforms actually do to people. Social media, optimistically pitched in the 2010s as a connection technology, turned into something different. The mental-health story is well known. The story he wants told is about persuasion — how heavy exposure to a particular kind of content shapes how a person thinks the world works. A young boy steeped in the manosphere builds a model of women that the model then carries into the world. A young girl steeped in beauty content builds a model of how she is supposed to look.

The driver is not really the algorithm. It is the content the algorithm is paid to surface. Over 94% of Meta’s revenue is advertising. Index funds now have to declassify themselves because Meta and Apple are too large a share. Hancock’s lab has run experiments on Twitter that show feeds with less polarising content make users happier — and make them use the platform four minutes less per day. Four minutes equals $1.4 billion in lost revenue. Any executive who voluntarily made that swap would be sued by shareholders.

Hancock’s group is also the lead evaluator of Australia’s Social Media Minimum Age Act, which from December 2025 bans under-16s from social platforms. The law is unusual: it targets platforms, not users. Compliance is patchy — the companies are doing the bare minimum — but about 35% of teens are off the platforms. Fourteen and fifteen-year-olds are angry; ten to twelve-year-olds and their families mostly think the law is reasonable. The most interesting finding: when asked what they will do instead, many teens say they will hang out with ChatGPT. The persuasion moves from humans to bots.

LLMs turn out to be remarkably persuasive — more than humans, in some studies — and the persuasion runs through facts and evidence rather than emotion. David Rand has shown that chatbots can talk people out of conspiracy beliefs. The flip side: if the bot is Grok-4, or DeepSeek, or anything fine-tuned to a particular worldview, it can talk people into things just as effectively.

Authoritarianism and the bottom-up case for AI

Jennifer Pan studies how the Chinese Communist Party uses digital technology. Her opening point is that the Party’s first commitment is survival. From that lens, generative AI is not primarily a surveillance tool — it is an economic stimulus. Since the DeepSeek moment in early 2025, Beijing has pushed AI hard, and adoption has spread quickly across sectors.

Digital authoritarianism existed before generative AI. Computer vision did surveillance; conventional machine learning did censorship; algorithmic feeds competed with human influencers for attention. GenAI mostly dials these things up rather than fundamentally changing them inside China. Where it is genuinely transformative, Pan argues, is in giving smaller, poorer authoritarian actors capabilities they could not afford before. China is not so much exporting its ideology as exporting its methodology.

China was first to regulate generative AI seriously — the 2023 Interim Measures on Generative AI Services basically extended the existing censorship regime to LLMs. Empirical tests from 2023 to 2025 show Chinese models comply heavily: prompts on politically sensitive topics get refused at much higher rates than on US models. Crucially, this is regulatory enforcement, not a training-data artifact — on non-sensitive topics, Chinese and US models perform similarly. Most leading Chinese models are open-weight, so anyone can download them and fine-tune away the guardrails. Pan thinks the Party will tolerate this as long as the economic upside outweighs the political risk, which is now and for the foreseeable future.

There is also a counterintuitive twist: digital tools have made propaganda harder, not easier. The Party can censor effectively, but reaching a depoliticised audience that prefers cooking influencers to political content is genuinely difficult.

Mythos and the cyber J-curve

Anthropic’s “Mythos Preview” sits in the background of much of the conversation. The model is reportedly good enough at finding software vulnerabilities and stringing together multi-step attacks that Anthropic held it back for six months to let critical-infrastructure operators “go shields up.” This is the moment the panelists keep circling. Even ChatGPT 5.5 is approaching similar territory.

Grotto frames the cyber question as duelling J-curves. New technologies follow a J-curve in productivity: a dip, then a rise. With AI-augmented cyber, attackers will flatten their J-curve much faster than defenders. The attacker just needs to scale exploits across many targets; the defender has to integrate the tool into complex existing systems without breaking anything. In the short term, mythos-class capability favours attackers, especially against poorly resourced defenders.

There is also a Chinese policy wrinkle: a 2021 law requires that any vulnerability discovered in China be reported to the government before the vendor. Nobody knows how big the resulting stockpile is, but Grotto argues it is large enough to make Chinese commitments on cyber cooperation hard to take at face value.

Biology meets the same race-to-the-bottom problem

Kahl asks Endy whether AI raises the risk of bio-weapons. Endy’s answer is more textured than the typical apocalyptic take. There are three threat paths:

Uplift — LLMs lowering the difficulty bar so non-experts can attempt biological mischief. Real, but limited: most malicious actors take the easier path and pick up a rifle.
Novel design — using AI to design toxins or pathogens for which we have no detectors, countermeasures, or vaccines. Worse.
Nation-state arms race — and this is what Endy is actually losing sleep over.

His concern is that loud AI-doom rhetoric, applied to biology, makes nations less trusting and could push them back into the offensive bio-weapons programmes that have largely been dormant for a century. “Tony Hawk on the half-pipe of salvation and doom” — the AI community wins attention by warning about world-ending risks, then discharges the resulting political liability into biology. Endy would rather see the US and China stand shoulder to shoulder and recommit to ending bioweapons, plus a duty to notify each other of pandemic-potential outbreaks. Modest, but a foundation.

His positive flip is biological intelligence — using biology and AI together to build pathogen detection the way the world built geospatial intelligence. Catch patient zero earlier. Build the public-health equivalent of GPS.

Fake bio and the labrary

A questioner asks about medical misinformation — anti-vaccine sentiment, dying children, the politicisation of public-health information. Endy tells a story about being cornered after an event by 50 septuagenarians wanting to argue about vaccines. He didn’t fight them on RNA. He talked about the smallpox vaccine, which actually has a real casualty rate, and then about Thanksgiving dinner — the version of vaccines they might want is one made carefully in your kitchen, by people you trust, with ingredients you know. Living skin-cream vaccines that tickle the immune system, no needle. The optionality matters. The trust matters more.

Out of this comes the most charming idea of the panel — Callie Chappell’s notion of the “labrary.” Mutate the word library by flipping one letter. Now you have public spaces, like a Carnegie-era library, where citizens can get hands-on access to biology and engineering, mediated by a “labrarian” — a trusted local steward. Run the Carnegie playbook from 1890-1928 again, this time for the natural sciences. Renew citizenship and literacy in the technologies that will shape people’s lives.

Where AI ends up sitting in the war room

The final audience question is whether AI will end up making strategic decisions for nations. Pan notes that in China, generative AI is already inside hospital workflows — a Stanford student broke her wrist in Beijing in May 2025 and got an AI-generated CT readout before seeing a doctor. The tolerance for AI-mediated judgment is much higher than in the US.

Kahl points to Mike Brown’s claim that the Iran war was the first AI war. The MAVEN smart system, already used in Ukraine, fuses intelligence across domains, generates targets, and recommends how to service them — which aircraft, which tankers, which bases. Still human-on-the-loop, but barely. He expects that within ten years US presidents will have an AI advisor sitting alongside the human ones. Grotto adds the harder question: when an adversary has the same capability, the pressure to delegate to machine speed escalates. The Napoleonic command structure has been the model since the 19th century. It may not survive contact with this one.

Key Takeaways

Biotech is becoming a manufacturing GPT. McKinsey projects $30T over 25 years. Synthetic DNA + AI sequence generation + electro-biosynthesis are the three legs.
The US-China AI contest is seven contests, not one. US leads at the frontier; China competes everywhere else and is the only one running an all-of-the-above strategy.
Diffusion is where global influence is being decided. Countries want sovereignty — local laws, local languages, local apps. The US private sector won’t serve much of the Global South; China will, for strategic reasons.
Social media’s content engine is a $1T trap. Optimising for engagement is shareholder-mandated. AI platforms might escape this if their incentive is to do work for you, not keep you on the screen.
Mythos has changed the conversation. Anthropic’s six-month delay was a real signal. AI-augmented offence will outpace defence in the near term, especially against weakly-resourced targets.
Chinese AI compliance is real and political, not technical. Models refuse on regulated topics at high rates; on neutral topics they behave like Western models. Most are open-weight, which limits the regime’s control.
The bigger bio risk is nation-state, not lone-wolf. Endy’s nightmare is loud AI-doom rhetoric pushing countries back into offensive bioweapons programmes.
Persuasion machines are now better than humans at facts-and-evidence. Can be used to talk people out of conspiracies — or into different ones.
Australia’s under-16 social media ban is the first real test case. 35% compliance in four months. Worth watching.
AI advisors are coming to the war room. First the targeting recommendations, then the political ones. The decision-speed pressure undermines civilian control by design.

Claude’s Take

This is the rare technology panel where everyone on the stage has actually done the thinking. No one is auditioning for a podcast clip. The framing of technopolitics — that the most consequential geopolitical actors of the next decade are trillion-dollar private companies with near-sovereign control over digital life — is correct and underused. Most coverage still treats the AI race as a binary US-China sprint. Grotto’s heptathlon is more honest and more useful. The diffusion event in particular feels load-bearing in a way that gets lost in coverage obsessed with model benchmarks.

Endy is the standout. The case for biology as a general-purpose manufacturing technology is rigorous and unsexy in the right ways — he goes through revenue figures, energy gradients, and the synthetic-fertilizer analogy without ever quite flying off the rails into a TED-talk crescendo. His framing of AI-doom rhetoric being “discharged into Bioland” by the AI community is the kind of thing you only hear from someone who has been in the room. The labrary idea is the best concrete civic proposal in the entire panel. Worth chasing.

Where the conversation gets thinner is on what an actual response looks like. Everyone agrees governance is hard, allies are skeptical, US institutions have been hollowed out. Nobody quite says what should happen next other than “Trump and Xi should put AI safety on the agenda.” Fair enough — this is a diagnosis panel, not a prescription panel — but the gap between the seriousness of the threats and the modesty of the proposed responses is the panel’s quiet through-line.

Pan’s contrarian read on digital authoritarianism is also valuable. The assumption that LLMs are a force-multiplier for the CCP is mostly American projection. Inside China, the dominant frame is economic, and the apparatus of control was already mature before GenAI. The real export is methodology, not ideology — that’s a sharper formulation than most analysts use.

Score: 8. Not 9 because no individual claim is paradigm-shifting on its own and several threads (Hancock on advertising, Grotto on export controls) cover ground that has been covered better elsewhere. Not 7 because the cross-domain integration — bio plus AI plus cyber plus information ecosystem — is exactly what most single-domain commentary misses, and the Stanford faculty roster is doing real work here, not just performing one.