Tomorrow

The forward view

Where this is going,
and who says so.

Theses, not predictions. Each one carries the signals that would confirm it, the strongest argument that it is wrong, and the researchers making the case in public.

Theses
15
Cited voices
18
Horizons
2026 → 2032+
Back to 1943
Horizon

Near — next 18 months

2026 — 2028 · 4 theses

Close enough that the first evidence is already visible in pricing pages and changelogs.

2026 — 2027high2 cited

Agents that finish multi-hour work unsupervised

The bottleneck moves from model quality to reliability over long horizons: checkpointing, verification and rollback rather than raw intelligence.

Tool-use protocols standardisingComputer-use models shippingEval suites shifting to task completion

First visible signal · Pricing that shifts from per-token to per-completed-task

Strongest case against · Long-horizon reliability may not be an engineering problem at all. If error compounds per step, a 99% step accuracy still fails most hundred-step tasks, and no amount of checkpointing fixes a base rate.

Watch pricing model changes
2026 — 2027high1 cited

Frontier-grade capability at commodity prices

Open weights plus distillation keep dragging the price of last year's frontier towards zero, while the true frontier gets more expensive. The market splits in two.

Open-weight models matching prior frontierInference cost per token falling ~10x/yearFree tiers absorbing more capability

First visible signal · Entry-price charts on Tomorrow flattening to $0 across whole segments

Strongest case against · Distillation depends on someone paying for the frontier. If frontier economics stop working, the cheap tier has nothing left to copy.

Watch entry prices in Tokenomics
2026 — 2028high1 cited

Small models move on-device by default

NPUs in phones and laptops make private, offline inference the default for everyday tasks, with cloud reserved for hard problems.

3B–8B models near GPT-4-class on narrow tasksNPU-equipped consumer siliconPrivacy regulation favouring local processing

First visible signal · Vendors advertising 'no data leaves the device' as a pricing tier

Strongest case against · Consumers have repeatedly chosen convenience over privacy. If the cloud model is meaningfully better, local inference stays a niche feature rather than a default.

Browse the on-device segment
2026 — 2030high1 cited

Post-quantum cryptography migration becomes mandatory

'Harvest now, decrypt later' makes Shor's algorithm a present-tense problem. NIST standards are published; the work is now a decade-long replacement of every key exchange in production.

ML-KEM/ML-DSA in browsers and TLS defaultsFederal migration deadlinesVendors advertising PQC readiness

First visible signal · Security tools on this radar listing PQC support as a pricing-tier feature

Strongest case against · Migration deadlines have slipped before. Without a demonstrated break, most organisations will treat post-quantum readiness as a checkbox rather than a rebuild.

Security tooling segment

Mid — 2 to 5 years

2028 — 2032 · 6 theses

Far enough that the mechanism is contested, near enough that the money is already moving.

2028 — 2030medium1 cited

Verification becomes the product

As generation gets free, provable correctness becomes the scarce good: formal checks, citation-grounded output and auditable agent traces.

Formal-methods startupsProvenance standards for mediaEnterprise procurement demanding audit logs

First visible signal · A 'verified output' segment appearing that did not exist before

Strongest case against · Verification only becomes a market if buyers pay for correctness. Historically most software buyers have paid for speed and accepted the defect rate.

See how we verify our own data
2028 — 2031medium1 cited

Continual learning replaces the frozen snapshot

Models that update from their own deployment experience without catastrophic forgetting would end the release-cycle rhythm this entire site is built to track.

Memory architecturesLong-lived agent stateResearch on sleep-style consolidation

First visible signal · Version numbers disappearing from vendor changelogs

Strongest case against · Continual learning has been ten years away for thirty years. Catastrophic forgetting is not a shipping bug; it is a property of gradient descent on a fixed capacity.

Watch version numbers in Changes
2028 — 2032medium2 cited

AI-native scientific discovery at scale

AlphaFold was the first instance, not the exception. Closed-loop systems that propose, run and interpret experiments compress research cycles in materials, chemistry and biology.

Self-driving labsAI-authored results in peer reviewNational compute programmes for science

First visible signal · A materials or drug breakthrough where the model is the named method

Strongest case against · AlphaFold worked because protein structure had fifty years of curated data and a clean success metric. Most scientific problems have neither.

Research institutions tracking this
2027 — 2032high1 cited

Energy, not silicon, becomes the binding constraint

Data-centre buildout runs into grid limits. Efficiency per joule becomes a first-class competitive metric and possibly a regulated one.

Utility-scale power deals by AI labsNuclear and geothermal contractsGrid interconnect queues

First visible signal · Vendors publishing energy per query alongside price per token

Strongest case against · Efficiency gains have historically outrun demand growth. If inference cost per useful task keeps falling an order of magnitude a year, the grid constraint recedes on its own.

Energy and cost per token
2029 — 2033medium2 cited

Quantum utility in chemistry and materials before general-purpose quantum

The first genuinely useful quantum advantage arrives narrowly — simulating molecules and materials that classical methods approximate badly — years before anything resembling a general quantum computer.

Logical qubit counts in the hundredsHybrid quantum-classical SDKs in productionPharma and battery firms buying quantum time

First visible signal · A quantum vendor publishing a per-logical-qubit-hour price list

Strongest case against · Classical methods keep improving too. Every claimed quantum advantage so far has been met by a better classical algorithm within a year.

Quantum vendors tracked here
2027 — 2031medium1 cited

Robot foundation models and the data flywheel

Whoever accumulates the most real-world manipulation data wins robotics the way web scale won language. Teleoperation fleets and simulation-to-real transfer are the two contested routes.

Open cross-embodiment datasetsVLA models shipped as APIsTeleoperation labour markets

First visible signal · Robot policies sold per-skill, the way models are sold per-token

Strongest case against · Language data was free and already on the internet. Manipulation data has to be paid for one hour at a time, which may cap the flywheel before it spins.

Model registry

Far — 5 years and beyond

2032 and beyond · 5 theses

Direction of travel, not a date. Held loosely, revised in public.

2030+speculative1 cited

A successor to the transformer

Attention is quadratic and memoryless by design. State-space models, retrieval-native architectures or something unnamed will eventually retire the 2017 formula.

State-space and hybrid models in productionContext windows hitting economic ceilingsNeuromorphic and photonic research

First visible signal · A frontier release whose architecture section does not say 'transformer'

Strongest case against · Attention has absorbed every challenger so far by adopting its good ideas. The successor may just be a transformer with different plumbing.

Model registry and architectures
2030+speculative2 cited

Embodiment — the same models, with hands

Vision-language-action models plus cheap actuators move general-purpose AI into physical work. The data problem, not the intelligence problem, sets the pace.

VLA model releasesHumanoid pilot deploymentsTeleoperation data collection at scale

First visible signal · A robotics segment on this radar with real pricing pages

Strongest case against · Atoms are not bits. Actuator cost, maintenance and liability do not fall on a software curve, and every previous robotics wave broke on exactly that.

Robotics companies on the radar
2029+speculative1 cited

Compute governance and international coordination

If capability keeps tracking compute, compute becomes the regulated substance — with reporting thresholds, export controls and possibly verification regimes between states.

Training-run reporting thresholdsChip export controlsModel registries in national law

First visible signal · A binding multilateral agreement, not a voluntary commitment

Strongest case against · Compute governance requires states that are currently in an arms race to accept verification. Nothing in the present geopolitics suggests they will.

Policy milestones on the timeline
2030+speculative

Humanoids cross the labour-cost line

General-purpose robots become defensible only when total cost per working hour undercuts human labour in a specific task. Chinese supply chains reach that line first, in logistics and inspection rather than homes.

Unit prices under $30kRobots-as-a-service contractsUptime and MTBF published like SLAs

First visible signal · A vendor quoting cost per robot-hour instead of a purchase price

Strongest case against · Humanoid form factor is a marketing decision, not an engineering one. Purpose-built machines beat general-purpose ones on cost per task almost everywhere.

Robotics companies on the radar
2032+speculative1 cited

Quantum and AI stop being separate roadmaps

Not 'quantum machine learning' as marketed today, but AI designing error-correcting codes and control pulses, and quantum simulation generating the training data classical models cannot compute.

ML-designed decoders beating hand-built onesQuantum-generated datasets for chemistry modelsShared national compute programmes

First visible signal · A frontier lab and a quantum hardware vendor shipping a joint product

Strongest case against · The two fields have incompatible timelines. Useful quantum hardware may simply arrive after the AI question has already been settled one way or the other.

Company lifecycle view

Scoreboard — calls we already made

A forecast page with no track record is a horoscope. These are dated calls, scored honestly, including the ones that did not land.

2 hit3 too early1 miss
Hit

Open-weight models reach the previous year's frontier quality

Open releases now trade blows with the prior generation of closed frontier models on most public evaluations.

Called 2023
Hit

Inference price per token falls roughly an order of magnitude a year

Cost for a fixed capability level has fallen steeply and repeatedly; the frontier price has not.

Called 2023
Too early

Energy, not chips, becomes the visible constraint on buildout

Power deals and interconnect queues are now routine news, but silicon supply is still the binding line for most buyers.

Called 2024
Too early

Agents handle unsupervised multi-hour work in production

Coding agents run for hours in narrow, well-instrumented settings. General unsupervised work is not there.

Called 2024
Miss

A general-purpose quantum computer does useful commercial work

Supremacy demonstrations were real; commercially useful general-purpose quantum computation was not, and still is not.

Called 2019
Too early

Humanoid robots reach meaningful commercial deployment

Pilots exist in logistics and inspection. Unit economics against human labour remain unproven in public.

Called 2024

The past is the better predictor

Every thesis here is a continuation of a line that started with a formula. Trace it back through eighty years of groundwork.

Open the timeline