The supporting record: observations, unresolved forecasts, earlier arguments and historical turning points. Our five prediction countdowns are published on Horizon.
21 past turning points · 3 now signals · 7 next forecasts · review record ↗
Our 12 September OpenRouter check matched 37 of 39 tracked model IDs. Two were absent, and several open weight models had paid hosted rates. The model ID, endpoint and check date matter more than a loose product label.
The SWE-bench maintainers report that Multimodal v2 is open source with 480 tasks for local evaluation. A reproducible task set makes model claims easier to inspect. We have not run this benchmark or verified a new leaderboard result.
The MCP repository points to a 2026-07-28 specification schema. Versioned interfaces give implementers something concrete to check against. This observation is about the published artefact, not enterprise adoption.
By 11 October 2027, sandbox controls, action approvals and inspectable run records could be standard product features in agent platforms. These controls matter when a model can change files or call external tools.
Target:
Review. Retained as emerging. Codex documents sandboxing, approval policies and optional telemetry today. That is an existence proof in one product, not proof of market-wide adoption or immunity to prompt injection.
Resolution test. Use a named, published sample of agent platforms. Verify that each control can be configured and observed in a real run, including a denied action. Documentation alone does not prove the controls work. A majority must provide all three features.
Original April claim and revision recordBy April 2027, shared protocols could become the usual route for exposing tools, data and agent capabilities in enterprise agent systems. MCP and A2A solve different parts of that problem. Neither removes the need to design permissions or recover from failures.
Target:
Review. Retained as emerging. The maintained specifications and SDKs are concrete infrastructure. Their existence does not show that most enterprises use them. We removed the unsupported implication that adoption had already been measured.
Resolution test. Publish the sampling frame before judging adoption. Count named production integrations using a shared protocol, excluding demos and announcements. A majority of that sample must use a protocol for a substantial part of the integration. Report the sample limits.
Original April claim and revision recordBy 11 April 2027, browser and desktop agents could be useful in a limited set of repetitive back-office jobs while broad consumer autopilot remains unreliable. The hard test includes changed screens, interruptions and unexpected outcomes.
Target:
Review. Retained as contested. OSWorld provides an executable desktop task benchmark and documents revisions to its evaluations. It does not measure consumer adoption or prove that a particular production workflow is reliable.
Resolution test. Compare repeat runs of named narrow workflows with varied desktop tasks. Include changed UI, lost sessions and recovery. Report failures and human interventions as well as completed tasks. A polished demonstration is insufficient.
By 11 April 2028, local inference could become the default for some offline and privacy-sensitive app features, with server models handling other work. Whether a feature is private depends on the full data path, including tools and fallback calls.
Target:
Review. Changed from speculative to contested after checking concrete local deployment options. Apple documents both on-device and server models, and llama.cpp supports local hardware. Neither shows that most mainstream apps default to local execution.
Resolution test. Test named shipping app features with network access disabled and inspect their declared data paths. Record hardware support, fallback behaviour and task quality. A local model inside a feature that uploads its inputs does not meet the privacy claim.
By 11 October 2027, an open weight reasoning model could offer a practical alternative to the strongest closed APIs on both reasoning and coding at a lower total cost. The useful comparison includes retries, latency and hosting, not just a model score.
Target:
Review. Downgraded from emerging. The R1 model card reports comparisons with older closed models. It supports a capability claim at that release, not parity with the September 2026 frontier. Hosted API rates also do not establish the total cost of running open weights.
Resolution test. Compare open and closed models on the same published reasoning and coding test sets with the same task budget. Predeclare an acceptable quality gap and count all inference, retry and hosting costs. Without an independent like-for-like result, leave the forecast unresolved.
By 11 April 2027, agents could produce most routine scaffolding and straightforward refactors in teams that adopt them. Review effort and defects determine whether this is a gain. The volume of generated code does not.
Target:
Review. Downgraded from emerging. SWE-bench measures issue resolution in a defined evaluation. It cannot establish what share of production work agents perform. The original claim about most fast-moving teams lacked an adoption baseline.
Resolution test. Follow named teams across repeated production tasks. Record accepted changes, review time, escaped defects and the human baseline. Define routine work before counting it. Do not infer a majority of teams from one lab or from benchmark scores.
By 11 April 2028, large organisations may commonly keep several model families in production for different tasks, costs and deployment needs. A catalogue with many choices makes this possible. It does not tell us what customers actually run.
Target:
Review. Downgraded from contested. The checked OpenRouter catalogue demonstrates available choice, not enterprise adoption. We found no representative production-use evidence for the original claim that most large firms will run at least three families.
Resolution test. Use a published, representative sample of large organisations. Count named production model families, excluding evaluations, unused contracts and reseller duplicates. The original majority-and-three-families claim remains unproven unless that threshold is met.
Original April claim and revision recordSystems matching or exceeding human performance across most cognitive tasks, including ones outside their training distribution, without task-specific retraining. See debate →
AI systems that autonomously execute multi-step knowledge work across tools, queues and approval boundaries, owning outcomes end-to-end rather than assisting a human operator.
General-purpose physical robots (humanoid or otherwise) that are commercially routine, not demo-quality, with fleet-deployable reliability and workable unit economics.
Coding agents reliably owning the loop from ticket to monitored production deploy across large codebases, leaving humans mostly specifying, reviewing and steering.
AI tutoring and assessment displacing institutional course and credential delivery as the primary structure through which people learn and signal mastery.
The pattern is no longer “open follows years later”; it is “open absorbs frontier ideas fast enough to matter commercially”. Stable Diffusion proved that closed advantages can leak into public ecosystems, LLaMA broke the sense that only a few labs could play, and DeepSeek-R1 showed that open reasoning can arrive with real teeth. If open gets to 90–95% of frontier capability at a fraction of the cost, permanent practical parity is enough to change who wins distribution.
The frontier does not stand still while open catches up. Closed labs still control the best compute, the richest feedback loops, the strongest product distribution and the hardest-to-copy post-training tricks, which lets them keep moving the goalposts from text to multimodal, from answers to agents, and from fluency to reasoning. Open will be formidable in cost-sensitive niches, but the top layer of capability will remain structurally concentrated.
o1 made it plausible that test-time compute is not just extra verbosity but a new way to buy capability on hard tasks. Once models can spend inference budget deliberately, the product question changes from “how fluent is it?” to “how much thinking is this worth?”. That creates a genuinely new optimisation space for models, interfaces and pricing, and it is why reasoning now feels like a separate frontier rather than a mere feature.
A lot of the current reasoning story may be packaging rather than paradigm. Bigger inference budgets, hidden chain-of-thought and better scaffolding can lift benchmark scores without solving the messy problems users actually care about, such as reliability, context management and action in the real world. If the gains are costly, slow and narrow, “reasoning” could turn out to be an expensive wrapper around familiar techniques.
Routine, high-frequency AI use wants low latency, predictable cost and privacy by default. Once a model is good enough to summarise, transcribe, draft, organise and personalise locally, shipping every interaction to the cloud starts to look like a design relic rather than a necessity. Cloud will still do the heavy lifting, but local could become the default surface most people touch first.
The average user will not choose local over cloud on principle; they will choose whatever works best. The strongest models, freshest world knowledge, richest tool access and fastest product-improvement cycles are still cloud-shaped advantages, and platform owners have economic reasons to keep the smart layer centralised. Local will grow quickly, but mostly as a complement to cloud rather than its replacement.
When a field starts inventing shared protocols, permission layers, eval harnesses and operational controls, that usually means the pattern is becoming real. Agents that touch tools, data and workflows need governable plumbing, and that need does not disappear even if today’s demos are messy. The category may still be early, but the infrastructure demand looks structural rather than faddish.
Every platform shift produces a premature standards rush before anyone knows what the durable primitive actually is. Many so-called agent frameworks are still thin wrappers around brittle model calls, and the winning abstractions may end up bundled into the major model platforms instead of living as an independent layer. What looks like category formation could still be the noisy middle stage before consolidation.
Vibe coding lowers the hardest barrier for many would-be builders: getting from intent to working software. That means domain experts can finally ship internal tools without having to become full-time engineers, which is how real capability usually spreads inside organisations. If guardrails improve, the organisational effect could look less like chaos and more like the spreadsheet moment for software creation.
Most software cost arrives after the prototype: security, integrations, ownership, maintenance and change control. Vibe coding widens the front door, but it also makes it easier to generate brittle systems that create hidden review and governance debt. In companies, that may mean more shadow IT and more QA burden rather than a clean productivity windfall.
The useful definition is economic: a system is AGI when it can reliably perform most paid knowledge work at human level or better, across domains, without task-specific retraining. This framing is measurable, matters commercially and treats “general” as a threshold of breadth rather than a philosophical claim about understanding. Reasoning models plus tool use already point at this definition, even if the threshold is years away.
The economic definition smuggles in a philosophical claim and hides the hard part. Genuine general intelligence arguably requires durable goals, embodied experience, robust world-modelling and cross-domain transfer that current systems do not have even when they score well on benchmarks. A system that passes every exam and still cannot run a small business on its own is not general; it is a very capable text engine. The label matters because it shapes policy, investment and safety expectations.
Andrej Karpathy named a new way of building software by talking to AI and steering outputs conversationally. It marked the moment coding started to feel like directing and taste-making, not just writing syntax.
DeepSeek released an open reasoning model it said was on par with o1, with open weights and MIT licensing. It reset expectations on open models, reasoning economics and who could shape the frontier.
Google's Gemini 2.0 launch foregrounded agentic capability — multimodal reasoning, tool use, and long-running task execution — as the explicit framing for what the next generation of frontier models would compete on.
Anthropic open-sourced MCP, a standard for connecting AI assistants to tools and data. It gave the agent era shared plumbing instead of endless bespoke integrations.
Models began visibly spending more effort on reasoning before answering. It marked the move from fluent prediction to deliberate reasoning as the next frontier.
OpenAI's GPT-4o unified text, audio and vision in a single model with conversational latency. It moved the assistant interface from turn-by-turn typing toward real-time spoken interaction across modalities.
Anthropic's Claude 3 made the frontier model race feel genuinely multi-player. It ended the idea that one lab would automatically own the assistant layer.
OpenAI previewed a video model with unusually coherent scene generation. It made world-simulation feel plausible rather than gimmicky.
Eliezer Yudkowsky's intervention helped push existential-risk arguments into mainstream debate. It split the story of AI into two tracks, capability acceleration and safety confrontation.
OpenAI released a much stronger model with a visible jump in capability. It moved AI from novelty to serious knowledge work, coding and professional use.
Meta's LLaMA escaped into the wider world and sparked rapid open model work. It broke the feeling that only a few labs could seriously participate.
A conversational interface put frontier AI in front of ordinary people. It turned AI from a sector story into a civilisation story almost overnight.
OpenAI released Whisper as open-weight automatic speech recognition matching commercial systems on many languages. It seeded a wave of locally-runnable transcription that did not need to call a cloud API.
Powerful image generation became openly available on ordinary hardware. It started the open model era in public consciousness.
OpenAI showed that human feedback could make models more useful and aligned. It changed the winning formula from just training bigger models to training and then aligning them.
DeepMind solved protein structure prediction at near-experimental quality. It showed AI was not just a media tool but a scientific instrument.
AI coding help moved into everyday developer workflow. It turned software creation into one of the first mass human-plus-AI production loops.
OpenAI showed text-to-image generation as a general capability. It opened the public imagination to generative AI beyond text.
OpenAI showed a giant language model that could do many tasks from prompting alone. It shifted the field towards general-purpose foundation models instead of narrow task systems.
Attention Is All You Need introduced the transformer architecture. It became the core design behind modern language models and much of generative AI.
DeepMind's system beat one of the world's best Go players. It broke the assumption that intuition-heavy human domains were still safe from machines.
● quiet week
No changes to any lane in the last 7 days. The
shift log is derived from git log -- src/data/horizon/ by the horizon bot.
Filter the map by what you actually care about. Picked themes sync to the URL, so you can bookmark or share a slice.
The horizon map runs on the same pipeline as the rest of the site. The bots propose, we decide. See how it works.