{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreicddzt4psaxwc54lrnkcsjtzeyddxrtfmicpbt6nef5g5226mejg4",
    "uri": "at://did:plc:5opbpi2nomj4y3d5kpwamkrd/app.bsky.feed.post/3mmt2yelan6d2"
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreicio6rssv5hl5gk4zpbztteeqs7rmeotneimkjgjt6rhajquusiyi"
    },
    "mimeType": "image/jpeg",
    "size": 527894
  },
  "description": "Artificial General Intelligence, or AGI, is one of those terms that is both overused and underdefined. Depending on who you ask, it means human-level intelligence, economically useful autonomy, recursive self-improvement, scientific superintelligence, or simply “the next thing after today’s chatbots.”\n\nA useful working definition is this:\n\nAGI would be an AI system that can acquire new skills efficiently across a broad range of unfamiliar tasks, rather than merely performing well on tasks it has",
  "path": "/what-achieving-agi-would-mean-beyond-bigger-models-and-longer-context-windows/",
  "publishedAt": "2026-05-27T08:16:57.000Z",
  "site": "https://corti.com",
  "tags": [
    "arXiv",
    "Claude",
    "Google DeepMind",
    "OpenAI"
  ],
  "textContent": "Artificial General Intelligence, or AGI, is one of those terms that is both overused and underdefined. Depending on who you ask, it means human-level intelligence, economically useful autonomy, recursive self-improvement, scientific superintelligence, or simply “the next thing after today’s chatbots.”\n\nA useful working definition is this:\n\n**AGI would be an AI system that can acquire new skills efficiently across a broad range of unfamiliar tasks, rather than merely performing well on tasks it has been heavily trained, prompted, tuned, or scaffolded to handle.**\n\nFrançois Chollet’s framing is especially helpful here: intelligence is not just task performance, but **skill-acquisition efficiency** under uncertainty and limited prior experience. (arXiv)\n\nThat distinction matters, because today’s AI progress is real, useful, and accelerating, but it is still largely built around specialization.\n\n## The Paradox: Useful AI Is Specialized, While AGI Is General\n\nIn practical engineering, the best AI systems are not “general” in the abstract. They are specialized.\n\nA production-grade AI solution typically needs:\n\n  * Domain-specific grounding data.\n  * A well-defined task boundary.\n  * Carefully designed prompts or instructions.\n  * Tool access.\n  * Retrieval over trusted sources.\n  * Evaluation datasets.\n  * Human review loops.\n  * Guardrails and monitoring.\n  * Integration into existing workflows.\n\n\n\nThis is true whether we are building a coding assistant, a customer support bot, a medical triage helper, a legal document analyzer, or a Copilot Studio RAG agent over SharePoint documents.\n\nThe hard part is rarely “ask a smart model a question.” The hard part is **getting the right context, constraining the task, selecting the right tools, validating the answer, and integrating the output into a reliable workflow**.\n\nThat is why the current wave of AI engineering is less about replacing software architecture and more about extending it. We build systems around models: retrieval pipelines, agent harnesses, tool routers, function calls, evaluators, policy layers, and observability. The intelligence is not only in the model; it is in the complete system.\n\nThis is also where skepticism is useful. Current AI is often framed as something that can look magical while still being, at its core, a computational trick: powerful pattern completion, not necessarily understanding in the human sense. I think that critique is worth taking seriously. Modern LLMs can be astonishingly capable without proving that they possess robust, general intelligence.\n\n## The Context Problem Is Still a Real Bottleneck\n\nOne of the biggest practical limits today is context.\n\nA model cannot reason about what it cannot see. For enterprise AI, this matters enormously. Real work lives in codebases, ticket systems, design documents, specifications, architecture diagrams, emails, SharePoint libraries, telemetry, incident timelines, Git histories, and organizational memory.\n\nContext windows are getting larger. Claude, for example, documents context windows up to one million tokens, and Anthropic describes that capacity as the total space available for conversation history plus generated output. (Claude)\n\nThat is impressive — but it does not eliminate the problem.\n\nLong context creates new engineering challenges:\n\n  * **Selection:** What should be included?\n  * **Compression:** What can be summarized without losing critical details?\n  * **Freshness:** Which source is authoritative now?\n  * **Attention:** Can the model reliably use the relevant part of a huge context?\n  * **Cost:** How often can we afford to send massive context?\n  * **Latency:** How fast can the system respond?\n  * **Evaluation:** Did the model use the right evidence or merely produce a plausible answer?\n\n\n\nThe deeper problem is not just context size. It is **context management**.\n\nA larger window is like a larger desk. It helps, but it does not automatically organize the work. AGI, if achieved, would need something closer to durable working memory, episodic memory, source awareness, goal management, and the ability to decide what information matters for a task.\n\nThat is very different from simply increasing the token limit.\n\n## Specialization Is Not a Failure of AI, It Is How Work Gets Done\n\nThere is a temptation to view specialization as evidence that we do not have “real AI” yet. I think that is the wrong conclusion.\n\nSpecialization is how intelligence becomes useful.\n\nHuman experts are specialized too. A surgeon, software architect, physicist, lawyer, mechanic, and product designer all use general intelligence, but their value comes from applying it through deep domain-specific models of the world.\n\nThe same pattern applies to AI systems.\n\nFor example, AlphaFold was not a general chatbot. It was a specialized scientific AI system aimed at protein structure prediction. Yet it had an enormous impact: DeepMind says AlphaFold has predicted more than 200 million protein structures, covering nearly all catalogued proteins known to science, and made them available through a public database. (Google DeepMind)\n\nThat is a key lesson: **some of the most transformative AI systems may not look like AGI at all.**\n\nThey may be narrow, deeply optimized systems that solve problems humans care about.\n\n## So What Would AGI Actually Bring?\n\nIf narrow AI and specialized AI systems are already useful, what would AGI add?\n\nThe answer is not “a better chatbot.” The real promise of AGI is that it could reduce the friction between **problem, hypothesis, experiment, and implementation**.\n\nToday, innovation is bottlenecked by scarce human expertise, time, coordination, and iteration speed. AGI could change that by acting as a general-purpose cognitive partner that can move across disciplines without needing to be rebuilt for every domain.\n\n## 1. AGI Could Become a Universal Research Collaborator\n\nA true AGI system would not merely retrieve papers or summarize documents. It could formulate hypotheses, design experiments, critique methods, connect ideas across fields, and update its approach when results contradict expectations.\n\nThat matters because many breakthroughs are interdisciplinary.\n\nBattery chemistry, climate modeling, drug discovery, robotics, chip design, cybersecurity, materials science, and synthetic biology all require crossing boundaries between domains. Human institutions are often bad at this because expertise is siloed. AGI could help bridge those silos.\n\nOpenAI has described AGI’s potential in terms of accelerating scientific discovery, increasing abundance, and acting as a force multiplier for human ingenuity and creativity. (OpenAI)\n\nThe important point is not that AGI would magically produce truth. It would still need verification. But it could massively increase the number of plausible paths explored.\n\n## 2. AGI Could Change Creativity from Output Generation to Concept Exploration\n\nToday’s generative AI is already good at producing artifacts: text, images, code, music, video, diagrams, summaries, and prototypes.\n\nBut creativity is not just production. Creativity involves:\n\n  * Framing the problem differently.\n  * Combining distant concepts.\n  * Detecting hidden constraints.\n  * Generating alternatives.\n  * Testing taste and usefulness.\n  * Iterating toward something meaningful.\n\n\n\nA current model can help brainstorm. A more general system could help explore design spaces.\n\nFor software engineering, this could mean going from “generate this function” to “explore five architectures, simulate likely failure modes, compare operational trade-offs, generate prototypes, test them, and explain which one is most robust.”\n\nFor product design, it could mean going from “write a feature spec” to “identify unmet user needs, map the competitive landscape, design experiments, generate UX variants, and reason about adoption risk.”\n\nFor science, it could mean going from “summarize known mechanisms” to “suggest new mechanisms, propose experiments, and estimate which ones are worth testing first.”\n\nThat is a different category of creativity: not just artifact generation, but **search over possibility space**.\n\n## 3. AGI Could Compress the Innovation Loop\n\nA lot of innovation is limited by cycle time.\n\nThe loop looks like this:\n\n  1. Understand the problem.\n  2. Gather context.\n  3. Generate hypotheses.\n  4. Build a prototype.\n  5. Test it.\n  6. Analyze results.\n  7. Refine the approach.\n  8. Repeat.\n\n\n\nToday’s AI can help with parts of that loop. AGI could potentially operate across the whole loop.\n\nIn software, this could mean an AI system that reads the product requirement, understands the existing codebase, proposes a design, implements it, runs tests, debugs failures, updates documentation, creates a migration plan, evaluates security implications, and asks for human input only at meaningful decision points.\n\nIn research, it could mean an AI system that reads literature, identifies gaps, proposes experiments, writes simulation code, analyzes the data, and surfaces unexpected findings.\n\nIn business, it could mean an AI system that connects market signals, customer feedback, product telemetry, and engineering constraints into actionable strategy.\n\nThis is where AGI could become economically transformative: not because it produces one brilliant answer, but because it makes iteration dramatically cheaper.\n\n## 4. AGI Could Make Expertise More Widely Available\n\nIf AGI works, its largest impact may be democratization of expertise.\n\nToday, access to expert reasoning is unevenly distributed. Large organizations can hire specialists. Small teams often cannot. Individuals often lack access entirely.\n\nAGI could make high-quality assistance available for:\n\n  * Education.\n  * Medical research support.\n  * Legal navigation.\n  * Software development.\n  * Scientific discovery.\n  * Accessibility.\n  * Entrepreneurship.\n  * Public-sector services.\n  * Engineering and manufacturing.\n\n\n\nThis does not mean replacing professionals. In high-stakes domains, professionals remain essential. But AGI could raise the baseline capability of everyone working with complex information.\n\nThat could be as significant as the internet, but more active. The web gave us access to information. AGI could give us access to interactive reasoning over that information.\n\n## 5. AGI Could Shift the Value of Human Work\n\nIf AGI can perform many cognitive tasks, the value of human work shifts.\n\nLess value would come from producing first drafts, boilerplate, routine analysis, or repetitive knowledge work. More value would come from:\n\n  * Choosing the right problems.\n  * Setting goals.\n  * Defining values and constraints.\n  * Making judgment calls under ambiguity.\n  * Validating outputs.\n  * Building trust.\n  * Owning accountability.\n  * Understanding human context.\n  * Creating meaning.\n\n\n\nThis is not a small transition. It would affect organizations, labor markets, education systems, and professional identity.\n\nThe optimistic version is that AGI gives humans leverage. The pessimistic version is that it concentrates power. Which path we get depends less on model capability alone and more on deployment, governance, access, incentives, and safety.\n\nOpenAI’s mission statement explicitly frames AGI as something that should benefit all of humanity, which reflects the scale of both the opportunity and the risk. (OpenAI)\n\n## Why AGI Is Not Just “More Tokens + More Parameters”\n\nThe industry often talks as if AGI will emerge from scaling: bigger models, bigger datasets, bigger context windows, bigger compute clusters.\n\nScaling clearly matters. But AGI likely requires more than scale.\n\nA credible AGI system would need capabilities such as:\n\n  * Persistent memory.\n  * Reliable abstraction.\n  * Transfer learning across domains.\n  * Grounded world models.\n  * Tool use with feedback.\n  * Long-horizon planning.\n  * Self-correction.\n  * Causal reasoning.\n  * Uncertainty awareness.\n  * Robust evaluation of its own outputs.\n  * Alignment with human intent.\n  * Safe behavior under novel conditions.\n\n\n\nCurrent systems approximate some of these through scaffolding. Agents can call tools. RAG systems can fetch context. Evaluators can judge outputs. Memory systems can persist facts. Planners can decompose tasks.\n\nBut stitching these pieces together is not the same as general intelligence. It is system engineering around a powerful model.\n\nThat does not make it unimportant. In fact, this may be the path: AGI may not arrive as a single monolithic neural network. It may emerge as an architecture — model plus memory plus tools plus planning plus evaluation plus environment feedback.\n\n## The Enterprise View: AGI Will Not Remove Architecture\n\nFor enterprises, AGI would not eliminate the need for architecture. It would increase the importance of architecture.\n\nEven with AGI, organizations will still need to answer:\n\n  * Which data is authoritative?\n  * Which actions may the system take?\n  * Which decisions require human approval?\n  * How are outputs audited?\n  * How are failures detected?\n  * How is confidential data protected?\n  * How are regulatory requirements enforced?\n  * How do we measure business value?\n  * How do we prevent automation from amplifying bad processes?\n\n\n\nAGI would not make these questions disappear. It would make them more urgent.\n\nThe better the AI, the more important the control plane becomes.\n\n## The Real Breakthrough: From Automation to Autonomy\n\nCurrent AI is mostly automation with a conversational interface.\n\nAGI would mean something closer to autonomy: the ability to take a goal, understand the situation, acquire missing knowledge, select tools, make progress, recover from errors, and adapt to new constraints.\n\nThat is the threshold to watch.\n\nNot whether an AI can pass a benchmark.\n\nNot whether it can write a convincing essay.\n\nNot whether it can generate code.\n\nThe meaningful question is:\n\n**Can it reliably make progress on unfamiliar, underspecified, multi-step problems in the real world?**\n\nThat is where general intelligence begins to matter.\n\n## Conclusion: AGI Would Be a New Innovation Substrate\n\nAchieving AGI would not mean that specialization becomes irrelevant. Quite the opposite.\n\nThe most useful AGI systems would likely combine general reasoning with specialized tools, domain knowledge, retrieval systems, simulators, evaluators, and human oversight.\n\nToday, we specialize AI systems because that is how we make them reliable. Tomorrow, AGI could make specialization easier to create, faster to adapt, and more widely accessible.\n\nThe near-term bottleneck is context: getting the right information into the model, at the right time, with the right constraints. The longer-term breakthrough is not just bigger context. It is systems that understand what context matters, can acquire missing context, and can reason across domains without being rebuilt from scratch.\n\nIf AGI is achieved, its greatest contribution may not be replacing human creativity. It may be expanding the surface area of creativity itself.\n\nMore people could explore more ideas, test more hypotheses, build more prototypes, and connect more domains than ever before.\n\nThat is the optimistic case for AGI: not an oracle, not a god machine, and not merely a bigger autocomplete system — but a general-purpose engine for accelerating human imagination, experimentation, and discovery.",
  "title": "What Achieving AGI Could Mean: Beyond Bigger Models and Longer Context Windows",
  "updatedAt": "2026-05-27T08:17:31.782Z"
}