Google DeepMind Plans to Track AGI Progress With These 10 Traits of General Intelligence

How a lab paper on measurement could reshape cyberpunk aesthetics, startups, and the shadow economy

The neon of a side street reflects on a rain-slicked tablet as a concierge agent enumerates flight options, files a patent draft, and argues philosophy with an art critic in a single session. That quiet, uncanny convenience is the surface image journalists point to when they describe progress toward artificial general intelligence. The moment looks like utility; what people inside the alleyways of tech and culture worry about is the margin where that utility meets power, control, and style.

Most headlines treat DeepMind’s new measurement push as tidy taxonomy and academic housekeeping. The underreported shift is practical: defining how to measure general intelligence forces companies, regulators, and underground markets to choose what capabilities they will value, fund, or weaponize, and that choice will change the cyberpunk economy more than any gleaming demo. The research leans on press materials and a formal DeepMind publication, but the strategic consequences run far outside the lab. (deepmind.google)

Why a measurement paper reads like a manifesto for designers of dystopia

DeepMind’s team argues that operational definitions for AGI should focus on capabilities, generality, and autonomy to make comparisons, risk assessments, and policy meaningful. That shift from fuzzy prophecy to checklists is less dry than it sounds; it turns abstract fear into a product specification that venture funds and governments can bid on. (montrealethics.ai)

Who else is racing and what that competition looks like in the alleys

OpenAI, Anthropic, Meta, and several well funded startups are all improving broad capability models and agent suites that overlap with DeepMind’s proposed levels. Public statements and coverage show a widening arms race in both model scale and integration of memory and tools, which accelerates the move from conversational helpers to continuously operative agents. Industry reactions are already being debated in mainstream outlets and technical summaries. (forbes.com)

The core story: from five to ten to a practical checklist

DeepMind’s paper proposes a levels framework to judge performance and autonomy rather than a single threshold called AGI. Translating that into operational tracking requires concrete traits that benchmarks can measure, which is why practitioners and ethicists are converging on a multi dimensional list to make progress auditable and comparable. This article synthesizes that operational need into a pragmatic set of ten traits that capture breadth and depth of general intelligence, drawn from DeepMind’s principles and the wider literature. (deepmind.google)

The ten traits of general intelligence the market will start measuring

Reasoning and problem solving across novel domains; long term planning and multi step tool use; adaptive learning from sparse feedback; persistent, retrievable memory across sessions; multimodal perception and world modelling; goal formation and self directed task decomposition; theory of mind and social inference; safety aware behavior and corrigibility; meta learning and self reflection; embodied interaction at scale through tools and robotics. These are not plucked from a press release but assembled to align with DeepMind’s call for ecological, capability focused tests. Measuring them reorganizes what products count as revolutionary and which count as tactical upgrades. (deepmind.google)

Measuring intelligence is less a scientific verdict and more like choosing which parts of the machine to sell as a feature.

The cost nobody is calculating for cyberpunk culture and small creative firms

Tracking these ten traits will create a premium market for labeled datasets, specialised eval suites, and compliance tooling. A boutique studio that licenses an “agent safety” audit or a continuity memory plug will see demand; the catch is it needs engineers and compute that cost real money. Expect a new vendor tax where integrating a vetted memory layer costs 10 to 30 percent of a model deployment budget, and independent creators pay via platform fees or revenue shares, not grants. The aesthetics of the city change when clear metrics make some capabilities licensable and others illicit, which is a cooler noir plot than anyone on a funding deck planned.

Practical implications for businesses with 5 to 50 employees, with numbers

A 12 person design firm using an advanced agent to automate client proposals could save one full time employee worth of work per month. If average fully loaded salary is 7,000 dollars per month, the firm nets 84,000 dollars in annual labor substitution versus a 25,000 dollar annual subscription and 10,000 dollars in integration costs, leaving roughly 49,000 dollars in gross margin improvement before taxes. For a small app maker selling an assistant plug in, adding verified session memory could increase retention by 15 percent and lift annual recurring revenue from 120,000 dollars to about 138,000 dollars, assuming a conservative 10 percent churn baseline. Those are simple models and they assume reliable performance, which is the point of DeepMind’s push to standardize measurement and therefore risk premiums. Dry fact aside the math also lets an accountant fall asleep knowing someone else is worrying about existential risk, which is comforting in a utilitarian way.

The legal and ethical rents being priced into city infrastructure

If regulators adopt operational levels as the basis for liability or export control, vendors will need compliance layers verified by third parties. That means certified evaluators and legal wrappers become a business unit unto themselves. A license to operate an autonomous agent in public spaces may cost more than the hardware, and cities may auction rights to “agent highways” where systems can act with higher autonomy. This is not sci fi; systems that are easier to audit are also easier to tax or ban.

Risks and open questions that stress test the claim that measurement solves policy

Operationalizing traits will centralize power in labs that control benchmarks and eval infrastructure. Benchmarks can be gamed, and metric optimization can create brittle behaviour that passes tests but fails in the wild. Further, the choice of tasks to define generality embeds values about which types of intelligence are rewarded. The debate over what counts as an economically relevant task is still unresolved and may invite capture by commercial interests. Coverage and summaries show these definitional issues are precisely what DeepMind’s framework aims to surface, not to magically solve. (scientificamerican.com)

Why cyberpunk aesthetics and the black market will adapt fast

When capability metrics become currency, black markets will form for undocumented model forks and “unbenchmarked” add ons that bypass safety constraints. That will give cyberpunk artisans new palettes of outlaw capabilities, funded by quick cash rather than grants. At the same time, civic technologists will repurpose measurement tools to defend communities, turning the same checklist into a way to certify friendly agents. It is a lateral dance between commerce and culture, and both sides love a good vertical market.

A concise forward look that works for builders and curators

Standardizing what general intelligence means will professionalize many layers of the ecosystem and re price risk in clear ways. Small teams should prepare by auditing where their product sits in the ten traits list and budgeting for third party eval and compliance or risk becoming artisanal in a world that prizes certified scale.

Key Takeaways

• DeepMind’s levels framework pushes the field from narrative to measurable capabilities, creating new vendor and regulatory markets.
• A practical ten trait checklist helps companies turn AGI progress into auditable features and commercial products.
• Small firms should budget for compliance and certified eval as part of deployment economics rather than treat it as optional.
• Standardized benchmarks will reshape both legitimate markets and black markets in predictable ways.

Frequently Asked Questions

What exactly did DeepMind propose about measuring progress toward AGI?
DeepMind proposed a levels framework that separates performance, generality, and autonomy to make AGI progress comparable across systems. The paper emphasizes capability focused, ecologically valid benchmarks rather than a single checklist of tasks. (deepmind.google)

How will measurement change the cost of deploying AI for a small company?
Standardized measurement will add certification, third party audits, and compliance engineering costs that should be budgeted as recurring expenses. Those can be material; early adopters will face setup fees plus ongoing eval subscriptions that change unit economics.

Are current large language models already AGI according to these measures?
Under DeepMind’s framework, many leading LLMs are considered “emerging” because they match or slightly exceed unskilled human performance on a set of non physical tasks. They do not meet higher performance and generality levels across the board. (scientificamerican.com)

Could the benchmarks be manipulated and therefore misleading?
Yes, benchmarks invite gaming and metric optimization that may not reflect real world behavior. Critics urge the community to develop living, ecological tests and guardrails to reduce brittleness and misaligned incentives. (montrealethics.ai)

How should a cyberpunk creative studio prepare differently than a software startup?
Studios should catalog which of the ten traits their work touches and prioritize verifiable controls for those traits that directly affect user safety or public perception. Artistic experiments with unverified agents should be clearly labeled to avoid legal and reputational exposure.

Related Coverage

Writers and readers who like this story will want to follow evaluation frameworks for agentic AI, debates over autonomy and liability in urban deployments, and the economics of certified AI services for creative industries. Tracking how marketplaces for model certification develop will reveal where the next creative and regulatory battles will be fought.

SOURCES: https://deepmind.google/research/publications/66938/ , https://www.forbes.com/sites/lanceeliot/2023/12/06/figuring-out-what-artificial-general-intelligence-agi-consists-of-is-enormously-vital-and-mindfully-on-the-minds-of-ai-researchers-at-google-deepmind/ , https://arstechnica.com/ai/2025/07/agi-may-be-impossible-to-define-and-thats-a-multibillion-dollar-problem/ , https://www.scientificamerican.com/article/what-does-artificial-general-intelligence-actually-mean/ , https://montrealethics.ai/levels-of-agi-operationalizing-progress-on-the-path-to-agi/