ASI Research and Development Lab: A Three-Budget Test

An ASI research and development lab is an organization whose declared goal is building systems that exceed human performance across most cognitive work. The label is cheap to adopt and expensive to earn. This article argues that you identify a real one by what it spends on verifying and constraining its own systems, not by its mission statement.

AI Overview

An ASI research and development lab is an organization that states artificial superintelligence as its goal, meaning systems that greatly exceed human cognitive performance in virtually all domains. No agreed threshold or test defines when that goal is reached, so the label cannot be checked from the outside by capability alone. The Three-Budget Model in this article splits lab spending into capability, verification and governance. On that measure, most self-described superintelligence labs are scaling labs with a mission statement.

Key Facts

What Counts as an ASI Lab, and What Doesn't

Definitions come first, because the term is used loosely.

Superintelligence. Nick Bostrom's Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014) defines it as any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest. That is a statement about performance, not about architecture or parameter count.

AGI. The working meaning is systems that match human ability across most cognitive tasks. Labs word it differently. OpenAI's charter describes highly autonomous systems that outperform humans at most economically valuable work. The wording differs, so the thresholds differ.

Powerful AI. A marketing phrase for capable models. It carries no threshold at all.

The problem is that none of these has an agreed measurement. A lab can say it is building ASI and no one can falsify the statement, because no benchmark defines the finish line. That is why the label tells you little by itself.

Three kinds of organization use the language:

1. Frontier product labs. They sell models and APIs, and they say superintelligence or AGI is the long-run aim. OpenAI, Anthropic and Google DeepMind fit here.
2. Single-purpose ASI labs. One stated goal, no near-term product. Safe Superintelligence Inc. announced exactly this in June 2024: one focus, one product, and insulation from management overhead and product cycles.
3. Corporate superintelligence units. A division inside a large company. Meta announced Meta Superintelligence Labs in 2025, and Mark Zuckerberg framed the aim as personal superintelligence.

Membership in a category says nothing about whether the work is verified. That is the gap the rest of the article addresses.

How an ASI Lab Differs From a Frontier Product Lab

The textbook difference is horizon. A product lab answers to quarterly revenue and release cadence. An ASI lab claims a horizon measured in research breakthroughs.

A lab with a product on a quarterly cadence is structurally biased toward shipping. Revenue targets create release dates, and release dates compress the time available for evaluation. That is incentive structure, not a claim about intent.

The money is also made elsewhere than in the superintelligence story. As the AI commoditization cliff argues, models alone are a weak moat. The orchestration layer is where much of the durable value sits. A lab that needs revenue while chasing a long horizon is pulled toward building products on top of its models, which is the product-lab structure it may claim to avoid.

What the Work Inside an ASI Lab Actually Is

Strip away the framing and the work falls into a handful of programs.

• Scaling. Kaplan et al. (arXiv, 2020) found that language-model loss falls as a power law in model size, dataset size and compute. Hoffmann et al. (Chinchilla, arXiv, 2022) showed that model size and training tokens should scale roughly together, and that a model trained on about four times more data than Gopher outperformed much larger, undertrained models.
• Algorithmic efficiency. Getting more capability per unit of compute.
• Post-training and reasoning. Reinforcement learning and test-time compute, which let a fixed model spend more computation per answer.
• Agents and tool use. Models acting over many steps, which raises the cost of getting evaluation wrong.
• Interpretability. Understanding what a model is computing, not only what it outputs.
• Evaluation and red-teaming. Measuring what a model can do and whether it can be misused. See model evaluation.
• Automated AI research. Using models to accelerate the lab's own research.

Rich Sutton's essay "The Bitter Lesson" (2019) is the intellectual anchor for the first item: general methods that leverage computation tend to win over methods that encode human knowledge. The field has since shifted from pure parameter scaling. Test-time compute and parallel scaling add new axes, covered in ParScale and Scaling Is Changing Shape.

Notice the asymmetry. Four of the seven programs make systems more capable. Two of them, interpretability and evaluation, make capability checkable. Agents and automated research raise the stakes on both. A lab's mix across these tells you more than its slogan.

The Three-Budget Model

The Three-Budget Model is a way to classify any lab by how it divides spending among three budgets.

1. Capability budget. Training compute, data, researchers working on scaling, algorithms and post-training. This is the visible budget. Epoch AI reports training compute for frontier language models growing about 5x per year since 2020 (90% CI 4x to 6x).
2. Verification budget. Evaluations, red-teaming, interpretability, and independent audit. It answers one question: how does anyone, including the lab, know what the system can do?
3. Governance budget. Staffing and authority to delay or stop a release. Money matters less here than power. A governance function with no veto is a communications function.

The model's central claim is contrarian: verification, not capability, separates an ASI lab from a scaling lab. A system you cannot verify cannot be called superhuman. It can only be called large. If a lab cannot show that a model outperforms humans across domains under independent testing, its superintelligence claim is an aspiration, not a result.

Which budgets are visible

An illustrative calculation

The following is illustrative. Every number is an assumption, not a measurement of any lab.

Assume a lab spends $1B on capability in year 0 and $50M on verification, a 5% share. Assume capability spending doubles each year, and verification grows 20% a year, which is a generous organizational growth rate.

By year 4, the verification budget would need to be about 7.7 times larger to keep its original share. The mechanism is simple. Capability compounds at the rate of the industry's scaling economics. Verification grows at the rate of hiring and process. The share decays as the ratio of the two growth rates, here (1.2 ÷ 2) = 0.6 per year.

The calculation is deliberately generous. It does not assume that verification gets harder as models improve, which is the likelier case. Stronger models need better evaluations, and evaluations for open-ended agentic behavior cost more than benchmark runs. The practical consequence: verification must be committed as a ratio to capability, not as a fixed number. A lab that publishes only the number has told you very little.

This extends the site's Verify-Cost Method, which prices the cost of checking an AI output for a user. The Three-Budget Model applies the same logic one level up, to the lab producing the system.

Scoring Public Labs on the Three Budgets

The rule for the table below: each cell holds a sourced commitment or "not disclosed." Nothing is inferred. "Not disclosed" means not found in the lab's own published material at the time of review, not that the practice does not exist. "Not verified for this review" means the source was not checked here. Status is dated to October 2026 and will change.

Read the table for what it shows. Three labs publish safety frameworks. One single-purpose ASI lab has the strongest mission language and the least published material. That is not evidence of bad practice. It is evidence that the signal you would need does not exist yet.

Case study: the superalignment compute pledge

In July 2023, OpenAI announced its Superalignment team and committed 20% of the compute it had secured to that point over four years to the problem of aligning superintelligence. In May 2024, the team's co-leads Ilya Sutskever and Jan Leike left the company, and the team was dissolved into other research efforts. Leike wrote publicly that safety culture and processes had taken a backseat to products. These details are publicly reported; the announcement page itself was not verified for this review.

The lesson is not that anyone acted in bad faith. It is that a governance commitment announced as a share of compute had no external enforcement, and it did not survive in the form announced. This is why the checklist below asks what happened the last time a commitment was tested.

The Compute and Capital Reality of Running One

Staying at the frontier is a capital problem first. The inputs are chips, high-bandwidth memory, power and financing, and each has its own constraint.

• Chips and memory: see the real cost of AI compute and the chip bottlenecks that define the moat.
• Power: why power is the bottleneck for AI covers the grid side.
• Capital: the compute cartel covers how silicon demand reshapes financing.

The relevance to the thesis is structural. Frontier training runs cost enough that few organizations can fund them from research budgets alone. They must raise capital or sell products. Both create pressure to release. So capital intensity pushes ASI labs toward the product-lab structure they claim to avoid. A single-purpose lab with no product has to keep raising money on the strength of its story, which makes disclosure about its verification practice more valuable, not less. See also AI compute.

Regulation and Disclosure: What Labs Must Now Tell You

Marketing disclosure is what a lab chooses to say. Mandatory disclosure is what law requires. Only the second is a floor.

EU AI Act. Regulation (EU) 2024/1689 sets obligations for providers of general-purpose AI models. Obligations under Article 53 cover technical documentation, a copyright policy and a summary of training content. Models presumed to carry systemic risk, with a training-compute presumption of 10^25 floating-point operations, face further duties under Article 55: model evaluation including adversarial testing, assessment and mitigation of systemic risks, serious-incident reporting and cybersecurity protection. The general-purpose model obligations began applying in August 2025, and the AI Office's enforcement powers over those providers follow from August 2026. For the business impact, see what the EU AI Act means for every AI company.

United States. Federal obligations on frontier labs are mostly voluntary commitments and company-published frameworks. A framework is a promise the company wrote and can revise.

Liability. Disclosure and liability connect. If a deployed agent causes harm, who pays? Who is liable when an AI agent fails covers the allocation.

The gap: even under the EU rules, evaluation results are reported to a regulator, not published. A reader outside the process sees a framework, not findings.

A Seven-Question Checklist for Evaluating Any ASI Lab

Run these in about 15 minutes on public material. Each answer is yes, no or not disclosed. Count only yes.

1. Is there a published threshold that pauses training or deployment?
2. Is there a named body or role with authority to enforce it?
3. Can that body block a release over the objection of the commercial side?
4. Are external evaluators given pre-release access to models?
5. Are evaluation results published, not only summarized?
6. Does the lab state a verification commitment as a ratio to capability spending?
7. What happened the last time a commitment was tested, and is the record public?

Reading the score: zero to two yes answers means a scaling lab with a mission statement. Three to five means a lab with partial controls. Six or seven is a lab whose claim you can test. No lab I reviewed earns seven on public material, and question 6 is the one that almost none answer.

Why It Matters

Investors. "Superintelligence" is priced as a premium. If the label cannot be verified, the premium is a claim about future capability with no measurement behind it. See AI valuations and the vertical AI thesis for where durable value may sit instead.

Operators. Building on a lab's models is vendor risk. A lab without enforceable governance can change its commitments, its terms or its availability on its own schedule. The case study above is the template.

Policymakers. Rules that reward the label reward the wrong thing. Rules that require evaluation disclosure reward verification.

Everyone. Mislabeling cuts both ways. Calling a scaling lab an ASI lab inflates trust in its claims. Dismissing a lab that does fund verification as hype misses the signal. The three-budget lens separates the two.

Limitations

• The scoring uses only public disclosures. A lab with strong internal practice and few publications scores poorly by construction.
• Superintelligence has no agreed measurement, so no score here measures progress toward it.
• The illustrative calculation is not empirical. Its growth rates are assumptions chosen to show a mechanism.
• The lab landscape changes within months. Frameworks are revised, teams reorganize and the table is dated October 2026. Check each lab's own page before relying on any cell.
• "Not disclosed" reflects what I could find in published material, not proof of absence.
• Conflict of interest: I, Liyam Flexer, run a site in this space. BestSuperintelligent and TheBestSI are my projects and have a commercial stake in how AI systems are evaluated. Treat the framing with that in mind, and verify the sourced claims independently.
• The checklist is a screen, not an audit. It cannot replace independent technical review.

References

• Nick Bostrom, Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014).
• Kaplan et al., Scaling Laws for Neural Language Models (arXiv, 2020).
• Hoffmann et al., Training Compute-Optimal Large Language Models (arXiv, 2022).
• Rich Sutton, The Bitter Lesson (2019).
• Safe Superintelligence Inc., founding announcement and site (June 2024).
• OpenAI, "Introducing Superalignment" (July 2023), and the OpenAI Preparedness Framework.
• Anthropic, Responsible Scaling Policy.
• Google DeepMind, Frontier Safety Framework.
• Meta, Personal Superintelligence (2025).
• Epoch AI, training compute and investment trend data.
• Regulation (EU) 2024/1689, the AI Act, Articles 51, 53 and 55.

Related

• Best Super Intelligence for Every Job: The Verify-Cost Method
• The AI Review Site That Refuses to Rank Anything Yet
• TheBestSI and the Road From AI Tools to Superintelligence
• Artificial intelligence coverage and the concepts of artificial intelligence and large language models

The Bottom Line

Judge an ASI research and development lab by what it spends on verifying and constraining its own systems, and treat the label as a claim to test. Capability spending is the easy budget to see and the easy one to grow. Verification and governance are the budgets that make a claim of superhuman performance checkable, and they decay in share unless committed as a ratio. Use the seven questions on any lab before you trust its claims, invest in it or build on it.