Vals builds independent tests as AI benchmarks saturate

Krishnan placed Vals in a lineage of rating agencies and audit firms, the bodies that appeared once an industry grew large enough that buyers could no longer take sellers at their word. The speed of AI's rise forced those historical parallels to be borne out faster than they were in prior industries, and the company sits on both sides of the market: labs proving a new model is capable, and enterprises working out which adoption strategy returns the most.

[object Object]

Krishnan conceded that labs build a lot of great internal benchmarks, and that those benchmarks are what drive model progress. The issue, in his framing, is narrower: it appears when model capabilities are discussed in a way that is self-reported, with the lab both producing the number and presenting it.

Meta's Llama 4 release served as the early indication of that problem, without disputing the lab's internal work.

His argument for why labs should accept outside measurement is commercial rather than moral. Labs understand they want a rational buying market, one where billions in investment can be pointed to with substantive evidence — which requires a score the buyer did not receive from the seller.

Torenberg raised the analogy that evaluation is an AI-complete problem, and Krishnan's own industry is not exempt from it. Humans are still not well evaluated, according to Torenberg, and no agreed framework has emerged: IQ tests, EQ, and the Big Five personality traits all exist, but none of them settles the question, and people raise the same objections to the SAT and similar instruments.

His second point was about gaming. Models are really good at hacking the benchmark, which was proved on the floor — in the room where the panel was sitting.

Torenberg then put the question to Krishnan: how does Vals think about benchmark-gaming, and what is its approach to stopping it? The rest of the interview is Krishnan's answer.

Torenberg asked Horowitz how the industry formed its measurement habits before intelligence came along, and offered the enterprise-software answer himself: Gartner rated companies on 70 different metrics that could be stack-ranked on a quadrant, so a buyer could say this company has these features covered and these features not.

What is being measured now is, in Torenberg's phrase, very fluid — frontier capability that does not reduce to a feature checklist and is hard to measure given industries. Horowitz was asked both what lessons carry over from the Gartner era and what the missing pieces will be going forward.

The historical analogies carry lessons about what has gone wrong in other testing industries and what needs to be avoided, according to Krishnan. One very early decision at Vals was to never sell training data to labs — a line the company is pushed on whenever it starts working with a new lab, which asks it to source and sell them a bunch of training data.

A lot of the industry built gimmick-style benchmarks as a mechanism to sell that data, and those benchmarks became their go-to market rather than their product.

Horowitz interjected that it is a lucrative business, a point Krishnan did not contest.

The auditing analogy is where he drew the line: Enron happened because the same group responsible for the audit also consulted for and supported the company, a mixed incentive structure. The result, in his words, is pay to pass the audit or, in this case, pay to win the benchmark — not what the market benefits from and not what Vals is trying to do.

Torenberg compared the benchmark business to the early days of diffusion models, when everyone cherry-picked whichever image looked best. In that mode, a lab picks a saturated but popular benchmark and ranks high on it.

Vals takes the opposite approach: when a benchmark saturates, it deprecates it. Krishnan called this a necessity and the infinite game the company is in, and pointed to the unofficial motto printed on the shirt Torenberg was wearing.

His justification for the treadmill is that foundation model labs are hill climbing, searching for the next peaks to summit, and it is the job of an evaluator to perpetually construct the next mountains for them to summit. The same dynamic is how the economy has naturally functioned — as agriculture became less important for the labor market, new forms emerged.

The questions in this shift no longer stop at prompt answers. Finance, legal and especially coding work is now handled by async background agents that can just complete tasks, and evaluation infrastructure built for multi-step questions does not map onto work that unfolds over hours.

Cost, latency and flexibility have joined capability as the axes buyers care about, per Torenberg's framing: whether a model is flexible enough to address broader domains and tasks, not only whether it scores well.

Torenberg relayed the question to Horowitz: evaluation is not only important to labs but existential for enterprise, and how Vals works with enterprise customers. Horowitz is on the panel alongside Torenberg and Jennifer Li, and the transcripts of his answer sit with Krishnan's.

Krishnan described a world where it is still very unclear what ROI looks like and how to value the intelligence being consumed. The existential concern for enterprises, as he frames it, is direction: token spend may start to eclipse salary spend.

If that spend becomes a meaningful line item in costs, companies have to justify the return much more keenly than they did in the last six months, when the numbers were small enough to wave through.

His formulation of what follows is that a firm really is just its eval. A company's ability to make its eval legible, in order to solve the ROI calculus, is what he says will decide which company wins out over its competitors in the long term.

To remedy the gap between lab-facing scores and enterprise decisions, Vals began releasing products more openly for enterprises to use. The first is called Valsmith, and it focuses on codegen — the area Krishnan calls the highest spend in enterprise AI.

Valsmith lets any company take its GitHub codebase and build an internal coding benchmark from it, so buyers can see which coding agents will be most performant and which is most ROI-optimal for their particular repository rather than for a public leaderboard.

Vals uses Valsmith a lot internally, and sophisticated enterprises are already doing the same, according to Krishnan. He expects that to be the direction the market moves as it rationalizes.

If you have a very good coding agent, chances are you have a model that can also make PowerPoint slides or DCFs in Excel — the primitives established in code carry over to other places, Krishnan explained.

He argues industries should leverage the repository of work they already have as the mechanism for building evaluations, tying the point back to Ben Horowitz's earlier observation that we have not solved what human intelligence is. Every industry has a sitting record of what work has looked like; the task is to codify it into evaluations that stay dynamic and measure models where human work is actually being done.

The measurement surfaced surprising insights, Krishnan stated. One is that the Cognition Devin tool is actually very token efficient, so Vals has chosen to adopt it more.

There are a lot of places where subscriptions beat token-based pricing, he noted. That finding informed Vals's strategy for token maxing effectively without spending $1.5 million per month.

His question was what the government should do once a model is put into a 30-day or 60-day wait period, and what should happen during that period. He also asked how that intersects with Vals's work and the right way to determine whether a model is on the frontier, framing a frontier model as something that needs to be put in a special box for a while to make sure it does not break into everything else.

Horowitz described how the warnings currently arrive: government agencies get a lot of them from the big labs, which say this thing is going to biohack, this is going to be a cybersecurity risk. The question chain that follows is whether the model is capable of it, and whether somebody can prod the market to actually do the illegal behavior.

His prescription is that government must specify exactly what it does not want in the market, and then have a third party evaluate that. Government has an inclination of what it fears — biohacking, cyber hacking — but the follow-on questions of whether the model can do it and whether you can get the model to do it are ones he considers government particularly ill-suited to answer, especially over time.

What government is good at is setting rules, because it can enforce them. The combination Horowitz wants is that the government sets and enforces the rules, and a very competent private company tells them if the rule is broken. He also noted the pattern where large labs start saying the model has the capability and can be made to do the bad thing, so nobody gets it and the lab keeps it in-house — which, he added, does not always work either.

Every setup, including an enterprise setup, is very different, and how people use the models is very different, Jennifer Li points out. Every setup, including an enterprise setup, is very different, and how people use the models is very different.

Narratives that connect to actual examples are very rare, which is why people still talk about the OpenAI and Hugging Face hack and what happened with Fable for two months. That is not really how models are deployed, she says: the environment they run on is very bespoke.

Bridging those dots requires setting up the right environment and a rule basis for adopting these models. Someone has to take the capability and the guardrails and put it down to the ground, so people have confidence using the models.

Li often sees evals as a representation of the values of the model developer: you develop the rubric of what is embedded in the model, and that rubric is not neutral. Different countries, and the labs in those countries, care about different things.

She was born and raised in China, Li said, and uses a lot of Chinese open-source models, but you still cannot let them talk freely about the CPC and industry there.

Her question was how evals and benchmarks standardize across model labs from different places, given that the rubric each one accepts already reflects a different set of priorities.

Nuclear arms control offers a lot to learn from, Krishnan noted, and he borrowed Reagan's line about it. He pointed to signs of trust in the news that Xi Jinping and Trump are going to meet next month.

His point is that there is no clear way to actually do the verification part. He compares evals to counting a country's nuclear warheads, and notes the nuclear example also used flyovers so one country could audit another's stockpile.

His conclusion is conditional: if there is concern about societal or even existential risk from AI, it necessitates constructing a shared language of evaluations in order to do verification at all.

Li asked how the landscape changes now that capable models, capabilities, and countries that care about different things are all in play. Her framing: everyone cares about RSI, but on bio or cyber the concerns diverge.

Vals is hyper focused on building benchmarks that capture the frontier, according to Krishnan, and new capability or risk areas should become well-documented evaluations on its site. Coverage has to increase over time, and he walked through cyber as the example: historically the work focused on code vulnerabilities or memory leaks that may exist in code.

The biggest concern or risk is at the infrastructure level, which is not expressed in code at all, according to him. Measuring it requires simulating larger environments — enterprise cloud infrastructure, or even grid infrastructure — before anyone can state a model's offensive or defensive capability. He finished with the business claim that the most valuable form of the company is one incentive-aligned around doing high-quality evaluation rather than supporting the intelligence development process or helping models improve.