After inference

Every breakthrough in AI over the past decade has followed the same pattern: find some data set or resource that can be made abundant and scale it.

GPT only became possible once we realized the internet could become a training dataset, making human knowledge effectively abundant. ChatGPT demonstrated that human preference could itself be scaled via RLHF, synthetic data and evaluation. More recently, reasoning models treated inference compute as something that could be allocated dynamically, allowing models to scale by "thinking harder" when the problem demanded it.

But what if the next resource can’t be made abundant so easily?

Two recent pieces point toward the same uncomfortable conclusion. Dwarkesh Patel argues that the next paradigm is continual learning (i.e. models improving from real world deployment rather than static datasets), but that this only works if you can make experience reproducible, parallelizable and cheap to iterate on (or as he calls it, “grindable”). Animesh Garg argues that deployment data behaves like an oil well. Early failures are high-entropy and informative, but routine successes provide little valuable information, and once a niche saturates, the well runs dry.

The venture industry hasn’t absorbed either argument. Nearly every humanoid company is racing to collect more teleoperation data, while startups like XDOF (freshly backed by Thrive, Spark, Lux and a16z!) and Rerun and Roboto are building the infrastructure to capture/clean/label and log/visualize and analyze (respectively) all interactions and deployments. The implicit assumption: if internet-scale data unlocked language models, world-scale experience will unlock physical intelligence.

But Dwarkesh and Garg are both pointing at the same problem, even if they're coming at it from different angles. More experience isn't the answer. The challenge is efficiently converting experience into learning.

Experience data is fundamentally different from internet data because, as Garg lays out, every additional experience has a real cost. Once every experience carries a price tag, the optimization problem changes completely.

For some systems, the environment can be fully indexed. These are closed world systems. And for systems operating in closed worlds, the optimization problem is coverage. We've seen this in motion with our portfolio company Motorica. Every motion capture session is expensive, requiring specialized performers, carefully designed sequences, and extensive cleanup before it ever becomes training data. As a result, a critical problem they needed to solve early on was understanding exactly which movements produce the largest jump in model capability across a new motion modality per dollar spent.

For other systems, the environment isn’t so easily understood. These are open world systems. These AI systems could be operating humanoid robots, autonomous vehicles, or enterprise agents. But what makes them similar is that the environments continuously generate new situations, new failure modes and new edge cases. The world keeps changing, so the model has to keep learning.

(The reality is a spectrum, not binary, but that doesn’t make for such a nice, pithy VC framework now does it?)

The obvious answer to the challenge posed by open world environments is simulation. If real world experience is so expensive, why not generate synthetic experience at scale? The short answer is that simulation works until it doesn’t. Unmodeled physics (e.g. dynamic friction, soft-body deformation) proves difficult to simulate reliably, for example. And more broadly, while world models are improving fast, the failure modes that matter most are the long-tail edge cases nobody anticipated. You can’t simulate a failure you don’t know exists. But the answer isn’t to abandon simulation. It’s to close the loop. The most valuable infrastructure won’t just push from sim-to-real, it will capture unexpected real world failures, pull that telemetry back into the simulator, generate synthetic variations of that exact edge case, and update the model. This “sim-to-real-to-sim” flywheel is the ultimate moat. (Angular has a recent investment on this thesis, in fact! More to come soon.)

Just reading this, it’s clear that we will need some sort of infrastructure that can help models learn from experience efficiently. 

That infrastructure is just now starting to appear. NVIDIA's ENPIRE project looks like a robotics paper, but it's really a paper about the infrastructure required to make reality learnable. They describe building systems for automatic evaluation, environment reset and parallel experimentation. This is the sort of scaffolding that will make real world experience “grindable,” to use Dwarkesh’s phrase.

But grindability is only the first problem. Before an AI system can improve from experience, it first has to decide whether an interaction is even worth learning from. Most deployments are routine. Only a tiny fraction contain genuinely new information, e.g. unexpected failures, successful recoveries, or entirely novel situations.

Once you've identified a valuable experience, the next challenge is figuring out how to incorporate that experience back into the system.

There are more pernicious problems here. For example, a model that learns from new experiences tends to overwrite what it already knew. This is called “catastrophic forgetting,” and there is no generally accepted production-grade solution yet. Whatever approach ends up working will require infrastructure to identify which experiences to learn from and how to update carefully enough that the model compounds rather than forgets. That stack doesn’t exist yet.

The first generation of AI infrastructure was built around inference and retrieval. Vector and graph databases retrieved context. Model routers selected the right model for the right task. Every layer of that stack assumes an existing, already trained.

The next generation of AI infrastructure is going to start after inference. Just as vector and graph databases gave LLMs context, we need new databases to index physical telemetry. Just as RLHF shaped texted alignment, we need some sort of “automated physical evaluation” (like ENPIRE’s vision-based success checks) to shape robot behavior. And so on. Rather than focusing on helping models answer questions more quickly or cheaply, we need infrastructure that helps models decide what to learn from those answers.

As an early stage investor, this sort of shift gets me particularly excited. Because, just as the inference stack wasn't built by the labs, I imagine this stack will be built by a different generation of companies, ones that see this opportunity more clearly and can build what the incumbents cannot.

We've spent the last three years building infrastructure for inference. The next decade will be spent building infrastructure for learning. And the companies that do it are probably getting started right now.

David Peterson

Could the future of software be fluid
How do we get the best of AI without losing the soul of software?

The future belongs to young missionary teams
Why it makes more sense betting on youth in the current moment

The AI-native enterprise playbook
Ten real-time observations on a rapidly evolving playing field

No more painting by numbers
It’s the end of the “SaaS playbook.

ENTERPRISE/TECH NEWS

Droning on. German drone maker Quantum Systems raised $1.2 billion at a valuation of roughly $8 billion, bringing in fresh capital to expand drone production and invest in software for autonomous systems powered by artificial intelligence (WSJ). This is the second major German drone investment in the past month, following a $500M investment led by Sequoia into Stark Defense (Sifted).

Strong shekel. Israeli tech companies are struggling with a very strong Shekel, which is making operations more expensive than budgeted. “A study by the Israel Growth Forum has found that Israeli tech employees are the most expensive in the world due to the strengthening of the shekel. Israel’s 400,000 tech employees have for some years been expensive in global terms, with their cost second only to that of Silicon Valley. The tech giants preferred to pay dearly for the Israeli worker, knowing that they would receive a better result than their counterparts in India and eastern Europe, while their price was still lower than tech workers on their home turf. However, according to a study by the Israel Growth Forum - an organization operated by Israeli companies such as Wix and Monday.com - the cost of Israeli developers has surpassed that of their US counterparts for the first time ever. This has been mainly due to the strengthening of the shekel against the US dollar, since most of the capital in Israeli tech is raised in dollars, while most of the expenses, including wages and options for local employees, are made in shekels.”

The enterprise backlash against AI lock-in begins in earnest. Palantir CEO Alex Karp gave an interview to CNBC that many described as unhinged but contained some interesting points. According to Dave Vellante in Silicon Angle, “Karp’s argument is that frontier model vendors (he didn’t mention Anthropic and OpenAI by name) intend to suck the knowledge out of enterprises and destroy the “alpha” companies enjoy through their proprietary data, processes and underlying business advantage.” Newstack ran a piece comparing Alex’s comments with recent statements by Mistral CEO Arthur Mensch. “It would be naive to ignore the commercial interests at play. Karp’s Palantir sells the deployment and governance layer; it benefits directly if enterprises treat models as interchangeable commodities. Mensch’s Mistral sells open-weight models and a training platform, and it benefits directly if enterprises distrust closed providers. Zoho’s Sridhar Vembu, who endorsed Karp’s position publicly last week, has his own reasons for wanting enterprises to own their AI infrastructure rather than rent it from Silicon Valley. But the fact that multiple executives across different market segments are saying companies need to own their data, maintain deployment flexibility, and not hand their competitive advantage to a provider who might become their competitor suggests the argument is resonating.”

HOW TO STARTUP

Towards open-source models. Jesse Zhang, CEO of Decagon, explained why open source will become a growing piece of the AI model landscape and - crucially - why its not yet there today. “And that's the resolution to the paradox. The reason open source share fell isn't that open source is losing. It's that enterprise AI as a whole is at the very beginning of the maturity curve. Last year enterprises stopped building and started buying, and thousands of brand new use cases spun up at once. New use cases run on frontier models, so closed share exploded. The 11% (open source) is a denominator problem: the pool of immature use cases is growing faster than the pool of mature ones. If that's right, then every use case being prototyped on a frontier model today is a future open source migration. As deployments mature, companies will do what we did: distill, fine-tune, specialize. The frontier labs will keep owning discovery. Open source will increasingly own production.”

Don’t forget to update. Arian Ghashgai of Earthing VC provided a strong case for consistent periodic updates from founders to their VCs: “The non-obvious reason why not doing this can actually have unintended negative fallout is that downstream VCs (e.g. multistage firms) often ask downstream VCs (e.g. pre-seed funds) about certain companies in their portfolios (i.e. they source from these funds), and it looks bad (for the company) to prospective investors if existing investors have no updates from the company to share (bad communication is not desirable)”

HOW TO VENTURE

Struggling with the bubble. A tweet by Madhav Chanchani summarized a key highlight in a conversation between Jack Altman and Everett Randle of Benchmark and Trae Stephens and Delian Asparouhov of Founders Fund: “Partners from two top US venture firms, Benchmark and Founders Fund, say they are uncomfortable with the current funding frenzy because it reminds them of 2021. “I’m very uncomfortable.....because it feels like we're getting back to a point where prices have become untethered from reality. It reminds me a lot of 2021,” says Trae Stephens of Founders Fund. Everett Randle of Benchmark adds: “The similarity I see with 2021 is that back then there was an almost unanimous sense of optimism because investors were making a tremendous amount of money..I think this duality is what defines a bubble.””

Blue Energy, GE Vernova, and Crusoe are advancing a gas-to-nuclear model that pairs natural gas with SMRs to deliver firm power for AI data centers while reducing nuclear financing risk.

Groundcover announced a major expansion of Agent Mode that lets artificial intelligence agents act on a team’s observability data across the development tools they already use.

Blue Energy
Head of Communications (DC, Hybrid)

FalkorDB
Core Developer (Tel Aviv, Hybrid)

Groundcover
Account Executive - East Coast (Boston)