The current landscape presents an unprecedented era for AI companies developing proprietary foundation models.
This era is marked by two key trends: established industry figures from major tech firms launching independent ventures, and renowned researchers with extensive backgrounds but unclear business objectives. While some of these emerging labs could evolve into giants akin to OpenAI, others might simply focus on groundbreaking research without significant commercial pressure.
Consequently, distinguishing which entities are genuinely pursuing profit has become increasingly difficult.
To simplify this, I propose a five-level continuum for any company developing a foundation model. This scale assesses ambition, not current revenue, focusing solely on the intent to commercialize rather than actual financial success.
Consider the following framework:
- Level 5: Generating substantial daily revenue.
- Level 4: Possessing a comprehensive, multi-phase strategy for immense financial prosperity.
- Level 3: Holding numerous promising product concepts, to be unveiled eventually.
- Level 2: Developing a rudimentary conceptual framework for a plan.
- Level 1: Prioritizing self-satisfaction over material gain.
Major players like OpenAI, Anthropic, and Gemini all reside at Level 5. The scale becomes more insightful when applied to the new wave of emerging labs, which possess significant aspirations but whose commercial intentions are less discernible.
Significantly, individuals within these labs often have the flexibility to align with any level. The current abundance of AI funding means business plans are not rigorously scrutinized. Investors are content to participate even if a lab primarily focuses on research. For those less driven by billionaire status, a Level 2 existence might offer more contentment than Level 5.
Challenges emerge from the ambiguity of an AI lab’s position on this scale, a source of much present industry contention. For instance, much of the concern surrounding OpenAI’s transition from non-profit status stemmed from its abrupt leap from Level 1 to Level 5. Conversely, Meta’s initial AI research could be classified as Level 2, despite the company’s apparent desire for Level 4.
Considering this framework, here’s a concise overview of four prominent contemporary AI labs and their respective standings on the scale.
Humans&
Humans& garnered significant AI headlines this week, partly inspiring the creation of this scale. Its founders present a compelling vision for future AI models, shifting focus from scaling laws to advanced communication and coordination functionalities.
Despite the favorable media attention, Humans& has remained reticent about how its vision will manifest into concrete, monetizable products. While product development appears to be their intention, the team refrains from specifying details. Their clearest statement indicates plans for an AI-driven workplace tool, designed to replace existing solutions like Slack, Jira, and Google Docs, while also fundamentally re-imagining how such tools operate. Essentially, workplace software for an evolving, post-software environment!
As someone whose role involves understanding these concepts, I admit I’m still quite perplexed by that final description. However, the proposal offers just enough specificity to categorize them at Level 3.
Thinking Machines Lab
Assessing this one is challenging! Typically, a former CTO and ChatGPT project lead securing a $2 billion seed round implies a clear strategic direction. Mira Murati doesn’t seem like someone to undertake ventures without prior planning, leading me to comfortably place TML at Level 4 in early 2026.
However, the past fortnight has brought developments. The exit of CTO and co-founder Barret Zoph received significant media attention, partly due to the unique factors involved. Additionally, at least five other employees departed alongside Zoph, many expressing apprehension regarding the company’s trajectory. A mere year after its inception, almost half of TML’s founding executive team has moved on. One interpretation is that their envisioned path to becoming a leading AI lab proved less robust than initially believed. On our scale, they aspired to Level 4 but discovered themselves closer to Level 2 or 3.
While insufficient evidence exists for a definitive downgrade yet, such a move appears increasingly imminent.
World Labs
Fei-Fei Li stands as a highly esteemed figure in AI research, particularly recognized for initiating the ImageNet challenge, which spurred modern deep learning advancements. She presently occupies a Sequoia-funded chair at Stanford and co-leads two distinct AI laboratories. Without listing her extensive accolades and academic appointments, suffice it to say she could easily spend her remaining career accepting honors and commendations. Her publication is also quite commendable!
Thus, in 2024, when Li revealed securing $230 million for World Labs, her spatial AI firm, one might have initially placed it at Level 2 or below.
However, more than a year has passed since then, a considerable duration in the rapidly evolving AI sector. During this time, World Labs has successfully launched both a comprehensive world-generation model and a commercial product derived from it. Simultaneously, tangible demand for world-modeling has emerged from the video game and special effects industries, with no major labs offering comparable solutions. This trajectory strongly suggests a Level 4 company, potentially on the verge of reaching Level 5.
Safe Superintelligence (SSI)
Safe Superintelligence (SSI), established by former OpenAI chief scientist Ilya Sutskever, appears to embody a quintessential Level 1 startup. Sutskever has meticulously shielded SSI from commercial influences, even rejecting an acquisition attempt by Meta. The venture lacks product cycles, and beyond the developing superintelligent foundation model, there seems to be no discernible product offering. Despite this non-commercial stance, he secured $3 billion! Sutskever’s inclinations have consistently favored the scientific aspects of AI over business, and all signs point to SSI being fundamentally a research-driven endeavor.
Nevertheless, given the rapid pace of the AI industry, it would be imprudent to completely dismiss SSI from the commercial sphere. During a recent interview with Dwarkesh, Sutskever articulated two potential catalysts for SSI to shift its focus: either “if timelines turned out to be long, which they might” or due to the belief that “there is a lot of value in the best and most powerful AI being out there impacting the world.” This implies that SSI could quickly ascend several levels on the scale, depending on whether its research progresses exceptionally well or encounters significant challenges.