Researchers launch Sequent, a startup built on the premise that AI alignment is off track

A group of AI researchers has founded a startup called Sequent, with a blunt public statement: work on aligning artificial intelligence with human values is not on track. It is one of the rare cases in which people from inside the system openly discuss its systemic flaws.

Why Sequent exists

The founders, who come from the UK AI Security Institute and several other research organizations, offer a direct diagnosis: the academic community and large commercial labs systematically underinvest in key areas of safety research. The problem is not a lack of safety statements, since every major player has them. It runs deeper: money and attention flow to areas where there is already a measurable result and a clear publication horizon, not to areas where no result will appear at all without investment.

The UK AI Security Institute is one of the few government institutes in the world created specifically to assess the risks of advanced AI systems. That its alumni are leaving to found an independent startup says something important: even within the institutional response to the alignment problem, people see structural limits that cannot be worked around from the inside.

Sequent says it will take a fundamentally different approach. The startup plans to support a "portfolio of underfunded research bets": directions that are high-risk, long-horizon or lack immediate commercial payoff. In essence, it applies venture logic to the academic safety agenda.

What lies behind the diagnosis

The phrase "alignment is not on track" is not a marketing slogan. Behind it is a specific thesis: the capabilities of AI systems are growing faster than our understanding of how they work and how to control them reliably. The gap between what models can do and what we know about them keeps widening.

The AI safety community commonly distinguishes several layers of the problem, which Sequent intends to attack:

• Interpretability — we do not understand how exactly models make decisions inside the "black box"
• Scaling — more powerful models behave unpredictably compared with smaller versions, and extrapolation does not work
• Goal specification — it is extremely hard to verify that a model is really optimizing what we intend, rather than a proxy metric
• Robustness — behavior under out-of-distribution conditions often differs sharply from behavior in training
• Agency — autonomous systems acting in the real world create fundamentally new classes of risk

FrontierCode and synthetic assistants

The same issue of Import AI carries two additional items. FrontierCode is a new benchmark for evaluating code generated by large language models. Its key difference from its predecessors: the tasks come from real production repositories rather than being synthetically generated for testing. This brings the evaluation much closer to real engineering practice and makes it harder to train a model specifically for the benchmark.

The second item covers experiments with "synthetic research interns". AI agents take over routine tasks in scientific laboratories: searching for and summarizing literature, initial data analysis and drafting sections of papers. Researchers are testing how far such agents can speed up the scientific process without lowering quality or introducing systematic errors into the conclusions.

What this means

The emergence of Sequent is a signal: part of the research community is convinced that the current AI safety mainstream is not keeping up with the scale and speed of the problem. Large labs do invest in safety, but primarily in directions compatible with commercial and regulatory goals. An independent startup with an explicitly declared mission can fill structural gaps that systematically go unaddressed.

Whether a small team can overcome the resource gap in competition with Anthropic or Google DeepMind remains an open question.