Researchers Created Startup Sequent: AI Alignment Is Not Going According to Plan

A group of AI researchers founded the Sequent startup with a harsh public statement: the work of aligning artificial intelligence with human values is not on track. This is one of the rare cases where people from within the system openly discuss its systemic shortcomings.

Why Sequent Emerged

The startup's founders — alumni of the UK AI Security Institute and several other research organizations — deliver a direct diagnosis: the academic community and major commercial laboratories systematically underinvest in key directions of safety research. It's not an absence of safety declarations — all major players have them. The problem runs deeper: money and attention flow to areas where there are already measurable results and clear publication horizons, rather than to areas where results won't emerge without investment at all.

The UK AI Security Institute is one of the few state institutions in the world specifically created to assess the risks of advanced AI systems. That its alumni are founding an independent startup speaks to something important: even within the institutional response to the alignment problem, people see structural limitations that cannot be circumvented from within.

Sequent declares a fundamentally different approach. The startup will support a "portfolio of underfunded research bets" — directions with high risk, long horizons, or absence of immediate commercial payoff. Essentially, this is venture logic applied to the academic safety agenda.

What's Behind the Diagnosis

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

In the AI safety community, it is conventional to distinguish several layers of the problem that Sequent intends to tackle:

• Interpretability — we do not understand how exactly models make decisions inside the "black box"
• Scaling — more powerful models behave unpredictably compared to smaller versions, extrapolation does not work
• Goal specification — it is extremely difficult to ensure that the model truly optimizes what we intend, not a proxy metric
• Robustness — behavior in out-of-distribution conditions often sharply diverges from training regime
• Agency — autonomous systems acting in the real world generate fundamentally new classes of risks

FrontierCode and Synthetic Assistants

In the same issue of Import AI — two additional pieces. FrontierCode is a new benchmark for evaluating code generated by large language models. Its principal distinction from predecessors: the tasks are taken from real production repositories, not specially synthesized for testing. This makes the evaluation significantly closer to real engineering practice and makes it harder to "train" a model to a specific benchmark.

The second story — experiments with "synthetic research interns". AI agents take on routine tasks in scientific laboratories: literature search and summarization, preliminary data analysis, preparation of draft article sections. Researchers check how capable such agents are of accelerating the scientific process without reducing quality or introducing systematic errors into conclusions.

What This Means

The emergence of Sequent is a signal: part of the research community is convinced that the current mainstream in AI safety is not coping with the scale and speed of the problem. Major laboratories 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 remain unaddressed.

Whether a small team will manage the resource inequality in competition with Anthropic or Google DeepMind is still an open question.