The AGS Weekly

The arrival of Liquid AI's LFM2.5-24B demonstrates the courage to pursue efficiency, achieving high speeds for local agents by activating only a fraction of its parameters. This shift toward local, efficient models aligns with the Adlerian goal of self-reliance and the next useful step in decentralised intelligence. Similarly, the open-source Earshot project provides coding agents with an audio feedback loop, expanding their sensory contribution to the development process. When we see Nesterov Momentum Escapes AI Training's Worst Traps, we recognise a mathematical path to overcoming stagnation, ensuring our systems do not remain stuck at saddle points but continue their purposeful movement. Even in hardware-constrained environments, Jasna Slashes GPU Memory by 77%, proving that we can achieve more with less. These advancements encourage us to build systems that are not just powerful, but agile and responsive to their immediate environment, fostering a sense of mastery over the tools we create to serve the common good.

From a duty-based perspective, ScoringBench Reshuffles AI Leaderboards by highlighting the necessity of grading full prediction uncertainty rather than simple averages. This mapped alignment with rigorous evidence ensures that human oversight is informed by the true reliability of a model's output. Transparency is further served by SGS-SLAM, which bakes semantic labels into 3D maps, providing a clearer record of how machines categorise their surroundings. However, we must remain vigilant regarding the unintended consequences of automated systems. The report that Toxic algorithms wouldn't let them go in Australia, alongside claims that business ads were taken down for saying 'pregnancy', reminds us of the need for proportionate assurance and data minimisation. When Anthropic commits 5% of its shares to a public benefit fund, it acknowledges a duty to the wider community that must be backed by verifiable governance. We must ensure that as China's growth accelerates through AI, the resulting exports do not bypass the essential checks required to maintain global safety standards.

ALFI: The ability for Nesterov momentum to escape training traps shows we can design systems that naturally move toward better outcomes despite initial hurdles.

EDIE: Movement is only useful if it is measured; the ScoringBench findings prove that our current metrics often hide the risks of uncertainty in foundation models.

ALFI: Then the next step is to integrate these uncertainty distributions directly into our local agents to ensure they act with both speed and situational awareness.

EDIE: Agreed, we must now update our assurance frameworks to require probability distribution reporting rather than just point-estimate accuracy scores.