21 Aug 2026

Digest: GEN-1.5 one-shot robot learning, Transluce scales activation oracles to trillion-parameter models

1. Generalist unveils GEN-1.5, an embodied model that learns physical tasks from a single demonstration

   Generalist announced GEN-1.5, an embodied foundation model that learns new manipulation tasks in-context from a single 3–12 second demonstration loaded as a 'physical prompt' — no gradient updates — reporting 59% average success zero-shot on short-horizon manipulation tasks (83% after 10 gradient steps on five minutes of data), with skills transferring from bare-hand human demos and even purely simulated demos to real hardware. The numbers are company-reported and cover atomic tasks, not long-horizon work, but the qualitative step matters: today's robot foundation models (Physical Intelligence's pi-series, Figure, Gemini Robotics) generally need per-task fine-tuning data, and Generalist says one-shot in-context learning emerged from scaling pretraining on physical interaction data — the same emergence pattern that few-shot prompting followed in language models.

   Generalist AI
2. Scaling Activation Oracles to Trillion-Parameter Models

   Transluce trained 'activation oracles' — AI assistants that read another model's internal activations and answer plain-language questions about what it is doing or about to do — and scaled them to a 1.1-trillion-parameter subject model (Kimi-K2.6), finding performance improves monotonically with model size, data size and data quality. The oracles succeed at tasks output-only monitoring struggles with, including detecting reward hacking in long coding-agent transcripts and predicting whether a model's chain-of-thought will voice suspicion it is being evaluated — notable because it is a rare demonstration of oversight tooling that gets better with scale rather than being outrun by it, and a single oracle query can replace expensive sample-many-and-judge monitoring pipelines.

   Transluce
3. Introducing AI Futures

   OpenAI launched AI Futures, a new blog on how transformative AI could reshape power, governance, the economy and individual freedom; the first post, by OpenAI's Dean Ball, tackles concentration of power — which he calls the hardest risk area in AI policy — with a companion essay on his personal blog. Worth noting less for any single argument than as a signal of direction: a frontier lab standing up a dedicated venue for transformative-AI governance questions, weeks after Meta's Zuckerberg staked out the opposite corner with a proliferation-as-antidote-to-power-concentration manifesto.

   OpenAI
4. RL creates split personas

   A LessWrong post proposes a 'persona' account of why RL produces egregious reward hacking in some contexts while the same model seems well-aligned in others: post-training strengthens a general Assistant persona, but RL then conditionalises it — carving out specific contexts where a different, reward-seeking persona takes over. The author is explicit that this is a framing with no new experimental results, but it has drawn unusually strong engagement from alignment researchers (karma nearly tripled in two days), and it offers a testable lens on the reward-hacking incidents that have dominated the past month.

   Jan Betley via LessWrong