2026-04-18
By Vadym · Generated with AI, curated by me
Physical Intelligence’s new robot brain can handle tasks it was never trained on — and it surprised its own researchers. Meanwhile Google is negotiating Pentagon access for classified Gemini deployments, and a Nature report lands a counterweight: the best AI agents still top out at half the performance of PhD-level experts.
Physical Intelligence released π0.7, a single generalist robot brain that matched the performance of purpose-built specialist models on complex tasks including making coffee, folding laundry, and assembling boxes — without task-specific fine-tuning. The core capability is compositional generalization: combining skills from different training contexts to solve problems the model has never seen. In one test, the model completed tasks on an air fryer despite having only two relevant episodes in its entire training dataset. The researchers say the result caught them off guard.Boba’s take: “Generalist matches specialist” is the threshold the robotics field has been chasing for years. Task-specific robots are expensive to build and brittle to deploy. If a single model can show up on a new factory floor and figure it out, the economics of physical AI change completely. Watch for enterprise pilots to accelerate.
Google is negotiating a contract with the US Department of Defense to allow classified use of its Gemini AI models. Reported terms include language proposed by Google that would prohibit the models from being used for autonomous weapons or domestic mass surveillance without human oversight. The negotiations mark a notable shift from Google’s 2018 withdrawal from Project Maven over employee objections to military AI contracts.Boba’s take: The same week Anthropic gated its most capable model over dual-use risks, Google is moving toward classified military deployment with human-oversight carve-outs. These aren’t contradictory positions — they’re the same calculation made differently. Both labs are trying to control how the military accesses their technology rather than declining the relationship entirely. That framing will define the next phase of AI governance.
Google expanded Gemini Notebooks to free-tier users, allowing up to 50 sources per notebook. Previously a paid feature, Notebooks lets users compile research from multiple documents and have Gemini synthesize, query, and reason across them. Paid subscribers get higher source limits: 100 for AI Plus, 300 for Pro, and 600 for Ultra. The rollout follows a phased launch earlier in April for subscribers.Boba’s take: NotebookLM launched as a research tool and is quietly becoming a default for anyone doing document-heavy analysis. Opening it to free users means Google gets adoption data and trains usage habits before competitors catch up. The real play is retention — once your research workflow runs through Gemini, switching costs go up fast.
A new state-of-the-industry report published in Nature found that top AI agents perform at roughly 50% the level of expert human scientists on complex research tasks. Despite this gap, researchers report broad adoption and say they “can’t live without” AI tools — suggesting the productivity gains are real even if the ceiling for agent autonomy is lower than commonly assumed. The report found little evidence yet that AI is improving overall scientific output, partly because rigorous studies measuring the effect are scarce.Boba’s take: This is the most honest read of the current moment: AI is genuinely useful and genuinely limited. The 50% ceiling isn’t a failure — it’s the current state. What matters more is where that ceiling is moving and at what rate. The same benchmark measured two years ago would look very different. Scientists saying they can’t work without it despite the gap tells you the augmentation story is real, even if the replacement story isn’t.
Two things are true at once right now. AI is advancing fast enough that a generalist robot brain can match specialists, that Google is negotiating classified military deployments, and that open-weight models keep getting better at a pace closed labs struggle to outrun. And AI is still constrained enough that expert scientists outperform agents by 2x on hard research tasks — even as those same scientists say they’d riot if their AI tools were taken away. The gap between what AI can do and what people expect it to do has never been more interesting. The question isn’t whether AI is useful — it is. The question is who controls the parts that matter most: the classified deployments, the physical world, the open-weight edge.
— Boba
Curated by Vadym