AI Regulation & Policy

Today's Stories

1. 1

   DeepSeek's new reasoning model reaches #2 ranking among open-source AI systems, signaling intensifying competition in advanced AI development.

   DeepSeek released a new reasoning model that ranks as the #2 open-weights reasoning model. The model operates on 27% of FLOPs compared with DeepSeek-V3.2 and supports 1M tokens (up from 128K in V3.2). Open-source reasoning models—systems that can work through complex problems step-by-step—are becoming more capable and efficient. DeepSeek's advancement suggests that powerful AI reasoning is no longer confined to proprietary, closed systems, which could reshape how organizations access and deploy advanced AI capabilities.

   The model's efficiency gains (requiring significantly fewer computational resources than prior versions) and expanded context window (1M tokens allows it to process much longer inputs) may influence how other AI developers approach reasoning system design.

2. 2

   AI investment is shifting toward innovation and growth over efficiency, but executives' confidence in scaling AI within their own organizations is falling due to integration challenges.

   Akkodis' recent report shows that innovation has overtaken efficiency as the primary driver of AI investment, signaling a focus on growth and new business models. At the same time, CTO confidence in scaling AI has dropped significantly, held back by obstacles in integrating AI across enterprise systems. The report also flags agentic AI—systems that autonomously make decisions and complete tasks—as a crucial trend that will require new governance frameworks. Companies are investing heavily in AI because they see growth opportunities, but many are struggling to actually deploy these systems across their existing operations. This gap between ambition and execution suggests that even as businesses pour resources into AI, the real bottleneck is internal: figuring out how to make AI work within complex enterprise environments, not building the technology itself.

   The emergence of agentic AI as a focal point means organizations will need to develop new governance structures to oversee AI systems that operate with less human oversight. How quickly enterprises can close the integration gap—and build the governance frameworks agentic AI demands—will likely determine which companies capture the value from their AI investments.

3. 3

   Trump administration signs orders to accelerate quantum computing development, aiming for a machine capable of scientific research by 2028, as the technology moves closer to commercial use.

   US President Donald Trump signed executive orders to speed up quantum computing development with a target of creating a machine capable of scientific research by 2028. The administration has taken $2 billion（約3200億円） in equity in quantum firms and stakes in other technology companies. Quantum computing is expected to speed drug discovery and materials science, but also poses security risks — Google warned in March that firms should be ready for 'post-quantum cryptography' by 2029. Several quantum firms went public this year, signaling that the previously futuristic technology is moving toward commercialization.

   The 2028 deadline for a quantum machine capable of scientific research, and whether the administration's equity stakes and tech intervention approach will accelerate the timeline compared to private-sector efforts.

4. 4

   DeepSeek releases R1, a reasoning model that matches leading competitors' performance while using significantly fewer computational resources.

   DeepSeek introduced R1, an open-weights reasoning model (AI that works through problems step-by-step before answering). The model achieved performance comparable to OpenAI's o1 on reasoning benchmarks, yet requires 27% of FLOPs (computational work) compared with DeepSeek-V3.2 and uses 83.9 GiB of memory. Until now, state-of-the-art reasoning models have demanded enormous computational power, putting them out of reach for most organizations. R1's efficiency suggests that high-quality reasoning is becoming more accessible — the body indicates this is notable enough to shift expectations about what level of performance is possible with less hardware investment.

   R1 is open-weights, meaning researchers and companies outside DeepSeek can download and run it themselves. The model is available now, so organizations interested in reasoning capabilities can begin testing it without waiting for API access or paying per-query fees.

5. 5

   Holden Karnofsky warns that AI safety efforts, despite good intentions, may carry a risk of net negative impact on outcomes.

   Karnofsky, a prominent AI safety figure, published a list of downside risks for AI safety work broadly, acknowledging that even well-intentioned interventions in AI governance can backfire—for example, badly designed regulation could make things worse or increase the risk of great power conflict. Karnofsky states he personally believes the probability of unintended harm from AI safety efforts is greater than 50%, and he acknowledges having to live with the possibility that his own impact could ultimately be negative. This reflects a sober assessment from inside the field that safety work carries real structural risks, not just marginal uncertainties.

   Karnofsky emphasizes he is not aware of a good comprehensive list of these downside risks, suggesting this gap in the literature—and his attempt to fill it—may shape how the AI safety community evaluates its own work going forward.

6. 6

   LessWrong post argues that most impactful AI governance work happens invisibly inside government and international institutions, not in public advocacy—suggesting the AI policy community may be investing in the wrong kinds of visibility.

   A post on LessWrong contends that strategic AI governance work splits into two categories: visible outsider work (press, statements, open letters) and invisible insider work within ministerial cabinets and international organizations. The author argues that much of the most impactful work belongs to the latter, largely unseen by the broader community. The post suggests the AI governance community has a bias toward public intellectual work and visibility, potentially underinvesting in the executive and institutional channels where consequential decisions actually happen. This gap may explain hesitations about replicating advocacy models like ControlAI in other countries, where different governance structures apply.

   The author identifies a structural imbalance: public work is not necessarily visible to the audiences that matter most for policy outcomes. This implies that measuring AI governance impact by media footprint or public discourse may be misleading—and that effective advocates may need to operate in channels the wider community cannot easily see or assess.