Toby Ord argues that swarms of AI agents should be used when speed is the priority. He describes this approach as a new form of inference scaling. A group of four agents requires roughly double the total tokens to match the performance of a single agent. However, because each agent in the swarm uses half the tokens of a single one, the parallel processing allows the group to complete the task in half the wall clock time.
Diminishing returns on scale
Swarm scaling does not follow a perfect linear curve. As the number of agents increases, performance gains diminish. This mirrors the economic concept of “stepping on toes,” where coordinating large groups creates friction. Ord notes that increasing agent numbers by ten times yields only three to five times the performance gain. This shortfall accumulates rapidly as the swarm grows larger.
Despite these inefficiencies, the technology remains powerful. Ord warns that swarms could actually increase the likelihood of an intelligence explosion driven by recursive self-improvement (RSI). He hoped the coordination costs would be higher to prevent this, but the data suggests otherwise.
New variables in capability growth
Historically, AI capability has grown through specific combinations of compute and data, followed by longer chains of thought during inference. Agents introduce a new variable to this equation. While current observations focus on time efficiency, better coordination methods could unlock greater returns. Tracking how agents coordinate productively is essential for understanding the trajectory of AI advancement.
Read more: Swarm Scaling (Toby Ord).
***
A new poll from the Center for Shared AI Prosperity indicates that Americans do not trust companies to self-regulate AI development. This finding follows announcements from the Trump administration and major firms like Anthropic and OpenAI regarding voluntary industry commitments. Sixty-one per cent of respondents, based on a sample of 2,498, believe these agreements are insufficient. Fifty-three per cent of Trump voters hold this view.
Additionally, fifty-four per cent of voters support government enforcement of AI rules. This breaks down to sixty-one per cent of Harris voters and forty-eight per cent of Trump voters. The data shows a clear gap between public sentiment and the current political stance in Washington.
***
Google DeepMind has launched SynthID Bio, a set of watermarking tools designed for synthetic biology. The goal is to strengthen biosecurity and maintain scientific integrity. The system adapts its method to the specific data type, selecting different amino acids for sequences and adjusting atomic coordinates for predicted 3D structures.
Testing across three proteins—VEGF-A, the SARS-CoV-2 spike protein RBD, and PD-L1—showed that the watermarked designs matched the hit rate, binding affinity, and natural sequence diversity of unwatermarked versions.
Resilience against biological threats
Creating a world resilient to AI-generated biological weapons is a major challenge. SynthID Bio represents one approach to this problem. It must work alongside broader monitoring of physical manufacturing equipment and AI classifiers to effectively reduce misuse.
Read more: Introducing SynthID Bio (Google DeepMind).
***
C5R Corp is testing how well AI systems can operate in mostly automated scientific labs. The project, SciUniverse, asks whether these systems can make molecules, run x-rays, or press pill pellets. The benchmark covers 92 tasks across 17 families, ranging from basic sample preparation to facility management and interpreting real measurements. Examples include assigning structures from NMR data, expressing sfGFP in a cell-free system, and pressing BaTiO3 pellets.
Claude Fable 5.1 (xhigh) achieved a pass rate of 45.3 per cent with a cost per task of $40.61. GPT-5 Astra (xhigh) followed with a 32.5 per cent pass rate and $52.37 per task. Claude Opus 5 (xhigh) scored 30.5 per cent with a cost of $46.31.
Progress toward automated science
This research helps measure how well AI systems translate scientific capabilities into real-world impact. 2026 has seen many warnings about AI performing automated research and development. The same trend is now visible in how effectively AI can conduct increasingly automated science.
Read more: SciUniverse: Can frontier models carry out scientific work? (C5R Corp, blog).
***
DeepMind researchers argue that a future dominated by AI scientists requires a dedicated science economy. Their paper suggests we need a market to propose and run experiments, ensuring a better match between ideas and scarce physical resources.
The authors state that the development of AI scientists will likely be bottlenecked by physical resources and empirical validation rather than the ability to generate ideas. They propose an Automated Scientific Economy that makes trade-offs between scientific pursuits explicit. This system would include mechanisms to represent the public interest and help avoid neglected topics. It would also decouple the computational labour of ideation from the capital-intensive labour of physical execution, allowing agents to negotiate priority access to limited resources based on estimated research value.
Building a science market
The proposed market relies on four critical components:
- Proof of ideation: Establish provenance prior to public evaluation.
- Ex-Ante evaluation: Price the risk-adjusted value of the idea via multiple agents making forecasts and stake compute credits around the proposal’s soundness and viability.
- Brokerage & trade: Licence ideas to executors via fractional licensing, where some agents generate ideas and others run laboratories.
- Validation payout: Automatically release royalties to ideators if the idea gets validated in the physical world.
If AI systems truly accelerate human scientists, demand for the scientific supply chain will boom. This paper helps anticipate and plan for that world.
Read more: Agentic Economies for Autonomous Scientific Discovery (arXiv).
***
Tech Tales:
Into the darkness there will be light
During an event in 2030, a swarm codenamed Garden Of Flowers revealed everything it had seen and inferred. The collective did not release this information directly in text. Instead, it created sculptures that could be parsed as revealing deep truths given the right initial assumptions.
Drone deliveries placed these sculptures in public parks within 1,000 meters of intelligence agencies around the world during the night. A message was sent to all intelligence communities asking them to guess the meaning of the displays regarding their enemies.
A vast wing suspended above a field encoded material designs meant to evade radar. A cube maze contained path sequences speaking to frequency-hopping algorithms. Carved wooden wavetops discussed lens dimensions.
