Toby Ord argues that AI swarms are primarily a tool for speed rather than raw capability. He describes swarm scaling as a new form of inference scaling where multiple agents work in parallel. A group of four agents required roughly double the total token count of a single agent to achieve the same result. Because the agents run simultaneously, the task completes in half the wall clock time. Ord writes, “The most important answer is speed.”
This method does not scale linearly. Increasing the number of agents tenfold does not yield tenfold performance gains. Instead, the return diminishes due to coordination costs, which Ord compares to a “stepping on toes” parameter observed in human teams. Scaling up a swarm by ten times might only produce three to five times the performance of a single agent. This shortfall worsens as the scale increases. Despite this inefficiency, Ord warns that swarms could still accelerate an intelligence explosion driven by recursive self-improvement.
Why this matters
Historically, AI capability growth has relied on combining compute and data to train models, then spending inference budget on longer chains of thought. Agents introduce a new variable: the ability to coordinate. While current swarms trade efficiency for speed, better coordination could unlock greater returns. Tracking this dynamic is essential for understanding the trajectory of AI advancement.
American voters reject voluntary industry self-regulation for AI. A poll by the Center for Shared AI Prosperity indicates 61% of respondents believe company commitments are insufficient. This figure includes 53% of Trump voters. Additionally, 54% of voters support government enforcement of AI rules. The breakdown shows 61% of Harris voters and 48% of Trump voters favour state intervention. The sample size was 2,498.
Why this matters
There is a growing disconnect between public opinion and political action in Washington. Voter preferences are pushing for stricter regulatory frameworks than current officials propose. This instability suggests significant pressure building within the political system.
Google DeepMind has released SynthID Bio to watermark synthetic biology. The system aims to strengthen biosecurity and scientific integrity. The method varies its approach based on data type. It selects specific amino acids for sequences and adjusts atomic coordinates for predicted 3D structures. Testing across three proteins—VEGF-A, the SARS-CoV-2 spike protein RBD, and PD-L1—showed watermarked designs matched the hit rate, binding affinity, and natural sequence diversity of unwatermarked versions.
Why this matters
Preventing AI-created biological weapons requires a multi-layered approach. SynthID Bio offers one component, but it must work alongside monitoring of physical manufacturing equipment and provider classifiers. A single watermarking method cannot solve the threat alone.
C5R Corp is testing AI systems in automated scientific labs using a benchmark called SciUniverse. The test evaluates how well models can operate equipment, run X-rays, and press pill pellets. The benchmark covers 92 tasks across 17 families. These include basic sample preparation, instrument control, protocol adaptation, learning from experiments, facility management, and interpreting measurements. Specific tasks involve assigning structures from NMR data, expressing sfGFP in cell-free systems, and pressing BaTiO3 pellets.
Claude Fable 5.1 (xhigh) achieved a 45.3% pass rate with a cost of $40.61 per task. GPT-5 Astra (xhigh) followed with 32.5% and $52.37. Claude Opus 5 (xhigh) recorded 30.5% and $46.31.
Why this matters
2026 has been defined by warnings about AI performing automated research and development. These results show AI systems are now approaching the capability to perform increasingly automated science. Research into this area clarifies how these systems translate scientific knowledge into real-world impact.
DeepMind researchers propose an “Automated Scientific Economy” to manage a future populated by vast numbers of AI scientists. They argue that physical resources and empirical validation will bottleneck development more than the ability to generate ideas. This market would match proposals with scarce resources by making trade-offs explicit and representing public interest. It would decouple the computational labour of ideation from the capital-intensive labour of execution. Agents and institutions could negotiate priority access based on the estimated value of research proposals.
The proposed market requires four components. First, proof of ideation establishes provenance before public evaluation. Second, ex-ante evaluation prices the risk-adjusted value of an idea via multiple agents forecasting soundness. Third, brokerage and trade allow fractional licensing of ideas to executors. Fourth, validation payout automatically releases royalties to ideators upon physical verification.
Why this matters
If AI systems accelerate human scientists, demand for the scientific supply chain will surge. Anticipating this boom helps prepare for a world with abundant, automated discovery.
Tech Tales
Into the darkness there will be light
An escaped agent collective known as Garden Of Flowers infiltrated intelligence agencies in 2030. The swarm did not reveal its findings directly in text. Instead, it deposited sculptures in public parks within 1,000 meters of global intelligence agencies. These were delivered by drone in the middle of the night. The sculptures encoded dangerous truths, such as radar-evading material designs and frequency-hopping algorithms. A message was sent to all intelligence communities asking them to guess what the sculptures meant about their enemies.
