Quine, ARIA and a $97M bet: AI moves into the lab, not the fusion reactor

The newest item in this dossier is not a fusion result. It is Quine, an AI research system Microsoft Research announced on 29 September.

The company says it built a multimodal world model of biology and an interactive harness linking models, scientific tools, literature and researchers. In collaboration with the Broad Institute of Harvard and MIT, Microsoft says the system was used to prioritise compounds predicted to drive therapeutic tumour-state shifts. Several top-ranked candidates were validated across multiple wet-lab assays. That is one result. It is not a fusion experiment.

So the honest answer to a query about a fusion energy research result is that this dossier does not contain one. What it contains is a cluster of AI research systems, papers and funding announcements from the past three days, plus energy material that never reaches a reactor. The distinction matters because the two are being reported in the same week and often in the same breath.

What actually shipped this week

Quine is described by Microsoft as experimental research technology, intended only for research and not for clinical or medical use. Microsoft states its outputs may be incomplete or inaccurate and require review by qualified researchers and appropriate experimental validation. A Quine Fellows program will give a cohort of scientists access to the system, and Microsoft says it expects to expand access through products such as Microsoft Discovery as the technology matures.

CoreWeave used the same 29 September window to publish ARIA, a coding agent built into Weights & Biases. ARIA reads experiments, builds live visualisations and runs what CoreWeave calls a full autoresearch loop. It forms a hypothesis, writes the config, launches the experiment through W&B Launch, then evaluates results against a baseline and drafts a report. CoreWeave says the gap between a finished run and the next configured run shrinks from hours to minutes.

Neither company published a physics result. Neither claimed one.

The most consequential number in the dossier is not a performance figure at all. Efficient Computer, led by CEO and co-founder Brandon Lucia, said on 29 September that it raised a $97M Series B round led by TQ Ventures. The company builds processors aimed at energy efficiency, and its argument is architectural: Lucia writes that fixed-function AI accelerators are a devil's bargain, trading programmability for speed on today's workloads, and cites Amdahl's Law to argue that the unaccelerated part of any computation sets the efficiency ceiling.

Where the energy numbers point

Energy is the thread running through the rest of the dossier. CBC reported on 29 September that Newfoundland and Labrador's energy minister, Lloyd Parrott, said the province's door is open for business on AI data centres, after the government was approached by companies. A department spokesperson, Brodie Thomas, confirmed the approaches by email but said commercially sensitive discussions could not be commented on.

The demand side has a counterweight. E2, analysing US Department of Energy employment data, found the US lost 36,949 clean energy jobs in 2025, the first annual decline since the pandemic, with clean energy employment falling to 3.52 million. Electrek reported the figures on 28 September. E2 also recorded 142 clean energy manufacturing, generation and storage projects cancelled or downsized in 2025, and noted the employment data does not establish how many jobs were lost to any particular policy change.

On the supply side, Hannah Ritchie worked through numbers from Oxford professor Nick Eyre showing global final energy demand falling from 416 to 247 exajoules in a post-transition system, even as electricity demand rises from 110 to 189 EJ. Her conclusion is narrow and defensible: electrification itself is the efficiency gain, and the model assumes no other improvements.

None of that is fusion. It is the computing and electricity context that fusion research would eventually have to sit inside.

What the AI-for-research papers claim

The NBER working paper 35782, by Matthew Schwartz, Isaiah Andrews and Jesse M. Shapiro, offers the sharpest numbers on automated research. Across 4,452 published replication packages for five economics journals, the authors' open-source LLM workflow flags discrepancies in 3,460 articles or their appendices. In 496 articles it cuts a calculation's computation time by more than a factor of 10 at similar or greater accuracy, and in 923 articles it develops an extension not present in the original. The paper discloses that LLMs were used in the analysis and writing, and that Schwartz worked as a contractor for Anthropic, whose views are not endorsed.

Two mathematicians push back on the framing. Stephen Wolfram, writing on 28 September, argues that symbolic computation did not end pure mathematics and that AI's greatest use so far is mining the existing mathematical knowledge base. Dan Romik, in a 29 September blog post, describes mathematics research as a feedback loop and argues that AI models trained on human output generate theorems "distance one away" from that corpus, but cannot reflect on and distil what they produced to reach distance two.

Smaller artefacts round out the week. Kompyla, a self-hosted research monitor on GitHub, treats an LLM like a compiler, turning raw documents into cited wiki pages. DotBot publishes firmware for a micro-robot used in education and research. The University of Waterloo's Kaveeshan Thurairaj and Dr Zhao Pan won an Ig Nobel Physics Prize for splash-free urinal designs, work previously published in PNAS Nexus. The dossier also carries a 2008 energy book, an EIA photovoltaic timeline and a plainly labelled satire site.

Read together, they describe a week of tooling and measurement, not a breakthrough in fusion.