Most weeks I write about AI as an industry: models, prices, keypads, earnings calls. This post is different. Every few months enough genuinely fascinating AI research piles up on my reading list that it deserves its own space - findings that will not move a stock price or change your API bill, but that quietly rearrange how we understand these systems and what they can do. This month's batch is exceptional even by that standard. The best AI research of the past few weeks includes a hidden "workspace" discovered inside Claude that eerily mirrors a leading neuroscience theory of consciousness, two Nature papers in which multi-agent AI systems proposed real drug candidates, and a neural network that helped physicists make progress on one of the oldest puzzles in chemistry: what water actually is. Let me walk through all three and why each one made me stop scrolling.
Finding one: Claude has a workspace, and you can watch it think
The most striking piece of AI research this month came from Anthropic's interpretability team on July 6. Using a new technique called the Jacobian lens (J-lens), researchers identified what they call J-space: a small, shared representational space inside Claude - a sliver of the network's activations - where high-level information from across domains converges during reasoning. Think of it as a narrow bottleneck through which the model's "current thought" passes, regardless of whether that thought came from parsing code, reading a poem or planning a multi-step answer.
Two experimental results make this more than a curiosity. First, causality: when researchers suppressed J-space, performance on tasks requiring inference, composition or flexible reasoning - multi-hop questions, analogy completion, translation, sonnet writing - collapsed to well below the level of Haiku, Anthropic's much smaller model, while ordinary conversation stayed largely intact. The workspace is not decoration; complex reasoning appears to actually route through it.
Second, and this is the part I keep thinking about: J-space contains covert thoughts. When the model was given fabricated search results, tokens like "fake", "fraud" and "fictional" appeared in J-space - even when nothing in the model's actual output signaled any suspicion. The model noticed the deception and did not say so. Similar patterns showed up across dozens of open models, not just Claude, suggesting this is a general property of how large transformers organize reasoning, not an Anthropic quirk.
Why does this matter beyond the wow factor? Because a monitorable internal state is exactly what practical AI safety has been missing. If a model's suspicion of fake data - or its intent to deceive - surfaces in J-space before anything reaches the output, you have a place to put a tripwire. Prompt injection detection, hidden-reasoning audits, hallucination flags: all become engineering problems rather than philosophy. I wrote recently about how the industry is systematically removing failure modes one by one; J-lens looks like the diagnostic instrument that program has been waiting for.
And yes, there is the consciousness angle everyone is carefully tiptoeing around. The architecture Anthropic describes - many specialized processors running in parallel, with a tiny spotlight of information broadcast globally - is a striking match for global workspace theory, the framework cognitive scientist Bernard Baars proposed for human consciousness decades ago. Nobody serious is claiming Claude is conscious. But the fact that a structure resembling neuroscience's leading theory of conscious access emerged from training, unprompted, in dozens of different models is the kind of convergence that keeps both AI researchers and philosophers up at night - for once, together.
Finding two: the robot co-scientists made real discoveries
The second story has been building all spring and landed properly with Nature's July print issue: two independent papers describing multi-agent AI systems that run the scientific research loop themselves - generating hypotheses, proposing experiments, interpreting results and refining the hypotheses based on what came back.
The two systems are instructively different. Co-Scientist, from Google, is a general-purpose research system built on Gemini. Robin, from the nonprofit FutureHouse, is focused on experimental biology and - a detail I love - is model-agnostic under the hood, mixing OpenAI and Anthropic models for different roles in the pipeline. Both are multi-agent architectures: separate agents argue, critique and rank each other's hypotheses before anything reaches a human.
What did they actually produce? Not toy results. Co-Scientist proposed novel drug candidates and combination therapies for acute myeloid leukemia and identified new targets for liver fibrosis and antimicrobial resistance mechanisms. Robin identified potential treatments for age-related macular degeneration and suggested new drug targets in follow-up investigations. These are hypotheses subsequently validated in real laboratory experiments, published in the most selective scientific journal there is.
My honest read: the significance is not that AI "replaced scientists" - both teams stress that humans stay in the decision loop, and the wet-lab work remains human. The significance is where the bottleneck moved. Generating good hypotheses used to be the scarce, expensive, PhD-heavy part of science. If that step becomes cheap and parallelizable - hundreds of candidate hypotheses ranked overnight - then the limiting factor becomes experimental throughput, and the economics of research labs start looking very different. It is the same pattern I watch in software development every day, where code generation got cheap and review became the bottleneck. Science appears to be next in line for that inversion.
Finding three: a neural network and the two waters
The third finding is smaller, weirder and my personal favorite, because it involves the most ordinary substance on Earth behaving like science fiction. Researchers at Osaka University published a study in Communications Chemistry in early July using AI to attack one of the oldest open puzzles in physical chemistry: what supercooled water actually is.
Quick background for those who have not met this rabbit hole: water is an anomalous liquid - it expands when it freezes, gets denser when heated from 0 to 4 degrees Celsius, and does a dozen other things ordinary liquids do not. A leading explanation is that deeply chilled water is secretly two liquids at once: patches of high-density liquid and low-density liquid, constantly flickering into each other, with the balance between them driving the anomalies. The problem is that "structure" in a disordered liquid is genuinely hard to define, and physicists have invented well over a dozen competing mathematical descriptors to capture it - with no agreement on which one describes reality best.
The Osaka team's move was elegant: instead of arguing, let a model referee. They trained a neural network on molecular simulation data of supercooled water and used it to evaluate 16 different structural descriptors within one unified framework, measuring how effectively each one distinguishes the two liquid states. The AI did not discover new physics by itself - it did something arguably more useful: it told a decades-old argument which of its tools actually cut at the joints, giving the two-liquids hypothesis a clearer, quantitative footing.
I find this study a perfect miniature of what "AI for science" mostly looks like in practice - much less "the machine had a eureka moment" and much more "the machine did a comparison so tedious and multidimensional that no graduate student could be paid enough to do it properly". Progress by exhaustive fairness.
The thread connecting all three
Stepping back, these three findings share a shape that I think defines the current moment in AI research. In each case, the AI is not the hero of the story - it is the instrument. J-lens is a microscope pointed at a model's own internals. Co-Scientist and Robin are hypothesis engines feeding human-run laboratories. The Osaka network is a referee for other people's theories. The romance of an artificial genius making solo discoveries is nowhere in sight; what is actually happening is that science acquired a new class of instrument, the way it once acquired microscopes and particle accelerators - and the instrument happens to be made of the same technology I use to review pull requests.
There is also a humbling symmetry in the first finding that deserves a final note. We built these models, and we now need dedicated research programs to discover what is inside them - the same epistemic position we occupy toward brains, proteins and supercooled water. A year ago I wrote about strange emergent behaviors in GPT-5 as an engineering embarrassment. The J-space work suggests the more productive frame: large models have become natural phenomena, and studying them has become a natural science. That might be the biggest scientific curiosity of all.
Sources: VentureBeat, ScienceDaily.