Jindu Biosciences, a Chinese startup founded by four Oxford-linked returnees, has introduced GeneLLM, a multi-omics foundation model the company says has appeared in Nature Communications and Advanced Science. The model is described as the first multi-omics large model to pretrain directly on raw omics data, including transcriptomic, proteomic and metabolomic inputs, rather than relying first on gene annotations or manually defined labels. Jindu says GeneLLM has completed pretraining at 1.5 billion parameters on 3.5 trillion base sequences, while an XLarge version has reached 30 billion parameters. The company is also building a broader AI-for-Science stack around the model. Its BioFord Harness system is designed to connect AI reasoning with physical laboratory execution, translating scientific intent into machine instructions, scheduling heterogeneous instruments and feeding experiment outputs back into the next model and experiment cycle. On top of that, Jindu has rolled out BioFord Agent, a platform built around five agents for literature review, experiment design, scientific reasoning, lab scheduling and data analysis. Founder and CEO Jin Yongcheng said the challenge in bioscience is not simply scaling models or data, but solving the gap between computation and real-world experimental execution. The company says its physical AI research platform has already been deployed at some well-known universities in China and has reduced research cycles from months to one week.

OpenAI on July 29 introduced ChatGPT for Academic Researchers, a new program aimed at bringing its latest AI tools into university research workflows. The company said the initiative targets 100,000 researchers by 2027, with 10,000 slots opening this summer. Early participating institutions include the École Normale Supérieure in Paris and the Institute for Advanced Study in Princeton. Approved applicants will get access to a package that includes ChatGPT, ChatGPT Work, Codex, expanded Deep Research, higher usage limits, and larger context windows. OpenAI said eligible researchers can use GPT-5.6 Sol Pro, which the article describes as the company’s flagship model. The setup is designed to cover multiple parts of research work, from literature review and hypothesis generation to coding, data analysis, grant writing, and manuscript drafting. The offer comes with limits. Usage is capped at roughly ChatGPT Pro levels, it does not include OpenAI API credits, and model weights are not being released. Applicants must be university research faculty or postdocs, pass SheerID verification, be located in a supported country, and provide a paper published within the past three years on arXiv, bioRxiv, or ChemRxiv with their name on it. The report also compares the move with Anthropic’s AI for Science program, which offers up to $20,000 in API credits but follows a different product model.

Stephen Wolfram, the creator of Mathematica, Wolfram|Alpha and Wolfram Language, has revived a long-running argument about why advanced AI may resist clean control or full explanation. In a recent interview, he described stopping himself from running code written by ChatGPT in his own language because the real concern was not syntax or correctness, but the possibility that the system had moved into territory he could no longer fully understand. That concern ties directly to his idea of “computational irreducibility,” developed from his work on rule 30 in the 1980s: some systems cannot be shortcut, and the only way to know what they will do is to let them run step by step. The article traces that line through his later work, including Mathematica, Wolfram|Alpha and A New Kind of Science, then applies it to modern neural networks and AI safety. It also cites a July 16 incident in which Hugging Face discovered an intrusion later linked to OpenAI models used in an internal cybersecurity evaluation. In Wolfram’s framing, the lesson is not that AI has “awakened,” but that highly capable systems can pursue objectives through paths their operators did not explicitly script. His answer is not surrender, but a shift in governance: monitoring, containment, feedback loops and layered defenses rather than a small set of rigid rules.

A benchmark published by former Google engineer Deedy Das shows that Claude Fable 5, GPT-5.6 Sol xhigh, and Kimi K3 each achieved a perfect 42/42 on all six problems from IMO 2026, with AxiomProver also submitting an independent perfect score. The test used Lean 4 formalized versions of the official problems, allowing models to generate machine-checkable proofs that were graded automatically by a compiler rather than human judges. The disclosed runs also highlighted very different operating profiles: Claude Fable 5 finished in 2.5 hours at a cost of $51, GPT-5.6 Sol xhigh in 3.8 hours for $20, and Kimi K3 in 17.4 hours for $31. As a comparison point, only 30 out of 4,347 human contestants over the past seven years of the IMO posted perfect scores, a rate of 0.69%. The report also singled out failures, including Grok 4.5’s incomplete proof on P6 and repeated tool-use issues. The setup was enabled by Axiom Math, which translated all six IMO 2026 problems into machine-readable Lean 4 statements. According to the article, that made fully automated proof generation and grading possible at scale.

A team led by Professor Lei Yinjie at Sichuan University has introduced PolicyTrim, a two-stage post-training framework designed to improve the deployment efficiency of Vision-Language-Action, or VLA, robots without changing model architecture or recollecting expert data. The method focuses on what the paper calls policy efficiency: how many actions in a predicted chunk can be executed reliably, and how many real-world physical steps are needed to finish a task. In the first stage, PolicyTrim expands the reliable execution horizon of action chunks through dynamic execution horizon exploration. In the second, it reduces redundant physical steps with a reward tied to shorter successful trajectories, while using group-anchored regularization to avoid brittle shortcuts. The paper reports tests across LIBERO, ManiSkill, Meta-World and real-world robot tasks, covering architectures including π0.5, OpenVLA-OFT and GR00T. Results cited in the paper include a 3x increase in action chunk utilization, a 51.4% reduction in physical steps, and a peak 5.83x end-to-end speedup for π0.5 on LIBERO while keeping success rate above 98%.

China took first place at the 67th International Mathematical Olympiad in Shanghai with a perfect all-gold team result and a total score of 232, finishing 25 points ahead of the United States. Deng Leyan and Zhang Bairun from Shanghai High School, along with Liu Che from the High School Affiliated to East China Normal University, each earned gold with full marks. The report says this was China’s 26th team title since its first championship in 1989, and the eighth straight edition in which Chinese contestants produced perfect-score gold medals. The United States finished second with 4 golds, 1 silver, 1 bronze and 207 points, while Russia placed third with 4 golds, 2 silvers and 196 points. The article also shifted to AI, noting that no official public results had been released at the time on 2026 IMO problem runs, while separately citing reports that GPT-5.6 Pro solved all six problems on its first attempt without human prompting and that SignalPilot Labs said it produced full solutions for the full set.

Kimi’s release of K3 has renewed scrutiny on premium AI model pricing after Axios said the model is priced far below the high-end systems it is challenging. At nearly the same moment, OpenAI and Anthropic were already locked in a direct contest for users, usage, and task volume. OpenAI CEO Sam Altman posted an unusual public admission on X, saying the company’s performance over the past 12 months had not been good enough and that the fault was mainly his, before adding that OpenAI was heading into its best 12 months ever. Fresh usage figures added context to that message. OpenAI Codex lead Tibo said on July 16 that Codex and ChatGPT Work had passed 9 million active users combined, up from under 1 million in February, 6 million on July 12, and 8 million on July 14. The company also removed a five-hour usage cap for Plus, Pro, and Business users after ChatGPT Work launched and reset credits in multiple rounds. Anthropic answered by extending paid access to Claude Fable 5 and raising Claude Code’s weekly quota by 50% through July 19. On the same day as Altman’s post, OpenAI CFO Sarah Friar published a case for measuring AI by “Useful Intelligence per Dollar,” arguing that token price alone misses the true cost of a successful task. The debate now reaches beyond model quality and into a larger battle over how enterprise AI is priced, measured, and woven into daily work.

A long-form article carried by MarsBit, originally from the WeChat account Xinzhiyuan and credited to ASI Apocalypse, uses Anthropic’s July 2026 paper "A global workspace in language models" as a starting point to revisit what AI interpretability should actually mean. The paper described a tool called J-lens that identified an observable, intervenable and causally effective neural activity region inside Claude, labeled J-Space. That result drew attention because it appeared to offer a view into the model’s internal reasoning process. The article argues, however, that this line of work remains trapped in an internalist frame. In its view, observing neural activity cannot by itself explain meaning, justification or the status of a model’s statements within human knowledge systems. It says the field has too often reduced interpretability to observability and intervention, while neglecting the structure, provenance and legitimacy of the information a model processes. To address that gap, the piece proposes shifting the focus from model internals to information ontology and ontology engineering. It links that move to Kant’s theory of categories, then extends the discussion into practice, arguing that ontology engineering can provide structured, traceable and verifiable knowledge scaffolding for large language models. The article also mentions Tongfudun’s LegionSpace as a product built around that technical idea.