Google’s Antigravity team has outlined a multi-agent orchestration framework called Teamwork and tied it to a set of results across theoretical research, systems engineering and open-source software. In the company’s account, seven mathematics and theoretical computer science findings were first produced by Gemini 3.1 Pro in Teamwork’s Long Proof mode, while three of those results were later fully reproduced by Gemini 3.7 Flash. Google framed that as the first time a Flash-tier model produced PhD-level mathematical research output. The post also described a systems project built with Gemini 3.7 Flash: a cycle-level, out-of-order RISC-V CPU simulator that booted xv6 into a shell and posted an average cycle error of 0.71% against BOOM hardware on unseen workloads after running more than 100 standard RISC-V benchmarks. On the software side, Google said Teamwork-generated optimizations for Eigen and ParlayHash were accepted by external maintainers and merged upstream. A central claim in the write-up is that the gain came less from scaling model size and more from how multiple agents were organized, challenged and validated during long-running work.

Elon Musk amplified a view from ARK Invest chief futurist Brett Winton, writing that 「The AI riptide is already underway.」 The discussion centers on a sharp rise in AI demand, with the article saying global AI inference token usage has increased about 25-fold over the past year and OpenRouter token volume is doubling roughly every 11 weeks. Cathie Wood argued that token consumption is growing exponentially and spreading across the economy, while frontier AI labs are seeing annualized revenue rise 5x to 10x within six months to one year. The article’s main claim is that AI infrastructure now offers unusually short payback periods and very high internal rates of return, drawing capital toward GPUs, data centers, and AI companies. It says this shift is not limited to venture funding. Debt markets and talent allocation are also being reshaped as builders of AI capacity can tolerate higher financing costs because compute remains scarce and profitable once deployed. The piece also states that Nvidia chips have become a form of collateral for debt tied to data center buildouts. According to the article, the pressure on traditional companies could show up in four ways: higher financing costs, weaker valuations and liquidity for non-AI stocks, more expensive debt refinancing, and a migration of talent toward AI-linked sectors. Musk replied to the broader argument with a short endorsement: 「You are right.」

OpenAI CEO Sam Altman said in a newly released podcast teaser that Astra’s ability to operate a computer "feels fully at human level," adding another data point to the race around advanced AI agents. The clip, which runs a little over a minute, features Altman answering a question about whether AGI has already arrived. His response was that it is close, and that Astra’s computer-use capability already feels comparable to a human’s. The report also points to fresh signs around Anthropic’s next model. A developer said Fable 5.1 has shown up in Amazon Web Services’ Bedrock API routing. According to the cited observation, requests that previously returned a 400 invalid-model error now return a 404 model-not-found response, which the article interprets as a sign that the identifier is recognized by AWS’ backend even if the model itself is not yet available. The source piece also reviews Astra demos and community testing of Fable 5 versus GPT 5.6 Sol, describing Astra as geared toward persistent agents rather than chatbot-style interactions. It cites examples including one-shot generation of a complete 3D world from a single prompt, a 56K-token reasoning run that produced a voxel pagoda in 38 minutes, and claims of long memory, multi-agent collaboration, and cross-tool desktop operation.

Anthropic is facing a proposed class action over its $200-a-month Claude Max 20x plan, after developer Karl Kahn alleged that the service delivered far less than the advertised 20x usage of the Pro plan. Kahn, who filed suit in the U.S. District Court for the Northern District of California on June 14, said he moved most of his work to Claude Code in April and found that a single five-hour coding session could consume 15% of his weekly quota. The dispute centers on how Anthropic applies limits across paid tiers. According to the report, Claude subscriptions are constrained by both a five-hour rolling window and a separate weekly cap, with either one able to stop usage first. Reference figures cited from Anthropic’s July 2025 email to subscribers and comments to TechCrunch show that the top-end Sonnet 4 allowance on Max 20x is 480 hours per week, versus 80 hours on the $20 Pro tier — a 6x difference, not 20x. The article also points to a broader transparency problem around token accounting and bug-related quota drain, while contrasting Anthropic’s approach with OpenAI Codex’s published weekly-limit framing and later quota reset after similar bugs were identified.

Lauren Tan, the engineer leading GrokBot at SpaceXAI, said she now runs more than 20 agents at once and saw them deliver more than 1,000 pull requests last month, with a target to double that figure in August. In a team workshop, Tan showed a GitHub contribution curve covering five months and more than 3,000 PRs, arguing that the jump did not come from blind faith in bigger models but from stricter verification and tighter repository constraints. According to Tan, the workflow can now cover the full chain from opening a PR and running validation to merging into main, often without her touching the code directly. She said the setup relies on multiple GrokBot instances running inside her open-source pstack system, along with commands such as /loop, /goal, and /swarm. She also described key components behind the process, including a control system called control glass, a feature map that tells agents how product surfaces are organized, and an earlier automated bug-handling agent named Benny. Tan said the larger shift is in how engineering work is organized. In her view, the value of human engineers is moving away from writing every line by hand and toward setting goals, defining permissions, and deciding what counts as correct code. She compared her role to that of a head chef: not cooking every dish, but designing the kitchen, assigning stations, and making sure the system holds.

A new case from physics is adding to the growing debate over AI’s role in frontier research. According to the source material, after a summer marked by AI systems tackling major mathematical problems, physicist Gavin E. Crooks handed Claude an open problem in stochastic thermodynamics and received a full solution within days. The reported result did not just answer one narrow question. It unified a range of previously published Detailed Fluctuation Theorem, or DFT, bounds by identifying them as low-dimensional projections of the same convex geometric object and by giving a complete hierarchy of moment constraints. The account places special weight on who posed the problem. Crooks is a prominent figure in nonequilibrium thermodynamics and statistical mechanics, known for the Crooks fluctuation theorem introduced in 1998–1999 while he was a graduate student at Berkeley. The source also notes his PECASE honor and his 2019 election as an APS Fellow. What stands out here is the process. Claude was reportedly used not just to explain a known concept, but to explore a difficult theoretical question, connect earlier ideas, and search for new mathematical structure. That shift, if sustained, could matter beyond this single thermodynamics result: it suggests AI is moving from handling problems humans already know how to solve to helping scientists investigate problems they do not yet know how to solve.

A research team led by Anima Anandkumar at the California Institute of Technology published a paper on Aug. 24 describing an AI-assisted approach to density functional theory, or DFT, that shifts a key computational step from cubic scaling to near-linear behavior. The work targets one of the longest-running bottlenecks in quantum chemistry, where simulation costs rise sharply as molecular or material systems grow. Instead of asking a model to predict the final answer in one shot, the team trained a Fourier Neural Operator, or FNO, to learn the forward map from input conditions to electron density, then inserted that model back into the standard self-consistent DFT loop. The paper reports complexity dropping from O(N^3) to O(N log N) for that step. In tests on drug-scale molecules outside the training set, the Kohn-Sham FNO posted a density error of 2.23%, versus 9.97% for a direct-prediction baseline; at 45 heavy atoms, the gap widened to 4% versus 41%. For large-scale validation, the researchers used magnesium dislocation structures. The paper says the new method converged on a system with 8,250 atoms and 82,500 valence electrons using a single NVIDIA B300 GPU. It contrasts that with a 2019 full DFT calculation on 6,164 magnesium atoms that used about 7,800 NVIDIA V100 GPUs on the Summit supercomputer.

A 108-page proof by Harvard mathematician Levent Alpöge and Claude claims to answer one of the longest-running questions in differential and complex geometry: whether the six-dimensional sphere, S⁶, admits a complex structure. The result, described in the source article as a direct construction rather than another attempt along older lines, builds a compact complex threefold X and then identifies it with S⁶ through calculations of the fundamental group, homology arguments, and a chain of classical theorems. The article says mathematician Qiaochu Yuan had GPT-5.6 Sol examine the manuscript. After an initial six-minute pass and another 15 minutes of review, the model reportedly found no flaw. Sol’s view, as cited in the piece, is that if the proof holds up, it may stand as the most important AI-assisted mathematical result so far. The report places this work alongside two earlier Claude-linked advances from July 20 and Aug. 10, and argues that the S⁶ case is different because the geometric object was constructed outright rather than found through search or recombination.