Chinese models crowded the leaderboards, OpenAI made an extraordinary claim about mathematical discovery, Europe began enforcing the AI Act, and two very different kinds of leverage failed spectacularly.
Welcome to This Week in AI, B2BNN’s new dispatch from inside the discourse. We are chronically online so you do not have to be. Each week, we will separate the macro movement from the micro drama and the policy decisions likely to outlast both.
The week’s big idea: There is no single AI speed limit
The acceleration and slowdown camps may disagree less about the destination than about the speed limit. Source: Nick Cammarata and Dean Ball on X.
The exchange above captures the American AI policy argument in miniature. Nick Cammarata observes that some people resisting an AI slowdown also expect progress to level off naturally, making a high regulatory speed limit feel harmless. Dean Ball replies that many signatories to the new Pacing the Frontier statement envision only a temporary slowdown, followed by progress at a rate much faster than today’s.
The statement itself recognizes that no company or country can slow unilaterally and calls for a U.S.-supported international effort. The surrounding debate remains strikingly American in its reference points. The revealing phrase is the rate of progress: useful shorthand that increasingly conceals several frontiers moving at different speeds.
The United States can influence the pace of American laboratories through infrastructure, export controls, procurement, regulation and capital. The global rate is emerging from several jurisdictions and several kinds of frontier at once. Alibaba and Moonshot are competing in coding. DeepSeek is attacking price. ByteDance is pushing generated video into something closer to a production system. Europe is deciding what must be disclosed when any of these systems reaches its market.
American centrism remains understandable. The United States still contains the largest concentration of frontier labs, capital and compute, and OpenAI’s new mathematics results may be the most consequential capability claim of the week. It has nevertheless become a poor map of the whole field. OpenAI’s release calendar is no longer the world clock for AI.
The major trend this week was diffusion: of capability, cost leadership, cultural production and regulatory power. There is no longer one frontier, one leader or one speed. And in a reality seemingly antithetical to the American mind, there may only be one competitor in competition.
The Macro
China is now an AI portfolio for national use, not a single challenger
For a while, the Chinese-model story in Western coverage was essentially DeepSeek. That frame is obsolete, as is the competition narrative.
On the August 1 Frontend WebDev Arena lab ranking, Moonshot ranked second, Alibaba third and OpenAI fourth. Z.ai, DeepSeek, ByteDance and Tencent also appeared among the top eleven labs. Six of those eleven positions belonged to Chinese companies. The Text Arena snapshot placed Alibaba’s new Qwen3.8-Max fifth and GPT-5.6 Sol fifteenth, although Qwen’s result was still preliminary.
The competition also spans several economic strategies. Moonshot released the 2.8-trillion-parameter open-weight Kimi K3. DeepSeek’s newly updated V4-Flash pushed the cost frontier down to $0.14 per million input tokens and $0.28 per million output tokens, according to Artificial Analysis figures reported by Reuters. Alibaba is scaling model size and long-horizon agents. ByteDance is advancing both general models and media generation.
“Can China catch up?” now obscures more than it reveals. The relevant question is which Chinese lab is leading which frontier, under which cost and access model.
And this really is the point: China is not competing on leaderboards, on any of these metrics, actually, it’s trying to deliver the best model to service the market operates in. It’s working to better Chinese society, which in the end is the most compelling competitive differentiator there is, and it not one any US AI company is aiming for at all.
Is OpenAI behind or ahead? Yes
The newest public leaderboards show OpenAI behind. GPT-5.6 Sol sat at number 15 in the August 1 Text Arena snapshot. On the WebDev lab ranking, OpenAI trailed Anthropic, Moonshot and Alibaba. The current Artificial Analysis Intelligence Index also places Anthropic’s best models slightly above GPT-5.6 Sol on its broad composite measure.
Then OpenAI announced that an internal version of its next major model, Astra, had produced ten results resolving or substantially advancing long-standing problems in mathematics and theoretical computer science. The set includes an explicit construction of a non-sofic group, a disproof of Connes’s rigidity conjecture and results on three Erdős problems. OpenAI says the discovery runs would have cost roughly $2,000 in total at Sol API prices. Humans prepared the manuscripts with the model, and Astra then formalized the arguments as machine-checkable Lean certificates. OpenAI published the paper, certificates and reasoning walkthroughs.
Independent mathematical assessment will still determine the significance of each result and whether every formal statement fully captures the intended theorem. The release is unusually inspectable, and it is already far more substantial than an unsupported benchmark claim.
So: OpenAI is behind on several new public rankings and ahead on some high-value research and coding evaluations. A leaderboard is a coordinate, not a crown.
Leopold Aschenbrenner discovered that a thesis is not a risk system
Leopold Aschenbrenner’s Situational Awareness fund fell 67 percent in July, sold most of its public-equity portfolio to Citadel and removed all leverage. It also remained up 80 percent for 2026 after its earlier gains, according to the investor letter reported by Reuters.
This was a brutal drawdown, not a funeral. It also does not disprove Aschenbrenner’s larger AI thesis. It demonstrates that a long-term technological thesis cannot protect a concentrated, leveraged portfolio from timing, liquidity and forced selling. Being directionally right about a decade and solvent through a month are separate disciplines.
The Micro
Qwen3.8-Max arrives at the frontier
Alibaba officially launched Qwen3.8-Max on August 3. The sparse mixture-of-experts model contains 2.4 trillion parameters while activating 95 billion for each request, supports a one-million-token context window and accepts multimodal inputs. Alibaba says it ranks fifth in Text Arena and second in Vision Arena. API access is available now, with model weights scheduled for release next week. In one company demonstration, it ran a 16-day autonomous coding project that built and iterated on its own agent framework. Alibaba’s announcement contains the specifications and benchmark links.
The parameter count supplies the headline. The combination of sparse activation, long-horizon execution and a promised weight release supplies the strategic significance.
MiniMax H3 goes open—and immediately gets distribution.
MiniMax has now publicly released the weights for H3, its general-purpose multimodal video modeli, only days after announcing it. H3 accepts text, images, video and audio within one context and generates clips of up to 15 seconds in 2K with native stereo sound. The downloadable model is already available through Hugging Face and ModelScope, while Picsart says it is bringing Hailuo 3 directly to its creators. That combination—high-end video generation, downloadable weights and immediate mass-market distribution—is the larger story. Seedance may be the spectacle release of the week; H3 may be the model that travels. It is another place where American centrism obscures what is happening: Chinese models are no longer merely approaching the frontier. They are pairing competitive performance with openness, lower prices and global distribution channels. MiniMax’s release announcement
Cursor had a dependency-stack week
Cursor’s public status history records repeated degradation between July 27 and July 31 across its IDE, cloud agents, automations and website. Some incidents involved Cursor’s own surfaces; others followed problems affecting models from Anthropic, OpenAI and xAI.
The pattern is more instructive than any one outage. A multi-model AI product inherits the availability problems of its providers and adds its own routing, state, tool and infrastructure risks. As agents become work infrastructure, fallback behaviour, retries and graceful degradation become part of capability itself.
The CoinKite failure reached an estimated $88.6 million
The COLDCARD hardware-wallet crisis began with a software integration error in random-number generation. The affected firmware could select a deterministic software fallback instead of the intended hardware source, sharply reducing the search space protecting wallet seeds. Block’s technical analysis explains how the failure entered the code path, while CoinKite’s advisory identifies affected users and fixed firmware.
Galaxy Research linked three waves of suspicious transactions to the flaw: about 1,367 bitcoin, worth roughly $88.6 million, swept from 4,585 addresses. That attribution remains an evidence-based assessment rather than a final finding. Crucially, installing corrected firmware prevents weak seeds from being generated in the future; it does not repair an existing seed. BleepingComputer summarizes the transaction analysis and the technical findings.
The AI angle remains unproven. Reporting has relayed CoinKite’s suspicion that an attacker may have used AI to find a flaw that an earlier AI-assisted review missed. No public evidence yet establishes that chain. The confirmed story is already serious enough: deterministic scaffolding failed at the exact point where cryptographic security required genuine randomness.
Seedance 2.5 turns clips into production units
ByteDance released Seedance 2.5 on July 31 with up to 30 seconds of synchronized audio-video generation in one pass, multi-round extension, timestamp-level editing and much larger reference sets: as many as 30 images, 10 video clips and 10 audio clips in one generation. The company’s stated goal has shifted from generating an impressive clip to completing a creative work. ByteDance’s release post includes demonstrations and access details.
That distinction matters. Longer continuity, precise revisions and persistent references move generative video toward an editable workflow. They also ensure that the copyright and likeness fights surrounding Seedance 2.0 will follow the technology into a more capable production environment.
The future is old enough for prestige television
Apple released the first teaser for Neuromancer, its 10-episode adaptation of William Gibson’s cyberpunk novel. Callum Turner stars as Case, with Briana Middleton as Molly. The series premieres January 22, 2027. Watch the official teaser.
There is something perfect about ending an AI week with a prestige adaptation of the book that gave popular culture much of its mental architecture for cyberspace, corporate power and disembodied intelligence. The future arrived, became infrastructure and is now getting an Apple TV rollout.
The Policy
The EU AI Act moved from framework to enforcement
On August 2, the European Commission’s AI Office and national authorities began enforcing major parts of the EU AI Act. The new transparency rules require chatbots and other interactive systems to disclose that users are dealing with AI. Deepfakes must be labelled, and generated or altered content must carry machine-readable markings that make it easier to detect. The Commission also opened complaint, whistleblower and downstream-provider channels. The Commission’s enforcement notice sets out what now applies.
The rollout is not quite the “full AI Act” moment described in some headlines. Requirements for Annex III high-risk systems have been delayed until December 2, 2027, while rules for high-risk AI embedded in regulated physical products move to August 2, 2028. Transparency and general-purpose-model enforcement are arriving first.
This is another place where American centrism produces bad analysis. Europe currently trails the United States and China in frontier-model development. It can still determine the conditions under which models, agents and synthetic media reach hundreds of millions of people. Capability leadership and regulatory power belong to different maps.
The Rumours
What the feed believes
“Washington is about to ban Chinese models.” The reporting is real; the ban is not. Officials have reportedly discussed federal procurement restrictions, security advisories, Entity List designations, and liability for American companies hosting advanced Chinese models. No blanket prohibition has been announced. Until an order or agency rule appears, this is direction of travel—not settled policy. Axios has the internal debate; Beijing has already threatened countermeasures.
“Qwen has already won.” Qwen3.8-Max is real, large and apparently competitive. A definitive new hierarchy is not. Launch-day tables always arrive before broad independent replication, and Alibaba’s promised open weights will not land until next week. For now, “Qwen is at the frontier” looks defensible; “Qwen is the frontier leader” remains an open question. Read Qwen’s announcement; watch the independent leaderboards.
“OpenAI has a counterpunch loaded.” Maybe—but there is no publicly announced product release to support the confident version of this claim. What is real is OpenAI’s attempt to move the argument from leaderboard position toward scientific usefulness: mathematical discoveries, longer-horizon reasoning and critiques of how models are evaluated. The rankings are evidence; the imminent comeback is still a rumour. OpenAI’s mathematics results; its ARC-AGI benchmark analysis.
“GLM-5.3 is coming very soon.” This one is not pure feed smoke. A developer working inside Z.AI’s official Java SDK created a glm-5.3 branch and changed a sample request from glm-5.2 to glm-5.3. A subsequent commit bumped the SDK from version 0.3.5 to 0.3.6. That is consistent with release preparation, but it does not establish a launch date. Z.AI’s current API documentation still identifies GLM-5.2 as its newest available model, with no GLM-5.3 announcement, pricing or model card.
The cadence behind the obvious anxiety is real: Kimi K3, Qwen3.8, DeepSeek V4 Flash and MiniMax H3 have arrived in rapid succession, with GLM-5.3 apparently waiting nearby. The comfortable assumption that Chinese labs remain safely behind is becoming harder to sustain. One of the posts driving the discussion.
What’s Coming Up This Week
The most consequential AI signal this week may not come from an AI lab: it may come from the yen.
The carry trade gets a stress test. A rare coordinated American-Japanese intervention sharply strengthened the yen, and both governments have left the door open to further action. The chronically online fear is that a sustained move forces investors to unwind positions funded with cheap yen. If that happens quickly, the selling may not remain inside currency markets; richly valued technology and AI assets could become sources of liquidity. Watch whether this remains an orderly repricing or becomes forced deleveraging. Reuters on the intervention.
The White House decides whether anxiety becomes an instrument. The question is no longer whether Washington is concerned about Chinese models. It is which mechanism—procurement restrictions, security guidance, an Entity List designation or liability for American hosts—might appear first. The deeper American-centrism question is whether losing default technical leadership will be translated into a national-security restriction. Watch for an executive order, Commerce Department action or a coordinated public warning from agencies.
Qwen faces the post-launch test. Qwen3.8-Max is available through an API now, giving independent evaluators a week to test Alibaba’s claims before the promised open-weight release. Watch performance per dollar, coding reliability and whether the model’s position survives testing outside Alibaba’s own harness. This is where a release becomes either a genuine platform shift or one more impressive benchmark table.
Vegas becomes AI’s split screen. Ai4 and Black Hat USA both run August 4–6. On one side of Las Vegas, companies will be selling autonomous agents and enterprise adoption. On the other, security researchers will be trying to break them. Watch for agent demonstrations, prompt-injection research, supply-chain vulnerabilities and evidence that deployment is moving faster than security.
The larger trend is that the AI race appears to be simultaneously a contest over capital flows, market access, national policy, openness and who gets to define the leaderboard. But only one “side” is really competing.
The week in one sentence
OpenAI can make historic mathematical progress while sitting fifteenth in Text Arena. Chinese labs can take six of the top eleven WebDev positions and still be described as challengers. Europe can trail in model development and lead in market rules. An investor can be right about AI and wrong about leverage. A hardware wallet can be air-gapped and fail through one deterministic software path.
The news is no longer one race. It’s everything everywhere, all the time.
