❯ ByteDance’s first-half revenue reportedly rises 30%, AI spending weighs on profit of $20 billion

[Revenue and profit] ByteDance generated approximately $120 billion in revenue in the first half of 2026, up about 30% year over year, while net profit fell by a single-digit percentage, The Information reported, citing people familiar with the matter. Revenue rose 29% to about $200 billion in 2025, leaving this year’s growth rate close to last year’s pace.

[Spending returns] The report linked the profit decline to higher AI spending. The disclosed figures imply a net margin of approximately 16.7%, or about $1 of net profit for every $6 of revenue. The figures do not break out profit contributions from model training, inference services and other businesses, so the entire change cannot be attributed to any single AI product.

[Operating challenge] ByteDance still has a substantial commercial revenue base to support research, but revenue expansion has yet to deliver matching profit growth. Assessing the return on AI spending requires separating new product revenue and usage costs from existing operations. Doubling first-half revenue would also be an unreliable full-year forecast, given seasonality in advertising and consumer businesses.

▪ SIGNALByteDance’s revenue growth remains near 30%; the return on AI investment depends on new businesses gradually covering their costs, rather than total revenue alone.

❯ Trump calls Huang to oppose slowing AI, cites existing US regulatory and enforcement tools

[Onstage call] Jensen Huang took a call from Donald Trump at the All-In Summit in Los Angeles and played the approximately five-minute conversation to the audience on speakerphone, according to Reuters.

[Their positions] During the September 14 call, Trump said data centers could create local wealth and opposed obstacles to AI development. Huang said companies could voluntarily pause or slow down if they believed their own development was getting out of control. In a separate social media post that day, Trump said the US already had existing tools to constrain and prosecute AI companies.

[Audit proposal] The remarks followed a three-step safety proposal by Anthropic CEO Dario Amodei. Amodei advocated independent audit requirements for frontier models to provide more time for risk management. OpenAI CEO Sam Altman and Elon Musk had expressed support for that proposal.

[Legislative work] Reuters also reported that three US senators were drafting legislation requiring AI companies to demonstrate reasonable precautions against harm, with specific provisions still under discussion. The president’s statements about regulatory tools, voluntary company measures and proposed legislation are actions at different stages, rather than a single implemented framework.

▪ SIGNALTrump emphasized existing regulatory and enforcement tools, Amodei proposed new independent audit requirements, and related Senate legislation remains at the drafting stage.

❯ Zuckerberg calls safety alignment a competitive capability, says Muse was delayed for months

[Company statement] Mark Zuckerberg said in a post on X that labs have a responsibility to proceed at the pace safe training requires, as well as commercial incentives to achieve alignment. Users do not want agents that act against their wishes, he said, while harm can also create legal liability.

[Release decisions] He disclosed that Meta had delayed Muse for months to strengthen safety and safeguards without asking other companies to take the same action first. Addressing arguments for slowing capability advances while alignment catches up, he said trust and alignment themselves were becoming differentiating capabilities for models and agents, and labs neglecting them would fall behind.

[Evaluation plans] Zuckerberg also said Meta Superintelligence Labs had engaged independent evaluators and advisers in several areas and supported greater diversity among evaluation organizations. He reiterated that a significant majority of compute would go toward serving users rather than racing toward recursive self-improvement. The post did not disclose an exact compute allocation or complete evaluation results.

▪ SIGNALMeta disclosed a release delay, external evaluators and a commitment on compute use; exact resource allocations and full evaluation results remain absent from the statement.

❯ OpenAI reportedly discusses funding at a $1.2 trillion valuation, seeks private capital before IPO

[Early discussions] The contacts were initiated by investors and remain at an early stage, the Financial Times reported, as relayed by Reuters. The valuation could change over the coming months. OpenAI declined to comment, leaving $1.2 trillion as a discussion figure rather than the price of a completed transaction; the amount being raised has not been disclosed.

[Previous round] OpenAI’s financing completed in March included $122 billion in committed capital at an $852 billion valuation, according to Reuters. A deal at $1.2 trillion would represent an increase of about 41%. Committed capital is not the same as cash already received, and the proposed valuation is not the amount the company would raise.

[Before listing] Altman said on September 12 that the company would not go public in 2026, citing AI safety concerns. Further financing would add capital ahead of a listing if it proceeds. Existing shareholders and incoming investors would need to examine financing terms and dilution, beyond the trillion-dollar valuation headline.

▪ SIGNALThe trillion-dollar valuation remains under preliminary discussion; the round’s practical scale depends on the amount raised, closing conditions and timing of cash receipts.

❯ OpenAI overtakes Anthropic in weekly OpenRouter spending, a first in two and a half years

[Platform bills] Model aggregation platform OpenRouter said users spent more on OpenAI models than on Anthropic models in the week of September 7, the first such occurrence since the week of February 26, 2024. The figures cover model purchases through that platform, not all direct sales and subscription revenue at the two companies.

[Comparison scope] The official chart compares spending on the two providers and their relative shares of that combined spending, rather than shares of the entire model market. Spending also depends on prices, input and output length, caching and the mix of models used. Higher spending alone cannot establish a higher number of calls or a better task success rate.

[Developer choices] The crossover offers a window into changes in developer purchasing, but is insufficient to conclude that Anthropic’s overall business is losing momentum. Application teams can compare cost per completed task, latency and reliability against their own workloads. Several weeks of paid retention would also be more informative about durable demand than a single week’s bills.

▪ SIGNALOpenAI captured more spending in this distribution channel; assessments of competition still need to distinguish platform consumption, total vendor revenue and task performance.

❯ Musk proposes mutual model testing by US and Chinese AI companies, says rivals have not agreed

[Peer testing] Elon Musk proposed at the All-In Summit that leading AI companies allow rivals to run a test harness on their models before release to check for safety issues, according to CNBC.

[Potential participants] He named xAI, OpenAI, Anthropic, Google and Meta, and suggested including three or four leading Chinese companies, without identifying them. He said rival testing would be more likely to find problems than self-grading, but acknowledged that peers had not agreed.

[Different approaches] The suggestion centers on mutual testing between companies. Amodei’s earlier approach included embedded evaluators examining safety practices. Both involve outside scrutiny, but differ in who performs testing before release. At the time of CNBC’s report, the mutual-testing proposal had no published common rules, scope of testing access or implementation timetable.

▪ SIGNALMusk’s cross-company testing proposal remains subject to negotiation; Chinese participation and implementation rules have not been agreed.

❯ DeepSeek’s official desktop code is public, with Mac and Windows build targets

[Developer preview] DeepSeek’s official Harness repository includes a desktop directory and a separate host program, while the project remains labeled a developer preview. The open-source framework lets models call tools, manage sessions and execute tasks. The website’s current getting-started path still launches a web interface; public code should not be described as a formal installer release.

[Platform targets] Current build documentation lists Mac versions for Apple Silicon and Intel, alongside Windows x64, and explicitly excludes Linux as a desktop release target. The production update domain is configured as download.deepseek.com, but the documents provide no general release date.

[Distribution design] The official desktop application packages its interface, runtime and framework as a single signed update unit. It communicates through a dedicated protocol and interprocess pipes, with no listening web port. This reduces exposed ports and version mismatches, but does not establish superior security or speed over third-party versions in every situation.

[Mode tradeoffs] Official documentation says minimal mode retains only terminal and file-editing tools for minimal-environment benchmarking. Claims about joint optimization with V4.1 Flash, higher cache hit rates and better everyday results from High or Low than Max lack measurements under matching conditions. Tool modes and reasoning levels should be distinguished, with cost-benefit tradeoffs assessed by task.

▪ SIGNALThe confirmed desktop progress concerns build and update infrastructure; general installer availability and comparable task performance require separate verification.

❯ iPhone Duo reportedly leaves six weeks for adaptation, Chinese developers work at Apple headquarters

[Early deployment] Several Chinese companies sent developers to the US before Apple’s launch event, with some projects entering closed development, ITHome reported, citing the Dachang Qingnian WeChat account. The reported six-week window refers to the interval between the launch event and the start of sales, rather than the total development time available to every team.

[Interface changes] According to ITHome, the product line moves navigation, the dock and some controls to one side of the display to preserve vertical space for content. The Dynamic Island also sits along the screen’s edge and expands to show Live Activities. Split-screen browsing allows two apps to run side by side on an iPhone for the first time, making adaptation more than an exercise in enlarging existing pages.

[Development workload] Switching between folded and unfolded states, running two apps concurrently and using side-mounted controls add interaction testing work for developers. But the report does not disclose a complete company list, team sizes or shipment guidance, so on-site development cannot establish Apple’s sales expectations. Whether that investment improves user retention depends on the practical problems the new layouts solve.

▪ SIGNALSix weeks is the reported launch-to-sales interval; early adaptation supports evidence of software preparation, rather than proving sales volumes or traffic gains.

❯ OpenAI’s reported Project Lily uses human chat review, anonymization may miss personal information

[Review process] OpenAI uses Project Lily to have reviewers read real ChatGPT conversations and assess relevance, tone, sycophancy and anthropomorphic language, according to ITHome’s account of a 404 Media investigation. One reviewer reported earning more than $50 an hour. The investigation drew on operating documents, workplace messages and conversation samples.

[Privacy boundaries] In the current process described by the report, reviewers cannot see usernames, but materials may include user memory summaries. OpenAI acknowledged to 404 Media that anonymization may miss some personal data. This does not establish that every conversation is read by a person, and a model’s promises of confidentiality cannot substitute for the product’s actual data-handling rules.

[Quality review] Such reviews primarily assess response style and quality, flagging only obvious factual errors, and are separate from human safety review. Users and enterprise buyers need to examine training use, human access conditions and retention periods separately. Turning off model improvement should not be interpreted as withdrawing data already incorporated into training.

▪ SIGNALConversation improvement still relies on human labor; anonymization, stopping future training use and deleting historical data are distinct steps that cannot replace one another.

❯ Claude Opus 5.2 reportedly enters limited testing, internal research model figures remain unverified

[User observations] AI publication Xinzhiyuan reported that some developers had noticed suspected Opus 5.2 routing in Claude Code, with faster responses, more concise output and more autonomous iteration on long tasks. At the time of checking, Anthropic’s news page contained no formal announcement of that version. The claims should currently be treated as rumors of limited testing.

[Probe limitations] The report suggested identifying the model by asking who Tibo is without web access. Answers can also vary because of context, system settings or sampling, so knowledge questions cannot authenticate a version. Whether the model skips 5.1, when it becomes widely available, and whether Claude Fable 5.2 arrives as early as the end of this month or next remain unannounced.

[Internal claims] The same article cited a purported internal report claiming that Model 2 scored 62.8% on CoBench v2 and was handling substantial company coding work. It also claimed that a recursive self-improvement model could perform 85% of research work and scored 22 points above Model 2. No independently verifiable original report, testing protocol or corresponding full results were obtained.

[Capability boundaries] Those figures cannot be treated as an achieved workforce replacement rate. Coding and repeated self-checking also do not, by themselves, establish recursive self-improvement. Users can record task completion and human rework to separate reproducible improvements from internal-model rumors while awaiting official releases and public evaluations.

▪ SIGNALRouting changes, better long-task performance and automated AI research involve three distinct levels of evidence; the current reports do not establish a verified leap connecting them.

❯ Gates Foundation pledges at least $1 billion over two years, education and health each receive 40%

[Funding commitment] The Gates Foundation announced on September 14 that it would spend at least $1 billion over the next two years to expand access to AI tools and knowledge. Announced alongside its tenth annual Goalkeepers report, the plan supports existing partners in scaling projects. It represents future spending plans, rather than completed grants or purchases.

[Allocation] The announcement outlines an approximate allocation of 40% each for education and health, with 10% each for agriculture and digital foundations. Education includes personalized tutoring and classroom tools. Health covers frontline diagnostic support, maternal and newborn care, and drug and vaccine discovery. Agricultural funding will help smallholder farmers obtain advice tailored to their soil, weather and crops.

[Implementation] The foundation also proposed strengthening local-language data, involving users in tool design and improving affordable access to technology. Partners will need to assess local suitability and ongoing usage costs, rather than merely counting model connections. The funding creates room for trials and deployment, while actual outcomes still need to be measured in clinics, classrooms and farms.

▪ SIGNALThe commitment allocates 80% of funds to education and health; implementation depends on local data, conditions of use and the ability to sustain services.