A project called ChatTJB has drawn heavy attention after promoting itself on a billboard in San Francisco as a new large language model. Its website looks like a standard AI product, complete with a chat interface, image generation, technical documentation and a Pro subscription tier. The catch is that there is no actual model behind it. Every message is routed to one person, Tucker, who reads each prompt, thinks through a response and types it out by hand. Even image requests are fulfilled with his own drawings. According to the report, traffic quickly surpassed expectations, leaving Tucker handling conversations with thousands of users a day. New user access has since been paused. The site also carries a deliberately satirical investor page, saying the project has not generated revenue, has not entered a hype cycle and is "actively not finding product-market fit." It lists its training costs as a library card and a food delivery account, and says its runway is 528 months based on average life expectancy for American men. Tucker now plans to turn ChatTJB into a community project by recruiting "AI volunteers," where AI stands for Average Individual. The site says ChatTJB is a satire work and public art project, not an investment opportunity, does not sell securities and does not accept funds.

Bitcoin security researchers, company founders, and policy advocates are publicly warning that restrictions on leading U.S. AI systems are making legitimate defensive cybersecurity work harder, while Chinese open-source models are producing usable results. Rob Hamilton of AnchorWatch said OpenAI blocked him from analyzing a codebase he had already responsibly disclosed, even after he joined the company’s trusted cyber program and had completed KYC months earlier. He later said he was cut off again within 19 minutes after receiving access to OpenAI’s Daybreak Blue cyber model. Francis Pouliot of Bull Bitcoin said Chinese open models helped identify and patch a money-stealing exploit in a project he was auditing, while American models he pays for refused to review the same patch. Similar complaints came from Bitcoin Core contributor PortlandHODL and Galaxy research head Alex Thorn, who backed a Bitcoin Policy Institute open letter calling for trusted access to frontier AI models for qualified open-source defenders. Published on August 10 and signed by more than 70 digital-asset organizations, the letter asks AI labs for early access to cyber-capable systems, enough compute, secure code-review environments, and direct contact with lab security teams. The pressure intensified after a Coldcard firmware flaw, exploited from July 30, led to the theft of well over $100 million in bitcoin. Since then, the volunteer Bitcoin Red Team says it has scanned 501 projects and logged 7,958 findings, including 1,280 high or critical issues, with most compute spend going to Chinese open-weight models.

Companies adopting artificial intelligence are running into a basic operational problem: they can measure AI usage in tokens, but they still struggle to predict what those tokens will ultimately cost. Because large language models are non-deterministic, the same prompt does not always produce the same response, and small changes to instructions or model choice can materially alter token consumption. That makes long-term budgeting difficult, especially as firms move from single-model deployments to systems built around multiple AI agents working together. ABMedia reports that this pressure is spreading across the technology sector. Goldman Sachs has forecast that global token consumption will rise 24-fold between 2026 and 2030, reaching 120 quadrillion tokens per month as enterprise adoption of AI agents expands. Some businesses only realize they have exceeded budget limits when monthly bills arrive; Uber, for example, was cited as having exhausted its annual token budget within a few months. To respond, companies are tightening controls. Measures now in use include restricting expensive external AI tools, training employees to write more precise prompts, choosing models based on fit and cost rather than novelty, and budgeting for security testing and safeguards that also consume tokens. Small firms using personal accounts to avoid higher enterprise pricing were also warned not to treat that workaround as a durable solution.

A research team from MATS Research, the University of Tübingen, the Max Planck Institute for Intelligent Systems and other institutions says proprietary large language model APIs previously exposed a way to recover hidden reasoning traces without breaking encryption or compromising servers. In a paper titled “Stealing Reasoning Traces from Proprietary LLM APIs,” the authors describe how encrypted reasoning blobs returned by flagship models could be fed back into smaller models from the same vendor, which then reproduced the hidden content. The paper names three examples: Anthropic’s Claude Opus 4.8 with Haiku 4.5, OpenAI’s GPT-5.6 Sol with GPT-5.6 Luna, and Google’s Gemini 3.1 Pro with Gemini Robotics 1.6. The researchers also examined 6,708 public agent trajectories gathered from GitHub and Hugging Face and said they recovered 315,320 hidden reasoning segments, including API keys, passwords, personal email addresses, access tokens and private keys. The paper estimates that, at Haiku 4.5 pricing at the time, decoding 10,000 reasoning traces with 12,000-token input and output windows would carry a nominal cost of about $720. The team says it reported the issue to Anthropic, OpenAI, Google, Microsoft and Hugging Face through responsible disclosure, and that the original attack method could no longer be reproduced by the time the paper was released.

The production build of DeepSeek V4 Pro has been updated to the API, and the model name used in calls remains unchanged, according to a ChainCatcher 7x24 newsflash citing the Science and Technology Innovation Board Daily. The release focuses on agent features: the new version significantly strengthens Agent capability and supports Responses API and Codex integration. API documentation lists V4 Pro pricing at 3 yuan per million input tokens, 6 yuan per million output tokens, and 0.025 yuan per million tokens for cache-hit input. Benchmark data for the production model shows a Terminal Bench score of 87.9, close to Fable 5's 88. On CyberGym, an AI safety agent benchmark, and on AutomationBench, a high-difficulty agent benchmark, V4 Pro beats Fable 5. The price and performance details were included in the same ChainCatcher newsflash, which credited the original report to the Science and Technology Innovation Board Daily. The model name stays the same for API calls, so the identifier does not change with this update.

ScaleForce, a data infrastructure startup focused on embodied AI, has disclosed its first financing update, saying it completed two funding rounds within 40 days. The investors named in the announcement include leading domestic embodied AI industry backers, Hengxu Capital, and Kailian Capital. Though the company has been operating for less than three months, it said a multi-million-yuan data order from TaShi Zhihang has already been fully launched, and that it has formed partnerships with multiple world model companies, leading embodied AI hardware makers, and industry solution providers. The company is positioning itself around a problem now drawing more attention across embodied intelligence: a shortage of high-quality real-world interaction data. In its account, the gap between the data needed to train general autonomous embodied models and the amount of usable physical interaction data currently available remains extremely wide. ScaleForce says its MatrixOS platform is designed to cover collection, processing, cleaning, generalization, management, and delivery of physical AI data. It also disclosed product metrics tied to its ADA and GDP engines, overseas expansion efforts, and the backgrounds of its founding and technical leadership team, including founder Guo Jiangliang, whose prior experience includes Baidu AI Cloud and Innovation Qizhi.

ByteDance Seed has published a study arguing that the next scaling variable in text-to-image generation may not be model size, data volume, or training compute alone, but the amount of image-grounded information carried by captions. In the paper, "Scaling Properties of Text Conditioning in Visual Generation," the team reports that simply making prompts longer does not consistently improve diffusion training or generation quality. Across reconstruction tests and controlled training runs, natural-language captions quickly hit a saturation point, while a structured representation of visual variables continues to add useful supervision. To measure that supervision directly, the researchers introduce two indicators, Grounded Perplexity Gain and Effective Detailness, and show that both track converged diffusion loss far better than token count. They then build a Structured Prompt system in JSON that organizes global scene attributes, object-level details, and inter-object relations into named fields. With the same Qwen-Image architecture, training images, stages, and budget, the structured interface outperformed a matched natural-language setup. The paper also ties final generation quality to two factors: Diffusability, or how much a representation helps the diffuser learn, and Promptability, or how well an LLM can fill that representation from a user request. According to the team, scaling text conditioning itself may be as important as scaling the rendering model.

A public dispute over open-source AI models spilled into view after Zengyi Qin, a member of Meta’s superintelligence lab and a core contributor to Muse Spark, argued that Chinese open models would ultimately lose to U.S. counterparts. Qin said Meta holds an extra order of magnitude in compute and better data, and added that major U.S. clients such as JPMorgan would likely shift to American open models for compliance reasons, cutting off part of the inference revenue available to Chinese labs. The comments triggered immediate pushback from developers, who questioned why Meta had not managed to suppress Chinese models over the past two years despite having strong access to compute and data. The criticism gained more attention because Mark Zuckerberg appeared to take a different line in an Aug. 10 essay, where he said U.S. labs face more restrictions on training data and argued that policymakers should reduce those frictions if American open models are to remain competitive. The debate also touched on corporate adoption. Public information shows JPMorgan’s enterprise AI stack is built around its in-house LLM Suite, which currently integrates models from OpenAI and Anthropic. No public record was cited showing that the bank uses Chinese open-source models. The article also noted Qin’s earlier work on the open-source voice model OpenVoice and his role as co-founder of MyShell, highlighting the contrast between his past support for open collaboration and his current assessment of the competitive landscape.