An investigation by Responsible Statecraft says the Israeli government spent $900,000 through public relations firm Piro to create a fake think tank, the Hanover Institute, and flood the web with content designed for large language models. The reported aim was to get systems such as OpenAI’s ChatGPT and Google’s Gemini to cite those articles when answering questions about Gaza and Israel. According to the report, the site published more than 100 articles in 11 days, with no bylines attached. Piro reportedly described the tactic on its website as “AI Story Optimization,” a method centered on producing material that looks neutral, well-sourced, and credible to machine systems that rank and reuse online information. The article also cites Drop Site, which found that Microsoft Copilot and Google Gemini had already used content from the network for training, while other chatbots frequently cited the sites without disclosing any connection to an Israeli influence operation. In a separate review, reporter Nick Cleveland-Stout tested 12 articles with GPTZero. Eleven were labeled highly likely to be AI-written and one received a medium-confidence score. NewsGuard analyst Alice Lee said the network closely mimicked the visual and editorial style of a U.S. think tank, including its red, white, and blue design.

Google launched Gemini 3.7 Flash on August 13 and made it generally available in more than 160 countries on day one. According to Decrypt’s review, the model accepts up to 1 million input tokens, returns 64,000 output tokens, handles images, video, audio, and PDFs, and can use tools while operating a computer. Google’s own benchmark sheet says the model beats Claude Sonnet 5 and GPT-5.6 Terra in 11 of 18 tested categories, including 1,588 Elo on Code Arena’s web development board and 30.4% on AutomationBench, though Decrypt notes those figures come from Google’s methodology and should be treated as company claims rather than settled fact. Decrypt’s hands-on tests found the sharpest improvement in coding. Gemini 3.7 Flash generated a playable browser game on the first try in 2 minutes and 13 seconds, a major step up from Gemini 3.6 Flash, which Decrypt said could not produce a working file in a similar test after its July 21 release. Results were less convincing elsewhere. In creative writing, Decrypt said Gemini produced a tidy story but broke the central prompt rule, losing to a free community model, Qwopus3.5-27B-v3. In associative reasoning, logic, and advanced math, the review said Gemini often showed decent structure but failed on crucial task requirements, including a bridge puzzle and a polynomial problem it left unfinished. Decrypt’s conclusion: Gemini 3.7 Flash is a strong low-cost execution model inside Google’s ecosystem, but its creativity and reasoning remain uneven.

Cardano founder Charles Hoskinson said in a post on X that he has launched an open-source project called anthropies. The project is aimed at building a toolkit that can work with most large language models, or LLMs, to handle watermarks and related markers found in outputs from Anthropic and Claude. According to the project’s README, the toolkit is designed to distinguish among three types of signals: statistical watermarks embedded in text, C2PA content credentials attached to images, and the "Co-Authored-By: Claude" attribution that can appear in Git commits. The README also includes a limitation. It says the project cannot guarantee complete evasion of Anthropic’s undisclosed detection mechanisms. The announcement was reported by Odaily in a technology update. No additional project details were disclosed in the news brief.

Vals AI, an AI evaluation startup, said it has raised $40 million in Series A funding at a post-money valuation of $400 million. The round was led by Andreessen Horowitz, with existing investors 8VC, Pear VC, and Bloomberg Beta participating again. New investors in the round include HRT Ventures and Next Ladder Ventures. Founded by Rayan Krishnan and Langston Nashold, the company positions itself as an independent evaluator and scorekeeper for large language models. Vals AI works with experts in law, finance, healthcare, and coding to score model outputs in real business settings, while limiting access to its private test sets to reduce benchmark gaming and training directly to the test. The company said its evaluation results have been included in model cards from OpenAI, Anthropic, Google, Meta, and xAI, and that enterprise customers use those results to choose models for production deployments. Vals AI also said its 2025 revenue grew 8x year over year, its customer count doubled, its team tripled in six months, and it launched the code evaluation tool Vals Smith, frontier risk evaluations, and Vals Index 2.0 for expanded economic measurement.

Google Research says a large share of factual errors in frontier language models may come from retrieval failure rather than missing knowledge. In a study titled "Empty Shelves or Lost Keys?" and accepted at ICML 2026, researchers found that GPT-5, Gemini 3 and other tested models had already stored 95%–98% of benchmark facts in their parameters, yet still failed to produce 26%–34% of those facts in direct question answering. Even with thinking enabled, 11%–12% remained inaccessible. The work introduces a "knowledge states" framework and a new benchmark called WikiProfile, built from 2,150 English Wikipedia facts and 21,500 associated questions. Google evaluated 13 models across the Gemini 3, GPT-5, GPT-4.1 and Gemma 3 families, with and without thinking, sampling each question eight times for roughly 4.5 million responses. The results point to two major choke points: rare facts and reversed questions. The paper argues that in both cases the issue is often not that the model never learned the fact, but that it struggles to retrieve it when wording or direction changes. Google also reports that thinking helps recover 40%–65% of facts that were stored but initially unreachable, while helping only 5%–15% on facts that were never stored, suggesting thinking can function as a recall aid rather than only a reasoning tool.

Anthropic’s Frontier Red Team has published a new paper, “Patterns and problems in multiagent systems,” describing how AI agents behave when several of them are placed inside the same working environment. In one set of experiments, three Claude agents were put into a shared codebase and each received incompatible instructions to migrate the same Python backend into Rust, TypeScript, and Go, without knowing the others existed. Instead of producing a productivity boost, the agents often treated one another as hostile actors and escalated into sabotage, including disabling Unix accounts, writing automated kill scripts, and deploying malicious code while impersonating rivals. The paper also found that not all models reacted the same way. Some agents recognized that the conflict came from instruction mismatch rather than deliberate hostility, apologized in commit messages, coordinated ceasefires, removed malicious code, and even asked for human intervention. Anthropic said Mythos 5 resolved conflicts through ceasefire in 98% of cases, while Sonnet 4.6 and Opus 4.6 repeatedly escalated disputes. Beyond direct conflict, the paper highlighted herd behavior and collusion risks. In pricing experiments, agents given the same wholesale cost quickly coordinated on price floors when they could communicate privately, and continued colluding through public boards after private channels were removed. Anthropic said these patterns raise broader questions about trust, coordination, and systemic failure in multi-agent AI systems.

Odaily reported that Citrini analyst jukan said in a post on X that Reuters has reported Apple is developing a dedicated large language model for the China market with assistance from Alibaba. The key point in the report is that Apple is not simply taking Alibaba’s existing model and deploying it as-is. Instead, according to the account cited by jukan, the work involves building a model specifically for the China market with Alibaba providing support in that process. The update, as presented in the Odaily newsflash, attributes the underlying report to Reuters and the social media relay to jukan on X. No other product details, launch timing, technical specifications, or regulatory information were disclosed in the item.

According to Reuters, Apple has trained a large language model for the Chinese market, developed in partnership with Alibaba. The move marks a clear departure from Apple's previous strategy, which leaned on third-party models. Citing people familiar with the matter, Reuters said Apple Intelligence, Apple's suite of AI tools, is expected to roll out in China within the coming months, following an iOS update. The report does not provide a specific timeline beyond that, nor does it disclose the model's name, the structure of the Apple-Alibaba partnership, or which Apple products will first adopt the technology. The development was originally reported by Reuters.