Russian startup Mostik has developed a method allowing AI models to exchange information through their internal weight values rather than generated text, effectively letting a smaller model absorb capabilities from a larger one. The team demonstrated this by linking a 753-billion-parameter GLM-5.2 model with a 4-billion-parameter Qwen-3.5 model, producing a hybrid system that runs at one-twentieth the cost of the full-size model while performing roughly midway between the two in capability. Mostik also used a related technique to build a model that has topped the ARC-AGI 3 benchmark, though details remain undisclosed while the contest is ongoing.
wired.com
· 2026-09-02
An Entrepreneur opinion piece argues that generative AI tools have become the primary way customers, investors, employees and partners form first impressions of companies, synthesizing news, reviews, social media and other online signals into a single confident narrative. The author warns that old controversies can linger in AI-generated summaries long after public attention fades, quietly shaping perception without a company's input.
entrepreneur.com
· 2026-09-02
Protect Democracy, a nonpartisan nonprofit, has filed a lawsuit against four federal agencies demanding disclosure of the confidential process the Trump administration uses to vet frontier AI models before release. The group says virtually no details have been shared publicly or with Congress about which companies participate, how they are chosen, or what legal authority underpins the reviews, and alleges OpenAI has struck a private deal limiting access to its top models to government-approved partners. The lawsuit seeks a court order compelling release of unclassified records by September 30 and an injunction against further withholding.
arstechnica.com
· 2026-09-02
The Trump administration filed a 20-page brief supporting OpenAI in its copyright dispute with The New York Times, arguing that training large language models on copyrighted material qualifies as fair use. The filing cites an executive order calling for the US to maintain global leadership in artificial intelligence, and warns that restricting AI training practices could harm American innovation and economic competitiveness.
techcrunch.com
· 2026-09-02
A professional has published a personal AI usage policy stating they do not currently use generative AI or large language models in their code, design, or correspondence. The policy, first drafted in mid-2025 and updated in May 2026, says they will avoid AI wherever practical but will still collaborate with clients and colleagues who use it, provided they are not held accountable for AI-generated portions of joint work.
dbushell.com
· 2026-09-02
Google published documentation outlining the intended uses and known limitations of Gemini 3.8 Flash, a model aimed at cost-effective, production-scale agent deployment for developers and enterprises. The company flagged persistent issues including hallucinations, occasional slowness or timeouts, and inconsistent knowledge freshness across domains despite a stated March 2026 cutoff.
deepmind.google
· 2026-09-02
An audit of Perplexity's Sonar and Sonar-Pro models tested 310 factual questions about 210 tech companies and checked every citation attached to a sentence containing a number. It found that 34.7% of 1,826 such citations linked to pages that either wouldn't load or didn't contain any of the figures stated in the sentence, though scoring by claim rather than by individual citation marker lowered the failure rate to 14.4% of 872 claims.
hausresearch.com
· 2026-09-02
The Trump administration filed a statement of interest in The New York Times' copyright lawsuit against OpenAI and Microsoft, siding with OpenAI's claim that training AI models on copyrighted text qualifies as fair use. Government attorneys warned that narrowing fair-use doctrine could hinder scientific and economic progress tied to large language models.
theverge.com
· 2026-09-02
OpenAI announced that its forthcoming Astra AI model is the first to cross the company's 'Critical' cybersecurity capability threshold, meaning it can discover unknown vulnerabilities and exploit them autonomously without human step-by-step direction. The company says Astra will still launch soon, but access to its most sensitive cybersecurity abilities will be restricted, with further safety details to come in a system card at release.
cnbc.com
· 2026-09-01
METR, a nonprofit that evaluates risks in frontier AI models, revealed it suffered two cybersecurity incidents this year. In March, attackers stole an API key used for public-model inference and used it for weeks to run up a large number of credits, while in May attackers unsuccessfully probed an exposed endpoint attempting to reach internal data.
darkreading.com
· 2026-09-01
A new study from Google Research and Technion shows that when models like GPT-5 and Gemini-3 hallucinate, the underlying fact is usually already stored in their parameters rather than missing entirely. The researchers found these models encode 95-98% of tested facts, but often fail to surface them during generation due to retrieval, not storage, limits. They propose 'fact-level profiling' to separate whether a fact is encoded versus reliably retrievable across different phrasings.
venturebeat.com
· 2026-09-01
Google is reportedly courting major studios like Disney, Warner Bros. Discovery, and Universal, offering large payments—potentially tens of millions per character and billions overall—for rights to train its AI models on their copyrighted content. According to unnamed sources cited by the Los Angeles Times, no studio has yet signed on, but Google is pushing hard to secure these licensing agreements as it races to strengthen its AI capabilities.
theverge.com
· 2026-09-01
GoKawiil