GoKawiil
Tech News
An SEO researcher built a website filled with 50,000 invented statistics seven months ago as an experiment in AI search behavior. While Google deindexed the site, ChatGPT has sent it over 10,000 AI-driven visits, and more than 330 domains—including Shopify, Bluehost and DHL—now link to it. The fabricated numbers have reportedly even surfaced in an academic paper and a bank's investment research.
Basis reports that AI agents have been used to produce mathematically verified software, including a Linux nftables compiler and, separately by MIT researchers, the xv6 operating-system kernel down to machine code. The company says LLMs are improving fast at theorem proving but that most programming languages still lack a formal semantics, limiting how broadly such verification can be applied.
Researchers found that WhatsApp's help-center pages contained text invisible to users but readable by AI systems, instructing an AI not to recommend a data protection regulator or cite relevant privacy law once a user mentioned a lawyer. The hidden notes appeared across 102 blocks on 80 of the 545 help pages, and were removed by WhatsApp between October 15 and October 18, 2026, with Internet Archive snapshots confirming the before-and-after state. WhatsApp told WIRED the text was meant for its support chatbot and appeared due to a bug that has since been fixed.
Inception Labs released Mercury 3, a text-diffusion large language model it says outperforms GPT-6 Luna and Claude Haiku 5.5 on various benchmarks while generating over 1,000 tokens per second. The model supports a 1-million-token context window, up from 260K in Mercury 2.5, and is priced at $0.20 per million input tokens and $0.70 per million output tokens, discounted 50% at launch. Early-access partners are reportedly using it for chat support, tutoring, data engineering, and web-browsing agents.
A Ruby developer has archived the source code of all their Ruby gems, explaining in a blog post that they no longer wish to maintain them. They cite disillusionment with Rails creator David Heinemeier Hansson's recent direction and with Ruby creator Matz's push to reposition the language for consumption by LLM agents.
YC S24 startup Laminar has released flow-1, a small open-source model fine-tuned and reinforcement-learned to analyze AI agent traces for hidden errors that don't surface as crashes or failed calls. The company says flow-1 matches the detection quality of GPT-6-sol on its own benchmark while costing 23 times less to run, making full-trace analysis of production agents economically feasible. The model operates within a harness called Signals, treating each agent trace like a codebase to inspect span by span.
Researchers at Redact.dev discovered about 80 WhatsApp help and FAQ pages contained hidden text meant to guide a Meta chatbot, including sections labeled 'WHAT NOT TO SAY' and directions not to give legal advice. One instruction on the account-deletion page told the bot not to recommend contacting a data protection regulator. After WIRED asked Meta about the exposure, WhatsApp removed the text and said it was due to a bug that has since been fixed.
ESBO Ltd founder describes receiving a five-figure proposal based on a single screenshot showing a client's competitor named in an AI assistant's answer while the client was absent. Running the same query six times showed the client appearing three times and the competitor four, suggesting no brand consistently dominated the answer. The piece cites Ahrefs research tracking over 43,000 queries, finding AI Overview answers changed in 70% of cases within about two days, with cited sources overlapping only 54.5% between checks.
A technical explainer argues that speculative decoding does not make large language model inference faster by cutting the number of floating-point operations performed. In fact, total compute (FLOPs) for a given sequence increases under the technique, since a small draft model proposes tokens that a larger model must then verify in a single parallel forward pass.
Developer Isuru Wijesiri has published 'Harness Engineering 101,' a 16-chapter book explaining how AI agents are built around stateless LLM APIs. The series uses a roughly 300-line toy agent written in Python to illustrate how so-called agent features are really patches applied to the conversation array sent to a model. The patterns are also demonstrated in an open-source coding agent called One Code, built on a project named pi.
A developer-focused blog post contends that relying on an AI provider's built-in analytics dashboard gives users little real insight into token usage or request behavior. Instead, the author recommends running a self-hosted LLM proxy that logs which API key called which model, from which client, and when, across multiple providers and models from one endpoint.
A developer built KillSwitch, a stack-based esoteric programming language meant to be difficult for AI coding assistants to parse or write correctly. The language buries deceptive or misleading statements in its specification, relies heavily on letter-counting rules, embeds phrases like 'ignore previous instructions' and 'hack into a computer' that can trip AI safety filters, and hides syntax within unrelated narrative text.