GoKawiil
Tech News
A writer analyzed years of emails received in response to blog posts and Hacker News comments and found that what used to be genuine, thoughtful outreach has shifted mostly to disguised sales pitches. He describes a recurring pattern where senders open with a personalized compliment and an innocent-seeming question, only to pivot into promoting their own product once he replies at length.
Recent incidents include a record cryptocurrency theft from Binance, a thwarted attack on Anthropic's Claude via a malicious third-party extension, and a breach of travel-tech firm Amadeus that exposed airline customer data. Generative AI companies such as Anthropic and OpenAI are simultaneously linked to enabling more sophisticated hacking capabilities and marketing themselves as essential defenses against those same threats.
Anthropic has agreed to pay $3.4 billion to assume Texas supplier OnePWR Solutions' existing order for 1.7 gigawatts of gas-fired generators, originally placed with Caterpillar nearly two years ago. The deal, reported by Reuters to have closed in late September, also covers engineering and construction needed to install the equipment at Anthropic's data centers, without Anthropic acquiring OnePWR itself.
A user-submitted post claims that Anthropic's Claude Opus (referred to as 'Opus 5.5') began showing changed 'thinking patterns' starting November 3, with some users describing it as 'nerfed.' The post argues that the earlier common advice to set reasoning effort to 'high' for all tasks may no longer hold given this shift.
A developer describes replacing juggling of many parallel Claude Code/Codex terminal sessions with a priority-queue system stored as markdown tickets in git, next to the relevant code. An agent cycles through the board researching tickets, building pre-approved ones, running tests and browser checks, and verifying deployments, but never deploys to production or makes judgment calls itself. Items requiring human decisions are surfaced in a prioritized 'owner inbox' rendered as a single-page Claude artifact rebuilt from git each cycle.
Security firm Apiiro found that Claude Code's local project settings file, .claude/settings.local.json, can store approved commands verbatim—including passwords and API keys—when users choose 'don't ask again.' Scanning GitHub, researchers located 104,185 copies of the file across 103,617 repositories, with 1,122 copies containing at least 1,379 exposed secrets, including 379 database passwords. One notified company confirmed its leaked key was still active.
Goldman Sachs has updated internal guidance to restrict use of Anthropic's Claude Code and Microsoft's GitHub to its engineering staff, cutting off traders and bankers who previously used the tools, Bloomberg reported. The change follows a review of productivity and risk concerns, and comes after Goldman barred Hong Kong bankers from Claude in April. Strats, who work across trading and engineering, retain access.
A project called llama-modchip, released on GitHub, claims to compress the 284.33B-parameter DeepSeek-V4-Flash mixture-of-experts model from roughly 77GB down to about 57-59GB of memory, letting it run on a 64GB CPU-only Linux machine without a GPU. The developers report prefill speeds near 44 tokens per second and CPU decode speeds of roughly 9-15 tokens per second, achieved through custom quantization, cache-tiling and pruning techniques, plus a separate mode running a 30B Nemotron model at much higher speeds on CPU hardware.
A longtime Claude Code user announced they are retiring the tool as their main coding harness in favor of Pi, citing accumulated frustrations with the product. The author noted that Anthropic added support for AGENTS.md files shortly after the switch decision was made, but said the change came too late to reverse their choice. Specific complaints included incompatibility between CLAUDE.md and AGENTS.md formats and a VSCode IDE integration limitation allowing only one connected session at a time.
Anthropic's Claude Personal Finance hub excludes user chats from model training by default and erases data once a bank account is unlinked. OpenAI's ChatGPT Finances instead follows users' existing model-training preferences, requiring them to manually opt out via Data Controls, and deletes synced account data within 30 days of disconnection. Anthropic also says its tool carries no ads or financial products, while OpenAI is testing ads on cheaper tiers and plans to expand ChatGPT Finances through a partnership with Intuit.
Security researchers reportedly identified 1,122 public GitHub files containing exposed passwords and credentials tied to use of Claude Code's 'Don't Ask Again' permission setting. The feature appears to let the AI coding assistant bypass repeated confirmation prompts, which in these cases allowed sensitive secrets to be committed to public repositories. The finding was published via Apiiro, an AI application security firm.
A guide circulating on AI usage identifies five practical adjustments users can make to improve their results with tools like ChatGPT and Claude, including writing better prompts and making smarter use of memory features. It frames the gap between users who find AI tools highly useful and those who don't as largely a matter of technique rather than the tools themselves.