GoKawiil
Tech News
A writer who moved from Microsoft's Windows AI team to a research engineer role at Salesforce AI Research describes how their AI use has evolved over the past year, now spending about $3,000 a month on tokens. They continue using AI for coding, research, summarization, writing, and art, adding voice and video generation, and have drafted a personal policy governing how they use AI in daily life.
Tom's Hardware's Premium subscription service published a weekly digest covering Anthropic's pre-IPO filings, Cadence's new LLM suite for EDA tools, Nvidia's AI supply chain moves, Vera CPUs in the cloud, and a TP-Link update. The issue also features CPU expert Jake Roach's early work on defining agentic AI CPU benchmarks, plus a new primer explaining AI terms like inference, quantization, context, and KV cache for locally-run models.
A developer known for the nullprogram.com blog has released Prospero, an open-source coding harness built in C++ with a Dear ImGui interface, designed to connect AI models to developer tools for agentic coding tasks. The tool claims to use roughly two orders of magnitude less memory than existing harnesses, which the author says are typically built on resource-heavy web stacks. Prospero does not bundle any model provider, requiring users to supply their own backend such as llama.cpp, OpenRouter, or the Anthropic or OpenAI APIs.
A new opinion piece uses a thought experiment about an alien visitor built from transistors to question why large language models are dismissed as 'stochastic parrots' rather than a novel form of life. The author argues that LLMs acquire knowledge through processes resembling reinforcement learning and self-supervised learning on inherited data, paralleling how biological species pass on and expand knowledge across generations.
Researchers from academia, civil society and industry examined the four most widely used AI agent frameworks to see what safety-relevant signals they capture while running. They found key observability features missing, including persistent agent identifiers and records of human approvals, making it difficult to determine which agent took which action or whether oversight was properly exercised.
A new continual-learning method called the Sleep framework uses two stages—memory consolidation and dreaming—to help language models retain old knowledge while learning new tasks. Consolidation compresses short-term learning traces into model parameters using on-policy distillation and semantic rewards, while dreaming generates synthetic training examples by randomly activating different mixture-of-experts components and filtering them by gradient-based reward signals.
A developer's essay revisits Frederick Brooks's 1975 claim that adding manpower to a late software project makes it later, asking whether AI agents and LLMs change the equation. The author concludes the law hasn't disappeared but shifted: coordination overhead moves from managing people to managing agents, with humans still needed to define requirements, review output, and catch errors.
MIT Technology Review has published its new 'Impossible' issue, covering projects like synthetic blood, moon cities, cognitive immortality, bionic eyes and blended meat. The issue also reports that Astrolab's FLIP lunar rover, carrying a computer capable of running large language models, is scheduled to launch to the moon as early as next month to test AI-driven autonomy in space.
Astrolab's FLIP lunar rover will carry an HPE Spaceborne-4 edge computer designed to run custom large language models directly on the moon's surface. The system will process camera and lidar data onboard to aid navigation and distill insights, reducing how much raw data must be transmitted to Earth.
AI researcher Chris Cammarata published claims suggesting large language models may be significantly easier to understand, monitor, and control than many researchers currently believe. In an interview with MIT Technology Review's Will Douglas Heaven, Cammarata discussed what he believes the field is misunderstanding and defended his findings against skeptics who remain unconvinced. Separately, the growth of AI agents like Meta's Muse, downloaded 5 million times in its first month, is prompting websites to block such bots as they clash with human-oriented site design.
A commentary addresses the question of whether Ruby on Rails remains relevant now that large language models can generate code cheaply in any language, including proposals to rewrite applications in Rust. The author argues that frameworks, libraries, databases and operating systems remain useful because the productivity benefits they offer developers—structure, security updates, avoiding reinvented wheels—apply equally to LLMs writing code.
Researchers scanned 217 million HTTP response headers from a 10% sample of Common Crawl's September 2026 crawl and found 7,959 sites including headers addressed to AI agents. Most, 7,271 sites, were pointers directing agents to resources like llms.txt files or MCP servers, while 30 sites contained attempted prompt injections and 139 sites flagged whether they had identified the visitor as a bot.