A Lab Left an AI Alone for 8 Days. It Rewrote Itself Into a Better Version.

AI is changing how modern teams grow.

The fastest-growing companies aren't always adding more employees. They're using AI to help existing teams accomplish more with less manual work.

From reporting and operations to routine business processes, Viktor helps automate repetitive tasks so your team can scale productivity without scaling headcount.

1  Weco AI published what it calls the first experimental evidence of recursive self-improvement, after leaving an AI alone for 8 days.

AIDE², Weco's agentic system, spent eight unattended days rewriting its own code. The result was an agent that outperformed the hand-tuned version Weco's engineers had spent two years refining. Recursive self-improvement, the idea that an AI can autonomously improve its own capabilities, has been a theoretical concern for years. Weco is claiming it has now been demonstrated experimentally rather than just theorized. Read the full blog post and form your own view on what the evidence actually shows.

2  IBM had its worst single day on record, dropping 25% after warning that clients are redirecting budgets toward AI hardware.

IBM warned investors that clients are shifting capital spending toward AI infrastructure - servers, storage, memory - at a pace the company did not anticipate. CEO Arvind Krishna told investors: "We did not anticipate the magnitude of the capex reprioritization." The selloff spread across the enterprise software sector, taking Oracle, Salesforce, and ServiceNow down with it. The pattern is the same one that has been punishing software stocks all year: every dollar flowing into AI hardware is a dollar not flowing into the products IBM, Salesforce, and Oracle sell.

3  PrismML compressed a 27-billion-parameter model small enough to run on an iPhone 17 Pro.

Bonsai 27B is a compressed version of Alibaba's open-source Qwen3.6 27B that PrismML shrunk from 54GB to under 4GB without gutting its performance. PrismML's CEO told CNBC that Apple is evaluating the technology, though no deal has been confirmed. Try a live demo here. The significance is practical: a capable 27B model running locally on a phone means AI inference with no network dependency, no API cost, and no data leaving the device.

Bonus read:  OpenAI's growing challenges are narrowing its IPO window according to the Wall Street Journal. Link is unlocked.

The essay. Demis Hassabis does not post often. When he published A Framework for Frontier AI and the Dawning of a New Age, the internet paid attention. The argument is that humanity is standing at a civilizational inflection point that he compares to the discovery of fire or the harnessing of electricity. He believes the decisions made in the next few months will shape what the next phase looks like. The essay is not a prediction document. It is a call to action from someone who thinks the clock is running.

The AGI question. The essay hinges on artificial general intelligence: an AI system that matches or exceeds human cognitive capability across domains. Hassabis believes AGI is probably only a few years away. He is careful to add that it remains a hypothetical milestone — no lab has achieved it, and timelines like his are predictions rather than certainties. The weight he gives the possibility, however, drives everything else in the essay.

His proposal. Hassabis wants a self-regulating Standards Body modeled on FINRA, the private-industry watchdog that monitors Wall Street. The body would work alongside AI labs and federal agencies to track the technology's progress, assess risk, and vet new models before public release — without requiring government bureaucracy to move at the pace of AI development. He wants the organization in place before the end of the year.

What makes this different. Most AI governance proposals come from regulators, academics, or critics outside the labs. Hassabis runs one of the most consequential AI labs in the world and is himself a Nobel laureate. His argument is not that AI needs to slow down. It is that the field needs a credible mechanism for tracking what is happening and catching problems before they compound. The proposal puts the responsibility inside the industry rather than waiting for government to impose it from outside. Whether that framing survives contact with political reality is the open question.

Write a task brief, not a prompt. A Sully AI researcher shared a technique called long-horizon prompting, which replaces conversational prompting with a pseudo-formal task brief that defines the goal, success criteria, constraints, and output format before the AI begins. The full skill template is on GitHub. The principle is that AI produces better results on complex tasks when it knows what done looks like before it starts.

Finding buried money. The founder of Ghostty used ChatGPT to surface $45K in overlooked invoices buried across three years of email. The post breaks down the exact strategy. If you have a large email inbox and outstanding receivables, the workflow is directly repeatable.

What to wear, decided by AI. An OpenAI employee built a tool that generates outfit combinations from your existing wardrobe with one commenter calling it one of the most practical personal AI use cases so far (4M+ views). The Claude skill behind it is on GitHub if you want to run it yourself.

Free Claude for teachers. Anthropic is giving all verified K-12 educators free access to premium Claude plans. Full details and eligibility here. Worth sharing with anyone working in education who has been paying out of pocket.

Three ways to design with AI. A viral post with 600K views walks through three distinct design strategies: leading with design skills and directing the AI, working piece by piece through a composition, and using a reference image as a starting point. The framework is practical regardless of which tools you use.

Travel-art illustration

David Joerg, Chess.com's AI/ML product manager, ran 30+ projects through Viktor in six weeks, from an AI phone-agent system to logistics and research, then onboarded six family members. Get Started for Free.