Tech • AI • Robotics • Game

VIDEO
ENFR
TodayPlay for freeBloxLeaderboardShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

OpenAI's O leak, Microsoft Copilot reset, Claude Sonnet 5.5 debut

AIMonday, September 28, 2026· 16 videos

Briefing

Audio player
0:00 / 0:00

OpenAI leak points to O

OpenAI appears set to center its September 29 Dev Day around a persistent agent called O, according to leaked ChatGPT Pro screens and code references. The product is described as an always-on assistant, with clues pointing to 63 languages, paid tiers, and infrastructure for long-running tasks that could continue for hours, days, or weeks. Separate references to AON suggest hosted agent orchestration, code execution, research, and possibly multi-agent coordination. Key questions remain unresolved, including permissions, memory scope, email access, and whether O will launch broadly or first as a premium feature.

Microsoft resets Copilot for work

Microsoft has repositioned Copilot from a bundle of features inside Windows, Office, and the browser into a dedicated workplace AI application. The new app emphasizes chat, document work, coding, and automation in one interface, while Autopilot agents are meant to operate across enterprise tools such as Word, Excel, and PowerPoint. The strategy directly targets shadow AI use inside companies by keeping sessions within managed corporate environments and tying access to enterprise administration. The move also lands as autonomous agents face rising scrutiny over safety, governance, and data control.

Anthropic launches Claude Sonnet 5.5

Anthropic has introduced Claude Sonnet 5.5 as a faster, cheaper replacement for Sonnet 5, with the company claiming outputs roughly 30% faster at up to 30% lower cost. The model is positioned as clearer in day-to-day collaboration while making a substantial jump in coding performance, especially on agentic software tasks. Pricing is designed to undercut Opus 5.5, including notably lower costs on cache writes, input, and output tokens. The release sharpens competitive pressure just as rivals push persistent agents and ultra-fast inference options.

OpenAI touts cybersecurity defenders window

OpenAI says a narrow defenders window has opened in cybersecurity, where frontier models can help organizations discover and fix software flaws before attackers industrialize similar capabilities. The company framed the advantage as temporary, arguing that open-weight systems are improving quickly enough to erase the lead if enterprises and governments move too slowly. GPT usage was cited at more than 1 billion weekly users, while customer demand has pushed cyber to the top of executive agendas. OpenAI also emphasized expanded safety staffing, higher safety-compute spending, and recent training pauses to examine emerging risks.

Figure Helix 2.5 lifts home robotics

Figure says Helix 2.5 has markedly improved household task execution for its Figure 03 humanoid in unfamiliar homes across the San Francisco Bay Area. In tests spanning roughly 30 homes, full-task success reportedly rose from 9% under earlier methods to 56% using the new Index training system. The approach relies on large volumes of human video showing everyday actions such as folding towels, tidying toys, and making beds. The broader claim is that robotics may be entering a scaling regime where more data and compute steadily reduce error rates, echoing modern model development in AI.

WCO claims recursive self-improvement

WCO says it has demonstrated a public example of controlled recursive self-improvement, with its AID system redesigning itself over 100 rounds in 8 days. A supervising agent repeatedly modified a lower-cost worker model, evaluated each redesign under a fixed budget, and kept only versions that improved hidden final scores. About 90% of attempts were rejected, but the surviving changes yielded seven sequential upgrades, culminating in Version 85. The company says that final version outperformed a hand-built baseline developed over 2 years on unseen tasks including Kaggle, AtCoder, and weather-forecast optimization.

Michael Saylor turns AI into finance

Michael Saylor is presenting AI less as a grand societal prophecy than as a practical engine for product design, capital structure, and software growth at Strategy. Facing a financing bottleneck while already holding roughly $30 billion in Bitcoin by 2025, the company reportedly used AI-driven modeling to help design new instruments and decision frameworks. The underlying thesis is that advanced AI will compress the value of routine human labor while increasing the value of scarce assets, networks, and distribution. That makes the story less about chatbots and more about how corporations use machine reasoning to restructure balance sheets and competitive advantage.

TypeSafe targets automation with Jev

TypeSafe is pushing Jev as a specialized model for workflow decisions rather than conversational generation, aiming at tasks inside automation stacks such as n8n. Instead of producing long-form text, the model returns structured judgments like scores, probabilities, and categorical choices for email triage, spam detection, support routing, and moderation. That narrower design is meant to make it faster, cheaper, and easier to integrate than general-purpose systems like ChatGPT, Claude, or Gemini for repetitive business processes. The release highlights a broader shift toward purpose-built AI components that optimize specific decision points instead of mimicking a universal assistant.

Videos covered

Previous briefings · AI