Tech • AI • Robotics • Game

VIDEO
ENFR
TodayPlay for freeBloxLeaderboardShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

OpenAI Dots and GPT-6.1 Soul challenge Microsoft Copilot, Claude 5.5

AITuesday, September 29, 2026· 28 videos

Briefing

Audio player
0:00 / 0:00

OpenAI unveils Dots and Space

OpenAI rolled out Dots, an always-on assistant system inside ChatGPT, alongside Space, a collaborative workspace for persistent AI work. The launch expands ChatGPT from a prompt-response tool into a multichannel agent reachable by SMS, phone, email and Slack. Dots are designed to monitor work, retain preferences, use connected apps and operate on a cloud computer with approvals for higher-risk actions. The company also introduced a new $500-per-month tier with higher limits, signaling a premium push around agentic productivity.

Leak hinted at ChatGPT Pro O

Before the formal launch, leaks had pointed to a $100 ChatGPT Pro plan listing memory, 100 GB of storage and an always on assistant. Separate configuration details suggested the name O, indicating OpenAI had been testing a persistent assistant identity rather than a standard ephemeral chat. While the leak left major questions unanswered on permissions, autonomy and retention, it accurately foreshadowed the company’s broader move toward durable agents. The episode underlined how aggressively frontier labs are racing to define the personal AI assistant category.

Microsoft resets Copilot for work

Microsoft relaunched Copilot as a dedicated workplace app, reframing it from a scattered feature set across Windows and Office into a centralized enterprise assistant. A key addition is Autopilot, which lets organizations deploy more autonomous agents under corporate controls. The strategy directly targets shadow AI by keeping usage inside company administration, compliance and identity systems. With Word, Excel and PowerPoint deeply entrenched in enterprise workflows, Microsoft is betting trusted distribution will matter as much as model quality.

Anthropic answers with Sonnet 5.5

Anthropic introduced Claude Sonnet 5.5 as a faster, cheaper and markedly stronger successor to Sonnet 5. The company says it is about 30% faster, can cost up to 30% less, and offers a 1 million-token context window with up to 128,000 output tokens. Benchmarks cited large gains in coding and agentic work, with one reported jump from 10.3% to 70.6% on an agentic coding measure. Early comparative tests also showed Sonnet 5.5 outperforming Sonnet 5, Opus 5.5 and Fable 5.1 in a complex game-generation task while remaining cheaper than flagship alternatives.

Cybersecurity race centers on defenders window

OpenAI, Standard Chartered and Sophos each argued that a narrow defenders window has opened in cybersecurity as frontier models help defenders reason about exploitable risk faster than attackers can operationalize it. OpenAI said more than 1 billion people now use GPT weekly and framed security as a top executive concern, especially in critical infrastructure. Standard Chartered emphasized moving beyond counting vulnerabilities toward mapping true exploitability across architectures, identities and trust boundaries. Sophos said AI-driven operations have cut average investigation times from 38 minutes to 89 seconds in many cases, while keeping humans in the loop for riskier actions.

Figure shows Helix 2.5 robotics gains

Figure said its new Helix 2.5 system sharply improved household-task performance for the Figure 03 humanoid in unfamiliar homes. Across roughly 30 Bay Area homes, full-task success reportedly rose from 9% under earlier methods to 56% using the new Index training approach. The company argues the result hints at a robotics scaling curve driven by larger video datasets and more compute, echoing the trajectory of large language models. The gains also intensify concerns around home safety, labor disruption and privacy as domestic robots become more capable.

WCO claims recursive self-improvement milestone

WCO said it has demonstrated a controlled form of recursive self-improvement, with its research agent AID redesigning itself over 100 rounds in 8 days. A supervising agent repeatedly modified a cheaper worker model, evaluated each variant under the same budget and kept only hidden-score improvements, rejecting about 90% of attempts. The surviving chain produced seven upgrades, and Version 85 reportedly beat a hand-built baseline refined over 2 years on outside tests including Kaggle, AtCoder and weather-system optimization. If validated independently, the result would add urgency to debates over evaluation, containment and the pace of autonomous AI research.

France and Washington politicize AI

Artificial intelligence moved further into mainstream politics on both sides of the Atlantic. In France, debate around the 2027 presidential race increasingly centered on sovereignty, public infrastructure, labor-market adaptation and the lack of concrete proposals from major camps. In the United States, a viral Saturday Night Live parody of Anthropic chief Dario Amodei and new White House outreach highlighted how AI safety has become a cultural as well as policy issue. A separate Trump-Xi summit in Washington reinforced the sense of a hardening US-China AI duopoly, with dialogue advancing faster than any binding global rules.

Videos covered

Previous briefings · AI