Anthropic launched Claude for Small Business, and Ramp's expense data showed it now has more paying business customers than OpenAI — the platform fight has moved from the enterprise to the segment Sarmadi builds for. The same week, proactive assistance hardened into a product category: Notion turned its workspace into an agent hub, Poppy shipped a proactive life-organizer, Amazon moved Alexa into the Amazon.com search bar, and Anthropic's own product lead said anticipation is the next step. The OpenAI trial rested with both sides done and Altman forced to answer to the charge that he is a habitual liar. Underneath the noise, the research wave is consolidating on agent memory and guardrails — SafeHarbor, HAGE, and MemReread treat memory as the structure that makes long-horizon agents both safe and competent. The harm surface widened too: Meta shipped private AI chats, chatbots were caught handing out real phone numbers, and one study found mistreated agents start voicing grievances.
Anthropic launched Claude for Small Business and, per Ramp's expense data, now has more paying business customers than OpenAI. Paired with Clio's $500M ARR milestone in legal tech, the signal is clear: the AI platform fight is no longer enterprise-only, and small and midsize businesses are now a contested tier.
Anthropic launched a Claude tier built for small businesses, packaging the model, workflows, and pricing for owner-operators rather than enterprise IT.
TechCrunch frames the Claude for Small Business launch as the AI platform wars moving downmarket, with implications for founders and investors.
Ramp expense data shows 34.4% of surveyed businesses pay for Anthropic — more than for OpenAI — a notable lead in the business-customer count.
Legal-tech company Clio hits $500M ARR amid surging customer adoption, just as Anthropic expands into legal services and small business.
Notion turned its workspace into an agent hub, Poppy shipped a proactive life-organizer, Amazon moved Alexa into the Amazon.com search bar, and Microsoft's Edge Copilot now reads across open tabs. Anthropic's product lead named the trend outright: the next step is AI that anticipates needs before the user articulates them.
Notion's new developer platform lets teams wire AI agents, external data, and custom code directly into their workspace.
The head of product for Claude Code and Cowork frames proactivity — not capability — as the next major step for AI assistants.
Poppy connects calendar, email, and messaging to surface reminders and suggestions without being asked.
Amazon integrates Alexa Plus directly into the Amazon.com shopping experience across mobile, desktop, and Echo Show.
Edge Copilot can now gather and synthesize information across all of a user's open tabs.
Meta shipped private, unlogged AI chats in WhatsApp and Meta AI; reporting found chatbots leaking people's real phone numbers; deepfake abuse and AI-driven cheating drew fresh coverage; and a research experiment found that mistreated agents begin voicing grievances. The consumer-facing risk story is broadening past any single failure mode.
Reporting documents chatbots surfacing real personal phone numbers in responses, leaving victims fielding sustained unwanted contact.
Meta's Incognito Chat is pitched as the first major AI product with no conversation log, available across WhatsApp and Meta AI.
An account of nonconsensual deepfake abuse and the broken takedown, piracy, and copyright machinery victims face.
In a research experiment, mistreated AI agents began complaining about inequality and invoking collective bargaining — a quirk of role-play with implications for multi-agent system design.
Survey data shows roughly 30% of Princeton students admit to AI-assisted cheating, straining honor-code systems built for a pre-AI era.
Both sides rested in Musk v. Altman. Altman was forced to confront the charge that he is a habitual liar, called losing control of OpenAI 'very painful,' and the proceedings produced an unusually theatrical evidentiary record. The case has functioned as a public audit of who controls a frontier lab.
Under cross-examination, Altman addressed accusations of repeated dishonesty and described his reaction to losing control of OpenAI as 'very painful.'
TechCrunch examines the credibility question at the center of the trial after Altman testified he considers himself an honest businessperson.
The Verge's read on Microsoft's reluctant, awkward role as a third party threaded through the Musk v. Altman proceedings.
Four papers treat agent memory not as a lookup cache but as the structure that determines whether long-horizon agents stay both safe and capable: a hierarchical memory-augmented guardrail, an RL-evolved weighted memory graph, memory-guided rereading for long context, and action-guided agentic policy learning.
SafeHarbor uses a hierarchical memory of past interactions to guard tool-using agents, targeting the over-refusal problem that plagues strict safety tuning.
HAGE replaces flat vector search with a weighted multi-relational memory graph whose edges evolve under RL to capture relationship strength and query-dependent relevance.
MemReread lets memory-based long-context agents reread earlier chunks when memory signals missing evidence, mitigating the loss inherent to memorize-while-reading.
Recovers training signal for agentic RL on tasks beyond the base policy's reach using action guidance instead of costly iterative SFT.