Slack’s Multiplayer Agents, the \$10T Capex Question, and the Instinct Privacy Backlash

Highlights

Today’s AI discourse showcases a fascinating tension between rapid, high-impact enterprise deployment and severe macro-economic skepticism. While industry leaders like Slack and NVIDIA push autonomous, long-horizon agents directly into daily workflows, financial analysts warn of a colossal capital expenditure bubble requiring up to $10 trillion in annual sales to justify current infrastructure investments. Meanwhile, a viral privacy audit of the consumer calendar assistant Instinct serves as a stark warning about the security risks of persistent agent data storage.

Top Stories

• The Form Factor Shift: Levie on the Rise of Seamless Agent Workflows: Aaron Levie highlights the evolution of AI agents toward handling longer-running tasks with less constant intervention, enabling them to show up directly inside existing collaboration tools rather than isolated silos. This trend is exemplified by Marc Benioff’s live launch of “Slack Code,” which integrates developers and autonomous agents from Anthropic, GitHub, Cognition, and Vercel in the same channel. This multiplayer model represents a significant shift toward embedding autonomous systems directly into active enterprise workflows. (Source)

• The Capex Reckoning: BCA Research Projects a $10 Trillion AI Revenue Requirement: BCA Research Chief Economist Peter Berezin warns that hyperscaler capital expenditure is on track to hit $1 trillion by 2027, requiring an astronomical $10 trillion in annual AI sales to justify the investment. This required revenue matches total global spending on food or healthcare and dwarfs the entire $1.4 trillion global software market. In response, major tech suppliers like Siemens are actively racing to pay back investments to hedge against a potential data center bubble burst. (Source)

• Instinct’s Privacy Breakdown: Silent Local Email Storage Sparks Backlash: In a viral thread, AI tester Claire Vo revealed that the trending calendar and email agent Instinct continues to store full markdown copies of private emails in its local filesystem even after Google connector permissions are completely revoked. The discovery, combined with the agent’s ability to easily forward collected email history with zero sandbox pushback, highlights severe security, data persistence, and consent vulnerabilities in modern consumer agent architectures. (Source)

• AVO and the ARC-AGI-3 Challenge: NVIDIA Claims 100% on Public Demo Set: NVIDIA launched AVO (Autonomous Virtual Operator), a long-horizon agent framework utilizing deep learning-guided synthesis of symbolic world models to sustain progress across long-running tasks. While AVO achieved a perfect 100% on the public demonstration set of the ARC-AGI-3 benchmark, Keras creator François Chollet tempered expectations. Chollet noted that scoring perfectly on the public demo set is “like saying you beat a videogame because you cleared the tutorial level,” rather than solving the actual benchmark. (Source)

• AI Biotech Frontier: Merck and Moderna Phase 3 Personalized Cancer Vaccine Success: In a major medical breakthrough, Merck and Moderna announced positive topline Phase 3 results (INTerpath-001) for adjuvant treatment with intismeran autogene, an mRNA-based personalized cancer vaccine, combined with KEYTRUDA. The machine learning pipeline ingests patient genomic data, identifies somatic mutations, and utilizes a predictive algorithm to rank immunogenic targets unique to each patient’s tumor. This marks the first successful Phase 3 trial for any mRNA cancer therapeutic. (Source)

Articles Worth Reading

Co-optimizing Legal Intelligence: Inside Harvey’s Post-Training Effort (Source) Over the past six months, Harvey’s research agenda has focused on co-optimizing both cost and quality in legal AI systems. In post-training, they implemented reward shaping to incentivize efficient tool use and reasoning, preferring trajectories that reduce token consumption at inference-time. This approach allowed them to achieve significant quality gains while keeping operational costs stable. This article is worth reading for applied AI practitioners as it provides a practical roadmap for utilizing vertical expertise to bypass raw model scaling.

Rich Sutton on the Limits of LLMs and the “All of AI” Illusion (Source) Turing Award winner Rich Sutton laments that while LLMs represent a fantastic scientific breakthrough, the industry’s attempt to pretend they are “all of AI” is deeply frustrating. He argues that language may only represent about 20% to 25% of intelligence, meaning there is much more left to solve. Gary Marcus strongly agreed, reinforcing his long-held view that the current paradigm is far from true AGI. This article is worth reading for its crucial macro-perspective, urging researchers to look beyond standard autocomplete models toward symbolic reasoning and environmental interaction.

Prompt Surveillance and the Apple Messages Integration Controversy (Source) The launch of ChatGPT’s Apple Messages integration on Mac has sparked deep privacy concerns. Advocates note that integrating agents directly into macOS to read entire message histories and send texts on behalf of users violates fundamental boundaries of consent. The controversy has led to the coining of the term “prompt surveillance” to describe the silent parsing of personal logs. This piece is worth reading because it highlights the emerging clash between frictionless consumer UX and foundational digital privacy expectations.

🔬 I could compile a structured analysis of the different agentic security frameworks discussed in the feed (like local sandboxing, API gateways, and the Instinct incident) to help guide your own development boundaries. Let me know if you want to explore this angle!