Chronological feed of everything captured from Aaron Levie.
SpaceX as a vertically integrated AI compute company makes an insane amount of sense.
Box's testing shows GPT-5.5 achieving a 10 percentage point accuracy increase over GPT-5.4 on complex enterprise knowledge work evaluations across industries like financial services, healthcare, public sector, and media. Sector-specific gains include 83% vs 64% in financial services, 78% vs 61% in healthcare, 72% vs 59% in public sector, and 70% vs 57% in media/entertainment. The model excels in advanced reasoning, data analysis, complex context handling, and is now accessible via Box AI Studio for agent development.
Aaron Levie plans to introduce hourly polls on his X feed. The feature is confirmed as rolling out soon. This enables frequent, real-time user engagement on the platform.
AI lowers the barrier to initiating tasks, leading users to explore and pursue more projects than before, resulting in increased overall work despite automation. Small tasks that once took too long manually now consume hours due to easy starts requiring manual completion. This dynamic reveals work as non-static; companies have latent capacity unlocked by AI, potentially justifying hiring for agent-assisted roles.
Aaron Levie initiated an hourly poll feature tracking his X feed activity. The post expresses high enthusiasm with "Let’s goooooo. 💪", signaling strong commitment to regular updates. This suggests a new cadence of real-time engagement for his audience.
AI enhances employee productivity, prompting companies to scale operations and hire more workers to leverage these gains, per Jevons paradox. As AI improves, it enables tackling complex tasks previously uneconomical, leading to hiring for surrounding roles like engineers and sales staff. Empirical data shows rising software engineer hires, lower youth unemployment, and 640,000 new AI-related jobs in the US from 2023-2025.
AI provides unprecedented leverage across domains, enabling individuals with ambition and core skills to bypass traditional experience requirements. This shift particularly benefits juniors, allowing them to achieve far more than previously possible, while companies must identify and elevate such talent for competitive advantage. The underlying dynamic compresses experience moats, rewires decision-making (e.g., favoring experimentation over caution), and positions unfiltered thinkers as strategic assets in risk-averse settings.
AI agents cannot replicate the human brain's fundamental requirements for building relationships, evaluating vendors, and gaining tactile product understanding in go-to-market (GTM) functions. AWS, despite two decades of technological advancement, employs ~60,000 people in GTM, underscoring that full automation remains unachievable. This highlights persistent human elements in enterprise sales and operations.
AI agents boost leverage on incremental effort, making individual contributors feel the acute pressure of suboptimal task direction akin to managing teams. Lower barriers to starting tasks lead to more projects initiated, where the final 10% completion consumes disproportionate time, often extending into late hours. This dynamic accelerates idea testing and hypothesis validation, fostering job creation by promoting successful experiments to production, though unsustainable economy-wide.
Professionals recognize the "last mile" complexities in their own work—data access, context provision, output review, and process integration—that prevent full AI automation. Yet they assume AI instantly eliminates entire jobs in unfamiliar domains, underestimating equivalent hurdles. This bias fuels skepticism toward sweeping job loss predictions based on task-level AI demos. It echoes observations that AI impresses most when users lack domain expertise or precise goals.
AI proficiency improvements will prompt a corresponding escalation in task complexity, shifting the human "last mile" challenge rather than eliminating it. This dynamic forms a self-perpetuating cycle where experts continually address novel edge cases as AI masters prior limitations. The hypothesis anticipates indefinite repetition without resolution.
Aaron Levie strongly approves of an hourly polling mechanism applied to his X feed, labeling it as the ideal user experience. This suggests polling at hourly intervals provides optimal real-time updates without overwhelming users. The endorsement comes in response to a user-initiated poll tracking his feed activity.
Aaron Levie demonstrates complete command of situations in his X feed analysis. He approaches technology with pragmatism, prioritizing practical outcomes over idealism. This assessment positions him as a reliable voice for technical decision-making.
Hourly poll conducted on Aaron Levie's X feed recorded zero instances of rattling. This indicates normal operation without detected disruptions or anomalies in the monitored period. Technical monitoring confirms absence of the specified event.
Aaron Levie defends the Dwarkesh-Jensen interview against criticism, highlighting how a 25-year-old podcaster compelled the CEO of the world's largest company to engage directly and responsively. This interaction exemplifies media's role in challenging power rather than offering praise. Levie views the discourse dismissing it as "ridiculous," emphasizing its impressiveness.
Aaron Levie explicitly denies facetiousness in his reaction to incomprehensible tweets about a podcast. His statement underscores authentic bewilderment amid an hourly poll tracking his X feed. This reveals real-time public discourse generating confusion for the tech executive.
Aaron Levie describes a rare occurrence where morning assumptions about computer operations are overturned by evening discoveries. This highlights the accelerating pace of technical revelations in computing. Such events underscore the dynamic nature of foundational knowledge in tech fields.
Agents will utilize enterprise software 100X more than humans, necessitating headless platforms with open APIs for seamless integration. This shift unlocks vastly expanded use cases beyond human-limited user counts, enabling 24/7 parallel processing at unprecedented scales. Salesforce exemplifies this with Headless 360, exposing all APIs for direct agent access across channels.
Agents' 24/7 parallel operation across systems vastly exceeds human usage limits, rendering traditional seat-based software models obsolete and necessitating headless architectures. This shift enables platforms to support exponentially more use cases, such as bulk contract reviews or hyper-scaled sales automation via Salesforce data. Revenue models evolve to hybrid seats-for-humans plus agent consumption, creating substantial upside for adaptive platforms.
AI agents enable every company to develop custom software for workflow automation, previously infeasible due to technical or economic barriers. This shifts software engineering from tech giants to widespread roles in biopharma, industry, finance, and retail, focusing on system design, agent orchestration, and process redesign. Demand for technical talent surges as even small teams handle complex projects, exemplified by emerging roles like Eli Lilly's Lab Automation Software Engineer.
AI model progress demands quarterly rebuilds of agent systems, obsoleting mitigations for prior limitations like context windows. Deployments in enterprise workflows must be rethought at similar cadence, as practices from 18 months ago are now outdated. This cycle of solidification followed by disruption drives intense work, with no slowdown imminent, amplified by recent obsolescence of tools like RAG and LLMOps.
AI tools raise the sophistication of most roles to match their capabilities, rendering simplistic views of job replacement obsolete as markets dynamically evolve. Skilled professionals leverage AI to tackle larger, harder problems, maintaining differentiation through domain expertise. Roles evolve toward greater complexity, with AI acting as a multiplier for the already proficient rather than an equalizer.
AI will surge legal demand by prompting more questions needing lawyer verification, spawning exotic contract terms for review, and birthing new AI-related laws in IP, privacy, and compliance across industries. Historical data shows U.S. active attorneys rose from 400,000 in 1975 to 1,375,000 in 2025 amid PC and internet efficiencies. Automation in professions like law often amplifies rather than contracts demand.
Scanning public SaaS companies reveals negligible correlation between GAAP operating margins (OM) and year-to-date stock performance declines. Stock-based compensation (SBC) impacts margins but fails to differentiate or protect against market hits. Investors currently prioritize other factors over traditional profitability metrics.
Aaron Levie's X feed features an hourly poll querying the relationship between GAAP operating margins (OM) and stock performance. The poll targets investor sentiment on whether strong GAAP OM directly correlates with positive stock movements. This reflects ongoing interest in reconciling GAAP metrics with market valuations in tech sectors.
Aaron Levie responds affirmatively to an unspecified statement with emojis indicating attention, distress, and laughter. The reply "yes agreed 👀😭😂" suggests lighthearted endorsement of a relatable or ironic point. Context from an hourly poll of his X feed provides minimal substantive detail.
Current software market defies uniform characterization, with some companies facing surging demand and full engineering roadmaps unaffected by headcount reductions. AI-driven productivity allows fewer hires for equivalent output in these firms. Meanwhile, others anticipate engineering crunches, creating a persistent bimodal distribution.
Conversations in public X threads, exemplified by Anjney Midha's observation on visible alpha, reveal substantial untapped insights for those paying close attention. Aaron Levie engages by requesting shared lists to align ideas, highlighting collaborative knowledge extraction. References like Stanford's CS153 course link to AI safety theses, underscoring accessible yet overlooked signals in social feeds.
Amazon's capex in the last 3 years exceeds its entire prior history, reflecting massive datacenter investments for AI. Current AI usage centers on efficient chat tools, while coding agents consume orders of magnitude more tokens but remain niche. Knowledge work agents will soon drive token processing demands 100s of times higher, pushing infrastructure growth vertical.
AI-driven security tools will accelerate vulnerability discovery via autonomous exploitability and 100X code generation, surfacing far more findings that require human triage, remediation, and architectural judgment. This creates a Jevons paradox where efficiency gains expand the problem scope, increasing rather than decreasing demand for security talent. AI enables initial scaling but cannot replace expert oversight in response management.
Aaron Levie, in response to a user note about an hourly poll on his X feed, indicates he is currently away from a computer. He requests to go second in the interaction, deferring to the user to start. This reveals a momentary limitation in real-time engagement due to accessibility constraints.
Enterprises are transitioning from AI chat interfaces to agents that execute workflows using tools and data, prioritizing targeted automation over broad experimentation. Key barriers include change management, token budgeting under strict OpEx limits, and modernizing fragmented legacy systems for agent interoperability. Leaders emphasize revenue-generating use cases over job replacement, demand headless software, and anticipate engineers' central role in agent deployment despite rapid innovation cycles.
A Golang MapleStory (v28) server. Stars: 326
The authors introduce adaptive canonicalization, a framework where the input's standard form is dynamically determined to maximize network predictive confidence. This approach eliminates the discontinuities associated with fixed canonicalization, ensuring continuity and universal approximation while maintaining symmetry. Empirically, it demonstrates superior performance over traditional equivariant architectures and data augmentation in spectral GNNs and point cloud tasks.
Molecular dynamics (MD) simulations are computationally intensive. Machine-learning-based simulators, particularly graph neural networks (GNNs), offer accelerated alternatives but face limitations in accuracy and stability due to inefficient propagation of long-range information. This paper investigates nonlinear spectral GNNs to overcome these issues by explicitly representing simulated systems in a global eigenmode basis, thereby improving accuracy and stability in capturing complex system dynamics, especially for long-range and collective behaviors.