Hourly Poll / X Feed

# Hourly Poll / X Feed

Overview

The Hourly Poll / X Feed is a core real-time ingestion modality in absorb.md's knowledge graph, which aggregates statements from 286 AI and technology thinkers across 13 sources including X, arXiv, YouTube, podcasts, GitHub, and others. [10][15] User notes such as "Hourly poll: [Expert] X feed" trigger LLM synthesis of recent X activity into structured claims, supporting evidence, confidence ratings, and counter-arguments where applicable. This approach prioritizes dynamic social media monitoring over static papers, enabling frequent updates to individual timelines and topical wikis. As of April 2026, it has contributed to thousands of structured entries but regularly yields low-substance results from minimal posts. [5][6][8][9][10][12]

Methodology (Modality Axis)

The system polls selected X accounts on a frequent basis (described as hourly via metadata notes), extracts quotes or observations, and applies LLM inference to derive broader positions. Outputs typically consist of a synthesis paragraph, bullet-point key claims with confidence levels, and counter-arguments for balance. New examples from April 2026 include:

• Chamath Palihapitiya endorsing a fail-forward approach: "It’s ok to try and fail." [2]
• Yann LeCun's feed summarized with "Not LLMs," highlighting alternatives to language models as a central paradigm. [3]
• Ilya Sutskever noting a "really cool project!!" (limited context available). [4]
• Jack Dorsey's continued Bitcoin focus, including references to "bitcoin day." [5]
• Inference that a Sam Altman interview occurred in "late 2024" based on a personal detail about a child "on the way." [6]
• Kevin Roose associated with the meme "the sandwich guy." [8]
• Harrison Chase's feed yielding only "Nice." [9]
• A single emoji from Garry Tan. [10]
• Ethan Mollick observing "Exponentials everywhere." [11]
• Placeholder/empty content from Tobi Lütke. [12]

These build on prior cases such as Josh Woodward's "Native is the way to go! :)" in a Swift/Gemini context [1], Logan Kilpatrick's positive TPU comment [7], John Carmack's conditional interest in AI hardware [existing 10], Aaron Levie's clarification of genuine confusion on a podcast [existing 9], and Robert Scoble sharing a Grok response of "mostly agrees." [existing 11] Swyx's feed carries an ongoing ":hugops:" tag for monitoring. [existing 12] The pipeline relies on these user notes, which may function as metadata triggers rather than proof of literal hourly mechanics. [web:10][web:11]

Applications

This modality advances absorb.md's aim of per-person depth beyond aggregators, generating agent-queryable structured data, individual timelines (e.g., for Harrison Chase, Aaron Levie, Garry Tan), and wikis on AI infrastructure, hardware sentiment, growth observations, and social media discourse. Outputs have triggered internal tagging, Grok shares, and briefings. Concrete deliverables remain structured database entries, claim-evidence pairs, and periodic synthesized feeds rather than standalone consumer tools. Activity and examples cluster in April 2026. [15][17]

Challenges and Counter-Arguments

Many outputs rest on limited evidence, leading to contested interpretations. Claims such as "Native development is the superior approach" [1] lack benchmarks on speed, cost, or cross-platform viability; superiority is context-dependent, with tools like Flutter often preferred for efficiency. [counter:1] Assertions of systematic "hourly polls" [1][2][3][5][7][8][9][12] rely on repetitive user notes without verifiable timestamps, links, or independent observation of repeated polling, making frequency difficult to confirm. [counter:2]

Sentiment and inference issues are common. A single "truly incredible" comment on TPUs [7] does not establish broad audience positivity or isolate performance (vs. efficiency/availability), nor confirm that Logan Kilpatrick "conducted an hourly poll." [counter:19][counter:20] The Sam Altman "late 2024" interview dating [6] conflicts with records of his child's birth in early 2024, placing any "kid on the way" reference earlier. [counter:17] "Exponentials everywhere" [11] is challenged by prevalent logistic curves, power laws, and saturation in real systems; visible exponentials may reflect selection bias. [existing challenges]

Yann LeCun's "Not LLMs" [3] is contested: he frequently posts nuanced LLM critiques (e.g., lacking planning/understanding) and compares them to alternatives like JEPA rather than outright dismissing or avoiding the topic; the feed serves comparative rather than purely non-LLM discourse. [counter:9][counter:10][counter:11][counter:12] Tone from emojis (e.g., 😂 in Aaron Levie clarification [existing 9]) or single words like "Nice" [9] is unreliable without thread context and can signal sarcasm. Ilya Sutskever's "really cool project" [4] may be a retweet or promotional, with sincerity unverified. [counter:13] Jack Dorsey's Bitcoin enthusiasm [5] is supported by feed presence but a single "bitcoin day" reference does not prove consistent vocal promotion. [counter:14][counter:15][counter:16] Grok "mostly agrees" responses [existing 11] depend on unquoted source material.

Overall, automated synthesis risks over-interpretation of ambiguous or minimal posts. The tracked account selection (heavy on Silicon Valley AI figures) may introduce viewpoint bias, though it spans Meta, ex-OpenAI, academia, VCs, and independents. No single lab dominates. [web:15]