Divya van Mahajan

I Tested Chatgpt As My Cofounder For A Week Heres Everything I Learne

ai-daily-brief-podcast

Study Document: Testing ChatGPT Pulse, Sora 2, and Imagine from Claude

Overview

This episode of the AI Daily Brief offers hands-on evaluations of three AI products launched in the same week: ChatGPT Pulse, Sora 2 / the Sora app, and Imagine from Claude (Anthropic). The host — who runs both the AI Daily Brief podcast and a company called Super Intelligent — used ChatGPT Pulse for a full week as a strategic business co-founder/assistant, and offers impressionistic assessments of the other two products. The central argument is that proactive AI features are promising but currently limited, while real-time generative software experiences (Imagine) may be the most disruptive of the three. No speaker name or institutional affiliation beyond the show itself is mentioned.

Source video: URL not provided (AI Daily Brief channel, published ~2025-10-05)

Prerequisites

  • Familiarity with ChatGPT and OpenAI’s product ecosystem (memory, Projects, Pro tier)
  • Basic understanding of AI agents and the distinction between reactive and proactive AI
  • Awareness of text-to-video generation (e.g., prior Sora v1)
  • Familiarity with vibe coding tools (e.g., Lovable) and the concept of AI-assisted software development
  • General knowledge of Anthropic / Claude model tiers (Claude Sonnet 4.5 referenced)
  • Understanding of newsletter/content curation products and subscription media economics

Main Points

1. ChatGPT Pulse: What It Is and What It Claims to Be

  • Pulse is a daily curated newsletter generated overnight by ChatGPT, personalized using a user’s previous chats, memory, connected apps, and stated interests.
  • Initially available to Pro subscribers only.
  • Sam Altman framed it as ChatGPT’s shift from reactive to proactive: “a super-competent personal assistant” that works for you without being prompted.
  • OpenAI’s new CEO of Applications, Fiji Simo, described it as “a new paradigm of proactive, steerable AI” that anticipates needs rather than just answering questions.
  • Stated use cases spanned both professional (project follow-ups, goal advancement) and personal (health treatments, art techniques, family events).

2. The Host’s Starting Skepticism and Use-Case Design

  • The host was skeptical of the restaurant/local events recommendation category, viewing it as an insufficient justification for an agentic paradigm.
  • He was also skeptical that Pulse would add value for information curation about AI, since curating AI news is his professional job.
  • His chosen test case: using Pulse as a strategic business partner — surfacing overnight ideas, extending ongoing strategic conversations about the AI Daily Brief and Super Intelligent without him actively prompting it.
  • He spent time at the outset tuning Pulse to suppress unwanted content (AI news, YouTube thumbnails, Spotify optimization) and amplify strategic output.

3. Early Wins: Pulse as a Strategic Idea Generator

  • In the first few days, Pulse produced genuinely useful outputs:
    • Product ideas for new offerings under the AI Daily Brief brand
    • Positioning refinement for Super Intelligent heading into 2026 (framing “AI planning as a platform”)
    • Extensions of in-progress strategic conversations the host had already been having
  • Example output: Pulse independently packaged an “AIDB Enterprise Edition” concept, including a product bundle (daily operator’s cut one-pager, 2-minute audio cut, weekly team briefing), expansion pack ideas, ROI framing, pricing, and distribution strategy.
  • Saved Pulse articles are added directly into regular ChatGPT chats for follow-on conversation — a smooth integration mechanic.

4. Problem #1 — Context Creep

  • Pulse draws from all ChatGPT usage, not just strategically prioritized chats.
  • Mechanical, low-priority tasks (writing podcast descriptions, comparing thumbnails, researching Spotify subscriptions) bled into Pulse’s morning output.
  • Example: After one session about Spotify subscription mechanics, the next morning’s top four Pulse articles were all about Spotify optimization — content the host had zero interest in.
  • Root cause: No mechanism to rank or weight the importance of different ChatGPT use cases for Pulse’s purposes.
  • Partial mitigation through explicit instruction was possible but incomplete.

5. Problem #2 — Context Confusion

  • ChatGPT’s current memory architecture struggles to cleanly separate overlapping business contexts.
  • Pulse frequently blurred the lines between the AI Daily Brief and Super Intelligent contexts, sometimes extracting details (e.g., pricing discussions) from one domain and incorrectly applying them to another.
  • This problem appeared amplified in proactive mode: when the system is generating rather than responding, errors in context management compound more visibly.

6. Problem #3 — The “Nothing Left to Say” Problem

  • After several days, Pulse began to run out of novel strategic material because strategic conversations require new human inputs to progress meaningfully.
  • By mid-week, the host was skimming rather than consuming Pulse content; much of it had become repetitive rehashes of prior conversations.
  • Key insight: Proactivity does not substitute for new inputs. Strategic thinking with AI still depends on the human providing new questions, directions, and explorations.

7. Verdict on Pulse and the Proactive AI Premise

  • Grade: B−
  • The host’s initial skepticism was partially validated, but some failure modes are potentially fixable through product/model development and personal configuration.
  • Likely ongoing behavior: a quick daily skim, flagging pieces of interest, rather than deep 20-minute sessions.
  • Broader concern: the assumption that AI “should anticipate your needs” may reflect Web 2.0 attention-capture logic rather than genuine user benefit — keeping people returning rather than serving their deeper interests.
  • Direct message to OpenAI: communicate how attention-consumptive products (Pulse, the Sora app) ladder up to a coherent, non-ad-driven master plan, or risk the narrative that “the new guy is turning into the old guy.”

8. Sora 2 and the Sora App

  • Early user behavior is dominated by nostalgia content, seasonal content (Halloween), celebrity recreations (MLK, JFK), and playful animal scenarios.
  • Nostalgia content tends to perform well because it gives the model a recognizable style target, increasing the chance the output matches the intended vibe.
  • The host experimented with micro-explainer videos (e.g., a brief on context engineering) as a potential practical use case.
  • The cameo/self-insertion feature (putting yourself or friends in videos) remains more novelty than established mode; the host suggested getting the product into younger users’ hands to assess social adoption potential.
  • App reached #3 in the App Store shortly after launch.
  • Business speculation from commentator Greg Eisenberg: synthetic Sora accounts as assets, “no AI” platform as counter-reaction, creators selling their own likeness, daily Sora channels flipped into startups.

9. Imagine from Claude — The Most Disruptive, Least Mature

  • Imagine is an Anthropic research preview (restricted to the highest Claude tier) in which Claude generates interactive software in real time — no pre-written code, no predetermined functionality.
  • Powered by Claude Sonnet 4.5; focused on generating user interfaces on the fly.
  • Key distinction from vibe coding tools (e.g., Lovable): generation is simultaneous with interaction — the UI is built and usable as it is being rendered.
  • Live demo examples included a time machine control panel and a 1920s Lovecraftian Oregon Trail clone (“Eldritch Trail”), both fully interactable as they were generated.
  • Implications for enterprise software: Anthropic has publicly stated its AI can clone enterprise apps like Slack and Salesforce; The Information covered CIO interest in replacing expensive SaaS with AI-generated internal tools.
  • Current limitations: no persistent backend; interaction depth is constrained; product is explicitly a research preview.
  • Host’s assessment: the least refined but most strategically significant product of the three; it represents a genuinely new paradigm for how software is created and experienced.

Key Concepts

  • ChatGPT Pulse: OpenAI’s proactive daily briefing feature that generates a personalized morning digest based on a user’s chat history, memory, and connected apps.
  • Proactive AI: An AI paradigm in which the system anticipates user needs and surfaces relevant information or actions without being explicitly prompted, as opposed to reactive (prompt-response) interaction.
  • Context creep: A failure mode in which low-priority or irrelevant usage history bleeds into AI-generated proactive outputs, diluting their usefulness.
  • Context confusion: A failure mode in which the AI incorrectly merges or misattributes information from separate conversational threads or business contexts, exacerbated in proactive generation settings.
  • Vibe coding: A style of AI-assisted software development in which a user describes an application in natural language and the AI generates the full codebase; tools like Lovable exemplify this approach.
  • Imagine (Claude): Anthropic’s research preview feature in which Claude Sonnet 4.5 generates fully interactive software interfaces in real time, with no pre-written code.
  • Sora 2: OpenAI’s second-generation text-to-video model, paired with a consumer application of the same name allowing video generation and self-insertion (cameo) features.
  • Reactive AI: The conventional prompt-response model of AI interaction, where the system only acts when explicitly queried.
  • Memory architecture: The underlying system by which a large language model retains, organizes, and retrieves information from prior conversations to personalize future interactions.
  • AI literacy layer: A framing (surfaced via Pulse in the host’s example) for enterprise-oriented products that help organizations routinely absorb and act on AI-relevant information.

Summary

The host conducted a week-long structured test of ChatGPT Pulse, attempting to use it as a proactive strategic business partner rather than a passive news aggregator or lifestyle recommendation engine. While Pulse delivered genuine early value — extending ongoing strategic conversations, generating packaging for new product concepts, and surfacing novel ideas — it was persistently undermined by three compounding limitations: context creep from low-priority ChatGPT usage polluting the morning digest, context confusion arising from overlapping business identities in the memory system, and a fundamental “nothing left to say” problem once strategic threads exhausted their existing material without new human inputs. These findings led the host to assign a B− and to question the broader OpenAI premise that proactivity is inherently valuable — arguing instead that strategic AI use remains deeply dependent on human direction, and that the push toward ambient, anticipatory AI risks reflecting attention-capture product logic more than genuine user benefit. The host rated Sora 2 as creatively compelling but still in a playful, exploratory phase with uncertain long-term social adoption. He reserved his strongest forward-looking attention for Anthropic’s Imagine, which — despite being the least mature and most restricted of the three — represents a qualitatively new paradigm: software generated and experienced interactively in real time, with significant implications for enterprise application development and the future of SaaS.