26 AI Employees → 9 Stronger Ones
Stream summary 74This stream in minutesWant to back the stream? The «داعم» membership on YouTube is EGP 60 a month: badges, emoji, and I answer your questions first on stream. Become a supporter
The gist
The stream documents broad practical work and discussion around building, marketing, and measuring AI-assisted products. Its strongest working thread was a Shopify app: reconsidering free-plan versus trial pricing, reviewing onboarding, code, and testing, and developing GA4 analytics that connect visibility, traffic, and orders. Deployment and test outcomes and the pricing impact were not settled, while Luqman viewed acquiring enough traffic as the main remaining challenge. Growth discussions covered inbound and outbound channels, demos, partnerships, advertising, separating traffic problems from conversion problems, demand validation and distribution, and possible later expansion into more languages, platforms, and markets such as Salla. They also included general opinions on bootstrapping, investment, when to raise funding, avoiding excessive early equity dilution, the trust barrier in government sales, and market-dependent YouTube monetization. The stream also explored smaller AI-agent teams with defined roles, dependencies, and human oversight; a permission-limited central executive agent; and voice-assistant tests that exposed latency, interrupted conversations, conflicting records, and incomplete integrations. Luqman compared coding tools and models including Codex, Claude, Gemini, and Grok, arguing that token prices and published benchmarks are insufficient without task-specific efficiency measurements and consistently framing his assessments as personal observations. Codex slowdowns and context degradation remained unresolved. The attempt to locate Gemini 4 was specifically a historical snapshot dated October 1, 2026 and did not establish its availability or final identity. Further discussion covered goal engineering, data and website security and AI’s use by both attackers and defenders, hardware constraints on local models alongside the opinion that local compute may matter long term, practical AI learning while continuing computer-science study, and whether AI weakens thinking or improves efficiency. Other topics included remote work and the possible advantage of lower-cost locations; general distinctions among work or study migration, asylum, and visits; non-personal opinions about financial independence, investment funds, and dollar-cost averaging; and workspace ergonomics through standing desks or walking pads. Luqman also discussed content and livestream strategy, including consistency versus repetition, streaming for accountability, useful standalone videos, and background-music copyright risk. The music work went beyond a general focus-music experiment: it covered tempo, layering, and Arabic maqamat, plus a Suno Studio workflow in which a musical scale is created with Python, uploaded, arranged in layers, and supplemented with sounds. This work and the assistant’s visual and vocal identity remained experiments rather than demonstrated successes.
180 moments from the stream
- 6:19:57
Deploying a Worker and an Install-Related Event
A brief live-work update mentions deploying a worker, completing a related step, and an install-related event. The short fragments do not provide enough context to identify the platform or implementation details.
- 6:23:33
Assessing Codex Latency and Deep-Audit Results
A discussion of unusually slow Codex execution on a task expected to finish quickly. The speaker said earlier deep audits usually surfaced critical, logical, or architectural issues, whereas this audit reported only smaller problems. The discussion also noted that reference context or specifications help distinguish a bug from intended behavior, while the exact evaluation mechanism remained unclear.
- 6:27:24
When Model Performance Feels Inconsistent with Benchmarks
A cautious observation that an unidentified model’s resource use and practical performance felt unusual relative to its benchmark results. The speaker did not call it bad, only inconsistent or difficult to interpret in use.
- 6:28:50
Improving an Ad with a Stronger Opening Hook
Practical feedback on an ad performing below average: continue the experiment to gain experience, then strengthen the opening hook if performance remains at the same level.
- 6:29:56
Why Per-Million-Token Price Is Not Enough to Evaluate a Model
A discussion of evaluating an unidentified model by actual task outcomes and total token consumption rather than input/output unit prices alone. One task reportedly produced an unsatisfactory result and had to be redone, illustrating how heavy token use can make a model expensive in practice even when its unit pricing appears reasonable. The conclusion remained observational rather than a complete controlled comparison.
- 6:34:41
A Read-Only Executive Coordination Agent
The demonstration shows an agent that can read and summarize company records but cannot send messages, publish, or edit. The proposed workflow is to interact with one central executive agent that relays information to other agents and reports back, instead of addressing each agent separately.
- 6:35:32
Fast Voice Interaction via the OpenAI API
The speaker discusses the character’s changed voice and says the OpenAI API is being used, making the live voice interaction feel responsive.
- 6:40:07
Customizing an Agent’s Avatar
An experiment in choosing a visual identity for the agent, preferring an original image made by the person being addressed over a recognizable NFT character, followed by discussion of editing the image and connecting it to an avatar model or tool; the final tool name is not certain from the captions.
- 6:46:47
Profile-image backgrounds and the desired level of realism
The speaker discussed background removal while considering consistency across an image set, suggested periodically rotating a profile image, and asked whether the intended character should look fully human or deliberately artificial.
- 6:51:01
Considering a global audience for a product
The speaker encouraged a participant to consider giving a project global scope instead of limiting it to one national market, explicitly presenting this as personal encouragement and opinion.
- 6:52:12
Balancing authenticity and visual polish in marketing
During the workflow, the speaker observed that replacing the image was unusually slow, without identifying the cause or tool. The image was then described as based on a real person with some facial details altered, while its polished look was intentional; the speaker endorsed conspicuous polish and strong contrast between skin tone and white clothing as an attention-grabbing marketing choice.
- 6:53:47
A Consistent Visual Ambassador for the Product
A discussion of using a recurring character or visual element as a brand ambassador throughout an application, with the observation that the background gives the design useful contrast.
- 6:56:31
Testing an AI Assistant’s Answers About Agent Activity
A practical test of whether an assistant could summarize another agent’s activity from logs and calendar data. The conversation exposed conflicting answers about completed versus scheduled work, after which the assistant clarified that prospect sources and selection reasons required more detailed research or CRM records that were not visible to it. This pointed to a need for better integrations and more precise answers.
- 7:00:51
Customizing a Conversational Assistant’s Voice
An experiment in directing the assistant to use a particular character voice, followed by lighthearted comments about its tone and interest in hearing how it performs in English.
- 7:03:58
Keeping Voice Conversations Active During App Navigation
A problem was identified where the voice conversation is interrupted when switching tabs or navigating between parts of the application, with a proposed fix to keep it active throughout navigation.
- 7:05:39
Reviewing a Professional AI-Assisted Website
The speaker was highly impressed by an audience member's AI-assisted website, describing the result as professional and marketable. He noted that using AI alone does not guarantee a strong outcome and encouraged aiming the project at a global audience.
- 7:13:10
Letting Codex Run in the Background Despite Its Slowness
The speaker noted that Codex was slow but could continue working in the background while he stepped away, with a plan to test the audio before ending the stream.
- 7:20:35
Holding Off on Judging a Project Idea
Asked whether a project idea was good, the speaker said its surrounding presentation looked good, but he needed to understand the idea more deeply before evaluating it.
- 7:20:57
Generating Ideas from Defined Criteria
In a separate response, the speaker advised defining specific criteria first, then asking AI to generate ideas based on them, emphasizing that the chosen criteria are the key factor.
- 7:21:21
An MCP and Plugin Marketplace Idea
The discussion raised the idea of building a large marketplace for MCPs and plugins, with the speaker requesting examples and competitors to better understand the space and compare existing offerings.
Got a question about something in this stream?
Ask in the group