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
- 0:41:20
Persistent AI Assistant vs. a Manually Configured Chat Session
In a live exploration, the assistant describes retaining work context, tracking tasks, providing updates, and helping with research, writing, translation, and organization. Compared with a session equipped with memory, tools, and scheduling, its stated distinction is mainly that those parts are already integrated; much of the concept could be reproduced through manual configuration.
- 0:45:27
Studying Computer Science to Understand AI Models
Luqman advises keeping up with AI developments and continuing in computer science or software engineering, with an emphasis on understanding how models work internally rather than only using them as tools.
- 0:46:43
Personal Command Centers Built with Graph Engineering
Luqman argues that graph-based systems are not a single ideal solution for everyone: each person can shape a command center around their own goals if they understand the concept and have a clear vision. He expects AI tools to increasingly help turn that vision into a personalized system, informed by experimentation and lessons from mistakes.
- 0:53:08
Distribution Can Be the Bottleneck Even for a Strong Product
The speaker argues that product quality alone does not guarantee fast growth: marketing, traffic acquisition, and distribution are critical, while investment may also help depending on the circumstances.
- 0:53:59
Treating AI Model Benchmarks with Caution
The speaker discusses anticipation for a new Gemini release while expressing doubts about published benchmarks, reflecting a preference for practical evaluation before accepting comparative claims.
- 0:54:46
Bank API Requirements Depend on Country and Business Registration
Answering a question about bank API integration, the speaker says requirements vary by bank and country and suggests that a registered business may make general banking interactions easier.
- 0:55:59
Reworking the Project’s Pricing and Monetization
The speaker says the project’s pricing structure, monetization strategy, and the relationship between features and plans are undergoing a major redesign and are still in progress.
- 0:57:31
The Shopify Project’s Vision: Help Merchants, Then Expand Beyond Shopify
The speaker describes a Shopify project that helps merchants appear on AI platforms. The longer-term vision is an independent platform beyond Shopify, presented as an ambition contingent on strong success or suitable investment.
- 0:58:51
Balancing Model Intelligence, Price, and Access
The speaker cautiously weighs stronger model intelligence against cost and plan availability, then notes that similar pricing and benchmark results make market comparisons confusing. Some model names in the captions are unclear, so their identities are not asserted.
- 1:00:53
Headline Pricing Is Not Enough to Compare Coding Models
Luqman argues that input and output prices alone can be misleading because models may consume and allocate tokens with different efficiency. In his experience, Codex and Claude are close in quality, so he looks for a metric that combines token cost with usage efficiency before judging their real value.
- 1:05:21
Reducing AI Tools’ Access to Company Data
Responding to concern about connecting an API to sensitive data, Luqman suggests defining a policy and restrictions that specify what the tool must not access. The discussion is brief and general, with no claim that particular controls were implemented or tested.
- 1:06:17
A Channel Documenting the Public Journey of Building and Selling AI Solutions
Luqman explains to a new viewer that the channel publicly documents his journey of building with AI and trying to enter the market, including both obstacles and successes. He also describes a community where developers, business owners, entrepreneurs, and students exchange knowledge.
- 1:07:36
Mixed Views on Grok and a Willingness to Reassess It
Luqman says his experience with Grok was disappointing despite its speed and possible value as a lower-cost option, while a participant offers a positive view. He responds openly and decides to revisit the experience in case he had not evaluated the tool fairly.
- 1:08:55
Reviewing an Unsatisfactory Practical Experience with Grok
Luqman recalls using Grok for about a week and attributes his dissatisfaction to its usage pattern, responses, and way of working. This remains a personal usage observation rather than a benchmark or a general verdict on the model.
- 1:13:04
Adjusting Model Effort to Usage
A general explanation of varying reasoning effort between medium and high levels according to remaining usage, while using the maximum and lowest settings only rarely.
- 1:13:39
Building an Agent Team with Shared Roles and Dependencies
An answer on organizing agents that use different harnesses for planning and coding. The proposed setup uses a shared system repository, a structured tracking file, a defined contract or purpose for each agent, a common goal, and a workflow whose dependencies require agents to exchange outputs. The speaker qualified this as an organizational concept and said he had not yet tested this system for coding, which he currently prefers to do himself with AI assistance.
- 1:18:56
Configuring Models and Harnesses per Agent
The discussion explains that each agent's setup can be changed, including its harness, model, and reasoning level, while token usage can be reviewed to help select an appropriate configuration.
- 1:21:54
Start With Your Own Vision Instead of Copying an Existing System
The speaker advises the questioner to define their own vision rather than copy an existing design, because that vision may produce a stronger result. He gives a general example of a workflow system handling research, visual generation, writing, and publishing, while noting that he may have overinvested in one part of it.
- 1:25:05
When Prompt Engineering Can Limit an Agent
The speaker offers the qualified opinion that some overly prescriptive prompts can do more harm than good by forcing a particular documentation or folder structure. In his view, stating the goal and allowing the agent to choose the method may work better over time—a shift from prompt-driven to goal-driven behavior.
- 1:26:45
Agent Delegation and Visibility Into Sessions
The speaker views a delegation tool or skill positively but is uncomfortable with unseen delegation through a CLI, preferring session visibility in a graphical interface. He mentions a prior experiment using computer-control capabilities for Codex to interact with Claude, while qualifying that the approach was slower and that his personal use of the delegation tool was limited. He reiterates that users should specify the desired outcome rather than prescribe implementation details unless they have strong engineering expertise.
Got a question about something in this stream?
Ask in the group