Giving My AI Employees a Brain
Want 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 centers on building a multi-agent automation system, especially a local pipeline that turns stream recordings into transcripts, captioned vertical clips, and reviewable assets before distributing approved content across platforms. Practical work exposed agents that stalled, exceeded task scope, or became difficult to coordinate, alongside device-load problems; the discussion therefore covered heartbeats, retries, end-to-end testing, phased activation, human approval gates, and local-versus-remote execution. Operational side topics included streaming hardware, distributing recording workloads, 4K capture, and copyright-safe music, without reaching a definitive diagnosis for the stream failures. A brain-like visual graph was explored for understanding agent relationships and status, though its practical value remained unresolved. Other proposed workflows included automated research-to-content production and Sentry-assisted bug triage, with limits on automatic changes and publishing. Audience questions covered getting started with AI-assisted coding, Codex, Claude, lower-cost alternatives, subscriptions, GUI versus CLI, and model selection; comparisons were framed as personal judgments rather than conclusive benchmarks. On product and marketing strategy, the speaker favored organic traffic, differentiation, and adequate data before judging pricing or polishing details, while discussing website clarity, client trust, open-source licensing, Arabic-support gaps, building an Arabic AI community, and AI’s uncertain effect on SaaS. Additional side discussions covered sleep debt and experimental stream schedules, as well as reported funding and valuation claims about Safe Superintelligence and the role of founder reputation; the captioned figures remained conflicting and were not independently verified. The stream also examined privacy, persistent memory, agent and superintelligence risks, and infrastructure concentration, carefully distinguishing speculation and reported stories from established facts. Overall, it documents varied useful designs, experiments, and open questions—not a completed or proven automation system.
118 moments from the stream
- 1:42:08
Comparing GPT and Claude for Computer Use
The speaker says GPT's computer control had recently worked better for him than Claude's, which he felt could take over the machine and make it difficult to use.
- 1:43:06
ChatGPT Ads Versus Organic Growth
The speaker discusses, with uncertainty, the availability and geographic eligibility of ads inside ChatGPT and asks whether viewers have seen them in Egypt. He suspects the cost per click may be high and prefers to address organic marketing and traffic before experimenting with ads.
- 1:46:26
Running Multiple AI Agents in Parallel
The speaker describes running two tasks in parallel: having GPT inspect a bot system and understand how its “employees” are organized, while trying another strategy with Claude for interaction around stream content. His near-term goal is to activate these agents and explore the workflow.
- 1:49:37
Preventing an AI Agent from Exceeding Task Scope
The speaker objects when an AI agent performs unrequested work and testing after being told only to build and wait for a file in a target folder to test the workflow. He describes a recurring problem where the agent advances a task beyond the requested stage without permission.
- 1:51:24
Phased AI Team Activation with Human Approval Gates
The speaker says making the whole system operational in one pass is unrealistic and proposes first understanding its configuration and relationships, then activating one department at a time. Automation can be broad, but sensitive content or posts made in his name should retain a human approval step.
- 1:57:05
An Automated Stream-to-Clips Review Pipeline
The speaker describes a workflow that watches a recordings folder, separates audio and video, transcribes the stream, selects potentially useful clips, reformats them vertically, adds captions, and reduces noise. Outputs enter a human review area, and only approved clips proceed to scheduling on video platforms. He agrees to try the workflow immediately, but this page does not contain the test result.
- 2:02:11
A Local Pipeline for Turning Streams into Clips
The speaker outlines a local workflow that picks up a stream file, separates its audio, transcribes it, and passes the transcript to GPT CLI to identify useful sections and command the machine to cut, caption, and render them. The results can be reviewed in a folder and scheduled for upload after approval.
- 2:04:32
Moving Gradually from Human Review to Full Automation
The speaker discusses keeping human approval at first, then potentially automating it after confidence in output quality develops. In his view, solving the harder video workflow could make automating posts and blogs easier, while the practical value of some cloud-computer components still needs verification.
- 2:07:35
Running, Polishing, and Fallback-Planning the Workflow
The speaker follows the workflow as it processes a file: Whisper handles transcription, then GPT is given the text to select clip locations before work returns to the local machine. Responding to a suggestion for a deadline and fallback, he says the system worked in an earlier test but still needs polishing, and favors building a fully automated process while starting with a smaller scope.
- 2:10:37
Finding Computer History Under Computer Use
The speaker points viewers to a Computer History option under Computer Use, following an earlier discussion of using it with screen captures and summaries and why it was not suitable for running directly on the live stream.
- 2:11:13
Persistent AI Memory and Privacy Boundaries
A discussion of a vision attributed to Sam Altman in which a model knows details of a user’s life and offers advice. The speaker was wary of continuous monitoring, questioned how computer activity could be understood without recording the screen or audio, and stressed the need to exclude private activity.
- 2:14:23
Caution Before Trying an AI Tool
The speaker preferred to wait until others had tried the tool and the AI community had discussed it, linking his hesitation to his personal unease with OpenAI’s philosophy. Separately, he mentioned an account attributed to Dario involving security and privacy concerns. These are attributed opinions and accounts, not test results.
- 2:15:14
Understanding Permissions Before Automatic Approval
The participants discussed an access request guessed to involve Python and transcription, while explicitly acknowledging that they did not know what the tool was actually doing. An approve-all option then appeared, and the speaker said permission bypass was enabled.
- 2:18:58
Automating Short Clips from a Transcript
The speaker explained a workflow where GPT reviews a transcript, selects relevant sections, and instructs a local machine to cut them. The specifications include limiting silence, synchronizing Arabic captions with speech, producing vertical output, and saving results to a review folder; processing was still underway at the time.
- 2:27:59
Why Investors May Back a Founder Before the Product
The speaker argues that investors may back a founder—not only a company—and relates this to the founder’s former OpenAI role and Safe Superintelligence Inc. He expresses skepticism about reports of an enormous valuation and fundraising at a stage with no publicly visible product. The captioned figures conflict, so no specific amount can be established from this page.
- 2:31:38
Discussing Safe Superintelligence’s Funding and Valuation
The speaker discussed reported financing and valuation figures for Safe Superintelligence, mentioning its September 2024 launch and a rapid subsequent increase. Because the captions blur funding and valuation, these remain speaker-attributed claims rather than independently verified facts.
- 2:32:23
Codex Plus or Claude Pro?
Asked briefly to choose between Claude Pro and Codex Plus, the speaker selected Codex Plus without giving further reasoning in this segment.
- 2:38:51
Moving Task Execution Beyond the Local Machine
The speaker explained that the current workflow delegates tasks such as preparing captions to his local machine, while he wants the AI system to execute them directly for faster completion. The demo was experiencing a stall, and he did not claim the alternative was already working.
- 2:41:12
A Camera That Auto-Zooms Using Board Markers
The speaker explained that the camera detects four black markers around the board and uses them for automatic zooming, while noting that he had intended to use the board more but had not yet done so.
- 2:41:52
A “Brain” Graph Connecting AI-Agent Work
The speaker asked for a graph connecting agents, departments, and workflows, with each task leaving a record that the graph could capture and link across the system. He stressed that the displayed version was an inactive visual sample and that the eventual goal was operational understanding, not decoration alone.
Got a question about something in this stream?
Ask in the group