Testing GPT Astra vs Fable 5.1
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 a practical, partly blind comparison of GPT Astra and Fable 5.1 using matched prompts for games, an SVG drawing, landing pages, and agentic tasks. Results were mixed: Astra often appeared stronger visually and in the controls of some games, while Fable sometimes delivered more challenging gameplay or more inventive design ideas. Speed varied by task and audience preferences split, so no clear overall winner emerged. Luqman discusses the limits of single tests, benchmarks, and model judges, arguing for evaluation grounded in domain expertise, bias reduction, critical-thinking and ideation tests, alongside efficiency, price, usage limits, and tool integration. The discussion also covers routing planning and execution to different models, sub-agent orchestration, documenting and compacting long coding sessions, preconfigured bots, video automation, backups, and changing developer roles. Claims about AGI, job displacement, and near-term end-to-end AI development remain attributed predictions or concerns rather than established conclusions.
104 moments from the stream
- 2:44:38
Task Decomposition and Sub-Agent Orchestration
Luqman and the audience discuss whether to give one agent a large task or separate planning from execution and assign detailed subtasks to specialized agents. He explains that modern agent systems may create sub-agents with separate contexts and coordinate them within one session, reducing the need to manage many sessions manually, while noting that the right approach depends on task size.
- 2:56:02
Preconfigured Bots vs. Building Custom Bots
Luqman explores an update that allows preconfigured bots to be added, including bots for SEO and AEO tasks, and compares them with bots he builds himself. In his view, the lack of direct integration between competing tools such as Codex and Claude can force an agent to control the computer, which he describes as slow and cumbersome; the practical question is whether a ready-made bot adds real value.
- 3:01:21
Backing Up Projects Locally and on GitHub
Answering an audience question, Luqman says he keeps project copies both locally and on GitHub, reducing the likelihood of losing both sources at once. He adds that he may strengthen security measures as the work develops, without claiming that the current setup provides complete protection.
- 3:08:45
Comparing Games Created by Astra and Fable
Luqman tried two versions of a delivery game and compared movement, jumping, obstacles, scoring, and increasing speed. After revealing their identities, his personal view was that Fable offered harder, more challenging gameplay, while Astra had stronger graphics and design; he stressed that the differences were small and the judgment subjective.
- 3:16:06
How Models Handle a Missing Input
In an audit test, Astra requested the missing file before proceeding, while Fable noted its absence and tried to continue with the available material. Luqman presented the two behaviors as a comparison without concluding that either was definitively better.
- 3:18:25
Using a Judge to Compare Two Answers Objectively
Luqman proposed putting two answers head to head before a judge that would cross-reference them, first define what ‘winning’ means, and then decide which was stronger in terms of strategy. This was a general evaluation framework, not a reported final verdict.
- 3:23:52
GPT Astra vs Fable 5.1: Graphics, Gameplay, and Efficiency
Luqman felt GPT Astra performed better on graphics while Fable was better at gameplay. He described the quality gap as small and emphasized efficiency as a deciding factor, while noting that he had not tracked usage closely enough for a firm assessment.
- 3:28:26
Comparing Two Audits with a Judge Model
Luqman described a test methodology in which two models deeply analyze the same material and a separate judge compares their outputs. The judge’s conclusion was not presented in this section.
- 3:29:34
Saving and Transferring Long Responses with Tables
The discussion highlighted the difficulty of moving long responses out of coding-assistant applications when they contain many blocks and tables. Manual copying seemed inefficient, and it was unclear whether the structure would remain readable; no confirmed saving method was reached here.
- 3:34:32
Testing an AI Model in a Domain You Know Well
The speaker argues that deep subject-matter expertise makes it easier to judge an AI model’s real capabilities. He proposes testing strategic reasoning, ideation, and digital marketing to see whether the model can surface worthwhile angles an expert has not previously considered.
- 3:37:50
Creating a Handoff Prompt Between Models
The speaker asks for a concise handoff prompt that gives another model the work’s background, current state, and next task so it can continue the research in parallel from the same point. Private project and financing details are omitted.
- 3:40:16
Recovering Repository Context and Reviewing Unpushed PRs
After losing track of the work state, the speaker asks the model to inspect what is happening and re-establish the current position, noting that several pull requests may not have been pushed yet.
- 3:41:05
A Changing Shorts Workflow and StreamLadder Quality Loss
The speaker says his previous workflow for producing short clips for platforms such as TikTok has changed substantially, and that he noticed StreamLadder noticeably reduces clip quality.
- 3:44:55
Plan implementation and verify visually before merging
The workflow turns unbuilt proposals into an implementation plan, then requests real screenshots of the result before merging.
- 3:47:39
Why model comparisons need more than one winner
An evaluation excerpt flags possible model-as-judge bias and says mechanical checks were run before either report was read. Luqman says the two reports found different issues, making joint use potentially more useful than declaring a universal winner.
- 3:52:36
A qualified view of what reaching AGI would mean
Luqman doubts a claim that AGI has been reached. In his view, AGI is more closely tied to critical thinking, ideation, creativity, and genuinely unexpected ideas than to design quality alone; he also finds the cited metrics unusual.
- 3:53:47
Build-test results depend on the task
Luqman recaps earlier tests involving a Flappy Bird-like game and a koshary game and page, describing the outcome between coding tools as relative and task-dependent.
- 3:58:31
Consistent skill use shapes coding-agent preference
Luqman says some coding agents appear to use skills more consistently than alternatives, raising his selection standard, while noting that he had not yet evaluated another option.
- 4:03:32
Astra and Fable 5.1 are close enough for efficiency to decide
Luqman says he cannot identify a clear winner between Astra and Fable 5.1: their philosophies differ, but the practical gap appeared very small. He therefore frames efficient work output as a useful deciding criterion while keeping the verdict tentative.
- 4:05:13
Testing Critical Thinking and Creativity in AI Models
The speaker says differences in design and debugging may be narrow between models, so he is more interested in testing critical thinking, ideation, creativity, and the ability to think outside the box.
Got a question about something in this stream?
Ask in the group