CHAPTERS
- 0:00 – 0:30
Why GPT‑5.5 is a big deal: powerhouse coding + a personal “hack test”
Claire frames the episode as a hands-on review of GPT‑5.5 after weeks of testing, teasing that it enables advanced coding feats she couldn’t do with prior models. She introduces her personal benchmark: getting an AI to help her “hack” a tiny retro PC-style Bluetooth speaker/screen.
- •Weeks of real-world testing of GPT‑5.5/GPT‑5.5 Pro
- •Claims: noticeably stronger for advanced coding and complex problem-solving
- •Episode promise: show concrete builds and a high-bar evaluation
- •Sets up the proprietary device “hack” as the ultimate intelligence test
- 0:30 – 3:01
Model availability, positioning, and the “intelligence tax” (pricing + ROI)
She explains where GPT‑5.5 and GPT‑5.5 Pro are available (ChatGPT and Codex, not API yet) and emphasizes OpenAI’s pitch: higher capacity and higher efficiency. She then lays out the steep token pricing and argues the cost can be justified when the model expands ambition, not just speed.
- •GPT‑5.5 and GPT‑5.5 Pro: in ChatGPT/Codex; not in API yet
- •OpenAI’s claim: smarter + more token-efficient
- •Pricing: $5/$30 per million input/output tokens; Pro $30/$180
- •ROI framing: speed gains vs. enabling bigger, previously infeasible work
- 3:01 – 4:34
The “intelligence overhang” problem in ChatGPT: do most users need this much brainpower?
Claire describes her uncertainty about GPT‑5.5’s value for typical ChatGPT use, arguing most everyday problems don’t require “super intelligence.” She predicts developers will benefit most, and calls for better product packaging to make the intelligence accessible for non-engineers.
- •ChatGPT context: hard to find tasks that truly require GPT‑5.5-level intelligence
- •Concept: “intelligence overhang” (more capability than typical demand)
- •Developer-centric model benefits vs. consumer/general enterprise usage
- •Request to viewers: share hard problems to test in ChatGPT
- 4:34 – 7:06
ChatGPT test: building a subtraction-learning app (and the cost of long “thinking”)
She tries GPT‑5.5 in ChatGPT by asking it to create an app to teach her child advanced subtraction. The model plans extensively (17+ minutes), produces a workable multi-module app, but she questions whether that level of deliberation is worthwhile for simple “vibe coding” tasks.
- •Prompt: build an app for two- and three-digit subtraction learning
- •Observation: GPT‑5.5 spends ~17 minutes thinking/planning
- •Output includes modules like mini-lessons, word problems, read-aloud
- •Verdict: ‘fine’ result; form factor and wait time may not fit casual use
- 7:06 – 8:08
Switching to Codex + GPT‑5.5 Pro: faster feedback loops and parallelized work
Claire pivots to where she feels GPT‑5.5 truly shines: Codex with the Pro model. She describes a dramatic boost in execution speed and autonomy, letting her run multiple tasks in parallel and finally tackle long-standing tech debt in the ChatPRD codebase.
- •Codex + 5.5 Pro as the standout setup
- •Immediate felt gains: speed/efficiency and less babysitting
- •Ability to launch many tasks concurrently
- •Sets up concrete engineering use cases from the ChatPRD codebase
- 8:08 – 9:39
Use case #1: turning a security scan CSV into grouped fixes (and validating via pen test)
She describes using OpenAI’s Codex security tooling to scan ChatPRD and then batching remediation work. By uploading a CSV of findings, she has Codex group issues thematically, propose changes, implement fixes, and then rely on human review and external pen test results for validation.
- •Run threat assessment/security scan; identify low-severity issues
- •Upload CSV to Codex; ask for thematic grouping + architectural review
- •Codex proposes and implements remediation changes in batch
- •Quality check: human code review + subsequent pen test comes back clean
- 9:39 – 11:11
Use case #2 (part 1): the ‘millions of chats’ legacy-format migration problem
Claire outlines a thorny data/tech-debt challenge: millions of stored chat records in inconsistent legacy formats due to evolving provider response schemas and attachments/tools. She explains why this is hard—semi-structured data, many edge cases—and how they’d been patching it repeatedly.
- •ChatPRD has millions of chats stored across legacy formats
- •Providers (OpenAI/Anthropic) changed response shapes over time
- •Data includes tools/attachments: complex, lightly structured payloads
- •Prior approach: repeated patches; new edge cases constantly appear
- 11:11 – 12:43
Use case #2 (part 2): one-shot migration + building a scalable validation harness
GPT‑5.5 Pro in Codex produces a migration solution covering ~98% of known edge cases. Claire then asks it to create a programmatic testing system (ideally a CLI) that replays threads through providers in a production-like test environment to smoke test and uncover remaining issues.
- •Codex builds a migration solution that covers ~98% of identified edge cases
- •Pull production-like data into a local/test environment
- •Request: programmatically test every thread against providers
- •Goal: scalable, team-usable CLI workflow for ongoing validation
- 12:43 – 15:45
Six hours of autonomy: subagent loop, minimal oversight, and error rates collapsing
She highlights an extended autonomous run: Codex works for nearly six hours without follow-ups, implementing smoke tests and iterating through validation. After running across ~2 million rows, only one edge case remains, and monitoring shows a sharp drop in errors—arguing AI coding can increase quality.
- •Prompt style: ‘I trust you’ + delegation to spawn subagents and iterate
- •Near-zero supervision: only a single approval in sandbox
- •Result: 2M rows tested; only 1 edge case missed
- •Observed impact: Sentry error rate drops significantly; quality narrative flips
- 15:45 – 17:46
The ultimate intelligence eval: hacking a proprietary Chinese Bluetooth screen/speaker
Claire introduces her long-running personal project: controlling a Divoom MiniToo Bluetooth device display programmatically from the terminal. Previous attempts with Claude Code, Opus, and GPT‑5.4 failed; the challenge is reversing proprietary Bluetooth messaging and bitmap encoding.
- •Device: Divoom MiniToo retro PC-style Bluetooth speaker + screen
- •Goal: bypass official iPhone app; control display via CLI/terminal
- •Difficulty: proprietary Bluetooth transport + unknown encoding/compression
- •Prior models/tools couldn’t solve it; months of manual effort
- 17:46 – 18:48
Reverse engineering workflow: packet sniffing, logs, and ‘desperate prompting’
She details the non-AI groundwork: enabling Bluetooth developer profiling on her phone, using a packet sniffer to capture traffic when the iPhone app sends images, and collecting logs. She feeds these artifacts to GPT‑5.5/Codex and asks it to do “anything” to make the display controllable.
- •Set up Bluetooth profiling/debugging on iPhone
- •Capture packets while using the official app to send images
- •Provide packet/log data to Codex as the core evidence
- •Prompt: infer protocol and implement a way to send images programmatically
- 18:48 – 21:52
Breakthrough: a CLI that drives the screen + a Codex notification hook demo
GPT‑5.5 succeeds—building a command-line tool that encodes/decodes bitmaps and sends them to the device. Claire demonstrates a notification hook wired into her Codex config so the device beeps/displays a message when Codex completes a task, cementing this as her standout ‘delight moment.’
- •Codex figures out bitmap handling and Bluetooth messaging
- •Build outcome: terminal-driven CLI that updates the device display
- •Bonus: notification hook integrated with Codex workflow
- •Live demo: device beeps/displays truncated ‘Your move’ message
- 21:52 – 23:36
Wrap-up: smart + efficient + autonomous—plus Codex ‘baked potato personality’ tweaks
Claire summarizes her verdict: GPT‑5.5 is extremely capable, efficient, and can sustain long autonomous work on complex engineering problems. She notes Codex’s default dull persona but shares a `/personality` tip to make it friendlier, then closes with a call for viewers’ hardest ChatGPT tests.
- •TL;DR: favorite ‘senior/staff engineer’ model for complex tasks
- •Strengths: autonomy, efficiency, durability on long-running loops
- •Tradeoff: expensive token costs (‘intelligence tax’)
- •Codex tip: use `/personality` to avoid the default dull style
