500x Engineer

Modern Web Guidance: I Tested Google's Free Skill in Codex

I tested Google's free Modern Web Guidance skill in Codex on an account-settings page for a made-up notes app. GPT-6.1 Sol received the same page request twice plain and twice guided. This episode pairs that small experiment with Google's separately reported evaluations.

What we found: both plain runs scored 2 of 10 and both guided runs scored 9 of 10 on our prewritten built-in browser-feature checklist. All four pages passed all ten Chrome click-through checks. Plain app.js files had 156 and 165 non-blank lines; guided files had 202 and 271. The guided code added browser features alongside support checks and fallback paths. Google's legacy-scan flagged one textarea-resizing pattern in each plain page and none in either guided page.

Google reports Codex improving from 54% to 82%, Claude Code from 57% to 88%, and Antigravity from 58% to 94% in its November 8 evaluation. Those are Google's results. Our trial used one prompt and one model, with Chrome-only interaction checks. We also cover default telemetry, the DISABLE_TELEMETRY=1 opt-out and the scanner coverage Google describes as roughly 20% of guides. My recommendation is to trial a small component, specify browser support and test the fallback behavior. Opinions are mine.

Full write-up: https://www.500xengineer.com/blog/modern-web-guidance-test

Join the free Skool community: https://www.skool.com/500xengineer

CHAPTERS:

0:00 Intro
0:52 What shipped
2:01 Google's numbers
2:56 My test
4:50 The code got longer
7:11 Here's what I'd do
9:10 The takeaway

SOURCES:

https://developer.chrome.com/blog/modern-web-guidance-v1
https://registry.npmjs.org/modern-web-guidance
https://github.com/GoogleChrome/modern-web-guidance
https://github.com/GoogleChrome/modern-web-guidance/releases/tag/v1.0.0
https://developer.chrome.com/docs/modern-web-guidance/modernization-workflow
https://developer.chrome.com/docs/modern-web-guidance/get-started

Local test evidence: the episode's lab/README.md and plain and guided source files, documented in the full write-up. One prompt, one model, two runs per condition; interaction checks in Chrome only. The marketplace installation and fallback behavior in Safari and Firefox were not tested.

DISCLOSURE:

The narration is AI-generated with ElevenLabs, using a clone of Elon's own voice. The short opening clip of Elon is an AI-generated avatar made from his likeness.

Can AI Find Security Bugs? I Tested Codex on My Own App

I tested GPT-6.1 Sol in Codex through my ChatGPT subscription on a tiny app with planted bugs and fake secrets. The review was read-only, with no execution or exploitation. This episode pairs that limited experiment with OpenAI's reported Linux campaign and Mozilla's record of Firefox fixes.

What we found: medium effort flagged six of eight planted patterns in 68 seconds; high effort flagged the same six in 117 seconds. The reviewer correctly explained why the discarded reset value and unused fake mail credential were harmless in the original code. After I revised the app, medium effort flagged eight of eight in 61 seconds. All three runs left the two retained safe decoys alone, and all three caught a real flaw in my answer key. The changed app and small sample limit what those scores and times establish.

OpenAI reports 948 unique Linux privilege escalation vulnerabilities counted through distinct patches. OpenAI's $359 average covers successful exploit development alone; its reported whole-campaign cost was $1,213,430 in API credits. We discuss the evidence, access restrictions and human checks behind those claims. My recommendation is an authorized read-only review followed by human validation, a verified patch and a regression test. Opinions are mine.

Full write-up: https://www.500xengineer.com/blog/gpt-6-cyber-ai-bug-hunt

Join the free Skool community: https://www.skool.com/500xengineer

CHAPTERS:

0:00 Intro
1:01 What OpenAI actually did
2:44 What it cost
3:42 Mozilla's receipt
4:59 The catch
6:02 My test
9:02 Here's how you can use this
10:28 The takeaway

SOURCES:

https://openai.com/index/a-thousand-fewer-paths-to-root/
https://openai.com/index/introducing-gpt-6-cyber/
https://deploymentsafety.openai.com/gpt-6-cyber
https://openai.com/daybreak/
https://learn.chatgpt.com/docs/security
https://www.mozilla.org/en-US/security/advisories/mfsa2026-118/
https://openai.com/index/the-defenders-window/
https://openai.com/index/patch-the-planet/

Local test evidence: the episode's lab/README.md, documented in the full write-up. One small app, three runs, one model; static review only.

DISCLOSURE:

The narration is AI-generated with ElevenLabs, using a clone of Elon's own voice. The short opening clip of Elon is an AI-generated avatar made from his likeness.

Where Do Claude Code Tokens Actually Go? I Measured 6.5 Billion

I run Claude Code all day on one Mac, so I read a full week of its local logs and sorted every token by where it went: about 28,700 requests and 6.5 billion tokens from October 30 to November 5, 2026.

What I found: 95% of the tokens were cache re-reads, because every step re-sends the whole conversation. Priced at Anthropic's API list rates (I'm on a flat plan, so this isn't my bill), the split flips: cache writes are 49% of the price, cache reads 39%, output 12%. The longest 10% of threads used 74% of all tokens. Tool output was only about 0.2% as new text, but by my rough estimate it came back about 68 times its size as re-reading. Then RTK's counter versus my logs, Ponytail's benchmark, and the order I'd do the fixes in: short sessions first. Opinions are mine.

Full write-up: https://www.500xengineer.com/blog/where-claude-code-tokens-go

Join the free Skool community: https://www.500xengineer.com/skool?utm_source=spotify&utm_medium=podcast&utm_campaign=ep10-claude-code-tokens

CHAPTERS:

0:00 Intro
0:59 The receipt
2:27 Which tokens weigh the most
3:37 Before you type a word
4:52 The long-session tax
5:57 The shell-output trick
7:56 Less code
8:51 Here's how you can use this
10:22 The takeaway

SOURCES: the full source list is in the YouTube description: https://www.youtube.com/watch?v=S9cOkVVRrKg

DISCLOSURE:

The narration is AI-generated with ElevenLabs, using a clone of Elon's own voice. The short opening clip of Elon is an AI-generated avatar made from his likeness.

AI Usage Limits Keep Changing. What Breaks If Yours Do?

In the last few weeks, three AI companies changed the rules my video pipeline runs on: OpenAI says ChatGPT and Codex will share one usage allowance from December, Elon Musk said serving Grok Bot had become too expensive, and Anthropic's new usage policy takes effect on November 12, 2026. So I mapped the 12 outside services my pipeline depends on, measured what one episode uses, and unplugged two of them on purpose.

Short answer: 4 of the 6 steps that make a video run on someone else's servers. One episode used 9,255 voice characters, one $0.86 twin clip and about 65,600 Codex tokens for three thumbnails, and Claude Code isn't logged at all. With the twin service "gone", a free local lip-sync model re-synced an old clip to a new line in 26 minutes on my Mac (SyncNet 8.69 against 6.92 for the cloud take, with the caveat that this model is trained with a SyncNet-style checker). A thumbnail with no new AI picture rendered in about three seconds. Then the checklist: map, log, one tested backup per step, pin versions, keep your files at home, read the change pages. Opinions are mine.

Full write-up: https://www.500xengineer.com/blog/ai-usage-limits-platform-risk

Join the free Skool community: https://www.500xengineer.com/skool?utm_source=spotify&utm_medium=podcast&utm_campaign=ep09-platform-risk

CHAPTERS:

0:00 Intro
1:07 Three companies, three changes
2:49 The map
4:05 What one episode uses
5:39 The drill
7:41 Reading the new rules
8:37 The checklist
9:34 The takeaway

SOURCES: the full source list is in the YouTube description: https://www.youtube.com/watch?v=9nYit7yvA2k

DISCLOSURE:

The narration is AI-generated with ElevenLabs, using a clone of Elon's own voice. The short opening clip of Elon is an AI-generated avatar made from his likeness. Some images are AI-generated.

Claude Code Security: What Can Your AI Agent Actually See?

Within five days, two software updates went hunting for API keys: fake builds of the Claude Code extension on Cursor's extension marketplace (Oct 29, 2026) and a fake cPGuard auto-update on Linux servers (Nov 2, 2026). My AI agents hold keys to my paid accounts, so I built a lab with fake keys to see what a bad tool running inside an agent's shell can reach, then audited my own pipeline and fixed what I found.

Short answer: with fake keys exported in the shell, the helper found 5 of 5 key names and could read every value, plus the .env file, the SSH key and the cloud keys. With nothing exported it found 0 names, but the files stayed readable. Inside a macOS sandbox profile it found nothing and couldn't connect out. Then five fixes in order of effort: stop exporting keys, scope keys with an expiry and a cap, turn on the Claude Code sandbox and its credentials setting, slow down updates, and write a rotation list. General security hygiene from my own setup, not a professional audit.

Full write-up: https://www.500xengineer.com/blog/claude-code-security-agent-keys

Join the free Skool community: https://www.500xengineer.com/skool?utm_source=spotify&utm_medium=podcast&utm_campaign=ep08-agent-keys

CHAPTERS:

0:00 Intro
1:02 Two updates, one week
2:07 What they went looking for
3:53 The lab
5:26 My own setup
6:56 Five fixes, in order of effort
9:07 Did it hit you?
9:48 The takeaway

SOURCES: the full source list is in the YouTube description: https://www.youtube.com/watch?v=ibZu577BMz8

DISCLOSURE:

The narration is AI-generated with ElevenLabs, using a clone of Elon's own voice. The short opening clip of Elon is an AI-generated avatar made from his likeness. Some images are AI-generated.

Can a 128 GB Mac Run DeepSeek's 233 GB Model? I Tried Anyway

DeepSeek V4.1 Flash needs at least 233 GB at its smallest build. My MacBook Pro has 128 GB. So I tried anyway, on the Mac I work on every day, with Docker, three Claude sessions and Chrome open. Local models (Qwen3.8 and DeepSeek in LM Studio, llama.cpp and Ollama) against Claude on three jobs from my own Shorts pipeline: write a script, fact-check a script, fix a bug, with the scripts judged blind.

Short answer: the 233 GB build did run, paged in from the SSD, at 0.42 tokens a second (one sentence in 65 seconds). Qwen3.8-27B at 4 bits (16 GB) wrote 30 tokens a second with the Mac quiet and 14 to 17 under my normal working load, and it is the one I'd actually use. In the blind test both Claude models scored above every local build; on the fact-check and the bug fix, every model got it right. Every number in the episode says whether it ran with the Mac quiet or under working load. This episode shows my own tests and promises no results.

Full prompts and the results table: https://www.500xengineer.com/blog/can-128gb-mac-run-deepseek-233gb

Join the free Skool community: https://www.500xengineer.com/skool?utm_source=spotify&utm_medium=podcast&utm_campaign=ep07-local-ai

CHAPTERS:

0:00 Intro
0:53 What 128 GB buys
2:20 The test bench
3:21 The 27 billion model
4:40 Big model, small appetite
5:43 DeepSeek
7:44 The blind test
9:23 What I would run
10:12 Try it tonight

SOURCES: the full source list is in the YouTube description: https://www.youtube.com/watch?v=oWbFupMUuEU

DISCLOSURE:

The narration is AI-generated with ElevenLabs, using a clone of Elon's own voice. The short opening clip of Elon is an AI-generated avatar made from his likeness. Some images are AI-generated.

I Let AI Agents Run My Content Business for a Month

For about a month I've let AI agents run my content business: they research, write, check and ship the work, a lot of it while I sleep. This episode is the receipts from my software factory, measured from my own logs on Nov 3, 2026: one agent that read 675 million tokens after I asked it to add a bumper to one video, a usage limit that stopped a whole fleet at once, an upload that "succeeded" 20 times and uploaded nothing, and 11 Shorts in a row that approved themselves. The thread through all of it is trust: scaling agents comes down to a strict playbook they follow every time, with checks that prove they did.

The five rules: one agent, one job; budget before speed; verify the result, never the exit code; a separate checker for every claim; gate the expensive stuff.

The numbers: 338 background agents (snapshot Nov 3, 2026, 11:46 pm ET) read about 3.66 billion tokens between Oct 23 and Nov 3. Dollar figures in the episode are what the same usage would cost at Anthropic's API list prices (Claude Opus 5.5); I'm on flat monthly plans, so I didn't pay these amounts. No company, repo or ticket names are shown, and the 3D render project is private. This episode shows my process and promises no results.

Build your own: zStack, the multi-agent setup from episode 4, is free and open source at https://github.com/simsketch/zstack . A premium course on how to use it is coming to the 500x Skool, and the free community has starter material now.

Join the free Skool community: https://www.500xengineer.com/skool?utm_source=spotify&utm_medium=podcast&utm_campaign=ep06-factories

CHAPTERS:

0:00 Intro
0:54 What a software factory is
2:50 The agent that wouldn't stop
5:01 Limits hit everyone at once
6:04 Success that wasn't
7:10 The gate that faded
7:54 The playbook
8:44 Build your own

SOURCES: the full source list is in the YouTube description: https://www.youtube.com/watch?v=rySjU1CgOxg

DISCLOSURE:

The narration is AI-generated with ElevenLabs, using a clone of Elon's own voice. The short opening clip of Elon is an AI-generated avatar made from his likeness. This episode's lead agent and both fact-checkers ran on Claude Sonnet 5.5.

How I Make AI Explainer Videos for About a Dime (Full Pipeline)

My laptop makes four YouTube Shorts a day. This episode walks through the whole pipeline behind them, station by station, in plain words, with an optional "under the hood" layer for engineers: the idea queue, research with Claude Code, a fresh AI fact-checker that can fail a script, the hook score, the voice (an ElevenLabs clone of my voice), scenes written as web pages and rendered frame by frame with HyperFrames, every animation tied to a spoken word, our 3D camera layer, sound and loudness, the automatic checks, approval on my phone (it approves itself if I don't answer within two hours), scheduling and cross-posting.

The worked example is "How an AI picks its next word". I made that Short by hand in a chat with Claude, outside the production line; the station-by-station visuals were rebuilt for this video. The Telegram screens are recreations of the bot's real messages, and the YouTube Studio screen is a mockup.

What it costs: the voice for one Short uses about 5 to 15 cents' worth of ElevenLabs credits (about 9 cents on average across ten recent Shorts, measured on my account Oct 31 to Nov 2, 2026). Those credits come from my ElevenLabs Creator plan ($22/month), and the AI side runs on my Claude subscription, also a monthly plan. Rendering and checks add nothing extra. No income claims: this episode shows my process and promises no results.

Make your own: the course Make Explainer Videos with Claude Code is in the 500x Engineer Skool (paid Premium tier) with a starter kit. Lesson 1 is free in Track Previews, and the free Set Up Your Computer course comes first. Claude Code needs a paid Claude plan.

Join the free Skool community: https://www.500xengineer.com/skool?utm_source=spotify&utm_medium=podcast&utm_campaign=ep05-pipeline

CHAPTERS:

0:00 Intro
0:46 The machine at a glance
1:35 The idea
2:16 Research and the fact-check
3:15 The script and the hook
4:04 The voice
4:38 Every scene is a web page
5:25 Every move on a word
6:00 The 3D camera
6:43 Sound and loudness
7:24 The robot checklist
7:51 The cover and the approval
8:57 Scheduling and sharing
9:39 What it costs
10:27 Make your own

SOURCES: the full source list is in the YouTube description: https://www.youtube.com/watch?v=wmL81xCYgv8

DISCLOSURE:

The narration is AI-generated with ElevenLabs, using a clone of Elon's own voice. The short opening clip of Elon is an AI-generated avatar made from his likeness. Course links point to a paid membership tier.

Microsoft Research ran 1,024 AI coding agents on one codebase with no boss agent, and no human approved a single merge. The rules allowed merges with zero required approvals and the CI merge gate was off. The score still went up, and the agents started reviewing each other's work on their own. I go through what Microsoft shipped (Agensh, MIT), how the leaderless loop works, the fine print behind the 1,024 number, whether you can run it, and what I built from it for real teams: zStack.

The fine print: the 1,024-agent result is one task (pandoc), one model and one reported result from a 6-hour budget on 16 machines. Going from 128 to 1,024 agents added about 4 points. The five-task average stops at 128 agents. On the newer GPT-6.1 model, a single agent (56.2%) already edged past the 1,024-agent run (55.1%) on pandoc, one reported result each. No cost figures, run counts or independent reproductions have been published. Benchmark numbers are Microsoft Research's own. Figures are as published on October 31, 2026. The GPT-6.1 figures come from the Agensh README, not the paper.

zStack (say "Zee Stack") is my agent skill for day-to-day engineering work: it routes a goal to a playbook and an execution shape, makes the agent cite the principle behind each decision, by default sweeps for sibling bugs after every fix, puts an evidence packet on every PR, pins its own merges to the verified commit, and merges on its own only after you raise its trust level, per repo and per kind of change.
- zStack core (free, MIT): https://github.com/simsketch/zstack
- zstack-fleet (the pipeline layer for many repos and accounts) is a download for Premium members of the 500x Engineer Skool, a paid tier. Full fleet waves also need the VIP orchestrator (a higher paid tier).
zStack install link: https://github.com/simsketch/zstack
No income claims: these are tools, and results depend on your own work.

Join the free Skool community: https://www.500xengineer.com/skool?utm_source=spotify&utm_medium=podcast&utm_campaign=agensh

CHAPTERS:

0:00 Intro
0:46 What Microsoft shipped
1:35 The leaderless loop
3:03 The fine print
4:37 Can you run it?
5:07 zStack: what I built from it
7:06 The verdict

SOURCES: the full source list (Microsoft's Agensh repo and paper, licenses, critiques) is in the YouTube description: https://www.youtube.com/watch?v=uqARyWAADtA

DISCLOSURE:

The voices in this video are AI-generated with ElevenLabs: Elon's narration uses a clone of his own voice, and the mascot Xy uses a stock ElevenLabs voice. The short opening clip of Elon and his on-screen stills are AI-generated from his likeness. zStack is my own tool; the fleet add-on is part of a paid membership.

DESCRIPTION:

Google's top-scoring model, Gemini 4 Carbon, is here, and most of us can't use it yet. Earlier this month it was a leak: Business Insider reported Google staff testing it, and testers on X said it had shown up on LMArena. On October 28 Google made it official. I go through what was reported, what was rumor, and what Google and Artificial Analysis actually confirm. I also cover Gemini 4 Argon, the model you can use today, what Carbon would change for builders, and how I'd get ready for the day it opens up.

Figures are as published on October 29, 2026. Benchmark numbers credited to Google are Google's own claims; the Intelligence Index and hallucination figures are from Artificial Analysis. Leaks are labelled as reported (Business Insider) or rumor (tester posts on X). The launch-day section is my opinion and makes no promise of income.

Join the free Skool community: https://500xengineer.com/skool

CHAPTERS:

0:00 Intro
0:37 The leak trail
1:55 What Google confirmed
3:06 Leaks versus launch
4:15 Argon, the baseline
5:01 What changes for builders
6:04 Ready for launch day
7:10 The verdict

SOURCES:

- Google: Gemini 4 Carbon, our next frontier model (Oct 28, 2026): https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
- Artificial Analysis Intelligence Index v5.0: https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index
- Business Insider: Google employees test Gemini 4 Carbon (Oct 9, 2026): https://www.businessinsider.com/google-employees-test-new-gemini-4-model-argon-barium-carbon-2026-10
- X: @Lentils80 tester post: https://x.com/Lentils80/status/2112770669549068684
- X: @LuminaBench tester post: https://x.com/LuminaBench/status/2112829559158800732
- X: @Lentils80 tester post: https://x.com/Lentils80/status/2113246123959668854
- X: @LuminaBench tester post: https://x.com/LuminaBench/status/2112830849771004187
- TestingCatalog: Gemini 4 Carbon early test results: https://testingcatalog.com/gemini-4-carbon-early-test-results
- X: Google DeepMind announcement: https://x.com/GoogleDeepMind/status/2115489356437651654
- Google: Gemini 4 Argon, built to solve your hardest problems (Oct 13, 2026): https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/
- VentureBeat: Google unveils Gemini 4 Carbon (benchmark table from Google's chart): https://venturebeat.com/technology/google-unveils-gemini-4-carbon-retaking-the-worlds-no-1-spot-on-updated-harder-benchmarks-but-access-is-limited
- Artificial Analysis AA-3D leaderboard: https://artificialanalysis.ai/evaluations/aa-3d
- Artificial Analysis: Gemini 4 Carbon model page: https://artificialanalysis.ai/models/gemini-4-carbon
- Gemini API docs: latest model (Argon pricing, 1M window): https://ai.google.dev/gemini-api/docs/latest-model?hl=en
- X: Logan Kilpatrick on Argon token volume: https://x.com/OfficialLoganK/status/2112767157746688480
- Google: Gemini 4 Argon announcement (Sep 30, 2026): https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
- Artificial Analysis AA-Omniscience: https://artificialanalysis.ai/evaluations/omniscience

DISCLOSURE:

The voices in this video are AI-generated with ElevenLabs: Elon's narration uses a clone of his own voice, and the mascot Xy uses a stock ElevenLabs voice. The short opening clip of Elon is AI-generated from his likeness. Product names appear as text only; no logos are used.

Watch on YouTube: https://www.youtube.com/watch?v=OICXHrLYR8s