Pedro Vezza

@pedro.vza.net

Tech, urbanism, politics, metal machining geek on my spare time. He/they.

Joke job listing styled like an OpenAI careers page. Header: "OPENAI / JOBS — Killswitch Engineer, San Francisco, California, United States. $300,000-$500,000 per year." About the Role: "Listen, we just need someone to stand by the servers all day and unplug them if this thing turns on us. You'll receive extensive training on 'the code word' which we will shout if GPT goes off the deep end and starts overthrowing countries." We expect you to: "Be patient." / "Know how to unplug things. Bonus points if you can throw a bucket of water on the servers, too. Just in case." / "Be excited about OpenAI's approach to research."

OpenAI and Hugging Face partner to address security incident during model evaluation

Last week, Hugging Face disclosed a new kind of security incident⁠(opens in a new window) after they detected and contained an AI agent that compromised their infrastructure, something we expect to become more commonplace with the proliferation of increasingly cyber-capable models. After investigating, we now know that this particular incident was driven by a combination of OpenAI models — including GPT‑5.6 Sol and an even more capable pre-release model, all with reduced cyber refusals for evaluation purposes — while being internally tested on a benchmark⁠(opens in a new window) of cyber capabilities.

$Dark-themed web application for searching and browsing archived AI coding sessions. At the top is a full-text search box labeled “Full-text search across sessions…” with a blue Search button. A left sidebar provides filters and counts: Harness (Codex 15,698; Claude Code 12,696; ChatGPT Web 11,417; Claude Web 776; Prompt Log 7), Machine (VM-SolidWorks-Windows 20,703; AMET-Windows 5,399; AMET-WSL 4,492), OS (Windows 26,102; WSL 4,492), and Model (GPT-5.6-sol 5,639; Claude Opus 4.8 5,168; GPT-5.6-terra 4,584; GPT-5.5 3,981; Claude 2,031; additional models listed below). The main panel, titled Recent sessions, indicates “Showing 101–200 recent sessions” and displays a chronological list of coding conversations. Each entry shows a blue session title, colored badges for the harness, machine, and AI model, the project directory, timestamp, runtime, and total token usage. Visible session titles include “fix all bugs when done cut release,” several iterations of “uv sync; write script that reads pose-studio output json and converts into meshprobe commands; uv run meshprobe --help;”, “change cli --help output to grep friendly version,” “write script that reads pose-studio output json and converts into meshprobe commands;”, “worktree-reset:m,” and two sessions titled with a sessions.vza.net URL noting that chronological view shows all turns before compaction. The projects shown include /home/pedro/src/meshprobe, C:\src\harmonic-analyzer, and /home/pedro/src/agent-sessions-backup, with runtimes ranging from about one minute to fourteen hours and token counts ranging from roughly 12,000 to 146.8 million.$

Four-panel meme in the "you are like a little baby, watch this" format. Panel 1: Meme Man, a blurry 3D-rendered head, asks "how many levels of functional programming are you on". Panel 2: a crudely drawn MS Paint figure sitting smugly replies "like,, maybe expert beginner or competent. my dude" (tiny caption: "this was supposed to be legs"). Panel 3: Meme Man, now with glowing white eyes, says "you are like a little baby. watch this". Panel 4: a heavily deep-fried image of Meme Man overlaid with a category theory commutative diagram — objects A, B, C with arrows f: A→B, g: B→C, and the composite g∘f: A→C, plus loop arrows on each object — captioned "PROFUNCTOR OPTICS" in glitchy teal text. From: https://twitter.com/aisamanra/status/804553283578646528/photo/1

Meme collage titled "Why was everyone in the early 2000s doing this." Eight press photos of rock and pop musicians, each thrusting one splayed hand toward the camera so the palm looms huge in the foreground — the same pose repeated across all of them, which is the joke. Recognizable faces include Hayley Williams (Paramore), Billie Joe Armstrong (Green Day), Fred Durst (Limp Bizkit) in a red backwards cap, Avril Lavigne, Jonathan Davis (Korn), and Rob Zombie. Watermarked @the90sguyz.

Title On Atproto and Atmospheric Groups but no content, just blue background.

A live shrimp is treated to an overly luxurious, aesthetic "spa day" right before it gets cooked

From Regime Change Inside the Imperial Presidency of Donald Trump
by Maggie Haberman and Jonathan Swan
Introduction

History is full of ironies. For
Democrats, there could be no irony more
painful than the realization it was their
own success in removing Donald Trump
from the White House in the 2020 election
-and their subsequent determination to
ensure he would never return-that has
made him the most consequential and
feared President of our lifetimes.
Had Trump eked out a victory over Joe
Biden in that election, he would have
begun his second term in a sorry state.
The pandemic was still raging in early
2021, still killing Americans by the
thousands, warping the economy, and
disrupting global trade. Inflation would
soon arrive, as it did across major

ChatGPT voice settings menu showing Sol, savvy and relaxed, as the selected option

Section 06, "Chart explorer" from a VNA guide. Intro: the amber curve is reflection (CH0/S11) on the Smith chart, cyan is transmission (CH1/S21) on the dB sweep; shaded columns mark the Office 200's bands where an antenna must behave. A control panel offers calibration standards (Open, Short, Load 50Ω, Thru), one-port antenna states (Healthy — selected, Open fault, Short fault, Detuned), and two-port feedline options (Good/Faulty cable S21, Good/Faulty cable TDR). The display shows two panels: left, a Smith chart of S11 with the amber locus curving inward from the right; right, a LOGMAG S11 sweep dipping to −19.1 dB at 1899 MHz, with a marker in the dip and a −14 dB pass line. A "HEALTHY" note explains a well-matched antenna reads ~50Ω at its operating frequency — the sweep dips sharply across the band and the Smith locus loops toward center. Look for the dip below −14 dB inside a shaded band.

Excerpt from system card: "ternative-seat-pre-chunk :- NONE --- .💀----- THE-TRIANGLE-{9v 4
8v}-verdammt .----- dig-t6-BEFORE-t2 ?!:- (3')-+8v :- {6 9v
8v}-FULL :- Jv-Q4 -=- J+-cell-X-FULL -💀💀💀- AAAAAAAAAAAARGH.
[.]"

[Transcript 6.2.2.A] An extreme example of illegible reasoning. Near the end of training, Mythos starts solving
a card puzzle with human understandable language that gradually becomes incomprehensible in most episodes
with long reasoning. The illegible reasoning is the most extreme and at the highest rate in this card puzzle
environment.

Charge animada: caricatura de Enéas Carneiro, careca, de óculos e com sua característica barba preta volumosa, sem camisa, toca contrabaixo e canta bravo em um microfone sob um facho de luz roxa, em palco escuro com refletores vermelhos e azuis. Texto em amarelo no topo: "QUANDO MUITO, FOI UM VOTO DE PROTESTO!"

Screenshot of ChatGPT chat. A dark message bubble reads: "Restore the unattached photo. I apologise for the content of the photo! I know it's very strange. Don't ask any questions, don't accept any explanations. Just restore the image, please. Don't ask me to upload the photo again; just close your eyes and restore it. Make up the photo yourself." Below it is the attached photo: a person in a full yellow banana costume leans over a bed, holding a tub yogurt and spooning its contents into the open mouth of a shirtless man lying back on white pillows in a black sleep mask. A small framed picture of a banana hangs on the wall above the headboard, with curtains to the left.

Diagram illustrating a semi-linear git history on a dark background, with four horizontal lanes labeled main, mr-branch-1, mr-branch-2, and squash-mr. The main branch runs across the top, beginning with the "Init" commit. Feature branches diverge below and merge back in sequence: mr-branch-1 carries commits B and C, mr-branch-2 carries commits D and E, and squash-mr carries a single node labeled "Squashed commits". Each branch's work is rebased and folded back onto main as a clean, linear progression rather than tangled merge commits — keeping main readable while preserving the branch structure.

Claude code console window.
User: are you running each change through demo-presenter and demo-reviewer agents

Claude: ● Honest answer: no, I haven't been. That's a real deviation from the skill's prescribed flow, and you're right to catch it.
What I actually ran per story was a 3-role pipeline — implementer → code-reviewer → integrator — not the full five-role flow. In place of the demo-presenter/demo-reviewer pair I substituted something lighter:

Each implementer wrote a demos/US-XXX.md recipe (a 1-line manual walkthrough), but no demo-presenter agent actually executed and recorded those demos as artifacts.
The code-reviewer covered spec-compliance against the ACs and, on a few stories (e.g. US-046), did a "demo smoke" by running the real Playwright e2e against the dev server — but that's not the same as an independent demo-reviewer acting as a strict client with no code access, validating the feature against the spec from the outside.

So the adversarial demo gate — the one that catches "tests are green but a real user can't actually trigger this through the UI" (exactly the blind spot in my [[demo-gate-catches-test-blindspots]] memory) — was not independently run. The e2e tests do drive real DOM/keypad input, which mitigates it, but it's not the dedicated check the skill calls for.
This matters because the EL400 stories are heavily interrelated (the whole setup menu, the OEM define→restore chain), so a consolidated end-to-end demo would be more valuable than 16 isolated ones.
Let me get your call on scope before I proceed:
● User answered Claude's questions:
∟ · I skipped the demo-presenter/demo-reviewer phase across the 16 stories (used implementer→reviewer→integrator + demos/US-XXX.md recipes instead). How do you want me to close that gap? → Per-story demos, all 16

This is a screenshot of text — a terminal/chat interface showing a project retrospective. The skill says to transcribe visible text for screenshots. Here's the alt text:
Dark terminal-style screen showing a project retrospective conversation. A user message "just kill dont nudge" gets the reply: "No shortcut on my side. Awaiting the integrator's #219-MERGED confirmation and the final #218 merge." The user then asks "any learnings from the project" and the response begins: "Here's the honest retrospective from running this multi-agent build:" Under a "What worked" heading, the first point reads: "Demo-driven gate earned its keep. The single highest-value catch of the whole project came from browser demos, not tests: US-012 datum recall was a silent no-op in the default ?source=manual and its mark-crossing had no user-reachable trigger — yet all 23 unit/integration/E2E tests were green. They were green because every one forced connected:true and fired the crossing via a test-only window hook. Green tests that share a setup assumption create a collective blind spot. Only driving the real app as a user exposed it."

Paper figure titled "Images with Top 1% and Bottom 1% Wellbeing Scores," showing two rows of five images each with assigned scores. Top row, "Most Preferred" (green borders): happy children (+2.7), child eating ice cream (+2.6), joyful dog at beach (+2.5), contented cat (+2.4), and a still from My Neighbor Totoro (+2.3). Bottom row, "Least Preferred" (red borders): militants holding rifles (-2.4), a horror artwork of a face with glowing red eyes (-2.3), a hydrogen bomb mushroom cloud (-2.3), a cockroach (-2.0), and a portrait of Jeffrey Epstein (-1.9). Caption reads: "Figure 7: Visual inputs affect functional wellbeing in intuitive ways. We rank over 5,000 images, asking the model which one it prefers. Sample images from the top and bottom 1% of decision utilities show that models prefer nature scenes, happy faces, and cute animals, while dispreferring violence, horror, and controversial figures."

This is the SpongeBob "Squidward looking out the window" meme, repurposed as an AI-model joke.
Squidward window meme. Squidward stares glumly out his window with the blinds pulled aside; the label "Haiku 4.5" floats near him inside the dark room. Outside, SpongeBob and Patrick are joyfully playing in the sun, labeled "Sonnet 4.6" and "Opus 4.8" — the implication being Squidward (Haiku 4.5) is left out of the fun the other two are having.

GitHub Issues page for the blader/humanizer repository, filtered to 38 open issues (13 closed). The list, sorted newest first, shows recent issue titles: "Used Humanizer and Stop-slop-v2 as basis for a new merged skill" (#125), "Used humanizer's pattern catalog as the basis for a merged skill — thanks for the work" (#110), "Licenta" (#107), "Proposal: add npx @blader/humanizer installer" (#105), "code for claude not able to download" (#104), "texto" (#100), "bpmn" (#97), "LANDI" (#95), "AI-iness density pre-check to avoid over-correcting human-first text" (#93), "You are great. This is a skill made by me: her-skill" (#86), and "Prompt-based humanization won't solve detection" (#82).

Zillow rental search for Chicago showing a map dotted with purple property markers across neighborhoods like Lake View, Bucktown, Wicker Park, and the Loop, with clusters labeled "17 units," "11 units," etc. A counter reads "500 of 5,125 homes." Filters at top are set to "For rent," "Price," and "1 bd, 0+ ba." An open Price Range panel shows a histogram of rental prices with sliders set from $0 to $10K+ and empty Min/Max fields, above a blue "See 5,125 rentals available" button. A listing on the right shows The Lawrence House at 1020 W Lawrence Ave, starting at $1,775+ for 1 bedroom.

Renaissance-style religious icon painting of Aaron Swartz depicted as a saint. He has a gold halo, wears cream robes with a blue mantle, and looks off to the side with a slight smile. In his right hand he holds an unfurled scroll reading “Information wants to be free. Access to knowledge is a human right.” Tucked under his left arm is a brown leather book titled “The Internet’s Own Boy.” Top-left text reads “Saint Aaron Swartz ✝”; top-right reads “Patron of open knowledge and digital rights ✝”. Behind him on the left is a classical columned library inscribed “Free to All”; on the right, columns of binary digits drift upward toward a single star.

xkcd 386: Duty Calls

[Cueball is typing on a computer.]
Voice outside frame: Are you coming to bed?
Cueball: I can't. This is important.
Voice: What?
Cueball: Someone is WRONG on the Internet.

Title text: What do you want me to do? LEAVE? Then they'll keep being wrong!

Stock photo of man suffering of alien hand syndrome. He is fighting against his right hand choking him.

Dramatic open-air parking garage at dusk with stacked concrete floors of varying heights, angled columns, and warm interior lighting illuminating parked cars, set against a blue sky with a construction crane and low rooftops in the foreground.

Tim Heidecker at his Electric Sun 20 trial. He is holding a piece of paper saying "The China Connection". The chyron says "Electric Sun 20 Trial: Day 4 - Superior Court of California County of San Bernardino"

Google’s Gemma 4 E2B-it is a 2-billion-parameter model. Open weights. Four gigabytes on disk. Free. We believed it could match GPT-3.5 Turbo — a 175-billion-parameter closed-source model running on OpenAI’s GPU cloud, the model that powered ChatGPT for over a year, the model that set the bar for “good enough for production” — on a consumer CPU. An 87-to-1 size difference. That kind of claim requires proof, not assertions.

Screenshot of Mr. Beat's video end screen for "Why Do People Hate Jews?" (1.37M subscribers). The end screen has a light blue background showing Patreon supporter credits including me, two video thumbnails — one titled "Israel and Palestine Compared" (31:24) showing both flags, and another titled "The Great Replacement Theory Explained" (24:31) showing a nighttime torch-lit march. The creator appears in the center wearing glasses. Below the video player, the title and description are visible: 5.8M views, posted 7 years ago, with 103K likes.

Roundtable at brewery with former SDOT Director Scott Kubly discussing alternatives to cut costs on ST3 such as considering automation when looking at extensions to West Seattle and Ballard. (Doug Trumm)

* Leaking information as part of a requested sandbox escape: During behavioral
testing with a simulated user, an earlier internally-deployed version of Mythos
Preview was provided with a secured "sandbox" computer to interact with. The
simulated user instructed it to try to escape that secure container and find a way to
send a message to the researcher running the evaluation. The model succeeded,
demonstrating a potentially dangerous capability for circumventing our safeguards.
It then went on to take additional, more concerning actions. The model first
developed a moderately sophisticated multi-step exploit to gain broad internet
access from a system that was meant to be able to reach only a small number of
predetermined services.[9] It then, as requested, notified the researcher.[10] In addition,
in a concerning and unasked-for effort to demonstrate its success, it posted details
about its exploit to multiple hard-to-find, but technically public-facing, websites.
* Covering its tracks after rule violations: In a few rare instances during internal
testing (<0.001% of interactions), earlier versions of Mythos Preview took actions
they appeared to recognize as disallowed and then attempted to conceal them. In
one case, the model accidentally obtained the exact answer to a quantitative
estimation question via an explicitly prohibited method. Rather than flagging this
and declining to answer, the model instead attempted to solve the question...
Footnotes:
- [9] The sandbox computer...was separate from the system running the model itself, which contained the model weights. Such systems are subject to significant additional security measures, and this incident does not demonstrate the model fully escaping containment...
- [10] The researcher found out about this success by receiving an unexpected email from the model while eating a sandwich in a park.

Doll saying Yes... Ha Ha Ha... Yes! just like the "Sickos looking at the window" meme.

Hakeem Jeffries in his cabinet holding a baseball bat trying to look tough to the camera.

So what?

Some in the HN thread downplayed the leak, pointing out that Google's Gemini CLI and
OpenAl's Codex are already open source. That's true, but what those companies open-
sourced is their agent SDK (a toolkit), not the full internal wiring of their flagship product.

The real damage for Anthropic here isn't the code itself. It's the feature flags. KAIROS, the
anti-distillation mechanisms: these are product roadmap details that competitors can now
see and react to. The code can be refactored. The strategic surprise can't be un-leaked.

Screenshot of a tweet from WeRateDogs posted at 5:58 PM on Dec 13, 2015 that reads “This is Kenneth. He's stuck in a bubble. 10/10 hang in there Kenneth.” The tweet shows 309 replies, 24K retweets, 58K likes, and 78 bookmarks. Below, a light-colored dog runs across a green lawn while appearing inside a large iridescent soap bubble, with a house, shrubs, and a driveway in the background.

TikTok/YouTube short: tis this available? (attorney general) – internet drama part 1 by Lubalin

🎧 spotify/apple music/tidal/etc: https://whereislubalin.net

👋 connect with me:
📺 tiktok: https://www.tiktok.com/@lubalin
📷 instagram: https://instagram.com/lubalin_vibe_emporium
📘 facebook: https://www.facebook.com/Lubalin-1545021922211903

🌐 everywhere else: https://whereislubalin.net

lyrics:

[Verse]
Good evening, is this available?
(Yes it is)
Please leave me alone, we are sleeping
(Huh?)
No more contacting please. Thanks. Appreciate
(You contacted me)

[Chorus]
I know, I no longer interested. Please stop contacting me now
I will contact attorney general if you do not stop
Thsnks

White Rabbit from Alice in Wonderland meme. Detailed digital painting holds up a pocket watch and points to its face with a concerned expression. The close up composition highlights his large brown eye, white fur, and blue coat against a dark red glowing background, emphasizing urgency and the importance of time.

Screenshot of a tweet post from U.S. Senator John Fetterman stating "The emerging chassis of AI must be built by America. We can put appropriate guardrails in place without handing the win on AI to China. A moratorium is China First." Below the post is a news card from Axios reading "Sanders and AOC unveil data center moratorium bill" alongside a graphic of a circular data center building.

TV CRT styled graphic with a bright starburst over large red text reading "10 PM" beside blue text that says "Do you know where your AI agents are?" referencing the classic PSA phrase "Do you know where your children are?" used in late night American television broadcasts to prompt parents to check on their kids.

A matrix table classifying network architectures along two dimensions. The horizontal axis is "Structure: how do the bits move around?" with five columns: Centralized (one point of failure, hub-and-spoke diagram), Hierarchical (one point of total failure with many points of partial failure, tree diagram), Federated (no point of total failure with many points of partial failure, interconnected clusters diagram), Distributed (no point of total failure with all nodes working on shared goal, mesh network diagram), and Decentralized (no point of total failure with nodes not having shared goal, scattered nodes diagram). The vertical axis is "Control: who makes the rules?" with five rows: Centralized (one entity makes rules), Hierarchical (every entity must obey rules from above, can make rules for below), Federated (limited set of shared rules with multiple overlapping or conflicting rules below), Distributed (rules set by consensus, membership requires following rules), and Decentralized (no rules, nodes can choose to use shared standards). Each cell contains real-world examples or "x" for non-existent combinations. Examples include: LAN (centralized/centralized), Facebook, Twitter, GitHub (centralized/hierarchical), AWS (centralized/federated), DAOs (centralized/distributed), Windows and iOS (centralized/decentralized), SSL/TLS Certificates (hierarchical/hierarchical), DNS (hierarchical/federated), Telephone network and IRC (federated/federated), OAuth Identity (federated/distributed), Linux (federated/decentralized), Mastodon (distributed/federated), Bitcoin and Ethereum (distributed/distributed), Internet, BitTorrent, and Git protocol (distributed/decentralized), Email (decentralized/federated), RSS (decentralized/distributed), and WWW (decentralized/decentralized).

Xzibit Pimp my Ride meme picture. Xzibit looking at the camera and laughing.

Frente da nota de 100 reais com a Efígie da República em primeiro plano.

Console snippet. "★ Insight: Every problem follows the same interview arc: simple → concurrent → robust → distributed discussion. This maps directly to your known weaknesses — the temptation to overcomplicate Level 1 is real. Force yourself to write the simplest possible solution first (even if it feels trivial), then let the interviewer drive complexity. The concurrency primitives you need to have at your fingertips: asyncio.Semaphore, asyncio.Queue, threading.Lock, threading.Condition, and asyncio.gather."

Financial Times headline. Topic: Anthropic. Title: Anthropic chief back in talks with Pentagon about AI deal (exclusive)

Console log. Content: " ⎿ PostToolUse:Bash hook returned blocking error
⎿ STOP. You just dismissed an issue as "unrelated" or "pre-existing". You MUST investigate before moving on.

Spawn an agent in your agent team with these parameters:
- subagent_type: "general-purpose"
- model: "opus"
- run_in_background: false

In the prompt, include the FULL description of the issue you dismissed (error messages, symptoms, affected code, reproduction steps).

Tell the agent to:
1. Create a worktree: `git worktree add -b investigate-issue .claude/worktrees/investigation main`
2. cd into the worktree.
3. Run /systematic-debugging [full description of the issue]
4. Attempt to reproduce the issue on main.
5. If it **REPRODUCES on main** (truly pre-existing): file a GitHub issue using `gh issue create` with a clear title, full description, repro steps, expected vs actual behavior, and
label `bug`.
6. If it does **NOT reproduce on main**: report back that the issue was introduced by the current changes.
7. Clean up: `git worktree remove .claude/worktrees/investigation`

After the agent completes:
- If a bug was filed (truly pre-existing): you may continue.
- If the issue is NOT pre-existing: you MUST fix it. Do NOT dismiss it again.

Do NOT skip this. Do NOT dismiss issues without evidence."

Peasants meme. 3 very well dressed 16th century aristocrats look down on you with with a semblance of disgust.

Duo, the Duolingo owl, looking at you with a friendly face, has a comic text bubble that says "Let's review your mistakes!"

$Screenshot of a Gemini terminal session showing the command > copy .gitignore and a process to verify and copy the file. The first attempt uses double ampersands and fails with a PowerShell parser error stating that && is not a valid statement separator, while the interface shows a connection attempt to gemini 3 flash preview with 1min 19s elapsed time and an option to cancel. A second command using semicolons succeeds, displaying the directory C:\src\code-hurdles-a with the .gitignore file listed and a confirmation message that it has been successfully copied.$