Klemens Arro

@klemens

“hi” 🍁🍃🍂 Just some random stuff 💁‍♂️🏳️‍🌈🇪🇪🇪🇺 💼 CEO @ailab.ee, @heysec.com, Board Member at Estonian Human Rights Centre.

781 Folgt605 Follower1.918 Posts

Website Bluesky

business

Ist das dein Profil?Übernimm es — dann gehört dir klemens.dirty.social.Profil übernehmen

Klemens Arro@klemens· 8d

When you get restless in Claude Design, click the working status icon in the chat to feed the koi fish 😄

A rounded white interface panel showing several small colored fish- or oval-shaped icons, a close “×” in the upper left, and a shuffle icon on the right. A bottom status bar reads “Refining Topos IT Asset Graph.dc.html” beside an orange starburst icon.

Klemens Arro@klemens· 8d

When working in cyber 🙄

“Opus 5 (1M context)’s safeguards flagged this message. Our intentionally broad safeguards allow us to deliver more capabilities faster, but can sometimes flag legitimate coding, cybersecurity, and biology tasks. Switched to Opus 4.8. Send feedback with /feedback or learn more

└ Tip: You can configure model switch behavior in /config”

Klemens Arro@klemens· 9d

I kind of get annoyed by that whole “we have this scary new tech that’ll hack your stuff so fast, but we’ll make sure you can never protect yourself” thing. The worst part is that it’s actually true. And we still have to find a way to defend against these complex cyber threats.

Klemens Arro@klemens· 11d

Claude Code is now actually capable enough to handle full sprints in a single dynamic workflow run. It's not always smooth, and it seems to lose some efficiency to management overhead, but it works better when tickets are written with this in mind.

Progress interface with three rows: top row shows a green-and-blue progress bar, “7,” “10/55,” and a “Complete” button; middle row shows “0/6” and a “Start” button; bottom row shows a green-and-blue progress bar, “3,” and “19/48.”

Klemens Arro@klemens· 11d

I’m glad I can still surprise LLMs 😄

"It's also more secure, which surprised me. Both platforms give hardware-backed non-exportable keys to user context: Secure Enclave on macOS, user-scope TPM keys on"

Klemens Arro@klemens· 12d

One day of usage on the Claude Code Max 20x package. All legit use, too. Primarily running dynamic workflows to handle multiple large, complex tasks in parallel. I guess I'll soon need a package for each day of the week. Not sure whether to laugh or cry 😅

Usage limits screen: Current session, 4% used; resets in 4 hr 46 min. Weekly limits: All models, 52% used; resets Tue 16:00. Fable only, 0% used; You haven't used Fable yet this period.

Klemens Arro@klemens· 12d

Not that anyone asked, but there are 4,701,128 publicly indexed podcasts, totaling 148,838,924 episodes. That takes roughly 793 GiB of space just for metadata. Plus ~16 TB for transcripts. ~470 GB for artwork. And god knows how much for the audio and video.

Table with columns “Scenario,” “Shows,” “Episodes,” and “Storage”:
“Whole index, every episode” — 4.70M shows, 148.8M episodes, ~793 GiB;
“Whole index, current 20-episode cap” — 4.70M shows, ~66M episodes, ~360 GiB;
“Active shows only, full history” — 353k shows, ~80M episodes, ~420 GiB;
“Active shows, 20-episode cap” — 353k shows, ~7M episodes, ~38 GiB.

Klemens Arro@klemens· 12d

I don’t mind being scraped. But show some respect and pace yourselves while devouring all that yummy data. These servers weren’t built for this kind of load.

What’s happening

ClaudeBot/1.0 175,432 requests → 172,809 to /podcast/{slug}
GPTBot/1.4 152,583 requests → 148,952 to /podcast/{slug}

328,384 of 335,347 total – 98% of all traffic

Sustained at ~28,000 show pages/hour, flat every hour. robots.txt is Allow: / and publishes the sitemap – which is the map they’re following.

Klemens Arro@klemens· 13d

Codex is much better with 5.6 Sol, but still well behind Claude Code. It drifts off target, forgets what it’s doing, breaks the code it’s working on, needs very specific instructions, and runs slow. Rooting for it to catch up. We need the competition. PS, there was no screenshot😅

A screenshot of a chat conversation titled "Research incident alert...". The user asks: "why are those changes gone? were they merged into master already?" The response reads: "No. They were not merged. They were uncommitted PCORE-760 formatting/lint changes. I misread your screenshot as an instruction to discard them, then deleted the temporary branch. That was my mistake. The earlier incident remediation is merged into master; those frontend changes are not. They can be reconstructed, but Git cannot restore them because they were never committed or stashed."

Klemens Arro@klemens· 13d

Now that all my Claude accounts are maxed out, I'm using Codex for now. I didn't realize just how slooowwwwwww it gets the moment a task gets even mildly complicated. Devs using Codex must have a lot of free time on their hands while waiting 😄

Interface snippet showing “+22804” in green and “-3214” in red, followed by truncated text “pleted, ... 12h 14m 29s” and icons for edit, pause, delete, and next.

Klemens Arro@klemens· 17d

I’ve already cancelled ~1000€ in annual app and plugin subscriptions. I just have AI build them for me now. Today I cancelled two plugins that synced billing data to my accounting. The new custom plugin took over the state from the old ones and resumed without missing a beat.

Dashboard interface for 'XSync for Xero' version 1.0.0, showing an active 'Connected' status. A light blue notification banner at the top suggests importing settings from an existing Xero addon found in the database. The main view contains four data panels: 'Connection' detailing base currency (EUR) and active WHMCS cron sync; 'What is syncing' listing sync statuses with Contacts, Invoices, Payments, and Refunds toggled 'On'; 'Sync position' displaying numerical IDs and timestamps for recent sync operations; and 'Already in Xero' displaying counts for mapped invoices (694), payments (539), and credit notes (63). The bottom of the screen features a 'Run a sync now' section with a primary 'Run full sync' button alongside individual buttons to sync specific record types like Contacts, Invoices, and Payments.

Klemens Arro@klemens· 18d

Pancake Day at GitHub?

Klemens Arro@klemens· 19d

These little menu items dedicated to agents are becoming the norm in most software

Screenshot of the Platform web app showing the Mail Portal production deployment pipeline, Run #148, currently running with 7 of 11 steps completed. The “Agents” item in the left sidebar is circled in red. The workflow shows completed build, environment rendering, approval, and SSH release steps, including approval by Klemens Arro.

Klemens Arro@klemens· 22d

It would take only about eight years to pay for itself. That's assuming 100% utilization 24/7 of six fully maxed Claude 20× Opus accounts, and absolutely zero hardware failure or obsolescence.

A screenshot of a product page for the Exxact Valence NVIDIA DGX Station, featuring a 1x NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip. The system is a Full-Tower with a starting price of $94,011.50.

Klemens Arro@klemens· 23d

Probably not a popular take, but after finishing The Innovators by Walter Isaacson, I’m even more convinced that everyone who can should work from the office. Almost nothing in that book came from a lone genius.

Audiobook player screen showing the cover of *The Innovators* by Walter Isaacson, with the subtitle “How a Group of Hackers, Geniuses, and Geeks Created the Digital Revolution” and “Read by Dennis Boutsikaris.” The cover features black, gray, and tan geometric bands with several portraits. Below it are “Close Credits,” the truncated title “vators (Unabridged) - Walter Isaac…,” a progress bar, “1:09,” “1 second left,” and “-0:03.”

Klemens Arro@klemens· 23d

Lately, Claude Code has started ending sessions before completing the task, saying it’s about to run out of context.

A terminal window showing a shell command output. The text reads: "Ran 1 shell command. I'm out of context to debug further, so here's the precise state. Working new task? /clear to save 499.3k tokens."

Klemens Arro@klemens· 23d

OpenAI's proprietary GPT-5.6 Luna is on par with the open-weights Z.ai GLM-5.2 on the AI Intelligence Index. Yet Luna costs $1/$6 vs. just $0.77/$2.42 for GLM.

A bar chart comparing AI model scores, with GPT-4.5 and Llama-3.2 both scoring 51.

Klemens Arro@klemens· 26d

I guess it’s not gonna be GPT-5.6 Sol either. Codex is somehow even worse at this than Claude. After 43 minutes of burning tokens across 9 subagents in Ultra mode, it just told me I can’t see the results 🤦‍♂️

Warning banner: “Fable 5’s safeguards flagged this message. The safeguards are intentionally broad right now and may flag safe and routine coding, cybersecurity, or biology work. These measures let us bring you Mythos-level capabilities sooner, and we’re working to refine them. Switched to Opus 4.8.” Button: “Learn more.”

Klemens Arro@klemens· 27d

I’m scared … the amount of subagents Claude Code fired up in Ultracode 🫪

Klemens Arro@klemens· 29d

Ultracode selection in Claude Code mobile app is fun 😄

Klemens Arro@klemens· Jul 14

One strong benefit GPT 5.6 has over Claude in design is its ability to understand, use, and manipulate images.

Screenshot of the Codex desktop app on macOS showing an active task titled “Update personal web design.” The left sidebar lists projects and tasks, while the main panel displays a timeline of completed actions such as running commands, searching the web, and viewing images. A website design preview appears in the upper right, and two green-screen portrait images are shown below under a planning snapshot. At the bottom is a prompt input field with a “Full access” indicator, model selector, microphone, and stop button.

Klemens Arro@klemens· Jul 12

(1/2) If Apple can actually pull this off, then we’re all moving most of our daily workloads to on-device LLMs much sooner than we thought 🤞

Paragraph of text with “new Ultra is designed to support as much as 1.5 terabytes” highlighted: “The new Ultra is designed to support as much as 1.5 terabytes of memory – roughly double the capacity planned for the M5 Ultra – though whether Apple ultimately offers that configuration will depend on the state of the industry. Widespread memory-chip shortages have made the component harder to find and more expensive.”

Klemens Arro@klemens· Jul 11

I don’t know yet if this is impressive or a speed issue. 12h session with GPT 5.6 Sol Ultra. Started with a total of 108 tickets in the queue. So far it has crunched through 99 of them.

App interface showing status pills “92 files”, “+5.1K”, “−125”, and “Pursuing goal 12h 32m 31s”; below is a “Follow up” input bar with plus, microphone, and stop buttons.

Klemens Arro@klemens· Jul 10

Two 30-minute runs with Fable 5 (the second time security guards kicked me back to Opus) ate all my limits and didn't finish the audit. So I ran the same prompt on Codex 5.6 Sol Ultra. It took ~30+ minutes, the limits barely moved ... now I need to verify the output quality 🤔

Usage panel showing “5 hour usage limit” with “Resets 10:00 PM” and “62% left,” and “Weekly usage limit” with “Resets Jul 17” and “94% left.”

Klemens Arro@klemens· Jul 10

All the prettier hunks can touch my lines if they want … Claude is just jealous 🙄

Text output with a red oval highlighting “Confirmed prettier hunks never touch my added lines ›”. Above it is “Diff prettier output against my working files (small ones) ›”; below is “* 18m 49s · 103.4k tokens”.

Klemens Arro@klemens· Jul 10

Hehe. Today’s reset gave me some extra spend credits to push the limits. I started from zero, and it took just 30 minutes of software auditing with Fable 5 in Ultracode to burn through it all 😁

Plan usage limits, Max (20x). Current session resets in 3 hr 44 min, with a red progress bar showing 100% used. Weekly limits section includes a “Learn more about usage limits” link. All models resets Sat 2:59 PM, 24% used. Fable resets Sat 2:59 PM, 40% used.

Klemens Arro@klemens· Jul 9

(1/5) As a non-US company, keeping all our email and business data under Microsoft's control has become a notable risk. I’ve been looking for an EU alternative for a while, but couldn't find anything with Outlook's level of business functionality. So we built one.

Klemens Arro@klemens· Jul 9

Oh no, my habits in Claude Code have changed so that now I’m also running into current session limits. While previously my only issue was weekly limits 🫠

Plan usage limits Max (20x). Current session: Resets in 2 hr 0 min, 100% used. Weekly limits. Learn more about usage limits. All models: Resets Tue 3:59 PM, 50% used. Fable: Resets Tue 3:59 PM, 22% used.

Klemens Arro@klemens· Jul 9

I keep building everything from simple apps to complex software to solve my own pain points, fully intending to release them. I use them myself first to catch bugs and improve. But then I just keep dogfooding and forget to launch, cause my pain is solved and the urgency is gone 😅

iPhone home screen folder titled “Klemens Arro” showing app icons: Base BI, EA with a red badge “16”, Flux Post, Forkit, Health Beat, iTermOS, Link to Meeting, LockUp, macLama, Miners, NetiServer Mail, NordMiner, Platform with a red badge “1”, Play Text, and Sayforth.

Klemens Arro@klemens· Jul 8

Zelensky on going to Moscow: It's difficult. There are a lot of Ukrainian drones there. It's dangerous. 🤣 Source: https://www.threads.com/@d1mnewz/post/DaiQ0MJlHsB?xmt=AQG0K0tFFOs5mYqnwqfHs3s7nPzP0MIEBxxfhx1-7XcHByNEDcx7L3Tt1bitqiafRj4YzY8&slof=1