Klemens Arro
LightNews LightNews
banner
klemens.bsky.social
Klemens Arro
@klemens.bsky.social
560 followers 750 following 1.4K posts
“hi” 🍁🍃🍂 Just some random moments & stuff from my life 💁‍♂️🏳️‍🌈🇪🇪🇪🇺 💼 CEO @ailab.ee, Board Member at @admcloudtech.com and Estonian Human Rights Centre.
🔗 https://klemens.ee
Posts Replies Media Videos
klemens.bsky.social
And here’s a Nano 🍌 Pro photo from my chill relaxing vacation.
A barefoot man stands smiling on sunny sand dunes while a huge, digitally added sandworm with a circular mouth full of sharp teeth erupts from the sand behind him, spraying dust into the air.
November 20, 2025 at 8:51 PM
klemens.bsky.social
Nano Banana Pro 🍌 was released. And here’s my test cat image.
A young man in a beige T-shirt sits laughing as if riding an enormous, unrealistically giant orange tabby cat on a woodchip path in a forested park, with trees and simple wooden structures in the background.
November 20, 2025 at 8:45 PM
klemens.bsky.social
I never realized how easy it is to build iPhone apps now with the Claude integration in Xcode.
Screenshot of a Mac desktop with Xcode in fullscreen split view. On the left pane, a chat assistant window shows instructions and Swift code snippets for adding “back online” notifications to a NotificationManager, including comments and a “Summary of Changes” section. On the right pane, the Xcode editor is open to a Swift file named Worker.swift in a project called “F2Monitor,” displaying a Worker struct, its init(from decoder:), encode(to:), and helper methods. At the bottom right, Xcode’s debug console shows JSON-like output while the app runs on an iPhone simulator.
November 20, 2025 at 3:59 PM
Reposted by Klemens Arro
techcrunch.com
The two companies will start in the UK this year, before expanding to other European cities in 2026 and the U.S. in 2027.
Uber Eats will use Starship sidewalk robots to deliver food in the UK | TechCrunch
The two companies will start in the UK this year, before expanding to other European cities in 2026 and the U.S. in 2027.
techcrunch.com
November 20, 2025 at 12:06 PM
klemens.bsky.social
OMG, yes! Get rid of these cookie banners!

www.theverge.com/news/823788/...
Europe’s cookie nightmare is crumbling
The seven-year nightmare is nearly over
www.theverge.com
November 19, 2025 at 7:48 PM
klemens.bsky.social
If Antigravity were to get web and mobile apps one day.
November 19, 2025 at 6:07 PM
klemens.bsky.social
It seems Claude has increased the deadline to use the free credit. I still have a (slim) shot at using it all.
Banner reading “Try Claude Code on the web, on us. We’ve added $1,000 in free usage credits to your account, applied automatically. Use before November 23. Read more,” with a partially filled blue progress bar below showing “68% remaining.”
November 19, 2025 at 5:56 PM
klemens.bsky.social
I have now used almost all models’ limits in Antigravity. I want to give someone money to get higher limits … but they don’t take money. As I understand it, even the Ultra package wouldn’t help here.
Cartoon image of Fry from Futurama with an angry, excited expression, thrusting a fistful of cash forward, overlaid with the bold meme text “SHUT UP AND TAKE MY MONEY!”
November 19, 2025 at 5:35 PM
klemens.bsky.social
I actually really like Google’s Antigravity. But it’s under so heavy load that I can’t actually use it even with any other provider’s models.
Like … Google … basically “the internet” … can’t keep up with the demand 🫠
Screenshot of a macOS desktop window showing a developer tool in split view. The left pane displays a coding assistant log with Laravel PHP commands, edited files (e.g., Organisation.php, User.php, AuthController.php), and a notice that the “Agent terminated due to error.” The right pane shows a formatted implementation document titled “User Authentication & Organization Management,” with sections for “User Review Required,” important notes, a warning about session configuration, and proposed database changes in bullet points.
November 19, 2025 at 4:30 PM
klemens.bsky.social
My first impression of Antigravity IDE is great. It feels simple, logical, and great to manage multiple dev agents. The integrated Chrome it uses is wildly awesome. Plus, Gemini 3 in it seems pretty good so far. I’m definitely going to give it more time.
November 19, 2025 at 1:43 PM
klemens.bsky.social
Seems like everyone is trying out Gemini 3. My Antigravity testing attempts are failing under the load.
Screenshot of a walkthrough comment panel showing two collapsed sections, “Thought Process” and “Thought for 6s”, followed by an error message in bold that reads “Error Agent execution terminated due to error.” Below it, smaller text says “Agent execution terminated due to model provider overload. Please try again later.” with “Good” and “Bad” feedback buttons at the bottom right.
November 19, 2025 at 11:44 AM
klemens.bsky.social
This looks like an interesting new agentic dev tool from Google. I’m going to test it out with a new project for sure 😋
Let the evolution of developers’ role into AI operators begin.

antigravity.google
Google Antigravity
Google Antigravity - Build the new way
antigravity.google
November 18, 2025 at 8:23 PM
klemens.bsky.social
I really like Claude’s thinking and research, as it’s not just producing words, but mostly actually gets the point I’m after. But it has way too small context and chat history to actually use it in most real cases I have needed it.
Screenshot of the Claude chat interface showing a banner that says “Claude hit the maximum length for this conversation” with a “New chat” button, a “Chat paused” message below, and various UI icons for actions, microphone, and audio at the top and bottom.
November 18, 2025 at 7:23 PM
klemens.bsky.social
Why is Gemini still missing MCP support? I think it’s the last one that doesn’t have it. Would really love it.
November 18, 2025 at 6:04 PM
klemens.bsky.social
Gemini 3 Pro is so far the best at converting PNG to SVG. The first attempt was almost spot on. The second didn’t go that well. But it is still better than any other model. And the SVG code is really clean.
Side-by-side comparison graphic of logo SVG conversions titled “Attempt 1” and “Attempt 2.” In Attempt 1 (white background), the original “ai·Lab” wordmark is shown above and the “SVG by Gemini 3 Pro” version below, which looks almost identical. In Attempt 2 (pink background), the original “heySec” logo with a yellow shield icon and clean text is shown above, while the “SVG by Gemini 3 Pro” version below is visibly distorted, with a warped shield, mis-shaped letters, and misplaced yellow dots.
November 18, 2025 at 4:42 PM
klemens.bsky.social
Cloudflare is coming back online, and now the amount of spam coming from my website is wild. I guess Cloudflare hasn’t gotten their WAF up yet.
Screenshot of an email inbox showing several unread WordPress messages titled “[Klemens Arro] Please moderate: …,” notifying about new comments on different blog posts.
November 18, 2025 at 2:58 PM
klemens.bsky.social
My first instinct is to get hyped on so good benchmarks. But then I remember that so far none of the benchmarks have reflected much of the reality. So I’ll keep waiting to test it in real.
timkellogg.me Tim Kellogg @timkellogg.me · 2d
Gemini 3 model card leaked

the URL is taken down now, was here:

storage.googleapis.com/deepmind-med...
Below is a faithful transcription of all visible entries: ⸻ Benchmark — Description — Scores Humanity’s Last Exam — Academic reasoning, no tools • Gemini 3 Pro 37.5% • Gemini 2.5 Pro 21.6% • Claude Sonnet 4.5 13.7% • GPT-5.1 26.5% ARC-AGI-2 — Visual reasoning puzzles (ARC Prize Verified) • 31.1% — 4.9% — 13.6% — 17.6% GPOA Diamond — Scientific knowledge, no tools • 91.9% — 86.4% — 83.4% — 88.1% AIME 2025 — Mathematics, no tools • 95.0% — 88.0% — 87.0% — 94.0% • A second line shows: 100% — — 100% — — MathArena Apex — Challenging Math Contest problems • 23.4% — 0.5% — 1.6% — 1.0% MMMU-Pro — Multimodal understanding and reasoning • 81.0% — 68.0% — 68.0% — 80.8% ScreenSpot-Pro — Screen understanding • 72.7% — 11.4% — 36.2% — 3.5% CharXiv Reasoning — Information synthesis from complex charts • 81.4% — 69.6% — 68.5% — 69.5% OmniDocBench 1.5 — OCR (lower is better: Overall Edit Distance) • 0.115 — 0.147 — 0.147 — 0.147 Video-MMMU — Knowledge acquisition from videos • 87.6% — 83.6% — 77.8% — 80.4% LiveCodeBench Pro — Competitive coding (Elo rating, higher is better) • 2,439 — 1,775 — 1,418 — 2,243 Terminal-Bench 2.0 — Agentic coding (Terminus-2 agent) • 54.2% — 32.6% — 42.8% — 47.6% SWE-Bench Verified — Agentic coding (single attempt) • 76.2% — 59.6% — 77.2% — 76.3% t2-bench — Agentic tool use • 85.4% — 54.9% — 84.7% — 80.2% Vending-Bench 2 — Long-horizon agentic tasks (Net worth, higher is better) • $5,478.16 — $573.64 — $3,838.74 — $1,473.43 FACTS Benchmark Suite — Internal grounding, parametric knowledge, search retrieval • 70.5% — 63.4% — 50.4% — 50.8% SimpleQA Verified — Parametric knowledge • 72.1% — 54.5% — 29.3% — 34.9% MMLU — Multilingual Q&A • 91.8% — 89.5% — 89.1% — 91.0% Global PIQA — Commonsense reasoning across 100+ languages • 93.4% — 91.5% — 90.1% — 90.9% MRCR v2 (8-needle) — Long-context performance • 77.0% — 58.0% — 47.1% — 61.6% • Second line: 26.3% — 16.4% — not supported — not supported
November 18, 2025 at 1:42 PM
klemens.bsky.social
We all thought the AWS incident was bad. Cloudflare was like: watch this
November 18, 2025 at 12:46 PM
Reposted by Klemens Arro
emiliusdraws.bsky.social
#3157 A helpful tutorial
A digital drawing of one of those drawing tutorials that says 'don't do this, do this.' The drawing is of an almost naked middle-aged man doing a flying kick. The 'Don't do this' one has a bunch of areas circled and a cross under it. The 'Do this' one is exactly the same but without the circles and a tick instead of a cross. The text reads 'Needlessly adding red circles will ruin your art' and a smaller bit of text at the bottom reads 'Bonus tip! Add a tick under all your art so people think you did it right.'
November 17, 2025 at 11:40 PM
klemens.bsky.social
I keep on breaking Claude with my research and analysis jobs 😬
Mobile app screen showing a ‘Research complete’ card with the text ‘0 sources • 706h 18m’ at the top. Below it is a sheet titled ‘Research Details’ with a close button on the left and a line indicating a step labeled ‘Creating my research plan…’ with an orange status dot.
November 18, 2025 at 7:12 AM
klemens.bsky.social
I fall into the camp that believes there will be a correction in AI companies at some point. The finances just don’t make sense from any way you look at them. But this won’t change what AI can do, and how it’s going to impact so many areas.
November 17, 2025 at 3:13 PM
klemens.bsky.social
Ironically, Claude Code builds a perfect MCP server for ChatGPT from a single prompt but can’t figure out how to build one that would work on Claude.
November 17, 2025 at 2:14 PM
klemens.bsky.social
Seems that coding with AI is not that expensive. I’ve been building all kinds of useful and useless apps with this credit. And still seems that I won’t be able to spend it all 🫠
Banner from Anthropic reading: “Try Claude Code on the web, on us. We’ve added $1,000 in free usage credits to your account, applied automatically. Use before November 18. Read more.” Below the text is a blue progress bar labeled “69% remaining.”
November 16, 2025 at 2:06 PM
klemens.bsky.social
It took SEVEN YEARS to see if an item is in stock 🤣 I had totally forgotten I even had ever made this order. At least I’m thankful they remembered and will refund it now.
Screenshot of an Amazon “Item Canceled” email. The message explains that the item was canceled due to lack of availability and a refund will be issued. In the “Details” section, the line “Placed on Monday, March 19, 2018” is circled in red, and a yellow “View or manage order” button appears below.
November 16, 2025 at 7:34 AM
klemens.bsky.social
This looks like a useful implementation. I assume this will soon have open-source alternatives, and most other code repository and documentation systems will adopt the idea.

codewiki.google
Code Wiki
codewiki.google
November 16, 2025 at 7:17 AM