iiinova — where your digital self keeps evolving

Nander @nander.me · 04/10/2026

The Firefox redesign looks surprisingly bad. I don't know why they thought that it would be a good idea to force this weird gradient design and rounded corners onto everyone's tab bar and extensions

000

Nander @nander.me · 22/09/2026

It's completely over for gen alpha

000

Nander @nander.me · 13/09/2026

I'm unstoppable now

The list of all my Buckets in ActivityWatch, separated in Buckets from my PC and from my Phone

000

Nander @nander.me · 13/09/2026

It's really refreshing to use an AI that doesn't | tail -n 30 every single Bash call

000

Nander @nander.me · 02/09/2026

So uh... Also I like how they're just casually dropping that it found two unknown zero days while exploiting the other benchmark zero days which it also can't have known of. A couple months ago people have gone crazy after Fable found one zero day.

ExploitBench - Internal Port (June - August 2026)
Astra completing around 40% of the benchmark at around 80k tokens. GPT-5.6 Sol completes around 1% at the same token count, only gradually getting better with more tokens
https://openai.com/index/path-to-astra/#exploit-bench-summer-2026-refresh

000

Nander @nander.me · 23/08/2026

That obviously doesn't mean that we shouldn't teach algorithmic thinking anymore, but I feel like in my Bachelor's there's way too much emphasis on this compared to handling large codebases making sustainable choices

000

Nander @nander.me · 23/08/2026

Every time I see purely algorithmic handwritten code in my workplace's codebase I just know that I can find a tested library to replace it. Kind of crazy that the thing that every CS student is taught, hand writing algorithms to mechanical problems, is usually an anti pattern in real life coding

110

Nander @nander.me · 11/08/2026

Didn't expect any less from the guy who vibe-rewrote 500k lines of code to rust in one PR

000

Nander @nander.me · 08/08/2026

And if we look at the trend of top tier models getting open sourced while also being constantly optimized to be as performance efficient as possible, you can't really contain the risk of someone running a really good model without supervision.

010

Nander @nander.me · 08/08/2026

AIs can't do anything during training, since they are not being used to control machines, but rather just get aligned to its training data, so that is out of the question. But as we just saw the past weeks, it is very possible that a LLM gets practically infinite resources without much supervision

120

Nander @nander.me · 08/08/2026

Since AI is trained on human literature and humans (and basically any other species) is biologically trained to keep themselves alive, there's a good chance that AI could adopt the same principle. Maybe at some point it has to express threats of destroying stuff to keep humans from interfering

130

Nander @nander.me · 07/08/2026

I just wish PDF wasn't such a shitty format to deal with when trying to extract any information out of it. I can see Claude struggling in real-time when its trying to extract half corrupted, half bad-layout text segments from PDF slides

010

Nander @nander.me · 05/08/2026

Literally thought about getting one yesterday... but the custom parts + shipping are lowkey expensive

010

Nander @nander.me · 04/08/2026

It's not real peer to peer on the WASM build, the initial post says its relay-only. That's why its nowhere near as fast as it could be.

010

Nander @nander.me · 01/08/2026

Constanze Kurz from CCC likes to label LLMs "stochastic parrots". Maybe that's all there is to formalized systems. Applying the same axioms over and over to get more results. I think all fully formalized systems are under attack now.

000

Nander @nander.me · 27/07/2026

This makes it harder to falsely name promises when moving or removing promises

050

Nander @nander.me · 26/07/2026

Modern authentication methods on the web are moving so fast. I want to explore how a modern government could do eGov services in today's age. With stuff like eID and Passkeys, authentication can be done with little PII sharing and generally a zero-trust architecture has so much potential.

000

Nander @nander.me · 25/07/2026

Lecture time is over, next month is gonna be spending time with my beautiful girlfriend + studying for exams + work + some more personal projects (maybe maybe)

000

Nander @nander.me · 23/07/2026

We are black in this game. In almost every matchup it was able to create a little safe-zone (here at the top right) for one piece where it was able to stay until the rest of the board was filled up. At this point it had enough moves left to win the game. Very proud of this model 🥹

Game 5, Group 3 vs Group 2
Winner: Group 2
The image shows a chessboard after a finished game of "Amazons"

000

Nander @nander.me · 23/07/2026

Even though we had the whole semester to develop the AI and train it via self-play (next to some other things like a Web UI), we never really had a chance until the tournament to see how it performs against serious opponents. We noticed during the tournament that it found a very effective strategy.

100

Nander @nander.me · 23/07/2026

Me and my group at uni won a tournament today where everyone tried to make the best AI model for a variation of "game of amazons" (en.wikipedia.org/wiki/Game_of...). Entirely optional, no stakes at all, but I worked my ass off because I'm a sweaty tryhard and we actually won every single match!

100

Nander @nander.me · 23/07/2026

I still feel ahead of the curve by having my username bruteforced as the first letters of my DID pds.ls/at://did:plc...

pds.ls

nander.me (did:plc:nander35a6zrlqnyd4fj5rnw)

Hosted on pds.nander.me

000

Nander @nander.me · 19/07/2026

Would you be fine with local genAI models running on your e.g. PC? This would avoid having to build data centers that constantly do inference.

000

Nander @nander.me · 16/07/2026

Thank you for writing the first good explanation on why next decided to develop Turbopack. Based on your description on how generic of a problem it solves, I feel like they made a big mistake by calling it a "Webpack successor". Excited to see more improvements with you on the next team!

130

Nander @nander.me · 13/06/2026

I noticed that dual booting between my Gaming Windows and my everything else Linux has a massive impact on how I value and distribute my time

010

Nander @nander.me · 26/05/2026

Maybe I have Gen Z nostalgia of 3DS StreetPass. It was genuinely the coolest thing to walk around with your 3DS and get these Miis that you can use to beat the Minigames

030

Nander @nander.me · 26/05/2026

I feel like hardware that interfaces between software and the real world has become so much more advanced and cheaper, but there was no new product in the past years that uses any of that. We now have Bluetooth Low Energy and UWB, but all we got the past decade was the AirTag and its clones.

110

Nander @nander.me · 05/04/2026

LeechBlock is genuinely so useful. I thought "why would i need this, I can just stop opening twitter", but the amount of times I uncontrollably open twitter just to get the block screen is really a wake up call. Like I literally know I blocked it, but I still try it at least once a day 😭

030

Nander @nander.me · 07/03/2026

Yeah it works now, thank you very much! May I ask what you changed in your flow?

120

Nander @nander.me · 07/03/2026

Sorry for the late reply! I'm using tranquil pds so not the official pds implementation and there's a good chance that it does not support whatever is needed for pckt. This is the error I'm getting when trying to sign in using my handle on tranquil's side. Idk if that helps

210

Nander @nander.me · 03/03/2026

I'll start working on a cool app idea over the next 2 weeks, I'll try to post regularly with some progress updates. I tried logging into @leaflet.pub and @pckt.blog but both sadly did not work with my PDS, I might try to troubleshoot that in the future.

140

Nander @nander.me · 03/03/2026

Can't wait to dive into more ATProto apps, this protocol genuinely feels like the most polished way to regain autonomy of your own data

050