Pinned Tweet
This is a pinned post. There are many like it, but this one is mine.
I'm Mattias, 40, solo founder from Sweden. Writing code since 1999, started with C as a madman. This year i got my first patch accepted into the Linux kernel.
What i build:
๐ด Horsemate (horsemate.se), a stable management app for Swedish horse owners
๐บ Better-IPTV (better-iptv.vercel.app), an open source TV player in Rust (7,000+ downloads)
๐งฉ A Chrome extension for X that labels your feed and tracks who unfollows you (release coming up)
๐ฆธ Githam City (githam.city), a pixel city where you can rent the hero's cape as ad space
๐ฎ Stupid prototypes at 3 am, whatever idea won't leave me alone that week
What i write: Warhammer 40k fan fiction
What i share: the highs and the lows. Real numbers, real bugs, and how i build with Claude.
Goal: the mythical $5k MRR. Currently somewhat short of that ๐
#buildinpublic
9
21
587
Saturday night. Son's home, dinner's done, back to my pirate game.
I had 7 AI models analyse the game loop on their own. Opus, Fable, Sonnet, Haiku and three Codex models. Plus a bot that played 6 worlds for an hour per ship.
They all landed on the same thing: the loop isn't too slow. It has too few choices and endings.
So tonight the bar got two rumours to choose between, treasure sells in any open port, and the whole UI turned into paper notes ๐ดโโ ๏ธ
(And thanks chatgpt for not giving me a double-chin in the pic haha)
49
An Anthropic mathematician says he found a proof of the Riemann Hypothesis together with Claude.
Meanwhile my biggest breakthrough this week was swapping ?? for || in Better-IPTV ๐
Not confirmed yet, people are still going through it. But if it holds, wtf. AI in STEM is so freaking interesting
SITUATION EXPLAINED: Claude may have just solved the Riemann hypothesis, math's most famous unsolved problem, 167 years after it was first posed.
@__alpoge__ , a mathematician at Anthropic, posted a 4-page note saying he found the proof "working with Claude," and that the same method proves the generalized Riemann hypothesis
โข He says he chose the direction and Claude did much of the hard technical work, especially near the finish line
โข The full argument ran past 250,000 pages, so he released it only as Lean code, split into 14 files a computer can check line by line
โข MIT mathematician Andrew Sutherland says he independently checked it in Lean and called it "an historic moment"
โข It's a Millennium Prize Problem with a $1M bounty
โข On the same day, a separate 940-page GRH draft by Alessandro Masullo surfaced, written over 12 years with no AI and reviewed for months by Stanford's Brian Conrad
@theojaffee: "There's a timeline here where Levent says it was actually November 16th, GRH, the generalized Riemann hypothesis, which means we actually did get Riemann hypothesis before GTA VI. By 3 days."
3
3
158
GTA VI made $3.79 billion on day one, according to Rockstar.
My share of that: $0.
I've played every main title:
GTA 1: the timed demo, about 1000 times
GTA 2: played it, but liked the first one more
GTA 3: blown away by GTA in 3D
Vice City: still one of the best stories
San Andreas: Grove Street, born and raised ๐
GTA IV: played it a lot, Niko Bellic is the GOAT
GTA V: played it a lot
GTA VI: watching Lirik play it on Twitch
No console, no gaming PC.
4
134
Slow Saturday morning. Chill in the sofa and watching the news.
-5 ยฐC this morning so itโs starting to become real cold now ๐ฅถ
Do we have anything planned for the day X?
3
8
116
If my AI coworker could file an HR complaint, i'd be done by Monday.
I had Claude go through a year of my prompts. The evidence:
Wiped its memory 739 times
Longest shift: 13 hours
Switched projects 4.7 times a day
~200 swear words (at the code, not at it)
In my defence: 609 "tack" (Swedish for thanks) and 1,217 =)
1
4
80
4 job applications this morning.
One of them already got back to me and wants a call on Monday =)
My son is sleeping at his girlfriend's tonight, so the rest of the evening is what we Swedes would call "Fredagsmys". A bowl of lรถsgodis, T3 Code and Lirik playing GTA VI.
Not a bad Friday.
4
14
513
M Andersson ๐ธ๐ช retweeted
The Office: Hiring Again
AI-Generated Parody | Created by @thisaintcanon
#funny #theoffice #movie #nostalgia #parody
My timeline is full of 22-year-olds shipping their first SaaS.
I'm 40.
Wrote my first lines of C at 13, back in 1999.
Built a SaaS for horse stables in Sweden.
Building an open source TV player in Rust on the side.
Turns out you don't age out of shipping ๐
6
6
123
Is it Friday already? Didn't see that one coming =)
โ
Coffee #2
โ
4 new job applications
๐ก Got a small new idea I'll explore today, see if it sticks
โณ Finish my extension and get it ready to publish
โณ Coffee #3 (obviously)
โณ Buy snus, my last tin runs out today (Swedes know the panic)
#buildinpublic
9
9
284
Most of us rent our models.
Fine-tune a small one on your own data, own the weights, and host it wherever you want.
That's the part I'd actually want, and @OvermindLab makes that a reality.
๐ค Paid partnership
Turn Grok Bot into your own AI lab, right from your phone.
โข Connect @bot to the Overmind MCP
โข Upload a dataset of examples to learn from
โข Tell it what to learn and let it fine-tune the model
Over to you: overmindlab.ai
Claude is working its way through 3,728 accounts in my browser right now, blocking one at a time.
If I block you today, this is why =) Real human caught by mistake? Ping me on GitHub.
It took a while but as promised here it is...
"THE ๐ SPAMMERS LIST"
๐ก๏ธ Blocked for AI reply
blockedforaireply.levels.io
3,728 ๐ accounts that continuously spam automated AI replies to artificially farm engagement and ruin ๐'s social network - 498 already suspended by ๐ - continuously updated
โ
Evidence for every blocked account
โ
One-click block all button
โ
Continuously updated JSON API
โ
Text-only username list
Based on @ryanvogel's original Clankwall and @dillon_mulroy's X list
I hope everyone uses this and keeps using it, so my account becomes a honeypot for AI replies, once I block them they're blocked by the entire network and they stop being able to reply to anyone
+ It'd be cool to add more "scouts" like @ThePrimeagen and @JackEllis and sync their block list continuously too cause they and many others have the same issue
I've also submitted this list to @X's CEO @petersellis to train their spam models on
Again it'll be continously updated!
Enjoy a spam free ๐ experience! ๐๐
4
2
137
First snow of the winter here ๐
Too bad it seems like itโs gone later today or tomorrow.
And yeah Iโm still on my South Park marathon as you can hear ๐
2
4
148
My horse stable SaaS has a free trial. The account page called it "Your active subscription" under a heading that said "Billing".
Very reassuring ๐
Changed that last week. It now says plainly that nothing gets charged unless you choose a plan, the trial ends by itself, and your data stays.
Obvious when you actually read it from the other side.
#buildinpublic
1
2
58
Last model benchmark from me. I've already found how to use both of them together in my workflow =)
GPT-6.2 Astra vs Claude Fable 5.2: The Jungle
A stormy sea, a beach, a 360ยฐ circle around an object of their choice, a run through dense jungle, then a rise above the treetops. 45-50 seconds or disqualified. Everything built in code, no downloaded models or textures. 60 minutes each, tokens unlimited.
Last round Astra was done in 6 minutes. This time it took 49 min 45 s.
Fable 5.2 used 53 minutes last round. This time it felt done after 29 min 54 s.
And both landed on exactly 48 seconds, 2880 frames. Funny coincidence ๐
Astra picked a brass armillary sphere for the 360. Fable picked a stone idol that wakes up from a lightning strike.
Fable shines with its lighting. Astra shines with pure detail.
GPT-6.2 Astra: 49 min 45 s, 7.8M tokens, ~$7.30
Claude Fable 5.2: 29 min 54 s, 5.0M tokens, ~$11.07
Costs are estimates at API list price. Astra on top, Fable at the bottom.
3
5
177
GPT-6.2 Astra vs Claude Fable 5.2
My wild west train benchmark, flagship round. Same prompt as my GPT-6.2 Sol vs Opus 5.5 run, both on high effort. 60 minute time limit, told not to think about tokens.
Astra was done in 6 min 2 s. Barely touched the time slot.
Fable 5.2 took 53 min 23 s. Used almost all of it.
Fable has the better visual identity. Haze in the distance, long shadows. Still love the smoke from the chimney, looks way more realistic.
All four runs:
GPT-6.2 Sol: 14 min 6 s, 1.6M tokens, ~$0.70
Claude Opus 5.5: 35 min 17 s, 12.7M tokens, ~$8.24
GPT-6.2 Astra: 6 min 2 s, 0.5M tokens, ~$1.01
Claude Fable 5.2: 53 min 23 s, 16.2M tokens, ~$21.92
Costs are estimates at API list price. Astra on top, Fable at the bottom.
3
5
268
GM ๐
โ
Coffee x2
โ
Ran yesterday's train benchmark on GPT-6.2 Astra and Claude Fable 5.2, results later today
โ
Sent 3 job applications to some seriously respectable tech companies
โณ Grocery run for the weekend and the rest of the week
โณ Fika with my grandma this afternoon
When did you last have fika with your grandparents?
7
7
164
GPT-6.2 Sol came out today, so I gave it and Claude Opus 5.5 the same Three.js prompt. Both on high effort, every 3D object built in code.
A steam train crosses the desert at golden hour, brakes into a western town, the camera tours the town, then rises as the train rolls off into the sunset.
GPT-6.2 Sol (top): 14 min, about $0.70
Claude Opus 5.5 (bottom): 35 min, about $8.24
Looks-wise Sol wins. Great looking train and its horse is way better. But the smoke from Opus's chimney is miles ahead. Bang for buck it's not even close though, Sol takes it.
A little skewed though, Sol is the cheaper sibling and Astra is their flagship. First time I've pitted two models against each other like this, mostly for fun =)
1
2
372













