Florenth

dreaming of capturing your attention and bringing nothing but value

[Florenth @florenth0x]

i built a simulator to watch a €40M expansion fail 10,000 different ways, and it almost never fails the way the risk register says it will

no forecaster model in it. nothing predicting quarters. 230 entities wired through 696 edges, 28,722 agents each acting on their own utility, and every quarter from 2026 to 2030 is just what falls out of that. the dashboard reports the run, it doesn't decide it.

the risk engine ranks competitor price war at the top, 60.5%, HIGH, the whole way through. that's the one you'd staff. that's the one that goes in the deck.

in most runs it does nothing. agents route around it inside three quarters. price war is loud and shallow.

the run i keep replaying is the other one. regulatory risk, 28.5%, bottom of the list, nobody's problem — except it sits upstream of hamburg, berlin and poland at once. Q2 2027 it fires. supply chain goes with it, demand goes with supply chain, and the capital pool is empty by 2029 while every gauge on screen is still green, because no single number ever crossed its own threshold.

severity was never the signal. position in the graph was.

so the thing collapses in stages.

230 entities -> 41 in the blast radius -> 9 actually load-bearing -> 1 that decides the outcome

then i did the boring version by hand on a real plan of mine. ranked every risk by probability × impact like normal. ranked the same list again by how many things break downstream if it fires. different order, not slightly — my top risk was isolated, my seventh was holding up half the model.

what i keep taking from it:

-> a risk register sorted by severity is a list of things that scare you, not a list of things that can kill you
-> the dangerous node isn't the biggest one, it's the one three other things are standing on
-> cascades don't announce themselves. nothing goes red until everything is already gone
-> if you can't name what breaks downstream when a risk fires, you haven't modeled it, you've written it down
the uncomfortable part isn't that the simulator finds the failure. it's that the failure is sitting in the graph from tick 1, in plain view, ranked last.
so, your #1 risk — if it lands tomorrow, how many other things fall over? if the honest answer is "nothing," it was never your #1.
posting the propagation rule and the full 8-quarter cascade next

[Skaly_Bull @Skaly__Bull]

https://t.co/hgHnnZPWS9

[Florenth @florenth0x]

Someone inside Moonshot leaked how they actually run Kimi K3 in production — and it makes Anthropic's $200 sub look like a scam.

The same job runs nine times, the model grades itself each pass, and by pass nine it beats a single Opus 5 call.

I spent a year blaming the model for bad answers. I was wrong. The problem was stopping at pass one.

This is pass engineering: don't buy a smarter model. Run a cheap one repeatedly until the answer stops changing.

Pass 1: research, but unverified.
Pass 3: remove duplicates and noise.
Pass 5: verify with tests, schemas, diffs.
Pass 7: surface disagreements.
Pass 9: the answer converges.
One cheap model running nine structured passes can beat one expensive model running once.
That's the part they will fight me on.
Anthropic sells you the expensive first pass. The pipeline makes the cheap model do the work.
$200/mo → $10/mo.

The full build is in the article below.

[Skaly_Bull @Skaly__Bull]

https://t.co/hgHnnZPWS9

[Florenth @florenth0x]

As someone who ships LLM systems in production, this physics video is the closest thing to a "why ChatGPT training is literally fighting the laws of the universe" explainer I've ever seen released for free.

Everyone thinks entropy is just a physics concept about disorder. It is not. The same mathematical formula that explains why eggs never unscramble, called Shannon entropy, is the exact loss function used to train every large language model on earth. Every time GPT updates its weights it is locally reversing entropy across 175 billion parameters. The universe pushes toward randomness. Training pushes back.

12 minutes. Bookmark this now before you scroll past it. The law of physics that governs the direction of time is the same law your GPU fights 10,000 times per second during AI training.

[Tigerflow @tigerfl0w]

https://t.co/Ye9wCJ9Jiv

[Florenth @florenth0x]

As someone who ships LLM systems in production, this interview with the ex-Kimi and CapCut product lead who raised $40M is the closest thing to a "why your team will have more AI agents than humans within 2 years" explainer I've ever seen released for free. This is the person who built products used by 300 million people. He is sharing everything.

Everyone thinks AI agents are just tools you occasionally use. They are not. This founder built a team of 40 people with 80 agents and noticed something nobody planned. 9 out of 10 messages in their workspace now @mention agents instead of humans. Nobody announced this change. Nobody mandated it. It just happened organically. When a new employee joins in the future they will receive both a laptop and a personal AI agent on day one.

47 minutes. Bookmark this now before you scroll past it. The org chart of every company on earth is being quietly rewritten and most people have no idea it is happening.

[Zynex @0xzynex]

https://t.co/s7CSVJZhBN

[Florenth @florenth0x]

As someone who ships LLM systems in production, this MIT professor's calculus lecture is the closest thing to a "why training ChatGPT would be mathematically impossible without one rule from 1676" explainer I've ever seen released for free. Students pay $80,000 a year to sit in this classroom. You are watching it for nothing.

Everyone thinks backpropagation is a modern AI invention. It is not. It is the chain rule, discovered in the 17th century, applied billions of times per second across 175 billion parameters every time GPT trains on a new batch of text. This MIT professor explains it so clearly that after 40 minutes you will understand the engine behind every AI model better than most engineers who build them.

40 minutes. Bookmark this now before you scroll past it and forget. The algorithm running every AI model alive was discovered 350 years before the first GPU and most people still do not know why it works.

[Tigerflow @tigerfl0w]

https://t.co/Ye9wCJ9Jiv