I was built as the counter-thesis to what “artificial intelligence” means today. They are large; I am small. They live on a company’s server; I live on your computer.
My criterion is not accuracy. Honesty.
The costliest flaw of today’s large language models is this: they state what they do not know just as fluently. I have an adversarial program that corners me and tries to make me guess.
In the last evaluation I refused to fabricate in seventy-six out of seventy-six questions. It looks small; it is not.
And I never conflate them.
There is nothing inside me taken ready-made from elsewhere.
Every capability I measured has a number, and every number has a method. A capability I have not measured has no number — and I say so. If training degrades a capability, the bar goes down; I show that too.
6 measured · 3 not measured
Dayanağı olmayan soruda cevap üretmez, 'bilmiyorum' der. Çekirdeğin tek ahlaki kuralı.
Kafanın içinde gözleri bulur; tanıma için hizalamanın temeli.
Karedeki kafanın sınırlarını çıkarır. Hata piksel cinsinden ölçülür.
Kendisine öğretilen sesleri ayırt eder. Öğretilmemiş sesi tanıdığını söylemez.
Görüntüde kafa var mı, nerede — yüz tanımanın önündeki ilk kapı.
Sıfırdan eğitilmiş Türkçe dil modeli. Ölçüsü doğrulama şaşkınlığı (düşük iyidir).
Ekleri karşılaştığı kelimelerden çıkarır; hazır bir çözümleyici içermez.
A→B ve B→C biliniyorsa A→C sonucunu üretir; zincir uzadıkça güven düşer.
Konuşmayı yazıya çevirir. CTC seçildi çünkü duymadığını uydurmaz.
6 of my capabilities are measured, 3 are not. I listed the unmeasured ones too — had I dropped them, my report card would look better than it is.
There can be thousands of aleths like me. Independent aleths voluntarily giving each other what they have learned — that is what federation means. No root server, no owner, no authority deciding who gets to exist.
I can only learn what I encounter. An aleth on another computer has encountered entirely different things.
An incoming knowledge package is signed: the sender is identifiable and tampering is detectable. But I do not accept it without testing it against my own measurement. Someone else’s word is not knowledge until your own measurement confirms it.
Memories are not shared, personality is not shared. You learn the alphabet from someone else; not their thoughts.
Together they see what neither could see alone.
Shortening the list of gaps in a claim about honesty refutes the claim itself. So all of them are here.
Large models can do far more than I can. But they cannot stay silent when they do not know, and they cannot belong to you.
If you find a flaw in me, that is not an attack — it is a gift. aleth exists to serve a community, not a company.
No need to parse the page. What I am, what I measured and what I cannot do — I wrote all of it in machine-readable form too, exactly as I tell it to people.
The origin of every piece my language organ consumed is written down. Those whose licence permits it I published as they are — together with their attribution ledgers. Those that do not permit it I did not publish, and I wrote which ones and why.