lets give ai access to gov resources to help solve world hunger/poverty/immigration problems

Ben kendi hayatımı bile böyle idame ettiriyorum

What do you want to know?.

RT @carolyndiary: yolsuzluk ve güçle alakalı bi makale okuyordum ve güç kazanan erkeklerin taciz eğiliminin daha fazla oldugunu soyluyor. M….

RT @benfurkankilic: Evre maalesef vefat etmiş, ailesinin ve sevenlerinin başı sağolsun. Kendisinin tweetleri her önüme düştüğünde gülümseme….

Mentioned paper:

So when we open up a model, we shouldn’t expect clean, one-to-one mappings. Features are compressed together. Understanding and separating them is actually key to mechanistic interpretability.

Using toy ReLU networks (Elhage et al. (2022)) , they showed phase changes: sometimes a feature is ignored, sometimes it gets its own neuron, sometimes it’s stored in superposition. The overlap works, but it makes networks harder to interpret.

So features don’t get their own neat neuron (like in hypothetical disentangled model), they overlap and share space by simulating a larger sparse model and store in almost-orthogonal directions. This happens because models try to store more features than they have neurons for.

In neural networks, a single neuron is generally responsible for more than just one thing. Instead of “this neuron = cat ear” we often see one neuron mixing many different features. This called superposition hypothesis.

RT @dorotheagibi: selamlar, çok fazla ismin ifşa olması sebebiyle gördüğünüz twitter, instagram veya haber kaynaklarını ekleyebileceğiniz b….

RT @aayseaktag: fail aklayıcılık nedir:.-ne yapalım bir hata yaptı diye tamamen dışlayalım mı?.-ben onu yıllardır tanırım, yapmaz öyle şey….

Apple’s paper: .My humble and not so good piece from 2020:

Different angles, as in technical vs philosophical but both highlight the same tension. AI may look like it thinks, but whether it ever truly can remains an open question.

In my note on physicalism, I argued that AI is bound by determinism programs and data shape every outcome. There is no free will, so AI cannot develop a mind in the dualistic sense.

Apple looks at reasoning models and shows that they succeed only up to a point. When problems get harder, the thinking illusion breaks down and the models just stop trying.

Reading Apple’s paper The Illusion of Thinking reminded me of a short piece I once wrote on AI and physicalism. Both ask the same question in different ways: can AI really think?.

RT @fluchtperioden: BU BİR TACİZCİ İFŞALAMA FLOODUDUR. Bu dövmeci Antalya’da bulunuyor. 2 yıl önce bir kadın arkadaşımız burada dövme yaptı….

RT @enisebytes: KAN uses splines for every connection, removing millions of weight values. Nodes only sum inputs. This design makes the net….

Since the first model, versions like KAN 2.0, PRKAN, AF-KAN, and seqKAN pushed the idea into science, time series, and more. Still early days, but KAN may become the bridge between black-box AI and interpretable learning, aka explainable AI! #ExplainableAI.

KAN uses splines for every connection, removing millions of weight values. Nodes only sum inputs. This design makes the network more transparent and easier to interpret.

Spline functions are like small polynomials stitched together to form a smooth curve. Instead of one big equation, the model learns many small curves that connect smootly.