Fatih⏩⤴️

A father cat with lots of kittens 🥰🧡 📍filmed in Gaza, Palestine 🇵🇸🐾 IG📸 : majdi_fathi

I wanted to share a bunch of my favorite hidden and under-utilized features in Claude Code. I'll focus on the ones I use the most. Here goes.

Israel has killed more journalists since Oct. 7, 2023 than the total number of journalists killed in WW2, Vietnam and Afghanistan combined

@Kallias_Lives @KaiserLoengramm What’s this?

@KaiserLoengramm Clearly, I need to start plugging my own favorite anime harder if I'm going to keep up.

Seeing as Japan X is now on everyone’s feeds, this is a perfect time to watch Legend of the Galactics Heroes, which is not only one of the greatest Space Operas ever, but will give you a better understanding of the rise and fall of governments than any PoliSci textbook I’ve seen.

Judge rules that 17-year-old Walid Ahmad likely starved to death in Israeli military detention, but closes the case anyway. Israel never charged him with a crime and still holds his body. Here’s how he died, as reported by his cellmate, 16-year-old American Mohammed Ibrahim:

Today we observe Mother’s Day across the Arab world 🌺 We honor and celebrate Palestinian mothers who are the spine of our steadfastness—who have maintained Palestine and its people for generations amidst dispossession, fragmentation, and ethnic cleansing. We honor all Palestinian mothers who have buried their children wrapped in white shrouds, who have endured unimaginable loss with unwavering grace, who are languishing in jails, and who continue to carry the weight of our struggle on their backs and forever in their hearts. Our mothers are the forefront defenders of our land and life.

Never forget or forgive.

A woman in Tehran weeps the destruction of her home, walking away with her cat wrapped in a blanket. Photo by Erfan Kouchari.

My personal guesses: - intelligence is tool use capacity. So when and how to use which tool to process data is the most critical skill - there is no general optimal intelligence: everything is perspective, domain and context dependent. From same conditions or context: two agents can make different but valid judgements - coupling the first two: a successful agent should use tools to adapt itself to different domains. It needs a meta tool.

Nice pipeline but this is response/logit level right (lets call it level2 )? not hidden state/activations/trajectory level (level 1). If it is level2, models collapse to their own reasoning in my experience so they are pseudo critical in their reasoning (only justifies what they know) and do offline exploration (only narrow part of their knowledge base). This is highly inflected in their training. Jepa tried to solve this but still not there

My AGI research began - and has continued for years - by taking three ideas seriously: 1. Popperian problem-solving 2. Deutschian explanatory knowledge 3. Universality

@ThePrimeagen Try these models when US is asleep, see the difference

Here you go and you are welcome: marginlab.ai/trackers/claud…

@seey_dev @che_shr_cat Yes. Jepa and alike methods create text templates that are hard to break. You over condition the model and steals its flexibility. I did not check but i am sure the model also loses reasoning capacity as well.

@che_shr_cat Linearity is a shortcut. The real bottleneck isn't data volume, it's the loss of emergent nuance.

1/ Next-token prediction is incredibly noisy. By forcing hidden states into locally linear "semantic tubes," you can match full-dataset fine-tuning accuracy using 16x less data. No complex JEPA scaffolding required. 🧵

@BecomingCritter This is how you put more syrup to your waffle

they don't want you to know the ideal packing of circles, it's taboo

@ThatchEffendi No they think them as we think about Israel. Spies every where, present in every court in the world, good at trade, very good at high war tech (naval warfare)

Do you think Byzantines thought about Venice the same way we think about Dubai

@breathMessi21 Definetely not Turkey

Which country is the most geographically gifted when it comes to geopolitics?

@ronalhwang Wow great work👏

New Paper💡 Have you ever heard of grokking, a sudden transition from memorization to generalization? People have attributed grokking to weight decay, Fourier structure, optimization regimes, phase transitions, numerical effects… These can shape the training dynamics, but they don’t answer the core question: "what determines which representation the model learns, and why it generalizes?" We argue the key is intrinsic task symmetries. Paper: arxiv.org/pdf/2603.01968

🔴Öldürüleceğini açıklamıştı! O röportajtan çarpıcı bir kesit; “Ben 5 Mayıs’ kadar hayatta kalabileceğimi düşünmüyorum. Ben ve kızım bir şekilde hayattan koparılacağız. Biz öldükten sonra adaletin sağlanmasını istemiyorum. Başıma bir şey gelirse bu karanlık yapı ve beni koruyamayanlar, sesimi duyup da susanlar sorumludur. Kamuoyuna sesleniyorum; intiharım asla söz konusu değildir. Başıma bir şey gelirse intihar süsü verilerek üzerinin örtülmemesini, peşine düşülmesini istiyorum.”

@MillionInt Isn’t compression/dimensionality reduction a learning process and hence sign of intelligence?

Most common failure case scenario I see among the smartest people in the world is treating inherently multidimensional objects as their single dimensional approximation. It’s a very common and very powerful heuristic, but it will often fool you in a world that is high ELO

@CynicalPublius If it was a too bad of a plane, Israel would not block US to give it to Turkiye

For years I have been reading about how the F-35 is a stupid, vulnerable boondoggle. Yet it appears to me that the system is flying over airspace covered by the most advanced air defense systems Russia and China have to offer, yet it is completely invisible, untouched, and accomplishing all of its assigned missions with zero aircraft losses. Any of you current or former US military fighter jocks have thoughts?

@ZeYanjie with that, all information technologies arc will end. Only minor improvements and few bursts in every 20-30 years. As it happened to electricity before.

Robotics/Physical world is AI's last challenge