🚫 NO AGI, NO ASI, NO SINGULARITY coming soon unless we discover new architecture ⚡ Why? - Transformers: too inefficient, no dynamic weight updates - KV-Cache: computationally heavy = no 100M token agents - Scaling laws: FAILING 📉 💡 RAG survives until new architecture

Fun post on pretraining with RL (obviously a very WIP idea, but some real experiments) tokenbender.com/post.html?id=a…

Now that the era of the scaling "law" is coming to a close, I guess every lab will have their Llama 4 moment. Grok had theirs. OpenAI just had theirs too.

🚀 Huge Update: GPT-5, GPT-5-Mini & GPT-5-Nano are now integrated into the llmlayer Web Search API! ✅ GPT-5: Unmatched intelligence for demanding tasks ✅ GPT-5-Mini: Optimal balance of speed & precision ✅ GPT-5-Nano: Lightning-fast, cost-effective performance Experience

i remember having a 1 hour discussion post class time with my stats professor about why we divide by n-1 instead of n for the sample variance, after i first learnt about this in my stat analysis class. the reason is shockingly elegant but still a little abstract unless you've

Excited to share our #RL_for_LLM paper: "Part I: Tricks or Traps? A Deep Dive into RL for LLM Reasoning"  We conducted a comprehensive analysis of RL techniques in LLM domain!🥳 Surprisingly, we found that using only 2 techniques can unlock the learning capability of LLMs.😮

Watching a movie and brainstorming with GPT-5 Pro, i ask a question , watch the movie, read the answer and ask another one. This is becoming a habit for me

Hype

Why can’t we deploy a python or nodejs backends with a simple deploy command, no docker, no git, just a simple deploy —ram=1 —cpu=1 —workers = 4 domain= some custom domain | is this hard ? Im thinking of building this

I'm planning to buy Cohere right after they acquire Perplexity. All I need is a modest $200 billion in funding—should be easy, right? Then, of course, I'll casually take down Google. No big deal. 🤷‍♂️

The ARC-AGI group just dissected the HRM inside out. Turns out the results are legit, but not related to the hierarchical "brain-inspired" stuff. They wrote an incredible article on it, and everything makes sense again. Kudos for the monster debugging job!

Why do i find GPT-5 thinking to be better than GPT-5 PRO with the right instructions

How can ai agents become more useful in long contexts if we do not solve the KV Cache quadratic computation issue. This is the main problem and everyone is trying to solve this by context engineering and memory, but this will not solve the core issue. Maybe we need a way to

🌅 Good morning, world! Exciting news: llmlayer now offers FREE credits for everyone to experience our powerful Web Search API firsthand. Try it yourself and see why we're confident you'll love the results! ⚡️

Night thinking !! AI agents heavily depend on context. Someone asked me an interesting question: Can we build an AI system that deeply researches a topic for a week—reading, analyzing data, similar to how a BCG consultant does before preparing an analysis? There’s lots of hype

LLMs are trained unintelligently. They’re brute-forced into shape, like cracking a password with infinite tries. It’s ugly and energy-hungry. Yet probably worth it to help uncover the true essence of intelligence. Which is likely to be beautifully simple.

The paradigm of AI subagents is going to be super interesting. There was probably some hope or belief that a universal agent would be able to handle everything you needed in a workflow by stuffing all the relevant context into the context window. But even with larger context

Today, I pruned 87.24% of Qwen 30B for a sentiment classification task while keeping 100% of its accuracy. This means we get to use big models on gpus with not that much RAM (potentially running models that would normally require an H100 on 3090 type gpus)! Imagine pruning