VectorQL

@Fried_rice @attentionmech Wasn’t it open source @grok

Claude code source code has been leaked via a map file in their npm registry! Code: …a8527898604c1bbb12468b1581d95e.r2.dev/src.zip

@kriyanshii @ycombinator Ruko apply karta hu

@VectorQl @ycombinator Aap bhi aa jaate toh maja aa jaata

got into @ycombinator startup school 🚀 i work on backend and infra - satellite data pipelines, kubernetes, distributed systems. contribute to ray and dagu. lately been tinkering with building things from scratch. if you're building something ambitious or just want to jam, hmu.🥤

@ibuildthecloud they got to keep selling the macbooks as well

Wait I've never paid attention to iPads. Why is this not just macos? It's like Mac OS, but you can't run a terminal. If you could put Brew on this and launch VMs this would be the perfect device for me.

@linguinelabs consider me impressed

Couldn't fall asleep so I made a better pretext Bad Apple - improved silhouette, added Japanese lyrics, found a better english translation. It's so beautiful

@SidBalli This is serious

Got publicly insulted by an on-duty cop for a broken number plate even after I accepted my mistake and was ready to pay the challan. Later noticed his own bike had an illegal plate. When questioned, he tried to run away while breaking multiple traffic rules.

@marijajankdqcx @threepointone Crazyy

craaazy

How Qwen3.5-9B beats frontier models. TLDR: It’s about the harness! The core idea is simple but powerful: instead of fine-tuning a model or calling expensive APIs, you wrap a smaller frozen model in smart infrastructure that makes it punch way above its weight. How ATLAS Works ATLAS is a three-phase pipeline for solving coding problems using a frozen LLM: Phase 1 - PlanSearch + Generation - The model generates multiple candidate solutions (typically 3) - PlanSearch helps the model reason through the problem systematically - It uses BudgetForcing to allocate computational resources efficiently Phase 2 - Geometric Lens (C(x) Energy Field) - This is the magic part. Each candidate gets a 5120-dimensional embedding from the model's self-attention - These are scored by C(x) (learned energy function) to pick the best one - Currently only +0.0pp performance contribution because the training dataset was too small (only ~60 samples). V3.1 is fixing this Phase 3 - PR-CoT (Process-Rewriting Chain of Thought) Repair - If the best solution fails, the model generates its own test cases - Then iteratively attempts to repair the solution - Real benchmark tests are only used for final scoring The Core Components 1. Patched llama-server - Running on K3s, providing both generation (~100 tokens/sec with speculative decoding) and self-embeddings for scoring 2. Geometric Lens (C(x)) - Energy field that evaluates candidate solutions based on geometric properties in embedding space 3. G(x) Metric Tensor - Currently dormant, designed to apply metric corrections to C(x)'s gradient signal 4. Sandbox - Isolated code execution environment for running tests Performance & Trade-offs On an RTX 5060 Ti 16GB: - 74.6% LiveCodeBench score with Qwen3-14B (frozen, no fine-tuning) - Cost equivalent ~$0.004 per task (just electricity for local GPU) - Compared to API models: DeepSeek V3.2 (86.2%, ~$0.002/task), Claude 4.5 Sonnet (71.4%, ~$0.066/task) The key difference: pass@1-v(k=3) means it generates 3 candidates, selects the best via Lens repair, then iteratively fixes failures. Not single-shot generation. You get frontier-model-like performance with: - No API keys - No data leaving your machine - One GPU, one box - Pay only for electricity The V3.1 version in development is targeting 80-90% LCB with faster throughput by switching to Qwen3.5-9B (smaller model, 3-4x faster with DeltaNet attention) and properly retraining C(x).

CSS is dead Frontend is about to change. Pretext side-steps the need for DOM measurements (e.g. getBoundingClientRect, offsetHeight), which trigger layout reflow, one of the most expensive operations in the browser. It implements its own text measurement logic, using the browsers' own font engine as ground truth (very AI-friendly iteration method).

@kriyanshii 🙈

@VectorQl chat with humans too

All I do now is chat with AI

@TechEmails Ok

Mark Zuckerberg texts Elon Musk December 13, 2024

@AntDX316 @atomic_chat_hq @grok but 8 tokens/sec is barely usable,no ?

@atomic_chat_hq @Grok Is this better than LM Studio and Ollama?

Google Turbo Quant running Locally in Atomic Chat MacBook Air M4 16 GB Model: QWEN3.5-9B Context window: 50000 Summarising 20000 words in just seconds.. You can do 3x larger context window, processing 3x faster than before!

Google Research unveiled TurboQuant on March 24, an algorithm that compresses LLMs' KV cache by six times and speeds attention computation up to eight times on GPUs, all without accuracy loss or model retraining. Apps like Atomic Chat now handle 50,000-token contexts three times faster on M4 Macs with 16GB RAM, enabling privacy-focused offline AI. Memory stocks dipped 2-3% amid demand fears, but analysts see it as profit-taking, predicting efficiency will boost overall AI use.

@garrytan Gstack is amazing

This is why on GStack I usually just run /plan-eng-review and /ship and it works

$20 plan is just not enough for anyone.

@Zai_org I want this on Vertex AI now

GLM-5.1 is available to ALL GLM Coding Plan users! z.ai/subscribe

@Prince_Canuma Legend has started coding

Coming to MLX 🚀

Future is here