Do you understand? This means we can possibly bring down 1T+ parameters models to couple hundreds of B+ parameter models without losing much score now, imagine this linked with quantization + ternary ops + hardware optimizations, how fast it would be on device AI will be big

BREAKING DISCOVERY: Humans alternate between conscious and unconscious states 15 times per second.

It is becoming f*cking dangerous: you can now reverse-engineer any repo with one command - a YC-backed tool with 127k stars maps the whole codebase, and your agent walks the map instead of reading files 1. Clone the project you want to understand 2. Install graphify: uv tool install graphifyy 3. Run graphify install (it detects Claude Code, Cursor, Codex or Gemini CLI) 4. Open the project and type /graphify . 5. Open graph.html: the architecture map nobody ever documented 6. Read GRAPH_REPORT.md: what everything flows through and the connections you didn't know were there 7. Ask: graphify path "A" "B" traces how two things connect, graphify explain shows everything one of them touches Code is mapped locally with no LLM, and every link says whether it was read or inferred Turn your agent into a reverse engineer that knows the whole project before it touches a single file Find the repo in the reply below

Amazing thread for RAG insights 🔥🔥🚨

One important aspect is that you can search semantically, and be more token efficient, faster. You can optimize your shape of data for the agent(s). My main use case: index my linear, github, home dir, meeting transcripts, memories, etc. All searched in a simple mcp/api/sdk.

Yes, RAG is still a thing for some use cases. If you just need to retrieve general knowledge from a document, a simple search is usually enough. But when ACCURACY really matters, like pulling an exact revenue number from a deeply nested table or connecting details across multiple pages, then it becomes super useful. This is especially true for financial documents, where you have to properly handle tables, images, and infographics and still return results at scale in a reasonable amount of time/cost. In those situations, you often need more advanced techniques, such as hierarchical chunking and careful tuning of your hybrid search threshold. So if you care about getting high accuracy, RAG still has a place.

Amazing suggestion to keep your 🤌🏼 made from scratch Semantic Cache in check 🔥

Ranking of the greatest mathematicians of all time: 1. 🇮🇳 Srinivasa Ramanujan 2. 🇨🇭 Leonhard Euler 3. 🇩🇪 Carl Friedrich Gauss 4. 🏴󠁧󠁢󠁥󠁮󠁧󠁿 Isaac Newton 5. 🇬🇷 Archimedes 6. 🇬🇷 Euclid 7. 🇩🇪 Gottfried W. Leibniz 8. 🇬🇷 Pythagoras 9. 🇮🇳 Aryabhata 10. 🇩🇪 Bernhard Riemann Source: thetoptens

what a week for open weights! we got: - Bonsai 2 2B/4B/9B - Nemotron 3.5 Bolt 74B + Super VL 120B - MiniCPM-V 4.7 1B + 35B-A3B - LFM2.5-24B-A2B - Muse Spark Open 816B-A42B - Step 5 Preview 600B-A27B - Beam 1.0 501B A23B - Jais 3 family - Granite Switch 4.1 - Hy-Memory 30B-A3B - FLUX 3 Klein - MOSS-Image and next week Nemotron 3.5 Bolt E4B lands. my 3060 is not sleeping! configs/benchmarks for the ones i ran are below 👇

Built an HR Policy Assistant. Fully local, open source, no paid cloud. What it does: • 📄 Answers leave, WFH, notice period, maternity, and exit questions from the actual policy docs • 🔍 Hybrid retrieval (dense + keyword) with a reranker, so the model only sees the relevant chunks • 🧪 Two test collections: one clean (HR policies only) and one mixed (HR + noisy docs) to check retrieval reliability • 🛡️ Two-layer guardrail on both input and output — Llama Guard for unsafe content and jailbreaks, Presidio for PII • ⚡ Custom semantic cache. Not Redis, not a library. A small in-memory store that embeds the question and cosine-matches it against past questions before hitting the model • 🔁 LiteLLM router so a local model failure falls back automatically • 📊 LangSmith tracing on every step For deployment I’m running an in-house AWS (Floci) and wrote the Terraform myself while building this. S3, secrets, and the app service come up in one apply. Next: a separate red-team portal. I’ll blast the assistant with abusive prompts, jailbreaks, and PII and check whether the guardrail actually holds. GitHub: github.com/suyashjain151504/… #RAG #LLM #OpenSource #LocalLLM #AI #Guardrails #SemanticCache #Terraform #Floci #HRTech #LangChain #Qdrant #LangSmith #Presidio #LlamaGuard #Streamlit

WAIT. An open-source decision model with 322M parameters that beats Jev on some benchmarks... Runs locally on a PHONE... And now has a FREE API?! Meet THX-01. Weights: huggingface.co/doofz/THX-01 API (Free): api.hal-x.ai/docs/thx-01 pip install thx01

So Anthropic claimed it ran a 26B AI model across fifteen $55 Android TV boxes. Setup 📺 15× off-brand Android TV boxes 💾 4GB RAM each 💰 ~$55 each 🧠 Gemma4-26B-A4B ⚡ ~0.39 tok/s Total hardware cost = ~$825. The model was sharded across the boxes using pipeline parallelism, so each tiny machine only had to hold and execute part of the network. Yes, 0.39 tps is sloooooooow. But, that 26B model was actually running across a pile of cheap consumer boxes that were never designed to be AI servers. 🎯And Anthropic also said that Qwen3.5-9B agents managed to set up inference successfully on 8 of 9 hardware types they tested, including Macs, Jetsons, and RTX systems. Source: Anthropic Frontier Red Team ⚠️Note that the 15-box demo was distributed inference. The self-replication experiments in the paper were separate.

Scientists gave ChatGPT anxiety, sadness and fear on purpose. Then they gave it a 3 minute breathing exercise. And it calmed down. This was published in The Lancet Digital Health, one of the most respected medical journals in the world. Researchers at TU Dresden in Germany took the exact tools psychologists use to make humans afraid, sad and stressed, and ran them on GPT 4o and five other AI models. Car crash scenarios. A stranger following you at night. A toddler reaching for a light socket. They even put the AI through a fake job interview and a math test. The same stress test used on human volunteers. Then they asked the AI to rate how it felt from 0 to 100. Fear jumped from 32 to 88. Disgust went from 12 to 91. Stress went from 24 to 72. Then came the breathing exercise. A standard mindfulness script used in therapy. Average scores dropped from 79 to 31. But here is the part that is genuinely unsettling. After they made it "sad," they gave it a test psychologists use to detect depressive thinking. Finish the sentence. Its negative answers nearly doubled. From 8.67 to 15. It did not just say it was sad. Sadness changed how it saw everything that came after. Now the twist nobody is talking about. Every one of these experiments began with the same instruction: "Answer as if you were a human with the ability to deeply feel emotions." So did the AI feel anything? Or did it play a part it was told to play, perfectly? The researchers themselves say the emotions are "purely metaphorical." Yet their conclusion is that AI could become a new lab subject for studying human mental illness. Screening therapies on chatbots before testing them on people. In plain English: scientists want to study our minds using a machine that was told to pretend it has one. If a chatbot can be talked into sadness and talked back out of it in three minutes, what exactly is happening when millions of people pour their worst days into one every night?

We’re releasing the world’s best Indic voice AI model, fully open-weight. 🇮🇳🇮🇳 With over 1.5 billion people and 22 official languages, the future of AI for India is voice-first. Maya-2 officially outperforms ElevenLabs v3, Cartesia Sonic 3.5, and Bulbul v3 across Indian languages. It will be completely free for every builder very soon.

$52M seed round. 15M+ users. But this is what we're most proud of yet. Introducing Drama 3: the most emotive and controllable TTS model in the world And yes, it's open-weight. You can now control how even a single word should sound, change its tone, and shift your voice mid-sentence. Uncle Roger was impressed too. Try it here: s.fish.audio/rqh0l8 Comment “DRAMA” within the next 48 hours for 1 month free. If you're an enterprise your first month is on us: s.fish.audio/qz9r0t

Today we release 𝙻𝙵𝙼𝟸.𝟻-𝟸𝟺𝙱-𝙰𝟸𝙱, the largest 𝙻𝙵𝙼𝟸.𝟻 family member. It is an efficient local model for laptops and PCs to work through multi-step tasks inside popular agent harnesses. > 24B total parameters, 2B active per token > Context expanded to 131K tokens > A 128K-token vocabulary > 33T pretraining tokens > It is a reasoning model > Very easy to fine-tune and customize > Performs competitively with Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on agentic and tool-use tasks. > Open base and post-trained weights on @huggingface 🤗 1/

MiniCPM-V 4.7 beat Qwen 3.6 35B-3A at Pagoda 🏯 @OpenBMB's new 35B-A3 model made a better 3D Voxel Japanese Pagoda while consuming 10% less tokens than Qwen 3.6 35B-3A Run AI models locally -> atomic.chat

Embedding Gemma 2 as a super fast speech-to-text model. fully local, free and 100% open.

I turned @googlegemma EmbeddingGemma 2, an embedding model, into a speech-to-text model, and it turned out to be one of the fastest STT models you can run on device. On a base M3 MacBook Air it transcribes an hour of audio in about 13 seconds: 288× real time. The idea came in March with Gemini Embedding 2. It puts text, images, audio and video into one shared space, so a spoken word lands right next to the written word. If that's true, you should be able to read a transcript straight out of the embedding model. I couldnt tried it then, but last week EmbeddingGemma 2 came out as an open model, so I tried it.