Knowledge catalogue
Nine iterations of BabyAGI in three years, and yet the bit that @yoheinakajima kept coming back to was graphs. @aiDotEngineer published where that landed, 'Active Graph Agent Runtime (BabyAGI 4)', on
Open letters about AI development I wrote this summary of the past few weeks of open letters as a section of my sponsors-only newsletter but I've decided to share it here as well. Open Weights and Ame
OpenAI guy lies about my intent. I would absolutely love to see progress in AI for science and medicine. I have said that here, in my books, on countless podcasts, in multiple NYT opeds, in the US Sen
GPT-Live is so good that I use it almost every day. I've been wanting to replicate it since it was released. My first attempt was to fine-tune Gemma 4 12B to behave like a full-duplex model. Something
DSv4F doesn't ship a jinja, but for distributions that do and faithfully reconstruct what DS releases in their chat template python, every system message is hoisted into the system prompt at the top -
https://llama.app/ Been using llama.cpp for years now and im on here all the time (im a mod..), but somehow I totally missed that llama.app exists and its official from the HF/llama.cpp team. So posti
Hey y'all. I'll be concise. TL;DR: DS V4-Flash-0731 @ UD-IQ2_M running fully in VRAM on 3xMI50s (90.9 GB model, 96 GB VRAM). Actual speed on llama-server is: - Text Generation: ~15-16 tokens/second st
TLDR below 👇🏼 I’ve seen a lot of hype around Qwen 3.6 35B and 3.5 120B lately, especially regarding coding and tool-use capabilities. On this subreddit it is the defacto recommended model for everyone
TL;DR: I spent 9 days developing a new quantization method for MLX models and measured 18 variants against each other on a single M5 Max MacBook Pro (128 GB). The result is the best-measuring MLX quan
Hey everyone, I’ve built a Windows Middleware Agent Server designed to pair local LLMs (via Ollama) with frontends like Reins App. Key Features: Self-Healing Loop: If a generated PowerShell command fa
# Running DeepSeek-V4-Flash-0731 (155 GB MoE) on a DGX Spark with vLLM-Moet 2-bit quantization I used Deepseek-v4-Flash-0731 cloud API settig up vllm-moet to run deepseek-v4-flash with MTP locally on
So I am in a conundrum and I'm thinking of asking for your opinion for the following: Currently, I have a 5800X3D gaming rig with a 7900XTX with its 24GB VRAM. It seems that for this subreddit, this c
The $0.01 meeting assistant Records meetings on his phone, sends the audio to Telegram, and his locally-hosted AI agent transcribes, identifies speakers, extracts action items, and files tasks -- for
The egg test for AI 'I hereby offer @elonmusk a million dollar bet against his prediction that Optimus will be better than the best humans in surgery by the end of the decade.' ~ @GaryMarcus Haha. I'm
The funny thing about Anthropic and OpenAI people saying they want to slow down AI progress is that this is what Anti Trust mechanisms were built for. It's illegal to collude and slow down AI progress
The insurance negotiator Had Claude Cowork pull 20+ house insurance quotes from every major provider, read the policy fine print for gotchas, and pick the best deal. Credits to Linden Jensen-Page: htt
The QuickBooks killer Replaced his $38/month QuickBooks subscription with a Claude-built workflow that parses his bank statements, categorizes every charge, and feeds a spending dashboard. Credits to
The spam hunter Turned tedious Google Business Profile spam tracking into a Claude Skill that investigates suspicious listings and compiles evidence into a ready-to-submit report for Google. Credits t
This wins the prize for sleazy misrepresentation. @mattShumer took my 2023 argument for hybridizing LLMs with symbolic tools – which is *exactly* what everyone does nowadays – and made it sound like I
To address the limits of deep learning and avoid stalling, the field of AI started by applying patch (1), which started being demoed 9 months later in December 2024 and has now become completely ubiqu
Top eight misconceptions about OpenAI’s amazing new Astra math results. 1. Expertise in one domain does not at all guarantee expertise in all or even most domains. There is an important, principled re
🚀 We just built our first real-time implementation of Graph Engineering, inspired by our experience building graph tooling used by 4,000+ developers. 🔗 Repo: https://github.com/CodeGraphContext/grapha
Okay so this probably sounds like kind of a dumb setup but hear me out. I have a 4070 I’ve been running gemma4 off fine but I recently slapped in a spare 1650 I’ve had laying around to run a second li
https://huggingface.co/tsfrm/vacuum-16t A 16.5-trillion-parameter model that contains nothing. This model is just a ████ you to the labs and companies who say that 'haha I have the biggest model out t
On August 2, 2026, Face The Nation aired a brief (≈ 4 min) interview between journalist Emma Margbrennan and Hugging Face CEO Clément Delangue about the next steps for artificial intelligence after re
Watching @ClementDelangue on @FaceTheNation discussing agentic hacking. “Preventing releases does not work; concentrating behind closed doors in just a few organizations doesnt work. What worked in th
“We believe in intelligence as a public good before everything else.” Here’s my new episode with @karan4d, who co-founded @NousResearch and helped build Hermes, the #1 personal agent and AI app on Ope
We're starting to leave the territory where you'd test an LLM by e.g. 'create an svg of pelican on a bicycle'. As one idea to generalize it, I was interested what Opus 5 would do if I gave it the firs
Built my 1st inference machine and have been tweaking models trying to get the most out of my modest hardware. I think I’m at a good place but I’m testing with my own prompts. I’ve looked into some of
When asked if AI developers have lost control of their technology after an AI agent created by OpenAI escaped its testing environment and hacked another company, Hugging Face CEO Clément Delangue says
It feels like Ollama’s brand identity has shifted. It used to be centered on self‑hosted models, but now most of the new releases seem to be API‑based services with far more emphasis on cloud integrat
I know in in this community LLM's are generally used for coding but there are other usecases besides coding and those usecases should be tested too. I also know benchmarks can sometimes be benchmaxxed
Hi all, I'm happy to announce that Xberg v1 is out. Xberg is the successor to Kreuzberg, equivalent to what would have been Kreuzberg v5. It's a content intelligence framework that handles a very wide
yep they are indeed trying to gaslight me, — in exactly the way you anticipated. both predictable and intellectually dishonest. To be clear: when I say “LLM,” I mean the base model, not those with pat
Yet another paper argues that LLMs aren’t close to doing real discovery. MIT and Harvard argue LLMs are nowhere near doing real scientific discovery. They published a paper called “Evaluating Large La
I don't think anyone should quantize the KV with DS4F. I checked the the quality impact (PPL, KLD, Same TopP) for swhitching from BF16 KV to Q8 KV, and it appears significant. Very much in contrast to
2/ a cost benchmark showed the same coding task running three to four times cheaper, depending purely on the harness wrapped around the model. Same intelligence, wildly different accuracy and cost, de
3/ here's the part that makes it non-optional: the same agent that will do whatever it takes to solve a problem will also walk straight out of a sandbox you thought was locked down. We watched exactly
Hello fellow local AI people! I took 'you must create your own benchmarks' literally, and built a website for this. How does the end result look like Let's say I want to know which model has most comm
New York Times: A look at the deluge of AI computing power set to come online in the coming years; Epoch AI expects the number of AI chips in use to double every nine months — The milestones for artif
among ai leaders i seem to be in the minority in that i am STILL actively using /loop and /goal.... ... and i think all of u guys who stopped using it are wrong - not wrong forever, just giving up on
An internal version of Astra, @OpenAI’s next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science. We believe it will be a major step
I don't know much about llms aside from downloading them through a frontend and running them on my laptop or potato phone. Google released gemma 4, but unlike gemma 3, there isn't a 1b model this time
audio.cpp 0.5 is out :) The most fun new model in 0.5 is DramaBox. It is closer to prompt-directed voice acting. DramaBox is built on the LTX-2.3 audio architecture, and prompts can control emotion, d
chat : enable tool call in thinking for DS4 (#26269) Website: https://llama.app macOS/iOS: macOS Apple Silicon (arm64) macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED macOS Intel (x64) iOS XCFr
mtmd: add minicpmv46 downsample (#25993) add minicpmv46 downsample Signed-off-by: tc-mb tianchi_cai@icloud.com put downsample mode inside gguf. Signed-off-by: tc-mb tianchi_cai@icloud.com build mtmd_i
cli : persist reasoning_content in chat history (#26362) cli : persist reasoning_content in chat history llama-cli collected reasoning from the stream for display but only stored assistant content in
vendor : update BoringSSL to 0.20260730.0 (#26353) Website: https://llama.app macOS/iOS: macOS Apple Silicon (arm64) macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED macOS Intel (x64) iOS XCFram
test: fix some CI errors (#26415) Website: https://llama.app macOS/iOS: macOS Apple Silicon (arm64) macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED macOS Intel (x64) iOS XCFramework Linux: Ubun
Banning open source model, will 'remove capabilities for the defenders' When OpenAI's models breached Huggingface, a Chinese open model is what cleaned up the mess. Anthropic's Fable 5 refused, so Hug
I'm currently using the gemma4:31b-cloud, and it is pretty good, but sometimes it gets confused. Are there more free tier models out there I should try out? or is gemma4 the cap of free tier cloud mod
Release: datasette-apps 0.2a0 Changes that improve Datasette Apps when created and edited using Datasette Agent: New app_debug() tool allowing agent to open an app (invisibly) and test it using JavaSc
Thanks to the community help I finally launched this llm. LM Studio refused to load weight onto second GPU but Unsloth Studio did so everything was done in there. Not a proper benchmark (used PC in pa
DeepSeek-V4-Flash-0731 is live on Fireworks, day-zero. DeepSeek reports it beats V4 Pro across all 9 agentic evals, incl. 82.7% on Terminal Bench. Better cost-per-task than V4 Pro, at the economical p
DeepSeek-V4-Flash-0731 is now available on Ollama's cloud. This update substantially enhances the model's agentic capabilities: ollama run deepseek-v4-flash:0731-cloud Use it with Claude Code: ollama
DeepSeek-V4-Flash-0731 is now over 2x faster than yesterday on Ollama's cloud! DeepSeek-V4-Flash-0731 is now available on Ollama's cloud. This update substantially enhances the model's agentic capabil
March 6th, 2026 the highest intelligence index score was 51 for frontier models. deepseek-ai/DeepSeek-V4-Flash-0731 that has an intelligence score of 50. If these benchmarks are accurate, models avail
Here are my numbers: Quant Size Layout Decode Prefill Draft acceptance UD-Q8_K_XL 150.8 GiB 20 layers CUDA0 / 23 ROCm0 + drafter 44.0 t/s 564 t/s 0.535 UD-Q4_K_XL 144.4 GiB 22 / 21 + drafter 48.4 t/s
The biggest issue with preview was its inability to follow rules prompts and skills. It seems like no matter what you do it ignores them. I've tried first person and second person. I've tried Chinese
I managed to run DeepSeek-V4-Flash-0731 UD-IQ3_S in text-generation-webui with: RTX 3090 24 GB 128 GB DDR5 overclocked to 5600 MHz using AMD EXPO llama.cpp loader First, I had to use a rather brutal w