KKY

@VictorTaelin 1. Construct a preference(DPO) dataset, you could make it by talking with an agent. Let the agent interview you with two answers, you choose you answer. 2. Train a model through DPO-lora on that dataset. Cheap but useful

@evilpsycho42 I don't think that helps, does it?

seriously, working with AI is MISERABLE for one and only one reason: having to re-explain the same thing "oh yeah this new session obviously doesn't know what proper case trees are, so let me explain it for the 5000th time in my life" I'm tired AGENTS.md doesn't solve this because it is impossible to fit the entire domain knowledge without nuking the context - it would be 1m+ tokens worth RAGs don't solve this, the agent won't search unknown unknowns SKILLs don't solve this unless I keep like a collection of 1750 skills with specific cuts of domain knowledge for each possible subset of my domain that I might need in a given chat, but that's a lot of manual work recursive LLMs or whatever don't solve this for the same reason, you can't dump a domain book and expect the AGENT will magically guess that it is supposed to search for a specific bit knowledge. unknown unknowns fine tuning doesn't solve this (OSS models suck and OpenAI / Anthropic gave up on user fine tuning) I honestly think a good product around fine tuning on your domain would be a major hit and an underdog lab should take this opportunity

@xwang_lk DeekSeek with Pi Agent works great for me.

When DeepSeek Code?

@thsottiaux Cybersecurity warnings happen even when participating in a Kaggle competition ... At least leave an option to summarize current session so users are able to continue their work with other models?

Never talk about goblins. Our latest blog is live. openai.com/index/where-th…

@1unyy @0xSero Pi Agent

@evilpsycho42 @0xSero Hey, quick question, what CLI are you using in that screenshot?

117.4M tokens for 2.24$ for a genius.

5.5 + 4.6 > 5.4 + 4.7 Agree?

@ZeroZ_JQ Pi Agent 兄弟，搭配一个ChatGPT plus可以起飞 x.com/evilpsycho42/s…

DeepSeek 搭配 claude code 好还是搭配 opencode 好？

Leverage a ChatGPT subscription (even the free tier) to fill in Pi Agent's missing web_search and image image generation features. Use it with DeepSeek or Qwen, buckle up, bro.

@nopmobiel @Prince_Canuma Interesting, what is the memory usage then context goes to 200k?

DeepSeek-V4-Flash-2bit-DQ works on my M3 max 128GB mac! And with 39 tok/sec quite speedy. Consumes 97GB peak. Thanks a ton for the mlx port and quant @Prince_Canuma #mlx #deepseek

@NuCode Really impressive, how is the experience with long context (200k)?

Mac（M5Max）速いしQwen3.6-27bと35b-a3bが両方同時に読み込めるのでClaudeCodeのopusとsonnetに27bと35bのモデルを分けてアサインした。快適になるなあ…

@runsonai I've recently switched to Pi Agent as well. An agent amazingly suitable for customization.

Very impressed with Pi (coding agent). It's like claude code barebones & customizable. Added custom plan mode: asks opus for plan, then have the option for codex to review it. A workflow I do a lot and now seamlessly integrated. Works w/ Qwen 3.6 as orchestrator.

@bigeagle_xd Humans are not good enough to judge LLM capabilities based on just a few prompts : )

glad to see oai proved that arena has problems

@0xSero Could you please share your setup of autoresearch for PI agent? A skill or extension?

I have heard your advice and will be personally optimising and using Qwen3.6-27B on my Framework. I think these 2 could go great together, 1 for vision and UI the other for logic.

@DavidOndrej1 @OpenRouter Most random providers seem to excel at only one thing, messing up the models. - wrong chat template - bad inference params - aggressive quantization - bad prompt cache strategy

We need more DeepSeek V4 providers on @OpenRouter

优雅的PI Agent + deepseek，加上基于openai oAuth的web_search, image_gen = 舒服

@runsonai Nice job! For local LLM inference, which device would you recommend? Mac or DGX spark?

Got Qwen 3.6 35B-A3B MoE running at ~65 tok/s (c=1) and ~121 tok/s (c=4) aggregate on my Asus GX10 (dgx spark). Model stack: • Target: Qwen/Qwen3.6-35B-A3B-FP8 - Drafter: z-lab/Qwen3.6-35B-A3B-DFlash • Spec decode: DFlash, 10 speculative tokens • Context: 200k - KV cache: bf16/auto, not fp8 Used vllm for this (see flags below)

@jun_song how could DeepSeek V4 flash run on a 96g mac? tiny context length & q3 quantized?

If you are using a Mac with 96GB RAM or more: Do not use Qwen3.6 27b. • Minimax M2.7 • Deepseek V4 Flash These two are much faster and smarter. Take advantage of the unified memory.

DeepSeek infra团队简直是神。 今天在PI Agent上体验了下DeepSeek v4 flash，ttft 基本在1秒左右，tps 100左右。6M input (96.8% hit cache), 48.2K output, 合计价格 ¥0.36。 简直神了 缓存命中太稳了，我开了两个sessons，所有的requests中，只有第一条消息和我在/reload 一个extension后的消息没有命中缓存（命中了才奇怪），可以说reuqest维度下缓存命中率100%。

@deepseek_ai You've always been setting the trend!

🔥DeepSeek Input Cache Price Drop! Effective immediately, the price for input cache hits across the ENTIRE DeepSeek API series is reduced to just 1/10th of the original price! Build more efficiently for less. 📌Reminder: The DeepSeek-V4-Pro 75% OFF promotion is still active until May 5th, 2026, 15:59 (UTC Time).