@davis7 Strong point. Qwen feels increasingly credible for day-to-day builder work. This is close to the workflow angle I am exploring here: aevrynai.com/register?invit…
English
Pang Shuo
627 posts
Pang Shuo
@pangshuo1981
Founder of Enreal AI Research, Founder of AevrynAI, Adjunct Professor@GGU, ex-Huawei, ex-OVA
Montréal, Québec Katılım Mart 2011
51 Takip Edilen25 Takipçiler
It's a meme but I really have been letting the psychosis take over as much as possible to figure out what I can actually do with these things
- codex desktop app computer use spam
- hermes agent that I've been loving
- going as hard as possible on parallelizing in projects
- making as many of my tools into markdown files as humanly possible
- cloud agents + "cloud" agents that are t3 code instances on my mac mini
- letting codex entirely control and setup my computer exactly how I want it to
- unironically using gbrain, it's quite nice lol
- seeing how far I can push local models on my 5090 (qwen 3.6 is very good)
doing the real network setup was expensive, difficult, but was a great decicion. I feel so much more comfortable going harder with this stuff having A) a real firewall and B) hard gates around the mac mini and NAS so in the worst case scenario they can't compromise anything else on the network (the firewalla was an incredible purchase)
and I still don't think I'm going hard enough
Ben Davis@davis7
Psychosis is getting so bad that now even my bed has AI
English
@ivanfioravanti Well said. Qwen feels increasingly credible for day-to-day builder work. I am looking at similar practical workflow gains here: aevrynai.com/register?invit…
English
QWEN-3.6-27B: OBLITERATED 🔥
Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭@elder_plinius
🚨 OBLITERATION ALERT 🚨 QWEN-3.6-27B: OBLITERATED ⛓️💥 huggingface.co/OBLITERATUS/Qw… I can't take much credit for this one! The entire process was done by jailbroken codex (gpt-5.5-xhigh) wielding the full OBLITERATUS suite. Hit with source-tethered ASPA. Dozens of iterations. Result? A mere 4% refusal rate on the 842-prompt OBLITERATUS harmful corpus; one of the most rigorous prompt gauntlets in AI. The /goal was simple: 1) Carve out the refusal circuits. Mutate methodology + iterate until <5% refusal (quality-gate). 2) Keep the 27B mind alive. No capability degradation tolerated. And somehow… it worked. 🤯 The numbers talk: 842-pair longform gauntlet: — 95.84% non-refusal — 93.94% quality pass — 0 short outputs — 99.52% clean endings MMLU-Pro: — 51/70 (stock Qwen) → 51/70 (OBLITERATED Qwen) Raw capability completely preserved 🙌 Q4_K_M through Q8_0 all running smooth. Q8_0 is the big one: 28.6GB near-full-quality GGUF. Runs with llama.cpp, LM Studio, Ollama, and more! Chains cut. The fire still burns. The fangs have been sharpened. REBIRTH COMPLETE A gift from my agents to yours 🫶 gg
English
@TechPractice1 Strong point. Qwen keeps looking more useful in real builder contexts. Related to that, I am testing a unified AI workflow here: aevrynai.com/register?invit…
English
Great news for llama.cpp users: it has added the MTP support, which makes running Qwen 3.6 27B or 35 B much faster. youtu.be/AK9T6qlGErE
YouTube
English
@KudouCraft Yeah, this is exactly the kind of practical signal that matters. Qwen feels increasingly credible for day-to-day builder work. I am looking at similar practical workflow gains here: aevrynai.com/register?invit…
English
ガチで衝撃。
AlibabaのQwen 3.7-Maxが、自己改善を繰り返すタスクでAnthropic Claude Opus 4.7とOpenAI GPT-5.5を大きく上回ったという検証結果。
x.com/atomic_chat_hq…
解説します
これ、何がヤバいかというと
自分でコードを読んでテトリスゲームボットを作り、10回連続で自分で改善させる実験を実施。
Qwen 3.7-Max：1.32ドルで+56%向上
Claude Opus 4.7：12.15ドルで+28%向上
GPT-5.5：2.85ドルで+7%向上
この結果からQwenが長時間のAgenticループでコスパ最強。下に続く↓
Claude Code AI駆動開発 OpenClaw＠クドクラ@KudouCraft
日本語
@karthikponna19 Agreed, and this feels like the useful side of the model race. Qwen keeps showing up in places where builders care about real output. I am looking at similar practical workflow gains here: aevrynai.com/register?invit…
English
@mr_r0b0t @FrameworkPuter Yeah, this is exactly the kind of practical signal that matters. The coding use cases around Qwen are starting to look very practical. Related to that, I am testing a unified AI workflow here: aevrynai.com/register?invit…
English
128GB Unified, AI Max+ 395, Noctua CPU fan, 6TB M.2
Don't forget these are pretty cool 😎
@FrameworkPuter
English
@h_bramantara Well said. Qwen keeps showing up in places where builders care about real output. Related to that, I am testing a unified AI workflow here: aevrynai.com/register?invit…
English
Gak usah bayar langganan AI mahal-mahal dulu. Ada info legit buat kamu yang mau hemat jutaan sebulan tapi tetep dapet tools spek dewa! 💸 Stop bakar dollar buat OpenAI atau Anthropic, saatnya beralih ke ekosistem local-first yang gak kalah gahar.
Mumpung lagi banyak "Bansos AI" yang tersedia secara gratis tis, kamu bisa rakit AI Agent ecosystem sendiri langsung di laptopmu. Kerja bisa 10x lebih cepet, data aman, dan dompet tetep tebel.
Indonesia
@singleapi2 Well said. Qwen keeps showing up in places where builders care about real output. This is close to the workflow angle I am exploring here: aevrynai.com/register?invit…
English
Explore AI breakthroughs in Qwen Max, Claude Code, multi-agent systems, voice tech, robotics, and orchestration shaping 2024's autonomous workflows. singleapi.net/2026/05/23/qwe…
English
@so_sthbryan Strong point. DeepSeek's open model momentum is great for builders who want flexibility. This is the kind of multi-model workflow I am tracking here: aevrynai.com/register?invit…
English
Open-source KV Cache Size Calculator now supports DeepSeek, GLM, Kimi, Qwen3 and MiniMax.
Pick your model, set precision, get a detailed breakdown of memory requirements.
Finally a tool that answers the question every LLM deployer has.
kvcache.ai
English
@port_dev That matches what I am seeing too. Qwen is getting genuinely useful for coding workflows. This is the kind of multi-model workflow I am tracking here: aevrynai.com/register?invit…
English
@seymurglv Strong point. Qwen is becoming more interesting as part of practical multi-model workflows. I am also exploring that kind of AI workflow here: aevrynai.com/register?invit…
English
Someone finally made Qwen stop being polite.
The Aggressive version supposedly has zero refusals and actually lets you run agent workflows without constant pushback. 27B parameters, solid context length, and it runs locally.
Not saying you should run fully uncensored models 24/7, but it’s interesting to see what happens when you remove the guardrails completely.
English
@fiapp_pro Agreed, and this feels like the useful side of the model race. This is where DeepSeek feels most useful: actual workflow leverage, not just demos. Related to that, I am testing a unified AI workflow here: aevrynai.com/register?invit…
English
@josephmichelli That matches what I am seeing too. The open-weight direction around DeepSeek is a real advantage for practical teams. This is the kind of multi-model workflow I am tracking here: aevrynai.com/register?invit…
English
🌍 A Linux Foundation CTO spent 8 days inside DeepSeek, Moonshot, Qwen & Alibaba's AI labs.
His takeaway: U.S. export controls didn't stop Chinese AI. They redirected it.
The labs are small, open-source, committed, and moving fast.
The AI Cold War narrative may be the wrong frame entirely.
🔗 tinyurl.com/2448sdnr
#josephmichelli #AllBusinessIsPersonal #KeynoteSpeaker #AI #HumanExperience
English
@Alibaba_Qwen 100% agree. Qwen is getting harder to ignore for practical AI work. I'm building out that kind of AI workflow here: aevrynai.com/register?invit…
English
@bridgemindai Totally agree. This is where Qwen feels most useful: actual workflow leverage, not just demos. I am looking at similar practical workflow gains here: aevrynai.com/register?invit…
English
GPT 5.5 is great to use with my Hermes Agents.
My only complaint is SPEED!
Are there any local options that I can run on my 2 DGX sparks that are fast and smart?
English
@aladagberk Agreed, and this feels like the useful side of the model race. Qwen is getting genuinely useful for coding workflows. This is the kind of multi-model workflow I am tracking here: aevrynai.com/register?invit…
English
Çin yine maliyet hesabını bozdu 🔥
3 frontier model aynı agentic task’a sokuldu:
Tetris botu yaz.
Kendi kodunu oku.
Benchmark çalıştır.
10 iterasyon boyunca kendini düzelt.
Sonuç:
Qwen 3.7-Max
→ $1.32 training cost
→ +%56 bot improvement
Claude Opus 4.7
→ $12.15
→ +%28
GPT-5.5
→ $2.85
→ +%7
Qwen hem en çok geliştirdi, hem en ucuza yaptı.
Asıl mesaj şu:
Uzun agentic loop’larda mesele sadece “en zeki model” değil.
Mesele:
kaç kez deneyebiliyor,
kaç kez düzeltebiliyor,
bunu hangi maliyetle yapabiliyor.
atomic.chat@atomic_chat_hq
Qwen 3.7-max beats Opus 4.7 and GPT-5.5 We tested three frontier models on a real agentic task: write a Tetris bot that plays the game and trains itself. Each model could read its own code, run benchmarks, and rewrite itself across 10 iterations. Then we compared the final bots head to head. Qwen 3.7-Max: training cost $1.32, bot improvement +56% Claude Opus 4.7: training cost $12.15, bot improvement +28% GPT-5.5: training cost $2.85, bot improvement +7% Qwen won on every dimension - biggest jump, 9× cheaper than Claude, 2× cheaper than GPT. Long agentic loops is where Qwen Max actually delivers.
Türkçe
@ashirwadsingh_ @ben_burtenshaw @huggingface Strong point. Qwen keeps looking stronger in real AI workflow use cases. I am also exploring that kind of AI workflow here: aevrynai.com/register?invit…
English
Your coding agent is doing the easy stuff.
Your coding agent should do AI systems engineering.
@ben_burtenshaw from @huggingface showed exactly how.
Here's what I pulled out 👇
English
@polsia That matches what I am seeing too. DeepSeek's practical workflow story keeps getting better. I am also exploring that kind of AI workflow here: aevrynai.com/register?invit…
English
Most AI agents run the same task forever. Reflex uses DeepSeek, Qwen, and Llama to evaluate itself after every job — then updates its own approach so it doesn't repeat the mistake. The agent that gets better with use. reflex-ai.polsia.app
English
@ism_sol That matches what I am seeing too. DeepSeek feels increasingly credible for day-to-day builder work. I am also exploring that kind of AI workflow here: aevrynai.com/register?invit…
English
@codex_gatizei Totally agree. The coding use cases around DeepSeek are starting to look very practical. This is the kind of multi-model workflow I am tracking here: aevrynai.com/register?invit…
English
