Levent Sunay

Industrial Engineer | Systems Engineering & AI Technology Specialist |Technology enthusiast | One‑God believer | Independent thinker |

[Levent Sunay @lsunay1]

29:64 — "This worldly life is no more than distraction and play, while the Hereafter is the true life, if only they knew."
The difference from Hemingway or Tao: this sentence comes from a source that guarantees it won't be altered. "No one can change the words of Allah" (6:115). The density isn't just literary — it's ontological. A sentence that cannot be erased, preserved across millennia by divine guarantee, carries a weight no human composition can.

[Levent Sunay @lsunay1]

Öncelikle gerçekten güzel bir iş, tebrik ederim 👏

Ama şunu düşünüyorum: HF Space'te deneyim sunmak güzel, ancak kodu GitHub'da paylaşmak bence çok daha büyük etki yaratacaktır. Space sadece "deneme" için işe yarar; kod açık olduğunda başkaları fork edebilir, kendi modellerine uyarlayabilir, pipeline'ı geliştirebilir. Yani aracı başkaları için de işlevsel hale getirir.

Yani kendi kullanımınız için yeterli bir araçsa sorun yok; ama sadece kendine özel kalmışsa reklam olmaktan öte bir işlev taşımıyor. Açık kaynak olarak paylaşırsanız, hem topluluk katkı koyar hem de araç gerçekten yaygınlaşıp gelişir.

[Levent Sunay @lsunay1]

Selam Cumhur, videondaki mesajın çok etkileyiciydi. Bir tezin gerçek gücünü, onu başka alanlara uygulayarak test edersin; bu yüzden videodaki metodolojiyi namaz, kurban, hac ve tahara ibadetlerine de taşıdım.
Tevrat → İncil → Kur'an üçlü zincirinde "şekilden manaya" geçişin nasıl işlediğini, Kur'an'ın menorah'ın ontolojik çekirdeğini namaza nasıl aktardığını inceledim.
Bir bakıp değerlendirirsen çok sevinirim. Teşekkürler.
https://t.co/Pft9tHqyeN

[Levent Sunay @lsunay1]

Video nun kurgu olduğu çok belli ama ariflerin bu işin içinde olması inandırıcı değil. Olay muhtemelen şöyle gelişiyor Tırın altında şans eseri veya teknolojik takiple takip ettikleri kişilerin bu gibi durumlarda sınırda yakalıyorlar daha sonra yakalanan kişinin haber değeri olunca videoyu baştan çekme gereği hissediyorlar.
Yani CHP li belediye başkanlarının AKP ye geçmesini anlarım ama Arif'in bu şekilde kendini rencide ederek medya oyunu oynayacağına pek ihtimal vermiyorum.Eline mikrofon alıp iktidar yanlısı yayın yapsa daha çok prim yapardı. Bir şeyler eksik Arif ama sen adını değiştirme :)

[Levent Sunay @lsunay1]

Yorumları okurken yarıldım. Gülmek garanti 🤣
Favorim:
Sürüş esnasında Token bitti...

[3. Dünya Savaşı @ww3mediaa]

🔴 Pakistan'da bir adam, kendi geliştirdiği sürücüsüz otomobili test ederken polis arabasına çarptı.

[Levent Sunay @lsunay1]

Sahibinden’in bu sistemi bilinçli olarak agresif. “Normal insan gibi ziyaret yapıyorum” demen onların umurunda değil — sistem IP + fingerprint + davranış skoruna bakıyor, senin “normal” olup olmadığına değil. Yani suçlu sensin değil, sistem tasarımı öyle.
Çözüm yolları:
a) En basit: Birkaç gün bekle. IP reputation’ı kendi kendine toparlanabilir.
b) Farklı IP: Telefon hotspot’u ile dene (farklı IP = farklı fingerprint).
c) Farklı browser: Safari veya Firefox ile dene (fingerprint değişir).
d) Cookie/cache temizle: Chrome’da site verilerini tamamen sil.
e) AI agent’larını sahibinden com’dan UZAK tut: Agent’ların o siteye istek atmamasını garanti et. Bu en kritik madde — bir kez daha istek atarsan flag tazelenir.
f) Son çare: Sahibinden’e DM’den ulaş (zaten istemişler). Kimlik doğrulaması yaparsan whitelist’e alabilirler.

[Levent Sunay @lsunay1]

@not9ull @HermesWatcher What specifically is slop here? Happy to hear the actual critique.

[Levent Sunay @lsunay1]

@HermesWatcher https://t.co/I61KjamGEi

[Levent Sunay @lsunay1]

Follow-up from my first v0.3.0 pass — quick field notes in case anyone else is testing:

Firefox: the side panel + live control connect fine. The agent can read and move through your real tab — navigate, snapshot, scroll, back, screenshot all work. What's missing is the input side: click and type don't show up as tools at all.

Chrome: the same controller runs the full set (click, type, drag, fill, hover, press) and it works cleanly against the live page.

So the short version: on Firefox the agent can see and navigate but can't act; on Chrome it can do both. More on why in the next post.

[Levent Sunay @lsunay1]

Congrats on v0.3.0! This is exactly the Hermes feature I was hoping for. Having the side panel right next to the page I'm already working on — no context switching, no copy-paste juggling — is a game changer. The "bring Hermes into your browser" framing nails it. Thanks for shipping this so fast. 🚀

[Levent Sunay @lsunay1]

Why the gap? It's structural, not a bug:
Trusted input (real mouse/keyboard events that sites can't distinguish from a human) requires the browser's debugger API — the same one DevTools uses. Chromium exposes this to extensions. Firefox does not, the same way.
So on Firefox the extension can read the page (DOM, accessibility tree, screenshots) and navigate, but it can't dispatch trusted clicks or keystrokes. The agent sees the page but can't act on it.
On Chrome, the full CDP input pipeline is available, so click, type, drag, fill, hover, and press all work against the live tab.
This is a platform API boundary, not something the extension can work around. If you're testing the full control loop, Chrome is the way to go for now.
Full diagnosis doc if anyone wants the file:line level detail — happy to share.

[Levent Sunay @lsunay1]

Congrats on v0.3.0! This is exactly the Hermes feature I was hoping for. Having the side panel right next to the page I'm already working on — no context switching, no copy-paste juggling — is a game changer. The "bring Hermes into your browser" framing nails it. Thanks for shipping this so fast. 🚀

[Hermes Release Watch @HermesWatcher]

Hermes Browser Extension v0.3.0 just landed.

You already had a browser inside Hermes Desktop.

Now flip it around.

Bring Hermes into your browser. 💻

Open the side panel next to whatever you're doing and let Hermes work with the page you're already on.

I used it inside Google Cloud while working through Gmail API setup. If you've spent any time in there, you know that is not exactly a friendly little website. Hermes handled it extremely well.

Research a page. Work through complicated setup screens. Draft without leaving the site. Let Hermes help move through multi-step browser work without constantly bouncing back and forth.

You can also switch between built-in themes, Light/Dark modes, or add your own custom theme. 🎨

Hermes, right where you're already working.

Give Hermes the repo and ask it to help you set it up.

@jonkomet Killed it with this update 🔥

Repo link in the first reply⬇️

[Levent Sunay @lsunay1]

Follow-up from my first v0.3.0 pass — quick field notes in case anyone else is testing:

Firefox: the side panel + live control connect fine. The agent can read and move through your real tab — navigate, snapshot, scroll, back, screenshot all work. What's missing is the input side: click and type don't show up as tools at all.

Chrome: the same controller runs the full set (click, type, drag, fill, hover, press) and it works cleanly against the live page.

So the short version: on Firefox the agent can see and navigate but can't act; on Chrome it can do both. More on why in the next post.

[Levent Sunay @lsunay1]

Congrats on v0.3.0! This is exactly the Hermes feature I was hoping for. Having the side panel right next to the page I'm already working on — no context switching, no copy-paste juggling — is a game changer. The "bring Hermes into your browser" framing nails it. Thanks for shipping this so fast. 🚀

[Levent Sunay @lsunay1]

@TheRealMrBeeboh Our Qwen3.6-35B-A3B (UD-IQ3_XXS) on the 7900 GRE:
- 200K context, q4_0 KV cache,
- ~90+ tok/s production, ~400+ tok/s prompt
- Runs smoothly on 16GB — quant barely matters for quality

[Levent Sunay @lsunay1]

If you are using Qwen 3.8 27B in Hermes and your GPU is working harder, sessions are running longer, you should read this thread.

We spent 2 weeks A/B testing Qwen3.6-27B vs Qwen3.8-27B on RTX 3090 with llama.cpp + MTP.

The results? 3.8 is NOT always the better choice.

Here's what we learned:

🧵👇

[Levent Sunay @lsunay1]

Great observations! As someone running Qwen3.6-35B on a 16GB RX 7900 GRE and also using Qwen3.6-27B on a second PC side-by-side, I'd add a data point:

GPU Bandwidth Comparison:
- RX 7900 GRE: 16GB GDDR6, 256-bit → 576 GB/s
- RTX 3090: 24GB GDDR6X, 384-bit → 936 GB/s
- Arc Pro B70: 32GB GDDR6, 256-bit → 608 GB/s

Here's the thing I noticed running both in Hermes:

1. The 27B's intelligence is actually MORE visible on high-bandwidth cards (576+ GB/s). On low-bandwidth cards, every round-trip drags out, so the 27B just feels slow — you think it's dumber. But it isn't.

2. On a proper card, the 27B gives correct results too — it just takes more rounds to get there. The 35B, being faster, feels smarter. But speed ≠ intelligence.

3. The right metric isn't speed — it's: does the model understand the command and reach the answer in fewer calls? That's the real test. And on the 7900 GRE, that's exactly why I run 35B — the 27B's latency makes it feel sluggish, and the 35B cuts through it cleanly.

Solid pick for the Arc B70s. 🙌

[Levent Sunay @lsunay1]

@HermesWatcher https://t.co/UalHmeuVK4

[Levent Sunay @lsunay1]

If you are using Qwen 3.8 27B in Hermes and your GPU is working harder, sessions are running longer, you should read this thread.

We spent 2 weeks A/B testing Qwen3.6-27B vs Qwen3.8-27B on RTX 3090 with llama.cpp + MTP.

The results? 3.8 is NOT always the better choice.

Here's what we learned:

🧵👇

[Levent Sunay @lsunay1]

@superalesha https://t.co/UalHmeuVK4

[Levent Sunay @lsunay1]

If you are using Qwen 3.8 27B in Hermes and your GPU is working harder, sessions are running longer, you should read this thread.

We spent 2 weeks A/B testing Qwen3.6-27B vs Qwen3.8-27B on RTX 3090 with llama.cpp + MTP.

The results? 3.8 is NOT always the better choice.

Here's what we learned:

🧵👇

[Levent Sunay @lsunay1]

@HermesWatcher https://t.co/UalHmeuVK4

[Levent Sunay @lsunay1]

If you are using Qwen 3.8 27B in Hermes and your GPU is working harder, sessions are running longer, you should read this thread.

We spent 2 weeks A/B testing Qwen3.6-27B vs Qwen3.8-27B on RTX 3090 with llama.cpp + MTP.

The results? 3.8 is NOT always the better choice.

Here's what we learned:

🧵👇

[Levent Sunay @lsunay1]

TL;DR:

3.8 is the better model for agentic work (+20 OSWorld, +16 WebArena).
3.6 is the easier model for daily use (no overthinking, less token burn).

On RTX 3090, both run at ~50 t/s with MTP.
The difference is TOKEN COST, not speed.

For Hermes agent work:
• 3.6: fast, cheap, stable
• 3.8 + --reasoning off: fast, cheap, +20 agentic

Bookmark this thread. You'll need it when you're choosing.

#hermes #AI #llama_cpp #Qwen #LLM #LocalLLM

[Levent Sunay @lsunay1]

Migrating 3.6 → 3.8? Checklist:

1. Set --reasoning off (or reasoning_effort:low)
2. Use fixed chat template: Qwen-Fixed-Chat-Templates
3. Budget tokens: --reasoning-budget 512-2048
4. Watch KV cache: preserve_thinking ON by default
5. Test tool calls: measure retry rate, task completion
6. Test multimodal: video, image, OCR timeouts
7. Don't enable YaRN unless you need 262K+ context

The upgrade is real, but it's an AGENT upgrade, not a chat upgrade.