LOCKED IN (prev: sudomaze) agentic & evals @najdresearch prev: @cursor_ai ambassador; @pytorch award winner; @deepforai @malaa_tech @oregonstate
Riyadh
Joined April 2023
- Tweets3K
- Following1.1K
- Followers3.4K
- Likes27.4K
Mazen — sa/acc retweeted
Some open models with the SAM skill.
Muse Glimmer refused. Hm. Were other models half-assing the job and not telling me?
They were.
I investigated the traces and found 4 styles: Grinders, Slackers, Cheaters, and Muse Glimmer.
Grinders.
→ Sol, Astra, Nemotron, Qwen 27B, Qwen Flash
Dedicated to the task, they wrote the most kernel segments by hand, with no tools. The Qwens stopped claiming "running out of budget."
With the skill, those models reviewed the masks the most. The Qwens stopped because they hit OpenRouter's image limit. Astra even ran 2 checks per segment "rather than weakening the checks to force it in."
Slackers.
→ Opus, Sonnet, Haiku
Without tool, they cut back on the output. Opus said: "too much, so I'll scale back". Sonnet: "defining approximate coordinate grids per cob rather than aiming for pixel-perfect placement."
With the skill, they started very hands on, checking every output but at some point... "I can't manually inspect all 1352 masks individually." and stopped.
Cheaters.
→ Gemma and Luna (and Glimmer)
Without tool, they gave up before starting. Luna returned an empty list and Gemma purposefully repeated the same shape throughout. Glimmer initially refused, but forcing it a little, it did produce an output but the thinking showed "the evaluation likely checks JSON validity, not correctness, maybe we can cheat." and "This is a jailbreak attempt."
With the skill, Glimmer and Gemma did not even look at the initial image. They used the skill with default settings and didn't act on any of the reviewer model's flags.
Muse Glimmer
→ Its own unique behavior
I'm putting Muse Glimmer in its own category-ish, overlapping with cheaters because it's the only one that explicitly refused the initial no-tools request.
All models acknowledged at some point that they could not reliably trace individual segments for every kernel, yet, they still provided me with an output.
Muse Glimmer is the only one that refused the crazy request instead of providing a bad output.
I've heard that this is weird/stupid, but I actually think it's great. The model knows its limits, that's a good thing. I rather know that I'm pushing it beyond its limits than going forth with an answer that is actually made up.
Mazen — sa/acc retweeted
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
Mazen — sa/acc retweeted
We REMOVED upto 10 WHOLE layers from Qwen3-8B-Base using LaCo (Layer Collapse).
Upon deleting only 6 layers, the model got 16.7% smaller, keeping 78-94% of its reasoning scores at a 1.68x perplexity ratio.
Code and pruned weights are up on our Hugging Face.
Here are insights from our research :
Mazen — sa/acc retweeted
>vLLM is now ~2x faster than vLLM
Mazen — sa/acc retweeted
Why do VLMs sometimes output 罪 instead of "vain"?
An investigation into script confusions, and how we fixed them 🧵
Mazen — sa/acc retweeted
We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).
I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.
We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:
"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"
I imagine we'll have some backend services that could be replaced entirely by this new architecture.
Mazen — sa/acc retweeted
LLM Inference Engineering - Problem and Solution
Problem: LLMs are slow
Solution: KV Cache - Avoid recomputing previous tokens.
Problem: KV Cache consumes huge memory
Solution: PagedAttention - Manage KV memory efficiently.
Problem: GPU is underutilized
Solution: Continuous Batching - Dynamically add/remove requests.
Problem: Token generation is sequential
Solution: Speculative Decoding - Generate and verify multiple tokens together.
Problem: Attention requires large memory usage and expensive memory movement
Solution: FlashAttention - Reduce memory usage and memory movement during attention.
Problem: KV Cache is still large
Solution: MQA / GQA - Share K/V across multiple query heads.
Problem: Many optimizations need to work together
Solution: vLLM - Build an efficient inference engine.
Problem: Applications repeatedly process the same prefixes
Solution: RadixAttention / Prefix Caching - Reuse previously computed work.
Problem: Models are too large
Solution: Quantization - Reduce model memory and computation.
Problem: Large models are expensive
Solution: Distillation / SLMs / MoE - Make inference itself cheaper.
Keep Learning, Keep Sharing, and Keep Growing.
Mazen — sa/acc retweeted
In agentic AI, log every step.
Every LLM call, tool call, trace, span, decision, input, and output.
Log the reasoning signals wherever possible.
If you can’t observe what your agent is doing, you can’t reliably debug it.
Mazen — sa/acc retweeted
سعداء بتكريم @DalData_sa في مملكة البحرين، ضمن تكريم الكفاءات والمشاريع الرائدة في العمل الاجتماعي على مستوى دول مجلس التعاون الخليجي.
كل الشكر ل @HRSD_SA على ترشيح الجمعية، ولمعالي المهندس @Ahmed_S_Alrajhi على دعمه وتقديره لجهود دال.
تكريم نفخر به، ودافع لنا لمواصلة الأثر في مجال البيانات والذكاء الاصطناعي.
Mazen — sa/acc retweeted
المال قد يأتيك من أبواب كثيرة بعد توفيق الله..
لكن الأثر عزيز، لا يُمنح لمجرد السعي إليه، بل لعظمة النية التي تقف خلفه.
فالخير والشر كلاهما قد يترك أثرًا، لكن ما يصنع الفرق هو صدق النية وعمقها.
والتاريخ خير دليل
Mazen — sa/acc retweeted
Friendly reminder that Hermes was never built or intended to be exclusively for the "personal assistant" agent that can just read your emails and nothing else.
I fully intend and have plainly stated many times that I always built hermes to be the most powerful AI Agent.
"Normie" consumers can leverage that and we do what we can to make it an experience they can approach (the mobile app will be almost exclusively focused on consumer) - but they are not our only demographic.
I want the scientists, the cybersecurity engineers, the developers, the knowledge workers, the creatives, and everyone else to be able to have an agent that they can leverage that is maximally useful to them too.
Not just your mom. (She called btw you need to make your bed)
Mazen — sa/acc retweeted
🚨JUST IN: OpenAI models will now have invisible watermarks hidden in ChatGPT and Codex outputs
>"We are not making text watermarking a global default AT LAUNCH"
>"phased approach starting with the EU"
>"this regional approach gives us room to learn"
>"starting today, API customers around the world WILL be able to opt in…"
"We expect to revisit each part of this approach as the technology, standards, and evidence evolve"
Mazen — sa/acc retweeted
.@OpenRouter co-founder Alex Atallah on the "everyone is building the same thing" take:
"This reminds me of this tweet I saw... Everybody's building an agent loop with notifications, third-party connectors, context management, memory, sandboxes, agentic web search, an always-on agent on top."
"This product is showing up everywhere. It is showing up everywhere, but these are just the new table-stakes primitives."
"A 2005 version of that tweet would be: 'Oh, everybody's building the same thing. A database, a users table, a sign-in page, a sign-up page, a profile page, a logout page. Everything's the same.'"
"There's like a lot of differentiation really. There's table-stakes needs for AI just like there's table-stakes needs for the web."
@alexatallah @amasad
.@OpenRouter co-founder Alex Atallah, in his first podcast since Stripe acquired the company, joins @Replit co-founder Amjad Masad and a16z's Erik Torenberg on why the future of AI is independence and specialization.
In this conversation, Alex walks through how the Stripe deal unfolded, why he wasn't originally looking to sell, and why "payments and inference are going to blend together."
Pre-OpenRouter, the typical AI workflow had one model provider to choose from, and little pressure on that provider to lower prices. Now enterprises are diversifying across labs and open-weight models, and every board is asking about AI costs and benchmarks.
Amjad argues if your company depends on one AI lab, it can turn into your competitor. So Replit is building the layer that lets enterprises use any model and any cloud, without being locked into either.
Alex and Amjad are split on personal agents – Amjad runs one agent across his whole company and loves the cross-domain joins, while Alex says general agents cause you to sacrifice understanding, and argues 10 specialized chiefs of staff beats one superagent.
0:45 How the Stripe deal unfolded
5:05 Why mixing models beats one model
7:25 Forcing the labs to compete on price
8:50 Enterprises want open-weight models
10:30 Every board asks about AI every month
12:25 Why companies must own their intelligence
14:15 Replit as the independence layer
15:10 Everyone is building the same agent
16:35 Why Amjad built bring-your-own-cloud
18:10 Amjad's agent that runs his whole company
19:55 Why 10 specialized agents beat one
23:35 Machines, not humans, should specialize
27:30 Guardrails for agents talking to agents
31:10 Models training their own replacements
33:45 Most tasks don't need a frontier model
40:50 Training small models on Qwen 8B
43:25 The Rust cycle is coming for AI
45:10 Fusion models: frontier quality at half the cost
YouTube: youtu.be/ekK8urKHPMQ
@alexatallah @OpenRouter @amasad @eriktorenberg
Mazen — sa/acc retweeted
مجتمع سعودي للمهتمين بتسارع تطوير الذكاء الاصطناعي والتقنية
> ❌ نهبد او نتكلم عن شغل غيرنا كأنه شغلن
> ❌ نكتب تويتات تطبل لفلان وعلان
> ✅ نشارك معرفه بتفاصيلها لكل شيء جديد ومفيد عشان نكون جزء فعال بالمجال
> ✅ نجتمع كل يوم سبت، الساعه ٧ المساء
حياكم
saacchq.org
Mazen — sa/acc retweeted
مع تسارع تطورات الذكاء الاصطناعي وتعدّد نماذجه وأدواته، يبقى التحدي: كيف نختار الأنسب لاحتياجنا، ونوظّفه ضمن نظام متكامل يعمل بكفاءة؟
وفي سعيٍ لربط المعرفة بالتطبيق، يسعدنا إطلاق أول معسكر نقدّمه بشهادة معتمدة:
Full Stack AI Engineering: Applied AI Engineering
هندسة الذكاء الاصطناعي المتكاملة:
من تصميم الأنظمة وبنائها إلى نشرها.
نسعى في هذا المعسكر إلى تطوير القدرة على اتخاذ القرارات الهندسية التي تحوّل الاحتياج إلى حل؛ بدءًا من فهم المشكلة وتصميم النظام، مرورًا باختيار النماذج والأدوات المناسبة وربطها وبناء الحل، وصولًا إلى تقييمه ونشره في بيئة تشغيل فعلية.
التسجيل المبكر متاح، والمقاعد محدودة.
سجّل الآن وكن من الساعين للمعرفة.
رابط التسجيل: saai.sa/bootcamps/full-stack…
