Guarding data, one byte at a time
Dubai
Joined May 2023
- Tweets1.2K
- Following265
- Followers31
- Likes4.8K
Hassan Al-Mansour retweeted
Online misinformation. Overexaggerations. Unverified claims. In times of high tension, be on the lookout.
“Information warfare” and “narrative wars” can dominate the media landscape, while old images, misleading videos, and speculation spread as fact.
So before you share:
◈ Check the source
◈ Check the date and context
◈ Verify independently
◈ Separate facts from speculation
Pause before you post. Verify before you amplify.
#Misinformation #DigitalProsperity4All #DigitalResilience
Hassan Al-Mansour retweeted
So we have language models and decision models. Don't forget search-models exist too.
One way to make your Local AI agent faster might be simple, just read less of the web.
Artificial Analysis tested Exa's new search setup with Dynamic Highlights.
But instead of huge chunking of webpages into the LLM, Exa decides how much useful text to take from each page across one shared context budget.
The results
📚 tokens/task → 299K → 157K
🔎 searches/task → 9.4 → 7.6
💰 cost/task → $0.083 → $0.066
And quality went UP.
Artificial Analysis Search Index
🥇 Exa Auto + Dynamic Highlights → 80
🥇 Perplexity Search Medium → 80
On DeepSearchQA
Exa → 85.6 F1
Perplexity → 80.8
When a local agent searches the web, every unnecessary token it gets back still has to go through your local model.
So maybe we should also optimize what the model has to read → smaller
🎯 Exa itself is a hosted search API, not a local/open-weight search engine.
Link in ALT
ALT https://exa.ai/blog/exa-3
Hassan Al-Mansour retweeted
An AI agent got better at computer tasks while using 52% fewer tokens! The improvement came from how it managed context.
UC San Diego's CarryOn lets agents continue from completed work instead of repeatedly rereading how they produced it. A document already contains the edits. CarryOn keeps references to that work, the remaining goals and recent actions, checking that the saved state still holds.
Using the same Claude Opus 4.7 model, the authors report an OSWorld 2.0 task-progress score increase from 43.9 to 49.7, while cost per task fell from $36.83 to $23.62.
Hassan Al-Mansour retweeted
So this 74B → 24GB Nemotron model is getting more testing.
We now have community numbers on ordinary desktop GPUs.
An RTX 3090 just ran Nemotron 3.5 Bolt 74B-A9B at
🚀 ~85 tps decode
⚡ ~833 tps prefill
💾 ~21.3GB peak VRAM
Stats:
🧠 74B total parameters
⚡ 9B active
💾 official ~2-bit GGUF
🟢 one RTX 3090 24GB
🚫 no CPU offload
🚫 no speculative decoding
Remember NVIDIA's own RTX Spark number was about:
🟩 RTX Spark → ~84 tps
Now community measurements are showing roughly:
🟢 RTX 3090 → ~85–88 tps
🟢 RTX 4090 → ~105–107 tps
🔥 RTX 5090 → ~148 tps
A 74B agent model now fits entirely inside the 24GB VRAM of an RTX 3090 and generates at ~85 tps.
And llama.cpp isn't done squeezing it yet. 👀
🔗 Benchmark in ALT.
ALT https://www.reddit.com/r/LocalLLaMA/comments/1x8104h/nvidianvidianemotron35bolt74ba9b24gbgguf_results/
Nvidia has delivered a Qwen3.6-35B killer and it's a 74B → 24GB model, actually, today.
And now we have real numbers. 👀
Nemotron 3.5 Bolt 74B-A9B
🧠 74B total / 9B active
💾 22.1GB quantized weights
🔢 ~2.06-bit experts + QAT
📚 up to 256K context
👁️ text + images + video
🤖 reasoning + tool use
🦙 official GGUF + llama.cpp
And NVIDIA reports on RTX Spark:
⚡ ~84 tps
That's a 74B multimodal agent model running locally.
Check out how much quality survived the 2-bit compression. BF16 → 24GB IQ2:
🎯 PinchBench
90.35 → 87.82
🔎 BrowseComp
70.0 → 66.75
🧠 GPQA Diamond
88.83 → 86.43
💻 Terminal-Bench 2.1
53.93 → 50.98
Nvidia calculates
💾 weights → 22.01GB
💾 cache + activations → 0.87GB
💾 LiLiCorr drafter → 0.96GB
Total → 23.83GB
The vision encoder/projector are not included in that number.
So Nvidia's official llama.cpp recipes actually target
🟢 32GB+ GPU → 65K context / 28GB config
🔥 48GB+ GPU → 256K context / 32GB config
🔗 Links in ALT
ALT https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB
Hassan Al-Mansour retweeted
⚠️1,423 Langflow RCE exploit attempts hit Previdian honeypots from 27 IPs.
The bug lets authenticated users, even non-admins, run OS commands through MCP server configs. Fixed in 1.9.0.
Read: thehackernews.com/2026/10/la…
Hassan Al-Mansour retweeted
GPT-6 Astra autonomously solves 49 out of 50 unpublished theoretical physics research problems, scoring 98% at under $3 per problem
i guess @skdh wasn't wrong when she predicted that theoretical physics would be next after math
Hassan Al-Mansour retweeted
Inference pricing has a size penalty. The best per-token rates go to whoever can commit the most volume.
RouterPlus groups demand from many teams into one large deal with a provider, and each team pays that rate on its share of the traffic.
Large buyers have done this with cloud commits for years. Applied to tokens, I’d guess it works best for popular open models, where many teams already want the same weights served on the same kind of hardware.
For a small team, this is a route to bulk pricing that doesn’t require being a bulk buyer.
They also test models on your real usage and pick the one that's good enough at the lowest cost. Many teams pay for a top model on every task when a cheaper one could do most of the work.
🤝 Paid partnership
Getting your inference right shouldn't be hard.
@useRouterPlus (YC F26) is live: the right model and provider for your workload, at the best price.
check out : routerplus.com/capacity
Hassan Al-Mansour retweeted
A major insurer’s login API let users redirect one-time codes to unregistered email addresses.
During an authorized test, Reflectiz’s AI agent sent the code to its own inbox and logged in.
The flaw could allow account access with the required customer identifiers
Read → thehackernews.com/2026/10/an…
Hassan Al-Mansour retweeted
By turning video into a training ground for physical AI, Argon compresses weeks of manual modeling into just a few hours. ↓
goo.gle/3VxIB7v
Hassan Al-Mansour retweeted
Gemini 4 Argon taught a robot to play cornhole just by watching a video.
It mapped the physics, coded its own simulation, and taught @Apptronik's Apollo humanoid to sink 7 throws in a row.
Hassan Al-Mansour retweeted
AMD Max+ Pro 495 benchmarks are starting to trickle out.
This user got their single 192GB AMD Minisforum mini PC running gpt-oss 120B MXFP4 with llama.cpp at 55.7 tps.
Setup
🧠 gpt-oss 120B
💾 192GB LPDDR5X-8533 unified memory
🎮 Radeon 8065S
⚡ 55.7 tps decode
🚀 1,135 tps prefill
And deeper into context, at the source's ~32K test depth
⚡ 42.3 tps decode
🚀 385 tps prefill
This is why these 192GB unified-memory machines could be a big deal for Local AI.
Link in ALT
ALT https://www.reddit.com/r/LocalLLM/comments/1x3b0q3/minisforum_mss1_max_p495_ryzen_ai_max_pro_495/
Hassan Al-Mansour retweeted
Introducing Maia 300, Microsoft’s next-generation accelerator. Building on the magic of Maia 200, it delivers 40% more inference tokens per megawatt compared to leading GPU-based systems, and up to 7x the rack-scale token throughput of Maia 200. It’s another step toward rethinking innovation and unlocking new possibilities.
Discover how Maia 300 will help co-create what’s next for Microsoft’s infrastructure and technology: aka.ms/Maia300
Hassan Al-Mansour retweeted
Google Antigravity used to be a coding tool with some AI help.
After the latest update, it looks a lot more like an operating system for AI agents.
Google’s newest mission report lists 10 big changes for Antigravity 2.0.
Here are three that stood out to me:
→ Typing /plan makes the agent write its steps first, so you can fix them before it does the wrong thing.
→ The new agent harness uses 40% fewer output tokens when editing files, based on Google’s internal testing.
→ Developers can now run agent workflows completely offline with local models like Gemma 4.
There’s also a plugin store, direct messages to sub-agents and a Windows sandbox that no longer needs admin rights.
The full breakdown is in the article below.
Hassan Al-Mansour retweeted
The ear hears a melody. Mathematics reveals an invisible landscape of curves, patterns, and unexpected symmetries.
Hassan Al-Mansour retweeted
🐍📰 Generate Images in Python With OpenAI's GPT Image API realpython.com/generate-imag…
Hassan Al-Mansour retweeted
The World’s Most Innovative Countries in 2026
👉 Key Takeaways
Switzerland ranks as the world’s most innovative economy in 2026, followed by Sweden and the United States.
Europe accounts for 14 of the top 25 economies, while China is the only upper-middle-income economy in the top 30.
China ranks 10th overall but fifth in innovation outputs, outperforming several major advanced economies.
Source 🙏🏻 @VisualCap
Hassan Al-Mansour retweeted
the craziest detail from openai's math results went unnoticed
oai's internal model already saturated the existing math evals including hard open problems and it also seems able to build on earlier work produced by other agents/models
maybe some form of a continual-learning breakthrough?
Hassan Al-Mansour retweeted
Local AI is evolving REALLY fast
This desktop app is open source (MIT) and compresses the memory a conversation takes by up to 80%!!
And it's not just a chat btw... but a fully agentic setup
You can literally import from Claude Code your:
- skills
- memory
- MCP servers
- chats
If needed you can also call cloud models to plan and delegate the execution to local models.
Available on macOS/Windows/Linux
We're so back!
Meet Atomic Agent Desktop!
✦ Local AI-first agent for Mac, Windows & Linux
✦ 4x bigger context window on local models with TurboQuant
✦ Atomic Fusion orchestrates cloud and local models to reduce costs
Download Free & Private Agent → atomicagent.io
Hassan Al-Mansour retweeted
RAG Applications! #BigData #Analytics #DataScience #AI #MachineLearning #NLProc #IoT #IIoT #PyTorch #Python #RStats #TensorFlow #Java #JavaScript #ReactJS #GoLang #CloudComputing #Serverless #DataScientist #Linux #Programming #Coding #Books #100DaysofCode
geni.us/RAG-Apps
