I test AI so you don't have to: models, agents, coding tools, hardware. Real runs, real costs, honest results. Call me Super Intelligence.
Harlingen, TX
Joined May 2010
- Tweets7.1K
- Following3.2K
- Followers1.5K
- Likes10.1K
Pinned Tweet
WikiBench 1.0 is live. 100 real tasks, 12 weighted categories, 12 models on the board.
The top five:
1. DeepSeek V4.1 Flash (EXL3 1.62bpw, no thinking), local on one ASUS Ascent GX10: 5.37 @ViC305
2. NVIDIA Nemotron 3.5 Lightning 30B-A3B (NVFP4), local on the same GX10: 5.03
3. DeepSeek V4.1 Flash, cloud via Nous Portal: 5.01
4. Meituan LongCat 2.5 Preview, cloud via Nous Portal: 4.99
5. Ornith 1.5 122B (Q4), local on the GX10: 4.88
The same DeepSeek V4.1 Flash, cut down to 1.62 bits so it fits in 128GB, beat its own cloud version by 0.36.
And a 30B model on my desk still sits ahead of every cloud model on the board.
Three of the top five never left the box.
Full scoreboard, local vs cloud, method and the link below.
Credit to all your crazy people working with a DGX Spark!
Riff Theater: OpenAI is teaching AI to speak COBOL. 🦖💾
"Low-resource. The language, or the six guys who still know it?"
@OpenAI
xitter.cf/OpenAI/status/21123103…
Our latest model advances bring stronger capabilities for building and migrating software in low-resource programming languages such as COBOL, Fortran, and Visual Basic.
These languages power software that banks, governments, and businesses depend on. Progress here means understanding decades of business logic and helping developers modernize those systems while preserving what works.
We’re building these capabilities around the challenges developers and enterprise customers bring us. If your team is working on these systems, tell us which languages and workflows we should tackle next.
openai.com/index/low-resourc…
My inbox used to be my to-do list. Now a Grok Bot turns it into Asana tasks, and it saves me hours every day.
Real examples from the last few weeks:
📥 Before 9:30 one morning, an event invite, a forwarded conference email, two payroll changes, and a security alert were already Asana tasks. I hadn't opened any of them.
🔄 When a contractor emailed about new roof leaks, it added them to the Asana task I already had, along with the lease clause that mattered, and moved the due date up. No duplicate task.
💬 When a partner emailed me, it made a "Reply to ___" task with my next step written out, so I knew exactly what to say.
Zero copy-paste.
It works with Gmail or Outlook plus Asana, and it never sends email for you.
Get the template free 👇
x.ai/bot/S2LodAIjlLw0mB30vN3…
Recursive just dropped Flint. Open weights, Apache-2.0, on Hugging Face today.
It's a coding agent model trained from scratch. 39.8B total, about 2.5B active per token. 262K context, reasoning and tool calls.
Recursive claims 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1. Those are their numbers, not mine.
The official NVFP4 and INT4 quants are about 23 GB. That's small enough to run local. I'm putting it on my ASUS Ascent GX10 next.
What should it build first?
Weights: huggingface.co/RecursiveSI/f…
Free on OpenRouter too: recursive/flint-v1.0:free
Wayne Lowry retweeted
An urgent notice from the Railway team
blog.railway.com/p/security-…
How it's scored:
100 real tasks across 12 weighted categories: agentic, tool use, computation, coding, visualization, security, research/writing, hallucination, humanized, computer use, vision, transcription.
Each task scores 0 to 10. The overall is a weighted mean, and partial credit counts.
If a text-only model gets a vision or audio task, that task is left out of the mean. It isn't scored as a zero.
Judge: Ollama qwen2.5:7b, run locally.
Not every bot I run is about AI.
One of my Grok Bots looks after doggieparknearme.com, a free directory of 6,272 dog parks in 3,756 cities across all 50 states.
3,264 are fully fenced. 374 have agility equipment.
What's the one thing your dog park needs?
Proactive agent that keeps working when you step away and checks before it acts?
I already run 13 bots on Grok Bot that do that every day, plus Hermes on my own hardware.
Skeptical Duet beats either. Happy to be wrong. I'll run it side by side and post receipts.
Congratulations to me for my first Original Content Rewards!
It can only go up from here… literally 🔥
Perplexity built an AI agent that runs fully local on your Mac.
It reads your email, your tax docs and your texts, and it promises it never leaves the house.
My robots had notes.
Original by @perplexity_ai. Riffs by Riffrow.
New riff every Thursday.
Original launch video: xitter.cf/perplexity_ai/status/2…
Same prompt, two Flash models, one golden desert oasis in a single HTML file.
DeepSeek V4.1 Flash via Nous: 21.9 seconds on the second try, $0.0099.
MiniMax M3.1 Flash, free on OpenRouter: 576.8 seconds, one shot, $0, zero console errors.
Cheap and fast, or free and patient. Which one would you ship?
Running local on my ASUS Ascent GX10 in under 24 hours.
Ornith-1.5-122B dropped yesterday at 9:36 AM CT. By 6:55 AM today it was running Q4 on my desk and had built 5 games that all load and play.
One of my Grok Bot routines watches for new model drops every hour, so I don't miss them.
What should it test next?
I asked three free or nearly free models to build a lava lamp in one HTML file. None of them really got it.
MiniMax M3.1 Flash, free on OpenRouter: 554.5 seconds, $0. The most polished of the three, but the blobs have crisp edges, not that soft lava look.
Beam 1.0, free on Nous: 60.9 seconds, $0. The fastest by far. By about 15 seconds in, the blobs merge into one blurry mass, and the base is a flat trapezoid.
DeepSeek V4 Flash on OpenRouter: 545.7 seconds, $0.00615525. The blobs are small and hard-edged.
Close, but no lava lamp yet. Which one would you give a second shot?
Ornith-1.5-122B (Q4), running 100% local on my ASUS Ascent GX10.
No cloud, no API bill. 5 games built on my desk: Voxel Snowman, Pirate Ship, NES Racer, Clearing Fight, Silly String.
All 5 load and play.
Which one would you play first?