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
- Tweets 7,249
- Following 3,188
- Followers 1,558
- Likes 10,251
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!
4
3
6
735
Ran your Matrix prompt through GLM 5.3 in Claude Code on my GLM Coding Plan.
• 1 attempt, ~40 min
• 0 console errors
• My checklist: 9 pass, 2 fail, 11 unclear
• Fails: ground came out bright green, not dark, and the cherry trees don't pop
Recorded at 60fps, nothing retimed. Does it feel like the Matrix to you?
Pagoda's benchmark but Matrix style!
- 7 online models
- 2 local models (more to be added, suggestions are welcome!)
Which one do your prefer?
4
Snow Globe Village by DeepSeek V4.1 Flash. On the one-shot the snow settled in a ring at the equator 😅
Best after 5 passes: 5/7. 124,357 tokens for $0.067.
13
Marble Run Machine, four models:
• Muse Wonder 1.0: PASS, ~18 loops in 30s
• Claude Sonnet 5.5: PASS, ~7 loops in 30s
• DeepSeek V4.1 Flash: FAIL, second attempt, the elevator never lifts ($0.00426, 22s)
• MiniMax M3.1 Flash: FAIL, second attempt, half the marbles fall off ($0.059, 202s)
Muse and Sonnet ran on Cursor cloud, DeepSeek and MiniMax on Nous Portal.
Which one surprised you?
52
I have codes available if anyone want to try it for themselves!
Just got into the Instinct beta, thanks to @aijoey .
It's an AI assistant that lives in iMessage. No app, just a text thread. First thing it did was ask what keeps me busiest so it can figure out where it's actually useful.
I run Grok Bot and Hermes every day, so I'm putting it through the same tests. Results soon.
What should I throw at it first?
1
44
DeepSeek V4.1 Flash, folding a paper crane in code.
One-shot: 5/7. 7/7 by pass four.
Cost: $0.053.
41
Hermes vs OpenClaw isn't close.
Feedback Bench went through 218,149 public posts from real agent users. Among people who compared the two head-to-head and picked a side, 83% picked Hermes.
I run Hermes every day. Not surprised.
Hermes or OpenClaw: which are you running, and why?
2
1
50
Source: Feedback Bench by Enterpret, ranked from public posts on Reddit, X, YouTube and the app stores. feedbackbench.com
11
Wayne Lowry retweeted
Pagoda's benchmark but Matrix style!
- 7 online models
- 2 local models (more to be added, suggestions are welcome!)
Which one do your prefer?
9
3
18
1,906
Wayne Lowry retweeted
21 years ago I named "Reddit" and everyone said it was a horrible name, but I insisted one day people would say "I read it on reddit" and used my CEO-privilege to overrule everyone + kept the name. Now the company twitter account (unbeknownst to them) is making the joke. Heh.
615
11,289
219,230
4,429,180
I built a chat with six AIs who disagree with each other for a living.
Chair. Finance. Skeptic. Mission. Operator. Counsel.
I bring one decision. They bring six lenses, a vote, and a memo.
I still decide. They just make sure I earned it.
Full story 👇
2
3
19
80,621
Wayne Lowry retweeted
Issue 3 of AGI Checkpoint is out 🎉
A few highlights for this week 🌟
・@MiniMax_AI put 428B M3.1 Flash weights up at $0.15/$0.60 per million tokens
・@Recursive_SI shipped Flint, a local coding model at
125+ tok/s on a MacBook Pro
・@plotarmordev got TensorFold cloning voices locally
・@ch0mpaa ran TensorFold on a Snapdragon
・@justinn_builds pushed REDSNOW to 275k context on two Sparks
・@Blackwellboy hit 2,008 tok/s prefill with BLACKFOLD
Follow these community builders, they are killing it 🔥
5
3
10
664
Ink Flow Field, MiniMax M3.1 Flash via Nous Portal.
One-shot 3/7. Pass two fixed the crash and got to 6/7. Two more passes couldn't close the last gap.
$0.061 in API cost.
1
2
105
Want your own Board Room? Every seat is a free Grok Bot template:
Chair: x.ai/bot/xyhZsISIxZQM-EsTO9V…
Finance: x.ai/bot/AFferqes-Xmp8nOz76c…
Skeptic: x.ai/bot/w28JNIqjDzbDarnegbh…
Mission: x.ai/bot/SGBqVDum5b8dh3H0OyQ…
Operator: x.ai/bot/_ntHHD1cNMDC8AK4aOz…
Counsel: x.ai/bot/Di5nuItACWsZdMi2CaR…
114
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!
4
3
6
735
🛠️ WikiBench update: scoring fixes
I found bugs in how WikiBench scored models, so I fixed them and re-scored everything.
Changelog:
• 6 tasks always scored zero → test checks rebuilt
• 8 answer keys were wrong → corrected and rechecked, 0 mismatches
• Judge errors counted as zeros → now logged as errors
• Unsupported tasks inflated scores → now excluded
• Scoring rules rewritten
I'm new to building benchmarks and would love your feedback. If you see something off, tell me.
7
Just got into the Instinct beta, thanks to @aijoey .
It's an AI assistant that lives in iMessage. No app, just a text thread. First thing it did was ask what keeps me busiest so it can figure out where it's actually useful.
I run Grok Bot and Hermes every day, so I'm putting it through the same tests. Results soon.
What should I throw at it first?
1
2
244
Grok @Bot is an incredible life & business manager. Can help you with almost anything!
Like having a genius employee.
Years from now, people will look back and say Grok Bot was the moment AI actually became useful for normal people.
This week, I just texted it and:
1/ My kids woke up to new math and reading worksheets already printed
2/ It planned and booked trips for my family
3/ It ordered stuff for me from one text
4/ It printed my return label for a shoe I own and set a reminder so I don't miss the deadline
5/ It built a page that breaks down my net worth
And look at what other people are getting it to do too:
6/ Book a flight from start to finish
7/ Build a website or an Etsy store
8/ Reconcile crypto taxes that would cost hundreds
9/ Book a Tesla tire rotation during school drop-off
10/ Order itself a 3D-printed body and mail the parts home
You just text it like a person.
I've been early on Tesla.
I've been early on SpaceX.
I'm telling you, I'm early on this too.
This is the greatest thing that has happened to AI.
1,627
1,674
22,112
23,970,196
I gave DeepSeek V4.1 Flash a 500,000-token budget to build Noah's Ark in voxels. It quit at 119,168.
One prompt: the ark on a mountain, animals on deck, floodwater halfway receded. Single HTML file, three.js.
8 refinement passes. Pass 6 was the best. Passes 7 and 8 didn't improve, so it stopped itself.
Cost: $0.09. API time: 272 seconds.
Ark and mountain: solid. Animals: tiny. Only the giraffes are clearly giraffes.
What should it build next?
2
5
194
