Developer | AI News & testing | Startups & SaaS | DM for Collabs ☄
Joined April 2025
- Tweets17.1K
- Following667
- Followers3.8K
- Likes58.5K
Testing what Gemini 4 Argon can do with procedural CAD geometry + custom GLSL shaders :
fig.1 F135-PW-100 Augmented Turbofan Test Cell478 turned & extruded vector solids solved via offline SDF ray-casting
11 spring-loaded modules
(3-stage Ti fan → 6-stage HPC → annular combustor → 15-flap Con-Di iris nozzle)
real-time WebGL Mach 2.2 afterburner plume, acoustic ignition wave & FLIR thermography
all in one HTML file +🔊
Gemini 4 argon just dissected a 1954 leica m3 in pure code
> 81 solid parts. 34 assemblies. re-shaded every frame in pure svg
> sub-frame motion blur on the explosion
> stroke the lever, unlock the 50mm f/2, explode to the brass bench
cleanest mechanical teardown a model has spit out so far. the part count and blur are both stupid precise.
all you need to build this :
xitter.cf/renproton/status/21118…
Opus 5.5 + /anatomy skill is f*cking insane for motion design but...
add /anatomy-reveal to turn Claude into premier motion design studio
• Real-size 3D models from a photo or name
• Motion-blurred reveal films, 1:1 to 16:9
• Exploded views you can take apart
send skills your Claude:
> typingmind.com/skills/annngu…
> skills.wheresryan.sh/anatomy
kimi k3.1 (kibitz) vs gpt 5.6 sol xhigh on the same web design prompt
left: kimi
right: gpt
kimi’s starting to close the gap on design tasks too.
If you're building solo or with a small team, deployment shouldn't slow you down.
Synteq Cloud Apps makes it simple to get your project online without the usual server setup headaches.
GitHub, CLI, or drag and drop it covers all three.
they're giving away $20,000 in platform credits.
comment "LAUNCH" on their post
🤝 Paid partnership
Building is fun. Deploying? Less fun.
Introducing Apps and Functions, DevOps handled so you don't have to.
Push to GitHub, deploy via CLI, or just drag and drop. We handle the rest.
To celebrate, we're giving away $20,000 in platform credits. Comment "LAUNCH" and we'll send it your way 👀
tested @evermind’s EverOS Cloud for agent memory. no custom rag needed.
hooked a support triage agent to ticket #409 (error 403 on sync) and dumped our sops + old tickets in. it searches memory before it starts, then saves the run after.
it turned the troubleshooting into a reusable skill on its own. next ticket it just pulls the skill instead of starting from zero.
same api for cloud or self-hosted. you can browse, rate, or delete everything in the memory bank.
works for coding agents, ops decisions, wearables, whatever needs context that sticks across sessions.
6 months of triple pro usage: everos.evermind.ai
🤝 Paid partnership
kimi k3.1 just showed up in the serving code
sglang added explicit support for moonshotai/kimi-k3.1. same tokenizer and chat format as k3,
plus one new thing: a medium thinking effort. one commit even references a deployment already running on deepinfra.
kimi k3 dropped july 16.
their last few gaps have been 50-60 days. that window is open right now. current k3 is still the open leader at 44 on artificial analysis, $3 / $15.
if the id sticks and it hits the arena this week, the next kimi is already in the stack. if the commits get cleaned up, it was just another early push.
Kimi K3.1 was recently spotted in 2 SGLang commits, one of which references a K3.1 deployment already running on Deepinfra
If we are getting K3.1 on the Arena this week, it might follow the same naming pattern as the previous Kimi models:
Kimi K2.5 - Kiwido
Kimi K2.6 - Kiwire
Kimi K3 - Kivine
Kimi K3.1 - ???
Adit_Yah ☄️ retweeted
MiniMax M3.1 Flash is actually insane.
I asked it to build this whale breach scene using Claude Code as the harness, and it cost just $0.25. That's roughly 8x cheaper than Astra and Sol.
I didn't expect MiniMax to be this good at 3D.
In my testing, it even beats GPT 6.1 Sol, and the result looks better than the others while costing less than DeepSeek Flash’s version.
What a comeback by MiniMax.
tried out @Aprilzz423’s skill with Gemini 4 Argon
to build a little F1 pit stop game. the fact that it comes with a hand drawn wobble slider and color toggles out of the box is so cool.
Skill link in comment .
Many asked me how to make these, so I turned it into a skill~
Colorfigures: Describe an everyday object, and Claude draws it hand-inked and isometric, with buttons that actually work.
You can also ask how something works and the skill will label the parts to explain how each part does its job.
Free and open source 👇
DeepSeek V4.1 Flash 1017 vs GEMINI 4 ARGON
after seeing some comparisons where DeepSeek was beating Argon, i had to try it myself.
Both models one shot this.
Argon took around 10 mins and used ~2% of my weekly quota.
V4.1 Flash finished in around 4 mins (free on Open Code).
DeepSeek team really cooked with this one, but i still don't think it's at Argon's level.
Then again, it's a Flash model, so it's honestly impressive.
prompt : Conversation LogBuild this image as a single-file HTML 3D scene using Three.js It's a floating donut-shaped voxel island with 7 world wonders: Christ the Redeemer, Great Wall, Taj Mahal, Petra, Colosseum, Chichén Itzá, Machu Picchu. Match the reference as closely as possible: same layout, same Minecraft-style voxel look, same colors, rivers between the regions, clouds, and the hole in the middle. Make it feel premium: soft shadows, nice lighting, slow auto-rotate, drag to orbit, scroll to zoom, subtle cloud drift and water shimmer. Keep it smooth at 60fps. Output the complete working file only.
After seeing a lot of results, DeepSeek is really back.
V4.1 Flash 1017 is hitting 77 on DeepSWE. Higher than GPT-6.1 Sol, Astra, and Opus on the same chart.
And it’s dirt cheap.
A flash model is matching or beating the ones people are still paying real money for.
Most people haven’t even tried it yet.
Look DEEPSEEK v4.1 flash 1017 cooked this in 3 mins
WTH, DeepSeek V4.1 Flash is outperforming Argon.
Is this real?
A flash model is beating the one Google is still gating for most people.
this is wild.
DeepSeek V4.1 Flash 1017 Max vs GPT 6.1 Sol Ultra
DeepSeek cost $0.44. Sol Ultra cost way more.
They took about the same time.
And the cheap one matched the expensive one.
This is getting ridiculous.
DeepSeek just dropped V4.1 Flash-1017 and it hit 77% on DeepSWE.
that’s higher than GPT-6 Astra, Opus 5, and most of the expensive models on the chart
And it’s sitting way over on the cheap side.
A flash model is matching or beating frontier stuff on real agent coding tasks while costing almost nothing.
This is the part that actually matters.
Most people are still paying $10 - $50 for models that aren’t clearly better on this.
Open weights + cheap + this score is a problem for everyone else.
🚀 DeepSeek-V4.1-Flash-1017 is here!
Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.
🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.
1/4
Fable 5.5 vs Opus 5.5
Legit used the exact same prompt on both.
Opus 5.5 is still in a league of its own , most AI models aren't even close.
And Anthropic is really cooking up Fable 5.5...
I seriously can't wait. xitter.cf/legit_api/status/21111…
ALT Opus 5.5