Content creator | AI researcher & builder | AI insights from 2030 | @beyond_xai
San Francisco
- Tweets 1,555
- Following 100
- Followers 49,142
- Likes 2,499
Andrej Karpathy predicted the future of AI once again:
“Everyone’s renting frontier models for jobs a local 3B model could do. Small Local LLM's are the future.”
this 18-page PDF breaks down Karpathy's case for running that core on your own machine.
the real question isn't "is the local model as smart?"
It's "which of my 1000 calls ever needed the smart one?"
19
133
846
58,384
This is insane…
this is the only GitHub repo you need to run local LLMs on your Mac/PC for any task.
github.com/0xSojalSec/LLMs-l…
explore it today, then read the step-by-step guide for setting up your first local LLM in the article below.
8
6
41
3,310
Codez retweeted
Like it or not, over 50% of social content will be AI-generated by the end of next year.
@higgsfield AI Influencer and Genjutsu let a small team create characters, animate them, and test digital "personalities" like ad creatives.
This changes the economics of building a media business, and Higgsfield is where the next generation gets built.
Calling it “AI slop” says more about a taste preference than the size of the opportunity.
Born in Higgsfield. All over your feed.
Introducing Higgsfield AI Influencer.
Create your own AI influencer and bring them into any trend.
Try now with up to 5 FREE generations on Higgsfield and in the ChatGPT extension.
Powered by Genjutsu.
Credits: @JeanPhilMadame
69
212
499
62,078
this is unreal…
I just found Anthropic’s Opus 5.5 prompting blueprint for motion design, and it’s insane…
17-page PDF with prompt techniques, motion & audio design workflows, and director prompts.
uploaded it to Claude and created CLAUDE.md with it - it’s a totally different level of motion graphics now.
send this PDF + the article below to Claude and turn it into an AAA motion graphics studio.
29
77
900
90,250
A Jev engineer just showed the 10 constraints that separate a real agent brief from one giant prompt dump
instead of piling everything into one instruction, the prompt works like an interface: each block owns one job, one constraint and one piece of context.
10 details stood out:
step 1 → lock one concrete output. here it’s a complete 75–90 second educational piece, not “make something about Jev”
step 2 → separate the learning goal from the deliverable so the agent knows what actually needs to become clear
step 3 → define the visual system before generation: materials, typography, texture, framing and logo behavior
step 4 → reduce the topic to one cause-and-effect relationship instead of asking the model to compress everything it knows
step 5 → map it into 7 scenes where every scene moves the explanation forward
step 6 → treat every scene as structured state: purpose, visual, motion, voiceover, sound and transition
step 7 → make transitions carry information. objects and diagrams should evolve into the next scene instead of resetting
step 8 → lock the output contract: storyboard, voiceover, sound plan, implementation, 1920×1080 and 24 fps
step 9 → define operating behavior separately: precise language, technical clarity and continuity over decoration
step 10 → keep the system boundary clean: Jev handles bounded decisions while Grok Bot stays downstream for generation
what matters here is constraint architecture.
strong prompts don’t necessarily tell the agent more. they leave fewer important decisions undefined.
↳ typesafe.ai/blog/introducing…
29
10
139
5,384
Opus 5.5 is insane for motion design…
if you’ve built the right structure of .md files for it.
this one screen with CLAUDE.md, MOTION.md, BRIEF.md, STYLE.md will turn your Claude into a production-grade motion studio.
send it to your Claude, then read the full guide on Motion Design with Opus 5.5 in the article below.
19
37
319
26,909
Andrej Karpathy predicted the future of AI once again:
“Everyone’s renting frontier models for jobs a 3B model could do. Small models are the future.”
this 18-page PDF breaks down Karpathy’s case for working with small LLMs.
the real question isn’t “Is the small model as good?” It’s “Which of my 1,000 calls ever needed a frontier model?”
And @thewebai just answered it for formal logic.
TwIL-LM3-Pro:
→ 3.6B params, on par with Qwen3-8B on formal logic
→ leads VibeThinker-3B on all 6 formal-logic tasks tested
→ 95.4% on BBH logic, 95% on SVAMP
→ 2.09 GiB in Q4, runs on CPU or 4GB VRAM
→ no API bill, no data leaving your machine
The secret isn't size. It's post-training.
PDF below. Model 👇
huggingface.co/webAI-Officia…
Half a million downloads in a month. Today, our open source family takes another step forward.
Thank you for the incredible support behind our first-generation models. We’re excited to introduce TwIL-LM3-Pro.
At just 3.6 billion parameters, it brings powerful reasoning to everyday computers, with quantized builds that run locally. No cloud required.
In our evaluation:
Formal logic: Highest recorded headline score among the small models compared—beating China’s VibeThinker-3B by 35% and Qwen3.5-4B by 24%, and Liquid AI’s LFM2.5-8B-A1B by 47%.
Broader reasoning: 95% on SVAMP and 64.1% on MuSR, the highest recorded scores among the small models compared.
BIG-Bench Hard’s logic subset: 95.4%, compared with VibeThinker-3B’s 61.1%.
We believe AI is entering a post-training era. The advantage will increasingly belong to companies with the best pipelines and those that can produce capable, personalized intelligence faster and more efficiently, then put it on devices people already own.
That’s what we’re building at webAI. And we’re only beginning to share what’s coming out of our lab.
Coming soon: Meridian, our family of frontier-class models built to run on device. Our most advanced models will be available through the @thewebAI application.
Join the waitlist as we expand access. Proudly built in Austin, Texas. 🇺🇸
39
301
1,824
140,640
SpaceXAI engineer:
"90% of people run one GrokBot and call it done, only 10% build teams of bots that talk to each other.
i'm have a team of 15+ GrokBot agents in a loop & graph. I have a Chief of Staff bot, PM bot, QA bot and 10+ workers - that's the new stack"
In a 18-minute demo, a SpaceXAI engineer shows how to build a team of GrokBot agents from scratch
worth more than a $500 agentic engineering course
skip Netflix today and watch this, then read how to build a fleet of GrokBot agents in the article below
7
5
66
7,287
Andrej Karpathy predicted the future of AI once again:
“Everyone’s renting frontier models for jobs a 3B model could do. Small models are the future.”
this 18-page PDF breaks down Karpathy’s case for working with small LLMs.
the real question isn’t “Is the small model as good?” It’s “Which of my 1,000 calls ever needed a frontier model?”
And @thewebai just answered it for formal logic.
TwIL-LM3-Pro:
→ 3.6B params, on par with Qwen3-8B on formal logic
→ leads VibeThinker-3B on all 6 formal-logic tasks tested
→ 95.4% on BBH logic, 95% on SVAMP
→ 2.09 GiB in Q4, runs on CPU or 4GB VRAM
→ no API bill, no data leaving your machine
The secret isn't size. It's post-training.
PDF below. Model 👇
huggingface.co/webAI-Officia…
Half a million downloads in a month. Today, our open source family takes another step forward.
Thank you for the incredible support behind our first-generation models. We’re excited to introduce TwIL-LM3-Pro.
At just 3.6 billion parameters, it brings powerful reasoning to everyday computers, with quantized builds that run locally. No cloud required.
In our evaluation:
Formal logic: Highest recorded headline score among the small models compared—beating China’s VibeThinker-3B by 35% and Qwen3.5-4B by 24%, and Liquid AI’s LFM2.5-8B-A1B by 47%.
Broader reasoning: 95% on SVAMP and 64.1% on MuSR, the highest recorded scores among the small models compared.
BIG-Bench Hard’s logic subset: 95.4%, compared with VibeThinker-3B’s 61.1%.
We believe AI is entering a post-training era. The advantage will increasingly belong to companies with the best pipelines and those that can produce capable, personalized intelligence faster and more efficiently, then put it on devices people already own.
That’s what we’re building at webAI. And we’re only beginning to share what’s coming out of our lab.
Coming soon: Meridian, our family of frontier-class models built to run on device. Our most advanced models will be available through the @thewebAI application.
Join the waitlist as we expand access. Proudly built in Austin, Texas. 🇺🇸
39
301
1,824
140,640
Andrej Karpathy’s Small LLM Setup PDF:
drive.google.com/file/d/1Aw6…
2
33
4,034