Pinned Tweet
This is what going solo at 40 feels like.
Iโm Aaron. 20+ years in tech. Ran a marketing agency for 8 years. Built apps and websites as a freelancer along with almost all of our agency systems.
In August, my business partner left our SaaS startup.
Now itโs just me, rebuilding ross South and Central America.
What Iโll share:
โ build lean with AI
โ run AI on your own hardware
โ live and work from anywhere
What are you building, or trying to get built?
Lets connect and grow together!
Iโm Aaron. 20+ years in tech. Ran a marketing agency for 8 years. Built apps and websites as a freelancer along with almost all of our agency systems.
In August, my business partner left our SaaS startup.
Now itโs just me, rebuilding ross South and Central America.
What Iโll share:
โ build lean with AI
โ run AI on your own hardware
โ live and work from anywhere
What are you building, or trying to get built?
Lets connect and grow together!
29
0
19
931
If this chart of Qwen4 27b is accurate, it's game over
Imagine having Opus 5.5 on your desk...
Imagine having Opus 5.5 on your desk...
0
0
4
573
Is this for real?
Qwen 4 is coming Now & Beats GPT-6, Fable 5.
- 27B variants land first.
- leak say the first samples are already approaching Astra/Opus 5.5-level quality.
- and Alibaba confirmed Qwen 4 will heavily use RSI (Recursive Self-Improvement) week ago.
- across the entire lineup Max, Plusget window: today or tomorrow may-be
- Open-source in November gets very interesting, after long time silent.
- The gap is closing,
1
0
3
50
โArenโt you scared to go to South or Central America?โ
The answer is a resounding no:
- Itโs no more scary than much of the US
- Thereโs no lack of McDonalds ๐
Of course it depends on what part you visit, but Panama is super safe and a fairly wealthy country with the canal.
t things is still less than the US, especially Uber rides and services so your US dollar goes further.
While itโs not as cheap as Colombia/Ecuador - itโs a great entry for any expat trying to adjust to life outside of the US with still maintaining a close proximity.
Whereโs your favorite place to travel or live?
The answer is a resounding no:
- Itโs no more scary than much of the US
- Thereโs no lack of McDonalds ๐
Of course it depends on what part you visit, but Panama is super safe and a fairly wealthy country with the canal.
t things is still less than the US, especially Uber rides and services so your US dollar goes further.
While itโs not as cheap as Colombia/Ecuador - itโs a great entry for any expat trying to adjust to life outside of the US with still maintaining a close proximity.
Whereโs your favorite place to travel or live?
0
0
2
60
I have to agree with Andrejโฆ
I trained my 2B model for just for 3 days and many of its specialized security code review tasks are already running better than Deepseek 4.1 Flash 1017
Smaller models makes much more sense from a standard hardware perspective bringing processing to edge devices.
I am lucky to have 2 DGX sparks with 256gb unified memory and a MacBook Pro with 48gb but the majority is working below 32gb of memory or vram.
A lot of the frontier models are a waste and overkill for many specialized tasks that businesses need.
The future is local AI
I trained my 2B model for just for 3 days and many of its specialized security code review tasks are already running better than Deepseek 4.1 Flash 1017
Smaller models makes much more sense from a standard hardware perspective bringing processing to edge devices.
I am lucky to have 2 DGX sparks with 256gb unified memory and a MacBook Pro with 48gb but the majority is working below 32gb of memory or vram.
A lot of the frontier models are a waste and overkill for many specialized tasks that businesses need.
The future is local AI
Andrej Karpathy:
โA 1B model can be much more effective than most current LLMs.
90% of current models waste a ton of capacity and money on remembering stuff that doesnโt matter.โ
here is his 5 core thesis about Small LLMโs:
1. the intelligence was never the knowledge. Strip keep the reasoning. Whatโs left is the cognitive core, and itโs small.
2. the core might fit in ~1B params. Maybe less. โPlenty of room at the bottom.โ Weโve been paying for a library when we needed a brain.
3. size stopped being the bottleneck. What makes an agent work is the tools it can call and how tight the feedback loop is, not parameter count.
4. intelligence lives in the system, not the weights. A raw LLM is a CPU with no OS. Small core + good harness > giant model alone.
5. you can own it. A model on your hardware is one no one can rate-limit, price-hike, or deprecate on you.
in 12 months the gap between teams shipping on-device and teams paying per-token frontier bills wonโt be capability.
TwIL an 8.8B model at 90%+ of frontier isnโt a benchmark flex.
Itโs the thesis Karpathyโs been repeating all year, made real.
study more on this below โ
1
0
0
71
If you want to follow my AI model training go here: nighttrain.ai
0
0
0
29
4-minute AI short film. Directed by Claude, rendered on my desk, $0 in API fees.
FLUX 3 Klein โ frames
LTX-2.5 โ video + SFX
Fish Audio Drama 3 โ voice
ACE-Step โ music
2x DGX Spark + a MacBook
FLUX 3 Klein โ frames
LTX-2.5 โ video + SFX
Fish Audio Drama 3 โ voice
ACE-Step โ music
2x DGX Spark + a MacBook
7
1
9
331
Honored to be on the list!
Check out Steve as well, building koru.do - an open source, local-first multiplayer AI workspace.
Check out Steve as well, building koru.do - an open source, local-first multiplayer AI workspace.
Sharing some of the builders I've connected with recently who are just starting out on ๐. Let's help each other grow.
@SofaVardazaryan, building apps and ScanCortex
@swhitf87, building Spreo
@joinearlynest, helping new apps find early users
@localhostnomad, building apps and local AI
@nocturnaljd11, building PacJobs in public
@tagtag193, building open source transactional email
@darrentian0, building Zeenly with agents
@KubalaChase, building Replio
@CobbleAppIphone, building a bite-sized life app
@heysahej, business operations and software
@PolDomingo_, indie hacking
@Developer__Rob, building ComingUp for iOS
@andreykuzmin342, building a visual .NET microservices tool
@nmaties09, building SaaS in public
@leo_granetto, building apps for athletes
@yarSaverin, building 10 public AI projects
@juanmcopia, building a first SaaS with coding agents
@PatrickALima98, building privacy-focused products
@RubanBhatia, building and sharing lessons
@rocker_hq, building an AI assistant for people and companies
2
0
3
92
Anyone else working on training models like this? I thought it was super cool.
x.com/localhostnomadโฆ
x.com/localhostnomadโฆ
Distilling DeepSeek V4.1 Flash 1017 into a 2B security model on 2 DGX Sparks, one night at a time.
Gap to the teacher: 11 โ 6 points in one night
Real CVE bugs caught: 2 โ 9
67 ms per answer vs 25 s
Retrains nightly at 3:45 AM EST. Follow the race ๐
1
0
4
250
DGX Spark + Jetson owners:
Time to start planning some physical AI projects!
What we gunna build?
Time to start planning some physical AI projects!
What we gunna build?
The era of System One Vision Model is here! Introducing VFlash: parsing the visual world in parallel at unmatched speeds for Physical AI.
Most embodied prediction tasks like object localization, robot trajectory planning, and dense 3D motion are fixed-length float vectors. For is always 6 tokens, a 10-waypoint robot arm trajectory is 54: we know these beforehand. Instead of decoding one token at a time, we can output the entire geometric shape in one go.
VFlash enables up to 12x faster action decoding and a 5.1x overall robot inference speedup on edge GPUs like the Jetson Thor and DGX Spark. Whatโs surprising is that it ends up even more accurate than the VLM backbone, with up to 61% lower 3D trajectory error. The intuition is that VFlash avoids the cascading errors of autoregressive decoding, where every coordinate is written in ink and every later point loses the chance to backspace.
Everything that moves will be autonomous, and everything autonomous will need to think on its own chip. The next inference revolution is physical reasoning at the speed of reflex.
And as always, itโs open paper, open source! Link in thread:
0
0
0
54
Distilling DeepSeek V4.1 Flash 1017 into a 2B security model on 2 DGX Sparks, one night at a time.
Gap to the teacher: 11 โ 6 points in one night
Real CVE bugs caught: 2 โ 9
67 ms per answer vs 25 s
Retrains nightly at 3:45 AM EST. Follow the race ๐
Gap to the teacher: 11 โ 6 points in one night
Real CVE bugs caught: 2 โ 9
67 ms per answer vs 25 s
Retrains nightly at 3:45 AM EST. Follow the race ๐
3
0
4
397
P.S. Every night it adds additional CVEs (security issues) and new synthetic code to increase the amount of data it trains on while improving accuracy each night.
Student is Qwen 3.5 2b
Student is Qwen 3.5 2b
0
0
0
34
Only 4 of these 16 AI chats are actually running.
My 2 DGX Sparks run 4 at a time. The other 12 wait in line.
โ All 16 tasks done in 37 s (66 s one at a time)
โ 108 tok/s combined
โ Each chat: 22โ44 tok/s
DeepSeek V4.1 Flash, fully local.
My 2 DGX Sparks run 4 at a time. The other 12 wait in line.
โ All 16 tasks done in 37 s (66 s one at a time)
โ 108 tok/s combined
โ Each chat: 22โ44 tok/s
DeepSeek V4.1 Flash, fully local.
1
0
1
121
