Pinned Tweet
7 years of cold calling. Every script, every tracker, every number I actually run.
Put it in a box. $49. No course. No coaching call. No mindset module.
The Cold Call Operating System. themeepingman.gumroad.com/l/…
a fast no costs you one call
"send me some info" costs you 3 weeks of follow ups
which one do you hear more?
2
18
The MEEPING Man retweeted
It's illegal to bring a plant into our country because it might upset the ecosystem but bringing in millions of people from the 3rd world is supposed to make it better.. righttt
Replying to @BLKMDL3 @wholemars
In case anyone is curious:
We’ve had a ~pixel-perfect world model since last year, which is critical for synthetic data generation.
The (much) smaller version of that in the car is a conditional mean-flow model, based on the recent sequence of Tesla Universal Vision Encoder features, which are the same ones that are plugged into the 3B encoder model that produces the driving actions.
We fork off from these UVE features, then run a separate ~1B diffusion model. This takes in the previous reality-view generation with added noise and then denoises it to the current visualization.
It then produces a small video chunk (8 images at 36 fps, at the resolution of the UI screen). This work is done by 1/2 of a TRIP engine every UVE frame at 18 hz and is spread over 4 cycles.
Currently, the latency is a bit high due to compute constraints, but we have ways to interleave newer UVE features along the diffusion transformer layers to cut roughly 35% of the latency.
unpopular opinion: "brand awareness" is what you call marketing you can't measure
one guy on the phone is very easy to measure
which one would you rather pay for?
22
The MEEPING Man retweeted
pirateface.co/
Public service announcement : it is already too late to suppress open-weight and open-source models.
This is a good thing.
We worked with @thinkymachines to make Inkling 2 Mini run fast locally:
- 210 tok/sec decode, 3620 tok/sec prefill on 96GB M5 Ultra Mac Studio
- 125 tok/sec decode, 1750 tok/sec prefill on 48GB M5 Max MacBook Pro
It built a native weather app 1.8x faster than Claude Sonnet 5.5 on an M5 Ultra Mac Studio and 11x cheaper (API pricing).
The looping architecture is ideal for local devices like MacBooks, which are memory constrained. Looping layers of the model increases inference-time compute (and intelligence) without increasing the size of the model.
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
The MEEPING Man retweeted
holy shit if this is true we are about to have a local compute renaissance
and it comes right as NAND prices are starting to drop... beautiful!
SSD is about to become the only limiting factor for local AI
Most frontier open models are ultra sparse MoEs
Qwen 3.8 Flash Next = 180B total, ~6B active per token (~3%)
GLM 5.3 Flash = 321B total, ~18B active per token (~6%)
DeepSeek V4.1 Flash = 552B total, ~16B active per token (~3%)
GLM 5.3 = 753B total, ~40B active per token (~5%)
Kimi K3.1 Pro = 2.8T total, ~104B active per token (~4%)
When ~95% of a model sits idle on every token, keeping all of it in VRAM makes no sense
Local AI has been evolving toward this for years:
2023-2025 - Layer offloading (Ollama, classic llama.cpp)
This chopped whole layers across GPU and CPU, leaving the GPU waiting on CPU once VRAM filled up
Oct 2026 - RAM expert cache (llama.cpp --moe-cache)
This shifted from whole layers to individual experts, caching hot ones in VRAM and pulling misses from RAM. Huge leap for decode, but still capped at the size of your RAM
Now - SSD expert streaming
Early tools like Colibri and Strata proved direct NVMe streaming actually works on consumer hardware. Instead of hitting a VRAM or RAM wall, your SSD becomes an active part of the engine
However most people hear "run from SSD" and assume 2 tok/s, which isn't entirely wrong. A few problems still exist:
- Basic caching kicks out experts you're about to need again
- Waiting on disk reads one layer at a time leaves the GPU and SSD sitting idle
- Getting speculative decoding to work well with SSD streaming is hard
- Long prompts are tough to keep fast without wiping out your hot cache
- Most early tools are experimental engines built for only a couple models
Making expert streaming feel fast takes solving all of those together
At Magnitude, we've been working on this for the past 2 months. Here's how we solved it:
- Smarter caching that keeps your session's most-reused experts in GPU memory
- Speculative lookahead that pulls upcoming experts off SSD before the GPU needs them
- Direct NVMe reads scheduled across layers so your SSD never sits idle
- Streaming prefill that reads each cold expert once without wiping your hot cache
- Built into a full engine that tunes itself to your chip across Macs and GPUs
All of this together means DeepSeek V4.1 Flash on a single gaming GPU at ~40+ tok/s...
We're excited to release it on Monday
GitHub here: github.com/magnitudedev/magn…
The MEEPING Man retweeted
For a while, SendCutSend had to turn away the big orders.
Someone would need 100,000 parts, and it would tie up a single machine for days, so the company couldn’t take the work. Jim Belosic (@jimbelosic) says fixing that, with more machines across more buildings, is what let the Fortune 50 start ordering.
Those bigger customers came from the hobbyists. The guys building weird projects in their garages turned out to have day jobs, and they did the selling.
“They’ll go fight for us and say, hey, I use this company. They’re amazing. We should use them at work.”
My guest today is Jim Belosic (@jimbelosic), Founder and CEO of SendCutSend, the company where you upload a design and get custom metal parts back, often the next day.
Jim started it in 2018 as a side project to cover the payment on one machine, which he bought to make better parts for the cars and motorcycles he worked on in his garage. He says it's now grown to almost 800 people and about $200 million a year in revenue.
We talk about why he keeps millions of dollars of inventory on the floor when every finance person tells him not to, why he wants more machines instead of faster ones, and why the money he raised this year hasn’t made him any faster.
Timestamps:
(00:00) Intro
(00:56) From Garage Side Hustle to a $200 Million Manufacturer
(05:19) Growing from Laser Cutting into Full-Service Manufacturing
(09:27) When Reindustrialization Became a Movement
(15:19) Stocking Millions in Inventory to Ship Faster
(19:28) From VC Skeptic to a 30-Day Series A
(27:05) What the Best Investors Ask on a First Call
(30:17) Running the Business with His Wife as President
(34:39) Why Money Isn't the Constraint on Growth
(39:01) What a Factory Needs from an Industrial Building
(45:11) Where AI Actually Helps a Manufacturer
(48:07) What to Look for When Walking a Factory Floor
(53:28) A $2 Billion Company Without a C-Suite
(1:01:11) More Machines, Not Faster Machines
(1:07:40) The Myth That America Can't Make Things
(1:11:03) Buying Companies to Skip Years of R&D
(1:17:03) Protecting Culture While Doubling Headcount
which is even funnier because I am actually 32...
17
The MEEPING Man retweeted
I don’t think I’m exaggerating when I say this: Inkling 2 Mini might be the inflection point for local ai.
1M context length, natively multimodal, only 20GB, AND as good as Qwen 3.8 27B
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
The MEEPING Man retweeted
I’m looking for the most cracked engineers to help kill Bambu Labs.
We’re building the best 3D printers, that run fully autonomously without any human in the loop from the ground up.
Work will involve working on hardware and building the printer, plus CAD experience to design and build it.
If you’re cracked and interested, hit me up
Gemini 4 Neon just made Thinking Machines' Inkling 2 Mini run almost 6x faster on AMD hardware, by writing and optimizing custom GPU kernels on its own.
An AMD engineer gave Gemini 4 Neon the Inkling 2 Mini GGUF and let it handle the entire process on a Strix Halo (8060S) GPU: writing kernels, profiling them, finding bottlenecks, optimizing and benchmarking.
It worked for 4.3 hours and made around 490 tool calls.
The final results compared to llama.cpp on ROCm:
Decode: 44.8 tok/s vs 23.8 — 1.88x faster
Prefill: 1,465 tok/s vs 233 — 5.9x faster
4-stream aggregate: 100.8 tok/s vs 54.1 — 1.86x faster
The generated kernels reached around 80% of the GPU's theoretical memory bandwidth.
An AI model autonomously writing and optimizing low-level GPU code to get this kind of performance is genuinely impressive.
The greatest LLM to come out of the United States.
Thinking Machines will rule over all the labs, I have no doubt about it
NEW: Inkling 2 Mini is the most efficient model we've tested:
- #1 on efficiency (cost vs. intelligence)
- #9 on intelligence (78.4) overall
- #1 on intelligence among open-weights models
- Lowest cost per correct answer ($25.80)
- Biggest accuracy gain from adding web search
It uses a looped transformer, which runs its middle layers three times for every token. That lets a 26B-parameter model reason more without generating more tokens, which keeps costs low. A model that size can't store as many facts, so it relies on search to fill the gaps, and that's where it gains the most.
The MEEPING Man retweeted
Austin, TX friends 👋
If you're in town for Thanksgiving next week and away from your family or don't have anyone to share a meal with. Come join me/us at my house.
Shoot me a DM if you want a plate and seat with your name on it.
The MEEPING Man retweeted
Boomers never made more than $60k a year and bought a new house for $120k and now that the pipes are rotting they want $800k
Lmfao as if
























