If you missed my talk at DefCon about running your own LTE network it's available on YouTube now!
youtube.com/watch?v=rEoIq_Z0…
Rick retweeted
The Air Force basically invented CNC because they got tired of machinists struggling to make complex aircraft parts by hand.
In 1952 MIT showed them a milling machine that could cut from numbers on a tape. Manufacturing never went back.
This is great if you are building in bio field
Excited to launch another edition of the Alive Prize. $100K to build what makes you come alive.
We started this in part because of our experience in school. It felt uncommon to find people doing things they were deeply, intrinsically lit up about. This is a pursuit of the opposite.
Howard Thurman said, “Don’t ask what the world needs. Ask what makes you come alive, and go do it. Because what the world needs is people who have come alive."
We want to help more people to find that feeling. We set aside $100,000 to fund individuals working on consequential problems.
Past recipients select and support new ones. Consider this a nudge to join them. The application takes ten minutes and closes on 11/25.
Thanks to @maxxyung11, @ladanuzhna, @RealJPGE, @romanelli_bio, @MLeshchiner, @sahirjaggi, @MattHdeJonge, @khawajasn, Julie Zhang, @FarhanKhodaee, @michaelbaum, @gustai, @hoomanrenezhad.
Apply here: aliveprize.com
1
32
Very good tutorial if you want to have a local stuff but imo turning off thinking can make it even more dumb than Gemini
Here's 7 steps to run an uncensored AI with no GPU (or a 6-year-old one):
every local AI post I make gets the same reply: "cool, I don't have a Mac Studio or a 5090"
FAIR, this one is for you... I tested it on a 2019 laptop with no dedicated GPU
By the end, you'll know how to:
- run an abliterated model on CPU and 8GB of RAM
- make it usable on a GTX 1060 / 1650 / any old card
- do it for $0, no API key, no account
So, let's go step by step:
1. Install Ollama
curl -fsSL ollama.com/install.sh | sh
Windows: download the installer from ollama.com, no WSL needed. Ollama runs on CPU automatically if there's no usable GPU. if you have an old Nvidia card it puts what fits in VRAM and the rest on CPU, no config
2. Pick the right size for your box (this is where people fail)
- 4GB RAM or 4GB VRAM: Gemma 4 E2B. ~3GB. dumb but runs anywhere
- 8GB RAM, or 6GB VRAM (GTX 1060 / 1660 / RTX 2060): Gemma 4 E4B. ~5GB. the sweet spot
- 12GB+ VRAM: skip this post, run the 12B from my earlier one guide
the mistake is pulling the 12B on 8GB and concluding "local AI is unusable". it was never going to fit (I see in comments how people arguing for it lol)
3. Pull the model
ollama pull hf.co/igorls/gemma-4-E4B-it-…
same author and method as the 12B heretic. ~5GB, runs in 5GB RAM. KL 0.055 from the original, refusals ~98% → ~21% on the hardest adversarial set with thinking on, near zero with thinking off
Gemma 4 is Apache 2.0, so no license homework on this one
4. Turn off thinking
ollama run hf.co/igorls/gemma-4-E4B-it-…
/set nothink
on a GPU thinking costs tokens. on CPU it costs minutes. and the abliteration targets non-thinking mode anyway
5. Keep context small
/set parameter num_ctx 8192
CPU speed is memory bandwidth, nothing else. every token re-reads the whole model from RAM. the KV cache grows with context and competes for the same RAM. 8k is plenty for chat, drafts, critique. 32k on CPU is self-harm
6. Three settings that make the difference (optional, worth it)
restart Ollama with:
OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 ollama serve
flash attention + q8 KV cache halves what the context costs in memory, with no real quality loss. on a 6GB card this is often the difference between "whole model on GPU" and "half of it spilled to CPU"
if Ollama runs as a service, setting this in your shell does nothing. on Linux: sudo systemctl edit ollama, add Environment="OLLAMA_FLASH_ATTENTION=1" and the KV line, restart
old card note: flash attention wants RTX 2000 or newer. on a GTX 1060 it may not engage. skip this step, it still runs
7. Check it's actually using what you have
/set verbose ← prints tokens/sec after every answer
ollama ps ← in a second terminal
ollama ps shows "100% GPU", "100% CPU", or a split like "60%/40%". if you have a card and it says 100% CPU, your driver isn't loaded. nvidia-smi should list the card. fix that before touching anything else
What to expect:
- CPU only, modern laptop: ~8-15 tok/s. reading speed, not typing speed
- GTX 1060 6GB / 1660: whole model on GPU, ~25-40 tok/s
- GTX 1050 4GB: split. faster than CPU, slower than you'd like
- a 2020+ Mac with 16GB: just works, ~20 tok/s
honest take:
- E4B is a 4B-class model. good for writing, chat, drafts, roasts, critique. bad for hard reasoning and agentic coding. don't plug it into Claude Code, you'll hate it
- 10 tok/s feels slow for 2 minutes and then it doesn't. you're not paying anyone
- if you can get 12GB VRAM used for $150, do it. the 12B is a different league
- there's no secret trick after this. the rest is buying RAM
*IMPORTANT: don't use it for anything illegal or harmful. that's on you, not the model
next: how I run the 12B on an 8GB card without it spilling to CPU
save it so you don't lose it
35
Rick retweeted
Wie eine Zertifizierung funktioniert:
> Firma macht seit 30 Jahren gute Arbeit
> großer Kunde verlangt plötzlich ISO 9001
> Berater kommt, Tagessatz vierstellig
> schreibt auf, was eh schon alle machen
> heißt jetzt "Prozess"
> Auditor kommt
> prüft, ob es aufgeschrieben ist
> nicht, ob es funktioniert
> Zertifikat hängt gerahmt im Flur
> nächstes Jahr Überwachungsaudit
> dann ISO 27001
> dann Nachhaltigkeitsbericht
> jetzt gibt es einen Qualitätsmanagementbeauftragten
> einen Informationssicherheitsbeauftragten
> einen Datenschutzbeauftragten
> und jemanden, der die Beauftragten koordiniert
Ein ganzer Wirtschaftszweig, dessen Produkt ein PDF ist, das bestätigt, dass es ein anderes PDF gibt.
Rick retweeted
Founders & friends, I could do with some amplification on this one if you're willing (also intros if there's someone you trust and like) 🙏🏻
If anyone wants a chance to be in the top companies this YC batch (picked by me). There is some allocation left in my rolling quarterly fund for F26.
You can see the companies I've backed in the past here:
h1nd.com/companies
You can see what founders and my network looks like here:
h1nd.com/testimonials
You can apply here and I'll be in touch same day:
h1nd.com/investments
Rick retweeted
After the FAA disaster we're all operating in a high-take environment, which is why I'm tapping the sign.
how.complexsystems.fail/
Rick retweeted
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
Rick retweeted
‼️ A hacker says an 8GB Nvidia crypto-mining card hides 80GB of memory. She unlocked it, and in her tests it outran Nvidia's A100 AI chip.
The CMP 170HX uses the same chip as the A100, with the extra memory locked off by fuses. Kate broke the lock by glitching its voltage with a homemade STM32 board that she says an AI model designed.
That switched on up to 126 compute units, more than the 108 on any A100. It hit 85.9 TFLOP/s in FP16 against 72.5 for a retail A100, with all 80GB usable.
Each card took 1 to 48 hours of glitching. One of her 10 only gave up 64GB.
Rick retweeted
Build your next $100M ARR startup with HeyPCB.
Rick retweeted
I have stage 4 cancer. I thought I'd fight it alone.
I was wrong.
Since mid October, ~20 strangers on X have stepped up to help me. All for free so far:
- An NYU Langone lab offered to run my tumor data through their system to sharpen my vaccine targets
- A U Michigan lab offered to test drugs on my actual tumor cells
- A biotech ran AI analysis on my tumor RNA
- A vaccine company offered to design a second vaccine to check the first one
Scientists. Engineers. People I've never met.
Cancer has been picking us off one at a time for all of human history. It counts on us fighting it alone, scared and quiet.
Not anymore.
We have AI. We have the internet. We have each other.
Cancer, we're coming for you, bitch. Buckle up!







































