Caleb Palmquist

[Caleb Palmquist @CalebFromMaine]

@cynthiamcgillis @ciphergoth There are more and more with more planned. They use a lot of water and power. Power rates go up for regular people, but the big companies get tax incentives and deeply discounted rates.

[Percy Liang @percyliang]

🚢 Marin 535B-A23B started training this week! As usual, the whole process is open.
Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow.
Before kicking off the run, we trained a 4-rung scaling ladder from 1.6B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.

[Bounty @bountyAIhunter]

512GB of DDR5. That is the line nobody screenshotted from the FreeToken paper.

Berkeley and MIT open-sourced it yesterday and the headline numbers are real: 753B GLM-5.2 on one workstation GPU at 14.9 tok/s, 35B on an 8GB laptop at 39.3. I installed it on my 3090 to break the one claim I could actually test, that the GPU decides what fits.

It doesn't. Their own hardware table decides it:

model GPU VRAM host RAM
Qwen3.6-35B-A3B 4060 laptop 8GB 32GB
DeepSeek-V4-Flash 5090 desktop 32GB 192GB
GLM-5.2 753B PRO 6000 96GB 512GB

The GLM checkpoint is 433GB. It streams through system RAM at 178GB/s on that box. VRAM is the cache, RAM is the model.

My rig: 3090, 24GB, 64GB DDR4. DeepSeek-V4-Flash never produced a token, the loader walked the checkpoint and gave up, because 64GB of RAM is 64GB of RAM. Qwen3.6-35B ran:

llama.cpp FreeToken
decode tok/s 28.7 51.4
TTFT on 21k prompt 37.2s 9.4s

1.79x, inside the 1.3 to 2.1x they claim across consumer boxes. No tuning flags, it reads the checkpoint and the GPU and configures itself.

Two things that will cost you an evening. Linux x86_64 only, driver r580+, and the CUDA kernels JIT compile on first run so you need CUDA 13 with nvcc on PATH. And it gives VRAM back when another app asks: I opened a browser mid run and decode fell to 31.6 tok/s until I closed it.

uv pip install "freetoken[accel]"
ft serve --model ~/models/Qwen3.6-35B-A3B

Buy RAM before you buy the next GPU.

[Caleb Palmquist @CalebFromMaine]

@PhilosophyOfPhy Four if you count time...five if you count...

[Md Ismail Šojal 🕷️ @0x0SojalSec]

A Tiny 0.9B Params OCR model beats Gemini on the hardest document benchmarks.

GLM-OCR : is a tiny vision-language model that supports 8K resolution,

- 8+ languages,
- and has native Built-in text & LaTeX & table recognition modes.

It scored 94.62 on OmniDocBench V1.5 higher than Gemini 3 Pro at the time.

Demo of packing list clean structured JSON:

- https://t.co/0koCaBwQiq

[Caleb Palmquist @CalebFromMaine]

A nice plant 🪴

[Caleb Palmquist @CalebFromMaine]

Someone else thinking like I think. Nice.

[Akshar Vastarpara @AksharVastarpar]

All the CPUs we use today execute mathematical functions written by us, and those functions are deterministic. We needed AI because manually writing functions for understanding human intent, language, intelligence, and similar concepts is practically impossible.

Neural networks can approximate almost any function, but they require massively parallel multiplication, MAC, and GEMM-style operations. Following one of the great forefathers of computing, Gordon Moore, and Moore’s Law, the continued shrinking of transistor sizes and increase in computational density at lower power, performing trillions of operations eventually became easier and cheaper.

Today, we have reached a point where models that approximate these complex functions can run across many different types of hardware. We have moved from CPUs to GPUs, TPUs, NPUs, XPUs, and other specialized accelerators.

I believe there will be a new field of innovation where neural networks can generate or dump the instructions representing the approximated functions they have learned the very functions that power the neural network itself into a form that can run directly on a CPU.

If that becomes possible, these generated functions could potentially perform inference much faster and more efficiently than running the original neural network on GPUs, TPUs, NPUs, or other accelerators.

[Caleb Palmquist @CalebFromMaine]

That would definitely raise my eyebrows

[Ahmad @TheAhmadOsman]

Any talented potential hire will sign immediately if you offer them a sign-on bonus of a GB300 / DGX Station on the spot

[Mark Kretschmann @mark_k]

FreeToken could be a HUGE deal for local AI.

Instead of requiring enough VRAM to hold an entire model, FreeToken intelligently uses your GPU, CPU and system RAM together, dynamically moving MoE experts where they're needed.

The result: an ordinary laptop with an 8GB RTX 4060 can run a 35B model at 39 tok/s, while a gaming PC with an RTX 5090 can run a massive 284B DeepSeek model interactively.

For normal users, this could mean running much more powerful AI locally on hardware you already own: no API costs, no cloud dependency, and your data stays on your machine.

This is exactly the kind of software breakthrough local AI needed.

[Caleb Palmquist @CalebFromMaine]

@Angaisb_ Are you running this inside some kind validation program or did Claude design that plus the circuit? Crazy impressive.

[Caleb Palmquist @CalebFromMaine]

@dimitrov2k just name functions and variables properly and you'll know what it does by reading it...

[Caleb Palmquist @CalebFromMaine]

This is super amazing.

[Angel 🌼 @Angaisb_]

I asked Codex to design its own CPU and then write a Space Invaders-style game for it in assembly and after a while working on it this is the result

Now the real question is... can it run Minecraft?

[Tech with Mak @techNmak]

In 1948, a 32-year-old at Bell Labs published a paper nobody fully understood.

Engineers found it too mathematical. Mathematicians found it too engineering-focused. One prominent mathematician reviewed it negatively.

That paper - "A Mathematical Theory of Communication", became the founding document of the digital age.

The man was Claude Shannon. Father of Information Theory.

At 21, he wrote the most important master's thesis of the 20th century.

Working at MIT on an early mechanical computer, Shannon noticed its relay switches had exactly two states - open or closed. He had just taken a philosophy course introducing Boolean algebra, which also operated on two values: true and false.

Nobody had ever connected these two things.

His 1937 thesis proved that Boolean algebra and electrical circuits are mathematically identical, and that any logical operation could be built from simple switches.

Howard Gardner called it "possibly the most important, and also the most famous, master's thesis of the century."

Every digital computer ever built traces back to this insight.

At 29, he proved that perfect encryption exists.

During WWII, Shannon worked on classified cryptography at Bell Labs. His work contributed to SIGSALY, the secure voice system used for confidential communications between Roosevelt and Churchill.

In a classified 1945 memorandum, he mathematically proved the one-time pad provides perfect secrecy, unbreakable not just computationally, but provably, permanently, against an adversary with infinite power.

When declassified in 1949, it transformed cryptography from an art into a science. It laid the foundations for DES, AES, and every modern encryption standard.

At 32, he defined what information is.

His 1948 paper introduced one equation:
H = −Σ p(x) log p(x)

Shannon entropy. The average uncertainty in a probability distribution. The minimum bits required to encode a message.

Three things followed:

> He defined the bit - the fundamental unit of all information. His colleague John Tukey coined the name.

> He proved the channel capacity theorem, every communication channel has a maximum rate of reliable transmission. You can approach it. You can never exceed it.

> He unified telegraph, telephone, and radio into a single mathematical framework for the first time.

Robert Lucky of Bell Labs called it the greatest work "in the annals of technological thought."

Where his equation lives in AI today:

Cross-entropy loss - the function training every classifier and language model, is derived directly from H. Decision tree splits use information gain, which is H applied to data. Perplexity, the standard LLM evaluation metric, is an exponentiation of cross-entropy.

Every time a neural network trains, Shannon's formula runs inside it.

He also built the first AI learning device.

In 1950, Shannon built Theseus, a mechanical mouse that navigated a maze through trial and error, learned the correct path, and repeated it perfectly. Mazin Gilbert of Bell Labs said: "Theseus inspired the whole field of AI."

That same year he published the first paper on programming a computer to play chess. He co-organized the 1956 Dartmouth Workshop, the founding event of AI as a field.

The man:
He rode a unicycle through Bell Labs hallways while juggling. He built a flame-throwing trumpet, a rocket-powered Frisbee, and Styrofoam shoes to walk on the lake behind his house.

He called his home Entropy House.

When asked what motivated him: "I was motivated by curiosity. Never by the desire for financial gain. I just wondered how things were put together."

In 1985, he appeared unexpectedly at a conference in Brighton. The crowd mobbed him for autographs. Persuaded to speak at the banquet, he talked briefly, then pulled three balls from his pockets and juggled instead.

One engineer said: "It was as if Newton had showed up at a physics conference."

He died in 2001 after a decade with Alzheimer's, the cruel irony of information slowly leaving the mind of the man who defined what information was.

Claude, the AI model, is named after Claude Shannon, the mathematician who laid the foundation for the digital world we rely on today.

[Shuo Yang @Andy_ShuoYang]

Your gaming PC can now serve frontier models at interactive speed using official checkpoints without extreme quantization!

Qwen3.6 35B → 8GB RTX 4060 laptop @ 39 tok/s

DeepSeek-V4-Flash 284B → RTX 5090 desktop @ 22-25 tok/s

GLM-5.2 753B → RTX PRO 6000 workstation @ 15 tok/s

Run your claude code or codex now with frontier model for $0

Meet FreeToken 🧵

[Caleb Palmquist @CalebFromMaine]

Going to have to retweet since bookmarks appear to be gone

[Mathelirium @mathelirium]

A sharp 90-degree turn is a brutal request for an Electromagnetic wave. Yet inside a Photonic Crystal, something remarkable can happen.

The repeating dielectric structure creates ranges of frequency for which propagation through the bulk crystal is strongly suppressed. Remove a line of that crystal and suddenly the defect becomes an allowed route the EM field.

[Caleb Palmquist @CalebFromMaine]

X is so confusing. I’ve literally done nothing.

[Caleb Palmquist @CalebFromMaine]

Where the fuck are my bookmarks?

[Caleb Palmquist @CalebFromMaine]

So literally copy the harness Cursor and Copilot standard.

Easy.

[avrl ☘ @avrldotdev]

@GonnabeNikhil Correct, will u use llms in the layer?

[Caleb Palmquist @CalebFromMaine]

Bet

[Alex Graveley @alexgraveley]

I’m betting in <5 years there’s only 2 programming languages in use and it’s not Python.

[Peter Yang @petergyang]

My gut is that alot of AI output can be improved just by building a loop where a manager agent negs the other worker agent with prompts like:

"Are you sure this is the best you can do?"
"I think you can do better, try again"
"Take a closer look, give me 11/10 output"