Alessandro Ferrari

MS@Harvard, BS@GaTech, Until Labs | 8VC fellow | prev @ BigHat Bio, Maverick Metals (YC S22), Biocentis

[Alessandro Ferrari @BioAlessandro]

Building a compiler + HSL framework to turn @__tinygrad__ kernels into VHDL, and synthesize the perfect FPGA for a given compute graph.

Tinygrad UOps -> KernelIR (my custom IR) -> Amaranth hardware modules

[Alessandro Ferrari @BioAlessandro]

I love when claude code jerks my whole UI because it finished writing a complete slop summary of the changes I am currently reading

[Alessandro Ferrari @BioAlessandro]

Guy who puts🇻🇦in his bio not because he is religious but because he likes claude

[Alessandro Ferrari @BioAlessandro]

@GaddipatiHarsha How about making a product that will receive upvotes? That feels like a better use of YC resources😂

[Alessandro Ferrari @BioAlessandro]

New X banner

[NASA's Kennedy Space Center @NASAKennedy]

The planet can spell your name – literally. 🔤🌍

This Earth Day, see your name written in landscapes captured by Landsat: https://t.co/kcP12dhsI2

[Alessandro Ferrari @BioAlessandro]

@javi_2326 @andrewgwils Oh don’t worry, I’m not a math or physics major. Unfortunate reality for me as well.

[Alessandro Ferrari @BioAlessandro]

@viennaCtrl_ @andrewgwils Everything is math

[Alessandro Ferrari @BioAlessandro]

@predict_addict @andrewgwils Because I’m a cs major and consciously chose not to do math and physics knowing it would have probably been better. Hindsight 20/20

[Alessandro Ferrari @BioAlessandro]

@adityaxprasad “Ex sweatshop worker, able to pull 20 hour days”

[Alessandro Ferrari @BioAlessandro]

@adityaxprasad Biohacking

[Alessandro Ferrari @BioAlessandro]

To settle the "buy GPUs vs rent" debate for side projects: once you buy that RTX PRO 6000, you will procrastinate and let it idle. If you're renting it for $2/hr, you will be more productive working on your side project than you will at your day job.

[Alessandro Ferrari @BioAlessandro]

@datavorous_ An agent orchestration harness with a CEO, engineers, and validator agents enabling devs to push 150kloc/day? Seems like the right next step if this kid wants to stop shipping weekend side projects and make a real impact on the world!

[Alessandro Ferrari @BioAlessandro]

@quantian1 Yea, the same way everyone can trivially setup FTP with a CVS system on top and recreate Dropbox

[Alessandro Ferrari @BioAlessandro]

PrisML team iteratively adding bits each couple weeks until they land on bf16 from first principles

[Sahin Lale @SahinLale]

Turns out adding 0 helps :)

Today we’re introducing Ternary Bonsai 🌳, a family of end-to-end 1.58-bit language models in 8B, 4B, and 1.7B sizes.

Ternary Bonsai 8B is within 5% of Qwen 3 8B at 9x lower memory.

Still tiny. Noticeably smarter

[Alessandro Ferrari @BioAlessandro]

The european mind (me) cannot comprehend that you would see an $18 flight ticket and your first instinct is buy all of them for $3400

Well played, well played

[Alex Kehr @alexkehr]

the american mind (me) cannot comprehend european airline flight prices

can i just book all 190 seats for $3400 and have a private 737 flight?

[Alessandro Ferrari @BioAlessandro]

I respect the A/B testing

[Jessica Paquette @barrelshifter]

corrupting the youth by having them fix random llvm bugs

[Justin Xia @justinqxia]

Renting H100s from runpod to write tinygrad bounties like a medieval peasant paying a tithe to his feudal lord for a meager plot of compute. I toil day and night, hoping my bounty harvest is enough to win the respect of the king and avoid starvation

[Alessandro Ferrari @BioAlessandro]

@itselouardi 😂have to keep it for intellectual honesty. I fucked up

[Alessandro Ferrari @BioAlessandro]

It's nice that we could get Bonsai-family support so quickly, but this is a bit disingenuous. I have never contributed to tinygrad so I am not in a position to critique this, however this implementation unpacks the 1bit weights as float16 and runs computations on float16 instead of running custom kernels on the packed weights, nullifying a lot of the benefits of the Bonsai architecture.

Q1_0 It is a packed 1-bit format: for each block of 128 weights, you store 16 bytes of bits and 2 bytes for a shared fp16 scale.

128 weights take 18 bytes total. If you unpack those same 128 weights into float16, that becomes 256 bytes (14x).

This is basically unpacking the "bit-based llm" in normal float16 and running calculations that way.

My understanding of that llama.cpp’s Bonsai support keeps the weights in the quantized Q1_0 representation and uses kernels that operate on that packed format, which is the whole point.

Again, I do not mean this as a shot at the implementation itself. Getting support working this quickly is genuinely cool. I might also be misunderstanding some parts of this, hopefully not too much, but would love to be corrected.

[the tiny corp @__tinygrad__]

Just merged an external PR for Bonsai-8B support (1 bit LLM). Because tinygrad has the correct abstractions, it was 5 lines. https://t.co/BLljWDANgq https://t.co/GlXWqPbYg5

[Alessandro Ferrari @BioAlessandro]

Yea, just replied as well, I was totally misunderstanding. For some reason I thought that the ggml loading to tensor wouldn't be fused with the rest of the code (not sure why I would think that) so the scheduler only saw the multiplication with the float16 d, and separately saw the rest of the model. The memory usage doesn't lie