Fabrizio Milo

LF angels investors (inception phase)
AI for Software Development at Scale.
I believe:
- English is the new programming language
- Code will eat the world

[Fabrizio Milo @fabmilo]

Created with @NotebookLM after discussing these topics with @Cyndesama @tensorqt @Niccolg92 and few others

[Fabrizio Milo @fabmilo]

Rent compute add Dstack And serve for fun or profit

[DailyPapers @HuggingPapers]

NVIDIA just released an optimized GLM-5.2 on Hugging Face

A 753B parameter MoE with 1M context,
quantized to NVFP4 for Blackwell GPUs—
nearly matching FP8 accuracy.

[Fabrizio Milo @fabmilo]

@katedeyneka I heard them too, seem just pyrotechnics 🧨

[will brown @willccbb]

something has definitely shifted in the past few weeks. seeing a huge uptick in large enterprises wanting to secure compute and post-train their own models in house, frequently on top of GLM-5.2. everyone is starting to understand how open source wins.

[Fabrizio Milo @fabmilo]

@rohanpaul_ai @sundeep Isnt this just “capacity” of a model?

[Fabrizio Milo @fabmilo]

@vytalow @ycombinator Interested

[Fabrizio Milo @fabmilo]

@quantbagel I wish I had that kind of problem

[Fabrizio Milo @fabmilo]

Surprised by this news. I wasn't bullish on modular after my experience with Swift-Tensorflow but now even less so. Congrats on the relative fast exit to Chris.

[Modular @Modular]

We’re excited to announce that Modular has entered an agreement to be acquired by @Qualcomm. The future of unified compute has never been stronger. Read the full announcement: https://t.co/FiQUL5CvNj

[Fabrizio Milo @fabmilo]

@eXist3nZ_89 @antirez

[Fabrizio Milo @fabmilo]

GPT 5.5 PRO made me smile. While asking to design new objective functions for LLMs he came back to me with one option labeled as "one of my favorite". Clearly an indication of human data underneath but for a second it gave a "human flair" as a response.

[Fabrizio Milo @fabmilo]

I was so excited for the good deal. Too good to be true.

[Fabrizio Milo @fabmilo]

@per_simmons_ Now we just need a MIT source code unreal engine system

[Fabrizio Milo @fabmilo]

They have support for quite a few models too
https://t.co/2po3VofREK
has anyone tried to fine tune one for your in-house model ?

[Charles 🎉 Frye @charles_irl]

The no-longer-secret ingredient is DFlash by @zhijianliu_ and @jianchen1799.

If you train a custom DFlash speculator on your data, you can get to lower latencies than any generic inference API can achieve.

That's the benefit of owning your inference!

[Fabrizio Milo @fabmilo]

Seems the whole inference market is trying to get off NVDA chips. What will be the operational friction to adopt non-off-the-shelves hardware?

[OpenAI @OpenAI]

We’ve designed and built our first AI chip: Jalapeño.

Designed from the ground up by OpenAI and brought to production with @Broadcom, Jalapeño is purpose-built for the LLM workloads powering ChatGPT, Codex, the API, and future agentic products.

Chips are foundational to the AI economy. Building our own expands our full-stack platform from products to models to infrastructure, and will help us scale intelligence, serve more people, and expand access to AI.

[Fabrizio Milo @fabmilo]

@__tinygrad__ C + Cuda

[Fabrizio Milo @fabmilo]

@bernhardsson @modal How would you add a pricing service on top of it? Let’s say If I want to sell my own trained llm

[Fabrizio Milo @fabmilo]

@ivictorialiao Cards against humanity is still one of my favorite :D

[Fabrizio Milo @fabmilo]

Huge if actually works

[alphaXiv @askalphaxiv]

Introducing autoresearch for arXiv papers

Change 'arxiv' to 'autoarxiv' in any paper URL

An agent deploys to resolve setup issues on the codebase, run a minimal reproduction, and estimate full replication cost. Read more below

[Fabrizio Milo @fabmilo]

@MrAhmadAwais We are at a point where llms with the right skill.md can create a full binary that serves just one model . Checkout @antirez ds4 project. What local hardware do you have?

[Fabrizio Milo @fabmilo]

@willccbb If you know how those equations are derived is all a hack of approximations