Fabrizio Milo
@fabmilo
Followers
835
Following
4K
Media
438
Statuses
2K
LF angels investors (inception phase) AI for Software Development at Scale. I believe: - English is the new programming language - Code will eat the world
San Francisco
Joined November 2009
Created with @NotebookLM after discussing these topics with @Cyndesama @tensorqt @Niccolg92 and few others
1
0
11
THE multi hop agentic benchmark is out. Talk is cheap, show me the run.
Today, we’re announcing the next chapter of Terminal-Bench with two releases: 1. Harbor, a new package for running sandboxed agent rollouts at scale 2. Terminal-Bench 2.0, a harder version of Terminal-Bench with increased verification
0
0
5
Great work by @Mike_A_Merrill and @alexgshaw on Terminal Bench 2.0 with @LaudeInstitute here at @databricks
0
0
4
The people crazy enough to think they can compress the world into a latent space are the ones who do
0
0
1
Just opening and closing claude code costs me $0.0005 is just me because of some of my settings ?
0
0
2
The Real Arena for AI Agents
0
0
5
I think Terminal Bench is the one true benchmark for multistep coding agent intelligence . Looking forward to attending the event.
Only a few days left until the 𝗧𝗲𝗿𝗺𝗶𝗻𝗮𝗹 𝗕𝗲𝗻𝗰𝗵 𝗠𝗲𝗲𝘁𝘂𝗽 at @databricks HQ! 🚀 📣 Join the #MLflow community this Thursday for an evening of demos, discussions, and new announcements from Terminal Bench 2.0. Hear from Danny Chiao on building high-quality agents
1
0
0
63 frames took 10 minutes to generate this Minecraft world model video from Open Oasis on my Apple M4, by just changing the Torch backend from CUDA to MPS. Can I make it go faster?
0
0
0
Catching up on world models? this is a good read to review the history that lead to current state of the art AI and have a mental framework to categorize the various projects.
1
0
0
Completed this nice intro RL course by @DeepLearningAI and @realSharonZhou it goes through the whole process of building a modern RL pipeline from training to production.
0
0
3
This explains why the other coding models have been deprecated. Definitely another powerful move in the coding space. Cerebras is unbeatable in speed and speed in the coding operational loop is important. I am actually curious to try this one.
Today, @cognition released SWE-1.5 – the world’s fastest coding agent, powered by Cerebras. SWE-1.5 achieves frontier-level coding ability, comparable to Sonnet 4.5 and surpassing GPT-5. Cerebras and Cognition engineers worked hand in hand over the past few weeks, training a
0
0
0
I think the main issue for independent vibe coders is cost, many have different accounts (openai / codex / openrouter) for various reasons and they want to utilize those credits before spending more money even if they have to do some extra manual work . If they are in big
Man, I think people using model-selector coding agents still don't appreciate the power of Amp's agent-oriented architecture. Here's a thread where main agent got it wrong, so I just asked it to use the oracle subagent and voila—bug fixed! https://t.co/lxsNB0ngww
1
0
0
Zero installation is huge . It feels like internet explorer era all over again
OpenAI Codex is now integrated directly in @code through the new Agent Sessions view - and can be powered by your GitHub Copilot subscription. Try it out now with VS Code Insiders and a Copilot Pro+ subscription. Happy coding!
0
0
2
Very strategic move from GitHub. They still own the entire workflow, how those files are changed is not their concern.
@github bringing all coding agents in one platform under one subscription with their new "Agent HQ" What will this unlock? Agent to Agent collaboration? Orchestration? I will find out more on the vision today as I have been given the opportunity to interview some leads
0
0
1
Interesting research from my university research lab on leveraging the property of “injectivity” of LLM prompts.
LLMs are injective and invertible. In our new paper, we show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space. (1/6)
1
0
3
MiniMax M2 is free until Nov 7th. Here is how you can configure claude code to use it.
@MiniMax__AI How to setup MiniMax-M2 in Claude Code, Codex CLI, Cline, etc. via compatible Anthropic API: https://t.co/TilKREMDww Took me some time to hunt this down. Thank me later.
0
0
0