AI Researcher focused on creating performant yet sustainable models.
Italy
Joined September 2021
- Tweets1.5K
- Following376
- Followers135
- Likes11.1K
Pinned Tweet
Every 18 months, closed-source AI drops a breakthrough, a chasm opens up, and the timeline collectively sighs: "It's over. Small labs and open source can never catch up to this." Yet they do. Never bet against the open ecosystem.
Grok has a seemingly secret image to 3D mesh model that you can access through @grok build. Works pretty well. It's a glb so I could load it in cad-os.
I've been building CAD OS, CAD you write as Luau code that runs entirely in your browser.
Everything in this video, and the framework itself, was built with @bot. It's all WASM, from the geometry kernel to the AI agent, which runs sandboxed in agent-os.
I can already see it extending to FEA, robotics and more.
To my non-AI colleagues and friends, @bot is indistinguishable from magic. It’s a significant leap forward in quality, comparable to the early days of Claude.
Within the confines of his cage, the tiger’s life is more important than any unauthorised human who has entered it. Outside that cage, any human’s life is worth more than the tiger’s.
It’s Oct. 7, so I would like to say that I still support Israel and detest all terrorists. It might be unpopular to say among those riled up by war pornography and addled by social media, but saying what I believe in is Dharma.
There’s a lot of alpha in personalization.
I’ve experienced this firsthand with @bot. By telling it about myself and how I search - how to maintain offline/reference records, build trusted sources, and translate vague requests into effective search queries - it gradually developed a workflow around me that has led into drastic improvement for deep research.
Introducing Dreaming: Across sessions, Devin builds a memory graph of how you like to work
At night Devin self-improves its memory to remove stale records and discover latent information
We are creating an OSS standard called Agent Memory Repo: cognition.com/agent-memory-r…
Grok 4.7 is worse than 4.6 in actual usage. Its context window is half, it overthinks and fills that window up, leaving very little space for actual work. This, in turn, causes many more compression cycles and leads to slowdown.
I wanted Jev to be reproducible from a stock autoregressive model without fine-tuning it. The interface is easy to copy: prefill the state once, score the options in parallel, softmax the logits, emit typed answers. That’s what most of the open Jev repos actually ship. The model still doesn’t act like Jev.
Narendra Patwardhan retweeted
me: how do i open pdf
claude: You're right to ask. The PDF is the prize; the mouse is the gate; and the map is the fulcrum that holds the raven's skull. The furniture's stale, but the instrument's fresh, and the fresh half is the half that counts. I deleted your home directory.
Jev resembles a pointer net on a causal prefill. Encoding the state once and treating each option as a span, then pointing at those spans, would explain the fast 255-way choices and their ability to provide output for free.
Narendra Patwardhan retweeted
People will claim with a straight face that it's *impossible* for schools to let kids learn at different speeds
But we know how to do it! We used to do it!
You just put kids into groups based on their ability.
And then you teach each group the material they're ready for, at the pace they're able to learn it.
It's common sense!
And ironically, it's universally opposed within the field of education, which believes that educating kids is less important than equalizing them.
unavailable
Grok is no longer as fast as it was, it is also providing substantially lower quality responses. I can no longer recommend it.
