clainstone (@cclainstone) header
clainstone (@cclainstone) profile picture

MSc Applied Math @polimi ⏳ Statistical learning by day | Local AI by night 2 x 5070ti | Ryzen AI 370 mini PC | MacBook Pro M4

Milan, Lombardy
Joined March 2022
Photo by clainstone (@cclainstone)Photo by clainstone (@cclainstone)Photo by clainstone (@cclainstone)Photo by clainstone (@cclainstone)Photo by clainstone (@cclainstone)Photo by clainstone (@cclainstone)Photo by clainstone (@cclainstone)Photo by clainstone (@cclainstone)Photo by clainstone (@cclainstone)
clainstone retweeted
Deepseek runs nicely in Claude Code, single click deployment. I want to release this repo soon.
26
1
123
5,347
‼️ Someone says they burned 8.9 billion tokens getting Claude Code to make a single strawberry in Blender, the free 3D modelling program. The creator's own write-up puts it at 2 weeks, 3,958 prompts and 17,453 tool calls.
98
92
23
1,630
119,103
clainstone retweeted
Software is both art and utility. AI is both an Industrial Revolution and a Renaissance. Factories are for utilities, at AIE Code I talked about how we can make better art.
Vibe coding a game is fun until you don't know what your own characters do. @trq212 on keeping creative control as models get better: readable specs, custom languages, and a software renaissance. The Sweet Side of the Bitter Lesson. Watch: youtu.be/xRJMiaUoD_8
54
59
3
1,000
49,938
clainstone retweeted
🚨 BREAKING: Jensen Huang is reportedly considering founding a brand-new semiconductor company 🚨 He could raise tens of billions on day one, would have full founder control, and could build fast on his own terms. With the relationships he's built over decades, he could easily plug into the #AI ecosystem. Imagine the demand and customer base lining up before his first chip even ships... @grok Could Jensen Huang launch a competitor to $NVDA, attract tens of billions in funding, and lock in major customers?
1,908
3,901
773
33,244
4,222,970
BREAKING: AMD stock, $AMD, crashes -13% in 30 minutes after Hunterbrook Media accuses the company of illegally selling chips to China. Hunterbrook Capital also disclosed a short position in AMD.
757
1,685
581
13,809
1,446,890
clainstone retweeted
an age of miracles
82
74
8
3,277
138,512
clainstone retweeted
Gemini 4 Neon is here 🚀 Almost 3x faster than Carbon with near-Carbon level performance and intelligence. We’ve been building so much with it and think you’ll love it. Rolling out across AI Studio, Gemini app, and Antigravity. Shipmas came early :)
28
18
2
499
16,067
Genie 4 is now available in Project Genie. It brings native audio, real-time editing, and a new agent to help you create interactive, virtual worlds.
231
235
192
2,441
218,624
clainstone retweeted
I tried it out. Sound on 🔊 RTX 5090, Nemotron 3.5 Transcribe: 59:39 of audio in 25.0 s, the whole hour in one request I ran the same hour on three machines. Turn the sound on. First you hear a meeting recording, and it tells four speakers apart. Then the three machines race. The 5090 takes the whole hour as one request. Word error rate there is 1.9 %. That's what NVIDIA's defaults are made for on 32 GB. For the race I cut the hour into 31 pieces. A 16 GB card can't take a full hour at once. 31 pieces, 59:39 of audio RTX 5090, 32 GB: 12.9 s Mac Studio M5 Ultra, 256 GB: 57.3 s RTX 5060 Ti, 16 GB: 77.4 s Word error rate: 1.5 % on all three In October the same 16 GB card needed 120 s for this hour, on an older engine build. Each machine runs a different quant, and the engine doesn't run the same way on all three. So it's not a pure hardware comparison. Quants: AD-NVFP4 on the 5090, AD-IQ4_NL on the Mac, AD-Q2_0_M on the 16 GB card, all by AtomicChat Method: wall time with model load, LibriSpeech test-clean, one run on the 5090 and the Mac, the better of two on the 16 GB card huggingface.co/AtomicChat/NV… Nemotron 3.5 Transcribe and NeMo-Speech.cpp by NVIDIA, quants and the 16 GB settings by AtomicChat @atomic_chat_hq.
Wow 😮 NVIDIA's Nemotron 3.5 Transcribe is the #1 trending model on HuggingFace.
3
2
18
1,015
clainstone retweeted
🚨 FABLE 5.5 Update: Drops this week. Anthropic's next Fable-class model has potentially surfaced as Silicon 2 inside Devin Build, carrying the old $10/$50 placeholder pricing. Other signals suggest partner testing is already happening, with early expectations pointing to significant gains in 3D modeling, SVG generation and web development. With Anthropic's IPO approaching, this week looks like an interesting opportunity to make a major move.
39
73
47
1,334
121,935
clainstone retweeted
Gemini 4 Hydrogen is live in Google Search globally 🚀 Our biggest model upgrade to Search yet brings: 🔹 Better understanding of your intent 🔹 Smarter query fan-out and linking 🔹 More visual, adaptive responses 🔹 Richer multilingual capabilities Swipe to see a full breakdown ⬇️.
135
239
53
3,530
277,624
clainstone retweeted
CONFIRMED: Gemini 4 Neon Releasing Today! Pricing: Input: $3 Output: $15 Cached input: $0.15 That's 25% cheaper than Gemini 4 Carbon
42
49
114
682
56,358
clainstone retweeted
if you’re running a networked Java application turn it off
166
181
40
3,953
178,758
clainstone retweeted
Introducing Qwen3.8-27B-pi-adapted 🤗 We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter. The results, measured as pass@1: 🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7% 💻 Terminal-Bench 2.1: 39.5% → 55.8% ⚡ 55–62% fewer tool calls per task This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛 Open weights: huggingface.co/harness-hub/Q… Methodology: huggingface.co/blog/harness-…
303
643
172
5,987
449,512
clainstone retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me. I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA. I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation. Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO. Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours. Jensen
60,713
121,650
24,466
1,172,964
58,244,832
clainstone retweeted
This is f*cking insane. World's top mathematician Terence Tao, at a workshop for his own AI math tool: "I would not just one-shot the proof... that doesn't teach you much of anything" In 48 seconds he explains how he actually uses AI on hard problems, and why most people are using it backwards One-Shot → Learn Nothing → Pen and Paper → AI Cleans Up His method: break the problem into pieces, understand each one yourself, then put it back together. "A lot of this you should be doing at a whiteboard" AI gets the nasty computation you don't need to understand. The thinking stays with you And he said it while showing off the tool. Tao is co-founder and chairman of SAIR, the company that built it A study in Turkish high schools found the same thing the hard way: students with unrestricted GPT-4 did 48% better on practice problems, then 17% worse on the exam once the AI was gone "Knowing when not to use them is part of the challenge" Everything below is free: • Tao's blog, where he logs his own AI experiments: terrytao.wordpress.com • Learn Lean, the proof language his tool runs on, through a free game: adam.math.hhu.de • Anthropic's free AI Fluency course: anthropic.skilljar.com Then read how the people building AI use it without letting it think for them, article below ↓
Karpathy says 75% of people have barely touched AI and only 0.2% have seen what it can really do here's what the people building it do that everyone else skips
13
47
2
386
87,383
⚡ Faster simulation means faster discovery. Introducing NVIDIA Rubin HPC, a new GPU in the NVIDIA Vera Rubin platform designed to accelerate simulation and data processing for science, engineering and analytics workflows. Delivering 3x the performance of NVIDIA Blackwell and with more than 50,000 CUDA cores, it accelerates the path from exploration to scientific discovery and industrial design. #SC26 🔗 nvda.ws/4C78FqM
75
60
10
573
28,044
d1-2B by @liquidai :Same model, three engines on a MacBook Pro M5 Max - MLX: 66 ms, 153 W - Core AI GPU: 67 ms, 120 W - Core AI Neural Engine: 197 ms, 37 W The Neural Engine is 3x slower (on M5 Max, I bet this is different on M4) and draws a quarter of the power. Same answers on all three. I had to use a video instead of live camera feed because ANE usage goes up as soon as you start the camera. Core AI export from the original weights, plus live GPU / Neural Engine / power charts: github.com/ivanfioravanti/ml… Download and have fun! Share your stats on other Apple Silicon machines! I'm now creating an iPhone version 😎
31
21
3
309
20,188
I did a PhD without the Internet.
"I graduated college without ChatGPT" is the new "I walked 5 miles to school"
303
4,228
336
56,127
1,044,209