Charly Wargnier ♨️ (@DataChaz) header
Charly Wargnier ♨️ (@DataChaz) profile picture

Deep in Agents, LLMs and automation · Open source at ♥ · Sharing AI gems daily to help you work smarter · Ex-data scientist @Snowflake | @Streamlit ·

📩 DM for Collabs
Joined January 2009
Photo by Charly Wargnier ♨️ (@DataChaz)Photo by Charly Wargnier ♨️ (@DataChaz)Photo by Charly Wargnier ♨️ (@DataChaz)Photo by Charly Wargnier ♨️ (@DataChaz)Photo by Charly Wargnier ♨️ (@DataChaz)Photo by Charly Wargnier ♨️ (@DataChaz)Photo by Charly Wargnier ♨️ (@DataChaz)Photo by Charly Wargnier ♨️ (@DataChaz)Photo by Charly Wargnier ♨️ (@DataChaz)Photo by Charly Wargnier ♨️ (@DataChaz)
❌ messy documents in. ✅ reliable, AI-ready data out. @UnsiloedHQ uses computer vision and dual-stream models to make it happen. ..and it just raised $7.5M from @felicis ↓
We’ve raised a $7.5M seed round led by Felicis so AI can read documents as easily as it reads databases. Finance, law and healthcare run on documents, and not databases like any other industry. Soon AI agents will read and write more documents than humans, and they’ll need to query them the way software queries tables. We’re building the computable layer that makes this possible. To do this at scale, we've brought together ML researchers and engineers from Google DeepMind, Cognition, and quant trading firms. We use computer vision and dual-stream models to turn complex multimodal documents into formats LLMs and AI agents can understand and reason over. If your team wants to turn messy documents into reliable, AI-ready data, book a demo: unsiloed.ai/book-demo
13
28
995
Charly Wargnier ♨️ retweeted
Okay, this is so cool @nvidia has just dropped a 4GB model designed for local voice assistants. Nemotron 3.5 Bolt E4B combines conversational AI with multi-turn tool use. The focus? Respond quickly, use less memory, and leave space for the application running alongside it. think voice interfaces, interactive characters, and embedded assistants. Commercial and non-commercial use supported.
17
4
33
4,108
Charly Wargnier ♨️ retweeted
❌ messy documents in. ✅ reliable, AI-ready data out. @UnsiloedHQ uses computer vision and dual-stream models to make it happen. ..and it just raised $7.5M from @felicis ↓
We’ve raised a $7.5M seed round led by Felicis so AI can read documents as easily as it reads databases. Finance, law and healthcare run on documents, and not databases like any other industry. Soon AI agents will read and write more documents than humans, and they’ll need to query them the way software queries tables. We’re building the computable layer that makes this possible. To do this at scale, we've brought together ML researchers and engineers from Google DeepMind, Cognition, and quant trading firms. We use computer vision and dual-stream models to turn complex multimodal documents into formats LLMs and AI agents can understand and reason over. If your team wants to turn messy documents into reliable, AI-ready data, book a demo: unsiloed.ai/book-demo
13
28
995
Charly Wargnier ♨️ retweeted
AI notetakers had one blind spot. the meeting room. @WisprFlow's new Notetaker just fixed that with a free iPhone app 👊
🤝 Paid partnership
Notetaker is live today on iPhone. Now, your meeting notes can go wherever work takes you, for free. Use Notetaker on desktop for virtual meetings and on iPhone when you’re meeting in person. Your notes stay together across devices, so you can pick up right where you left off.
1
12
33
1,469
AI notetakers had one blind spot. the meeting room. @WisprFlow's new Notetaker just fixed that with a free iPhone app 👊
🤝 Paid partnership
Notetaker is live today on iPhone. Now, your meeting notes can go wherever work takes you, for free. Use Notetaker on desktop for virtual meetings and on iPhone when you’re meeting in person. Your notes stay together across devices, so you can pick up right where you left off.
1
12
33
1,469
Charly Wargnier ♨️ retweeted
you don't need to learn agent frameworks this weekend. you just need this prompt and Claude Fable 5.2. thank me later.
12
5
65
7,098
Charly Wargnier ♨️ retweeted
.@ComposalAI just launched Sparks 🔥🔥🔥 multiplayer agents. with a cloud computer for each one. they run on the Claude, Codex or Grok subscription you already pay for. smart stuff.
Today we're launching Sparks: multiplayer agents that orchestrate your entire software factory. Each one gets its own cloud computer and runs on the Claude, Codex or Grok subscription you already pay for. And they coordinate amongst themselves: one catches a production error reproduced it hands it off to a dedicated reviewer and merges the fix.
3
6
32
2,722
Charly Wargnier ♨️ retweeted
a 0.5B speech recognition model you can run on your phone. Apache 2.0. 45 languages. Reson8 just open-sourced Rumble-2-Flash, and they say it beats models 20x its size.
Rumble-2-flash is open weights you can get it in all your favorite places. It’s better than models 20x it’s size small enough to run on your phone HF link below
9
14
38
3,316
.@ComposalAI just launched Sparks 🔥🔥🔥 multiplayer agents. with a cloud computer for each one. they run on the Claude, Codex or Grok subscription you already pay for. smart stuff.
Today we're launching Sparks: multiplayer agents that orchestrate your entire software factory. Each one gets its own cloud computer and runs on the Claude, Codex or Grok subscription you already pay for. And they coordinate amongst themselves: one catches a production error reproduced it hands it off to a dedicated reviewer and merges the fix.
3
6
32
2,722
a 0.5B speech recognition model you can run on your phone. Apache 2.0. 45 languages. Reson8 just open-sourced Rumble-2-Flash, and they say it beats models 20x its size.
Rumble-2-flash is open weights you can get it in all your favorite places. It’s better than models 20x it’s size small enough to run on your phone HF link below
9
14
38
3,316
Charly Wargnier ♨️ retweeted
3 ten-second videos. 228 million views. @maubaron just broke down the whole system with Claude Fable 5.2 🤯
3
13
33
3,076
Charly Wargnier ♨️ retweeted
THIS IS HUGE Google researchers just built an AI that learns while it “sleeps.” It updates its weights, stores new knowledge and practises on its own generated data. That tackles a stubborn problem. Training AI on new information can damage what it already knows. Bigger context windows help during a conversation, but they don’t automatically turn that information into lasting knowledge. Google Research’s Sleep framework splits learning into two phases: > Wake: Process inputs and collect short-term memories. > Sleep: Consolidate those memories into long-term parameters while preserving earlier knowledge. Then comes “dreaming.” The model generates synthetic examples, selects useful ones and trains on them to sharpen its capabilities. The reported results: • Near-perfect scaling on BABILong tests up to 10 million tokens. • Near-zero performance loss in sequential translation learning. • 79.2% on AIME 2024 and 69% on AIME 2025. “Sleep” is a training process inspired by biology. But the idea behind it is compelling. We keep giving AI more room to hold information. This research asks what happens when we also give it time to process that information. The next leap could happen after the conversation ends :)
14
8
34
4,096
Charly Wargnier ♨️ retweeted
this is a must-bookmark. Your next startup could start with this list. Every idea the @ycombinator team shared recently, curated here:
3
2
8
2,062
Charly Wargnier ♨️ retweeted
.@sama after watching old GPT-Sol crushing the new one
Replying to @ArtificialAnlys
GPT-6.2 Astra is flat with GPT-6 Astra, and GPT-6.2 Sol loses 3 points relative to GPT-6.1 Sol in the Artificial Analysis Coding Agent Index (Codex harness at max effort).
4
1
18
3,428
Charly Wargnier ♨️ retweeted
freelancers realizing they have till midnight to land clients before GTA VI takes everyone offline for a month:
4
1
11
2,037
Charly Wargnier ♨️ retweeted
the perks you can get with a student ID are just ridiculous.. if you're into tech, do yourself a favour and check them out:
8
4
26
4,151
Charly Wargnier ♨️ retweeted
Anthropic valuation: $1.4 trillion Samsung valuation: $935 billion Anthropic 2025 revenue: $4.6 billion Samsung 2025 revenue: $236 billion
11
6
54
7,245