iislucas (Lucas Dixon) (@iislucas) header
iislucas (Lucas Dixon) (@iislucas) profile picture

machines learn, graphs reason, identity is a non-identity, incompetence over conspiracy, evil by association is evil, expression is never free, stay curious

Paris
Joined January 2010
iislucas (Lucas Dixon) retweeted
Gemma 4 just hit 200M downloads in only 2.5 months! For context, total downloads across the entire Gemma family of models were at 100M when we launched Gemma 3. The community's acceleration is incredible. Thank you to everyone building with Gemma. Watch how developers are driving real-world impact:
106
173
37
1,714
229,207
iislucas (Lucas Dixon) retweeted
Diffusion language model speed is unreal. Here is DiffusionGemma vs Claude Haiku vs GPT 5.4-mini (low reasoning) racing to generate matplotlib figures. The models were prompted to generate Julia-set fractals, magnetic-pendulum basins, blackbody curves, Chladni plate modes, and hydrogen-orbital grids.
Introducing DiffusionGemma, our first exploration with open diffusion text generation models 🔥Generate blocks of text at a time 🤏26B MoE built on top of Gemma 4 ⚡️Up to 4x faster in popular consumer GPUs 🤗Apache 2.0 Excited to see what the community builds with it!
1
7
1
21
2,393
iislucas (Lucas Dixon) retweeted
Meet DiffusionGemma! An experimental open model that explores a fast approach to text generation, released under an Apache 2.0 license. Moving beyond sequential, token-by-token processes to generate entire blocks of text simultaneously. Here’s what’s new with DiffusionGemma: 👇
161
795
273
4,950
1,005,874
iislucas (Lucas Dixon) retweeted
Introducing Magenta RealTime 2 (MRT2): the live music model you can play as an instrument. MRT2 offers MIDI and prompt controls, and runs natively on a MacBook with <200ms latency. Open weights. Open source inference engine. Suite of apps and plugins. Hear what it can do and try it out for yourself below 🧵
22
92
34
500
114,877
iislucas (Lucas Dixon) retweeted
[1/5] Intelligent behavior relies on maintaining an evolving, dynamic model of the environment. But even frontier models sometimes fail at tracking state in dialogs, e.g., actual transcript from 4/20/2026:
4
12
4
64
10,596
iislucas (Lucas Dixon) retweeted
We're studying cubic surfaces at @GoogleDeepMind! Our first paper, among other things, resolves a 54-year-old arithmetic geometry question of Manin's; one attempted by Swinnerton-Dyer and the first author through the decades. A 🧵 about X³ + Y³ + Z³ + ζ₃T³ = 0 and AI for math:
12
80
4
453
65,571
iislucas (Lucas Dixon) retweeted
🏎️ gemma-webgpu: a zero-dependency, blazing fast Gemma 1B running entirely in your browser. Full vibe coded from my cell phone. 🔥 136.8 tok/s on M4 Mac (3.3x faster than transformers.js) 📱 101 tok/s on iPhone 17 (270M), 34 tok/s (1B) What we built from scratch: • 18 hand-written WGSL compute shaders with fused ops (fusedNormAdd saves 36 GPU dispatches per forward pass) • Q8_0 dequantization directly on GPU — higher quality than q4 AND faster • Range request streaming loads weights layer-by-layer (~44MB chunks), uploads to GPU, frees JS memory immediately. Peak heap: ~50MB even for the 1GB model • That streaming trick is what makes 1B run on iPhone. it never holds the full model in RAM 12KB gzipped. Zero dependencies. npm install gemma-webgpu
10
6
4
69
8,818
iislucas (Lucas Dixon) retweeted
I wanted to share something I built over the last few weeks: isometric.nyc is a massive isometric pixel art map of NYC, built with nano banana and coding agents. I didn't write a single line of code.
365
655
256
8,036
1,730,370
iislucas (Lucas Dixon) retweeted
Gemini Nano Banana Pro can solve exam questions *in* the exam page image. With doodles, diagrams, all that. ChatGPT thinks these solutions are all correct except Se_2P_2 should be "diselenium diphosphide" and a spelling mistake (should be "thiocyanic acid" not "thoicyanic") :O
319
914
244
8,729
3,020,941
iislucas (Lucas Dixon) retweeted
Mapping LLMs with Sparse Autoencoders pair.withgoogle.com/explorab… An interactive introduction to Sparse Autoencoders and their use cases with Nada Hussein, @shivamravalxai, Jimbo Wilson, Ari Alberich, @NeelNanda5, @iislucas, and @Nithum
8
56
1
500
44,451
iislucas (Lucas Dixon) retweeted
AI is everywhere, but the most impactful features start with a real user need. Ask how AI can add unique value to that experience, not just how to release an AI feature. Here are three questions to consider before you start building. 1️⃣ What’s the core user problem? 2️⃣ Does AI offer a unique solution? 3️⃣ What does AI take away from the user? Learn more: goo.gle/PAIR
22
17
3
82
10,432
iislucas (Lucas Dixon) retweeted
my coworkers have been subjected to this yap so I now continue here: this is genuinely one of the most useful papers to engage with for interp imo. easily one of my most referred to papers. if you have not read it, pls do. mandatory reading until the heat death of the universe.
6
23
1
396
28,051
iislucas (Lucas Dixon) retweeted
Veo holograms 🦝⚡️ Visualizing animal superpowers! Just discovered Veo 3's amazing ability to render 3d holograms. Virtual interfaces within the simulated world. 🔊 Prompts in 🧵
2
4
1
19
4,848
iislucas (Lucas Dixon) retweeted
We are hiring Applied Interpretability researchers on the GDM Mech Interp Team!🧵 If interpretability is ever going to be useful, we need it to be applied at the frontier. Come work with @NeelNanda5, the @GoogleDeepMind AGI Safety team, and me: apply by 28th February as a Research Engineer or Scientist
2
36
6
274
52,936
iislucas (Lucas Dixon) retweeted
We scaled training data attribution (TDA) methods ~1000x to find influential pretraining examples for thousands of queries in an 8B-parameter LLM over the entire 160B-token C4 corpus! medium.com/people-ai-researc…
1
17
7
123
27,202
iislucas (Lucas Dixon) retweeted
What a way to celebrate one year of incredible Gemini progress -- #1🥇across the board on overall ranking, as well as on hard prompts, coding, math, instruction following, and more, including with style control on. Thanks to the hard work of everyone in the Gemini team and elsewhere at Google! 🎊
Big news on Chatbot Arena 🔥 The new @GoogleDeepMind model gemini-exp-1206 is crushing it, and the race is heating up. Google is back in the #1 spot 🏆overall and tied with O1 for the top coding model! Highlights (improvement since gemini-exp-1121 in parentheses) - First place overall (2->1) - Tied with GPT-4o-1120 after style control (4->1) - Tied with O1 on coding leaderboard (3->1) - First place on hard prompts (2->1) Keep it up @GoogleDeepMind! The rate of progress is crazy. For analysis and to test the model, see below 👇
88
306
113
1,567
784,036
iislucas (Lucas Dixon) retweeted
🚨 New Paper 🚨 Can LLMs perform latent multi-hop reasoning without exploiting shortcuts? We find the answer is yes – they can recall and compose facts not seen together in training or guessing the answer, but success greatly depends on the type of the bridge entity (80%+ for country, 6% for year)! 1/N
7
51
8
205
47,529
iislucas (Lucas Dixon) retweeted
I’m so proud of the updated version of #MusicFXDJ we developed in collaboration with @jacobcollier, available today at labs.google/musicfx. Over the past year I’ve spent countless hours experimenting with our real-time music models, and it feels like I’ve learned to play a new type of instrument, in the spirit of what we started @GoogleMagenta to achieve. It was validating to hear Jacob echo these sentiments the first time he played an early version. With his help, we’ve added new controls, raising the ceiling on what players are able to do with it. Now anyone can experiment with this new instrument and share what they create with the world. I'll start things off with my most recent jam session...
It’s here! Thrilled to collab with @jacobcollier on our latest #LabSession exploring the magical possibilities of generative music and #MusicFXDJ. Watch full video: youtu.be/y7gKlzvg8xk Tune in 10/24 at 5 PM ET as Jacob livestreams a MusicFX DJ sesh on his YT channel.
7
34
10
190
42,592
iislucas (Lucas Dixon) retweeted
🧵Responses to adversarial queries can still remain latent in a safety-tuned model. Why are they revealed sometimes, but not others? And what are the mechanics of this latent misalignment? Does it matter *who* the user is? (1/n)
2
12
1
62
13,035
iislucas (Lucas Dixon) retweeted
We’re welcoming a new 2 billion parameter model to the Gemma 2 family. 🛠️ It offers best-in-class performance for its size and can run efficiently on a wide range of hardware. Developers can get started with 2B today → dpmd.ai/4d0MKEH
34
290
84
1,587
489,901