Founder at Jovyan AI - making ML teams 10x more productive
Paris, France
Joined September 2013
- Tweets950
- Following445
- Followers114
- Likes6.7K
Lam Dang retweeted
You already know Jev is fast and cheap. Our new arXiv draft shows, it makes the correlated mistakes as other flash-tier LLMs. The consequence: you don't gain much by cascading Jev's "low confidence" outcomes to an LLM! Explanation in thread. arxiv.org/abs/2609.29769
Lam Dang retweeted
Everyone is benchmarking Jev against GPT-5.6.
That's the easy comparison. Jev isn't an LLM, of course it's faster.
So I benchmarked it against the boring open-weight encoders found on @huggingface that have done zero-shot classification for years.
Jev wins on accuracy. It loses at the one thing it's named for. 🧵
Lam Dang retweeted
- Take a past Kaggle competition with text classification
- Compare your model with similar size LLM single token prediction
Otherwise it is just ads
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Lam Dang retweeted
I’ve been working towards AGI my whole life, and as we enter this pivotal moment, I’m stepping into a new role as Chair of Google DeepMind & Chief Scientist of Alphabet. This will allow me to focus on long-term strategy, and accelerating scientific breakthroughs, including leaning into my work at Isomorphic to help cure disease.
I’m excited that @koraykv will be stepping up to lead GDM as SVP, alongside @joshwoodward and our exec team. I could not be more excited and confident about our amazing next chapter! 🚀
blog.google/company-news/ins…
If you are feeling claude code is feeling dumper than before: add this to .claude/settings.json
```
"model": "opus",
"effortLevel": "high"
```
Also to avoid draining usage too fast, this work out quite well in CLAUDE.md
```
## Subagent Execution
When executing a plan that involves changes across multiple files, break it down into logical chunks of work.
Launch a Sonnet subagent to handle each chunk, each agent should ensure the code is tested and working .
Afterward, review the full implementation once more to ensure nothing was missed and no mistakes remain.
```
I fell into a rabbit hole reading the writeups from Kaggle’s Deep Past Challenge, where teams had to translate Old Assyrian into English.
Main takeaway: data quality beats modeling tricks.
Tiny official train set (1.5k pairs), messy OCR from academic PDFs, train/test mismatch, unstable leaderboard.
The best teams mostly won by:
-> rebuilding sentence pairs from PDFs
-> cleaning + normalizing hard
-> using byte-level models like ByT5
-> using LLMs for extraction, alignment, filtering, repair
being conservative at inference
Not much architectural magic.
Mostly better data.
Honestly it feels closer to real ML work than many Kaggle comps.
I wrote a longer breakdown here: open.substack.com/pub/jovyan…
Data cleaning is the number 1 pain point in data science. But do we agree on what it is? Why is it hard? And when it is necessary ?
Deep dive in this blog post
open.substack.com/pub/jovyan…
Lam Dang retweeted
How @cognition /@windsurf new SWE-1.5 model was built (probably), based on piecing together bits of information that was shared.
Base model by: @Zai_org possibly GLM-4.6
RL Training: @nvidia on 'thousands' GB200 NVL72
Inference: @cerebras at 950 tks/sec
Lam Dang retweeted
Ever wonder what it actually takes to train a frontier AI model?
YC's Ankit Gupta (@agupta) sat down with Nick Joseph (@nickevanjoseph), Anthropic's Head of Pre-training, to explore the engineering challenges behind training Claude—from managing thousands of GPUs and debugging cursed bugs to balancing compute between pre-training and RL.
They cover scaling laws, data strategies, team composition, and why the hardest problems in AI are often infrastructure problems, not ML problems.
01:05 – From Vicarious to OpenAI to Anthropic
06:40 – What pretraining is
11:20 – Why next-word prediction won out
16:05 – Scaling laws and the feedback loop of compute → models → revenue
21:50 – Building Anthropic’s early infrastructure
27:35 – Efficiency hacks and debugging at scale
33:10 – Generalists vs. specialists on the pretraining team
38:45 – Challenges of training across thousands of GPUs
44:15 – Working with new chips: GPUs vs. TPUs
49:00 – Pretraining vs. post-training (RLHF and reasoning models)
54:25 – The future of data quality and availability
59:10 – Where pretraining goes next
Lam Dang retweeted
As a researcher at a frontier lab I’m often surprised by how unaware of current AI progress public discussions are.
I wrote a post to summarize studies of recent progress, and what we should expect in the next 1-2 years:
julian.ac/blog/2025/09/27/fa…
Lam Dang retweeted
We were able to reproduce the strong findings of the HRM paper on ARC-AGI-1.
Further, we ran a series of ablation experiments to get to the bottom of what's behind it.
Key findings:
1. The HRM model architecture itself (the centerpiece of the paper) is not an important factor.
2. The outer refinement loop (barely mentioned in the paper) is the main driver of performance.
3. Cross-task transfer learning is not very helpful. What matters is training on the tasks you will test on.
4. You can use much fewer data augmentations, especially at inference time.
Finding 2 & 3 mean that this approach is a case of *zero-pretraining test-time training*, similar to the recently published "ARC-AGI without pretraining" paper by Liao et al.
Recently I asked data scientists on reddit : What do they hates the most in data science work. Here is the answer 🙈.
And you, what do you dislike in your day-to-day work?
In the last two weeks, I committed nearly 10,000 lines of code.
That might not seem impressive for a seasoned developer — but I'm not a developer. I'm a data scientist.
And most of what I've built, I've done for the very first time:
🧩 A ReactJS extension integrating Jovyan AI in Jupyter Notebook cells
🌐 A NextJS-powered website to launch the Jovyan AI waitlist
🔐 An authentication system with Supabase to secure Jovyan AI's backend (Python & JS)
Sure, my code isn't "perfect" by veteran developer standards — but it's clean, functional, and maintainable.
⚡ This is the incredible power of AI: One person can now seamlessly handle:
UX & UI design
Architecture with technical and security specifications
Development and testing
Launch & marketing
We are truly living in the golden age for builders and entrepreneurs.
What once required a full product team can now be achieved by just two or three people — faster, leaner, and more agile.
Exciting times ahead! ✨
#AI #ProductivityTips #DataScientists #buildinpublic #Entrepreneurship #SoftwareEngineering
🚀 Jovyan AI is live! 🚀
Two weeks ago, we quietly launched jovyan-ai.com and were thrilled to see a few dozen early users sign up to test it out. 🙌
Today, we’re officially online with the first release of Jovyan AI — your copilot for data science and machine learning, built directly into Jupyter Notebooks.
You can generate, edit, and debug code right inside each cell — no more copy-pasting into a separate interface.
✨ What’s unique about it?
Unlike other tools, Jovyan AI captures the entire notebook context, including inputs, outputs, charts, and images.
🧠 It’s not fully agentic yet, but we’re pushing the boundaries one cell at a time.
This is just the beginning — a small step, but one that already makes Jupyter more powerful and more joyful to use.
👩💻 If you’re using Jupyter for work, research, or side projects, join our waitlist — we’re expanding access soon! We are currently giving Gemini-2.5 Pro access for FREE for our beta testers (limited seats).
🔗 jovyan-ai.com
Lam Dang retweeted
Go go go @pyannoteAI, let's solve speaker diarization together!
DMs are open!
linkedin.com/posts/pyannotea…
🧑💻Which AI coding agents work best for data science & ML in Jupyter?
I tested them all:
❌ Cursor, Windsurf, Cline: No Jupyter support
😕 Gemini in Colab: Disappointing
⚠️ GitHub Copilot : Works, but limited
My surprising pick? Full analysis here 👉 tinyurl.com/ybb39xy
🧠💻 I recently realized something frustrating: Most coding assistants just aren’t built for data science and machine learning. They often: ⚠️ Generate tons of code before even looking at the data ⚠️ Struggle to work inside notebooks ⚠️ Focus on code logic but not the data
🧵1/3
So with @jonathan_ttu , we decided to fix that.
🚀 Meet Jovyan AI — your AI teammate inside your Jupyter Notebook.
✅ Fully integrated — no need to switch environments
✅ Built specifically for DS & ML workflows
✅ Helps you explore, reason, and build — with you in control
2/3
📍 Check it out and register here: jovyan-ai.com
Demo: youtu.be/q8eQyi-1B1k
We’d love your feedback — and we’re just getting started. 😉