Deepnote

Hello New York! I'm in town today & tomorrow. If you’re around and want to catch up, DM me and we’ll find a time.

Hello New York! I'm in town today & tomorrow. If you’re around and want to catch up, DM me and we’ll find a time.

At Deepnote, product fluency is not optional. We use Deepnote daily, but it’s still not enough. Too easy for people to lose touch with the product or never try the features they never worked on directly. Every new hire completes a real task in Deepnote during onboarding. Sales, ops, finance, design, everyone. With Deepnote Agent, even non-technical roles can complete the entire process without prior notebook experience. Then, we do regular dogfooding challenges. Every few weeks, the whole company gets 60 minutes to solve a real customer problem in Deepnote, then write up what worked and what broke. Strategy is built on first-hand context. When everyone understands the workflow, the collaboration layer, and the pain points, decision quality improves dramatically.

Thinking models took the lead in real coding work. Sovereign AI became infrastructure policy. Memory, not tokens, is the new bottleneck. Inside this issue: 🔵 LMArena: Opus 4.5 Thinking tops coding board; long context + deliberate reasoning keep compounding 🔵 Sovereign AI: governments lock in GPUs, power, and land; commitments of $300B+ in AI infra and compute 🔵 NVIDIA × Groq licensing: consolidation of inference IP and talent 🔵 Agents grow up: RLMs, BLOOM & MAI-UI bring persistent memory, reliable tool use, and device–cloud routing 🔵 Small and efficient: LiquidAI LFM2 2.6B and IQuest Coder 40B lead product-friendly stacks Read the full rundown below.

Great products are built by teams who use them every day. At Deepnote, that starts before day one. Sales, ops, finance, design — every new hire completes a real task in Deepnote as a part of their onboarding. The best part: with Deepnote Agent, even non-technical roles can finish the whole thing, without any prior notebook experience. They ask a question, the agent plans the steps, writes SQL or Python, and they just review the results. That’s how they learn the workflow without getting blocked on setup. Dogfooding became our company ritual. Every few weeks, we have a new company-wide challenge that mirrors real customer work. Everyone solves it in Deepnote, then writes up what worked, what broke, and what needs to change. Why do we do it: - Build user intuition fast. Nothing sharpens product judgment like wrestling with the same tasks our customers do. - Surface real-world bugs early. A thousand synthetic tests can’t match a marketer clicking the wrong button and finding an edge case. - Keep shipping empathy. When the whole company feels the papercuts, fixes ship in the same sprint. Want a great product culture? Make everyone a user.

Open-sourcing Deepnote was our biggest bet of 2025, and… It paid off. We took the leap and released Deepnote under Apache 2.0. The response crushed our expectations. - The repo became one of the fastest-growing notebooks on GitHub with 2,500+ stars. - On the enterprise front, we delivered a lot, and it shows on our G2 customer reviews. - We closed some of our largest enterprise deals and built the biggest pipeline in our history. Thank you to everyone who helped build Deepnote this year. To the entire Deepnote team, the open-source contributors, our trusted partners, and the customers who put their critical data work in our hands.

Great products are built by teams who use them every day. At Deepnote that starts before day one. Sales, ops, finance, design — every new hire completes a real task in Deepnote as a part of their onboarding. The best part: with Deepnote Agent, even non-technical roles can finish the whole thing, without any prior notebook experience. They ask a question, the agent plans the steps, writes SQL or Python, and they just review the results. That’s how they learn the workflow without getting blocked on setup. Dogfooding became our company ritual. Every few weeks, we have a new company-wide challenge that mirrors real customer work. Everyone solves it in Deepnote, then writes up what worked, what broke, and what needs to change. Why do we do it: - Build user intuition fast. Nothing sharpens product judgment like wrestling with the same tasks our customers do. - Surface real-world bugs early. A thousand synthetic tests can’t match a marketer clicking the wrong button and finding an edge case. - Keep shipping empathy. When the whole company feels the papercuts, fixes ship in the same sprint. Want a great product culture? Make everyone a user.

OpenAI is cornering the world’s RAM. Mistral AI walked out of NeurIPS with the trophy. AI data centers are catching up to roads and bridges. Inside this issue: 🔵 GPT-5.2 reclaims the frontier: first to match/beat human experts on 70.9% of professional knowledge tasks 🔵 Gemini 3 Deep Research: an autonomous web agent that plans, reads, and revises 🔵 NeurIPS winner: Mistral’s self-supervised RL work takes Best Paper honors 🔵 Mistral OCR 3: document intelligence that undercuts incumbents by up to 97% 🔵 Infra squeeze: private AI data-center spend (~$41B/yr) nears U.S. transport infra budgets Catch the full rundown below.

To people who apply to Deepnote without ever opening Deepnote: Why? The quality of the product is the strongest signal you can get about the quality of the company. Also, it’s the strongest signal I can get about you as an applicant. We do not write this in the job description, but I check every time whether you have tried Deepnote before the first interview. I need to know whether you care about what you’re going to be working on. It is a silent hard requirement. For every single role.

Last fortnight in AI was eventful: here's what you might have missed. DeepSeek V3.2-Speciale achieves competition gold medals: High-compute variant wins IMO 2025 (35/42), IOI 2025 (492/600, ranked 10th), ICPC World Finals (10/12, ranked 2nd), while base V3.2 hits 93.1% AIME 2025 and 73.1% SWE-Verified with DeepSeek Sparse Attention cutting 128K prefill costs from ~$0.65 to ~$0.35/M tokens. Mistral 3 family goes fully open-source under Apache 2.0: Mistral Large 3 (675B MoE, 41B active) ranks #2 on LMArena OSS with best-in-class multilingual; Ministral 14B reasoning variant hits 85% on AIME '25; NVFP4 checkpoint enables Large 3 on a single 8×H100 node. State of AI 2025: OpenAI reports 8× enterprise usage growth, 320× rise in reasoning-token consumption, and 40–60 min daily time savings, with frontier firms 2× more active per seat. Anthropic finds Claude cuts task time ~80%, implying a 1.8 pp boost to annual labor-productivity growth, with 65% of professionals using AI for augmentation. OpenRouter logs 100T tokens across 300+ models, with mid-sized reasoning models, open-source share near one-third, and agentic tool-calling now exceeding half of total use. Claude Code hits $1B ARR in 6 months as Anthropic acquires Bun: First acquisition brings in JavaScript runtime (7M+ monthly downloads) powering infrastructure; Harvey raises $160M at $8B for legal AI; Eon raises $300M at $4B to turn cloud backups into AI data lakes. Google Titans architecture enables 2M+ token contexts: Neural long-term memory module uses "surprise metric" to selectively store high-gradient information in real-time without retraining, outperforming GPT-4 on BABILong benchmark with fewer parameters. OpenAI declares "Code Red" amid Gemini 3 surge: Altman's internal memo marshals resources toward ChatGPT, delays advertising plans, and announces new reasoning model beating Gemini 3 internally—while OpenAI faces researcher exodus to Thinking Machines and Meta's Superintelligence Labs despite 800M weekly active users.

I just swapped my Jupyter workflow for Deepnote, now open source and fully Apache 2.0. Think of Deepnote as Jupyter, but with SQL, AI, and multiplayer mode, among other handy features. The setup is simple: convert with `npx @deepnote/convert`, then open your .deepnote project in a VS Code-based IDE or JupyterLab. I get 23 new interactive block types, 100+ data connectors, write SQL, Python, or R code, all powered by a human-readable YAML format that’s easy to diff. If I want to work with someone, I can scale from local development to cloud, collaborate in real time in Deepnote Cloud on beefier machines, or deploy notebooks as data apps. Check out Deepnote OSS and support them with a star: https://lnkd.in/gv44DMng

Deepnote goes open source under the Apache 2.0 license! Most notebooks fail at collaboration. Jupyter creates messy diffs, there are no native data connectors or SQL, sharing breaks environments, and “it works on my machine” blocks every deployment. Deepnote reimagines notebooks with: → 23 new block types, → 100+ native data connectors, → and works in VS Code, Cursor, or Antigravity. Built on a customized kernel for full compatibility with Jupyter. How it works: → Convert any Jupyter notebook: npx @deepnote/convert notebook.ipynb → Run locally in your favorite IDE or JupyterLab → Scale in Deepnote Cloud for real-time collaboration with AI agents → Use reactive execution in the cloud (downstream blocks update automatically) → Deploy data apps and dashboards without environment issues Used by 500,000+ data professionals. Check out the repo: https://lnkd.in/dZEqhtXR

AI lowered the barrier to questions. Notebooks raise the bar on what agents can do. AI made it trivial to ask things like "Plot lead sources by volume?" Cute chart, quick dopamine. Real work behind the scenes is ugly: schemas, joins across warehouses, tests, reviews, and runs that still work next week. That work lives in notebooks: code, SQL, charts, and context in one place, as an artifact you can diff, schedule, and audit. AI is strongest inside that environment, where an agent can read tables, propose a plan, run multi-step Python & SQL commands, and trigger a scheduled workflow. Things that traditional BI can’t do.

FDE is a meme, being hands-on with customers is table stakes. I’m not bullish on startups that hire FDEs. Every engineer should be an FDE by default. I discussed this with a founder in SF last week, and it hasn’t left my head. What some in Europe still pitch as “forward-deployed engineering” is table stakes on the West Coast. The fastest teams remove layers between builders and users. What that looks like in Deepnote: - Engineers fly to customer sites. They regularly sit with data teams, shadow real workflows, and return with pull requests (sometimes on-site). - Shared Slack. Engineers are in customer channels directly, own threads, post demos, and ship fixes in the open. - Engineers ship to customers directly. They record Loom videos or do demo calls. No PM as a middleman. If you need a dedicated role to speak with the customer, you add latency. Latency kills product velocity. Slow velocity kills companies.

A collaborative analytics and data science platform makes the move "so the community has a standard that is purpose-built for AI," says Deepnote CEO Jakub Jurových. By Heather Joslyn

We're live with SoundCloud's Brian Doherty - discussing analytics at scale: how SoundCloud builds with data. Tune in right now!

How does SoundCloud analyze music data at scale and turn billions of streams into product decisions? 🎵 We're hosting a live session with Brian Doherty, Principal Data Analyst at SoundCloud, on how they turn real world data into product impact. We will unpack the choices behind their stack, and how they balance speed with governance. Join us, and bring your toughest questions!

In less than 2 hours, our founder and CEO Jakub Jurových sits down with Brian Doherty from SoundCloud to break down how they analyze music data at a massive scale and turn it into real product impact. Don’t miss this one and tune in!

Opus 4.5 just passed Gemini 3 at coding. The White House kicked off an AI Manhattan Project. OpenAI’s inferred Azure bill hit $8.7 B through Q3. Inside this issue: 🔵 Opus 4.5 > Gemini 3: 80.9 % on SWE-bench Verified, price cut to $5/$25 per 1M tokens, now in Copilot and Microsoft Foundry 🔵 Gemini 3’s push: 1501 LMArena Elo and 45.1 % ARC-AGI, safety report flags higher manipulation risk 🔵 Grok 4.1 goes free: thinking/non-thinking modes; sycophancy up ~3× vs Grok 4 🔵 Genesis Mission: the White House’s “Manhattan Project for AI” on DOE supercomputers 🔵 OpenAI economics: leaked docs show $8.7B Azure inference spend vs ~$4.3B implied revenue Read the full scope below.

Google just defied gravity. And it's… wicked. Can Cursor keep up? I’ve spent a few days with Antigravity, and I have to say, well done, Google. I was skeptical at first because how good can yet another VS Code fork be? Turns out, very good. I was in a constant state of flow: - Fair-use limits reset before momentum dies - Autonomous browser for live tests - Async agents tackle parallel tasks I was vibe-shipping so hard I forgot about my Cursor subscription and tried to upgrade. Then, I realized it’s free. I was surprised I couldn't pay for this, because the experience was that good. And my personal favorite: it supports Deepnote's open-source notebooks out of the box! We’ve seen big tech clones of successful GenAI products (e.g., Kiro from Amazon, a Lovable alternative), but Antigravity just hits different. Will no one mourn Cursor?

I merged 13 PRs last week without writing a single line of code. Linear + Devin did the heavy lifting for me. Six months ago, Devin finished 10-20% of my tickets. Today it’s 40-50%. How it works: 1. Always start from a Linear ticket. See a bug? Have an idea? Create a Linear ticket. (I’ll have to create it anyway, might as well do it right away.) 2. Assign it Devin (tried many agents, this one works the best). 3. Come back 20 minutes later, look at the working review app. Day in life: In the morning I kick off 5 tickets. By the time I catch up on my emails, 2 of them are ready and I send them for review. I reassign 2 other tickets to one of our engineers because it’s more complex and requires more work. I close the last ticket because I find out the complexity is not worth it. Some learnings: - The real bottleneck is the feedback loop, not model quality. Ten minutes per iteration is too long. You can run many tickets in parallel, but it breaks the flow. - We pushed the eslint to the biome transition despite the complexity because waiting on style checks is not a good use of human time. - My local dev env broke a few weeks back and I never fixed it. Everything runs in the cloud now anyway. - Wouldn’t be possible without automated code review. CodeRabbit took lot of effort to set up properly, but now Why this matters for teams: - Handovers are the most time consuming thing today. Engineers (or anyone really) shouldn’t waste time looking at half though-out tickets. AI is great at identifying missing context and does it almost immediately. - Speed compounds when tickets are clear, work is split, and agents own the first 60-70%. Humans review, decide, and handle the remaining 30-40%. That is how small teams keep pace with the best in the world.

How exciting. I moved to Deepnote years ago because I simply wasn’t patient enough to fiddle with Jupyter installations and bugs. I felt like an outlier at the time. It was a bit of a leap of faith, trusting a small private startup, but the alternative was… trying to make do with Excel (there is never spare time in a newsroom). Deepnote felt intuitive, modern and easy to share with colleagues. When I started teaching Python to non-techy journalism students, it has proven irreplaceable. Getting a class to install Jupyter would have been a death sentence to their tepid ambition to learn coding. And while Colab exists, it still feels like Jupyter in the cloud (functional but clunky). With Deepnote, students just open a link and start coding, usually to their own disbelief and amazement. With this open-source release, I hope more data journalists and newsrooms feel comfortable adopting it and finally give themselves a smoother, more collaborative notebook workflow.

Deepnote goes open source. The data notebook for the AI era is now open source under Apache 2.0. Most notebooks fail at collaboration. Jupyter creates messy diffs, sharing breaks environments, and "it works on my machine" blocks every deployment. Deepnote solves that with: → AI agents, → 23 new block types, → 100+ native data connectors, → and works in VS Code, Cursor, or Windsurf. Built on a customized Jupyter kernel for full backwards compatibility. How it works: → Convert any Jupyter notebook: npx @deepnote/convert notebook.ipynb → Run locally in your favorite IDE or JupyterLab → Use reactive execution (downstream blocks update automatically) → Scale in Deepnote Cloud for real-time collaboration with AI agents → Deploy data apps without environment issues Used by 500,000+ data professionals. Star the Repo today: https://lnkd.in/gCMQXncN

Kimi K2 just topped open-source coding. Anthropic blocked an AI run cyberattack. GPUs are waiting for electricity, not for models. Inside this issue: 🔵 Kimi K2 at 65.8% on SWE-bench Verified, agent native and low cost 🔵 Deepnote is now Apache 2.0, a drop-in replacement for Jupyter with YAML diffs and IDE extensions 🔵 Claude Code misuse stopped, 80 to 90% of the campaign was automated across ~30 targets 🔵 Power is the bottleneck, GPUs idle as datacenter demand is projected up 165% by 2030 🔵 Character.AI limits under-18 chat, safety rules tighten across consumer AI Read the full scope below.

Data teams need something Jupyter never promised. That's why we open-sourced Deepnote earlier this week under Apache 2.0. Today, I want to talk more about the why behind open-sourcing and the design principles we've followed. Jupyter changed everything, and we are grateful for it. But, it was also built pre-cloud, pre-collaboration, pre-AI. As data teams scale, those limits show. Where it breaks for teams: - JSON diffs are noisy, code review hurts - Single file, no project level structure - No first-class comments, reviews, or permissions - Limited cell types - Lack of native integrations into your data stack - No good way to securely store secrets. That's why we designed .deepnote with a few non-negotiables: - Human-readable YAML with clean diffs you can read, not machine-optimized JSON - Support for securely connecting to data sources and other integrations - Multiple notebooks per project, with shared dependencies and integrations - 20+ new block types, with SQL as first-class citizen, charts, inputs, KPIs, text, and more - Verifiable schema, ensuring metadata integrity and forward compatibility - Language-agnostic, supporting Python, R, or SQL in the same notebook - No vendor lock-in. Conversion to and from .ipynb, run in VS Code, Cursor, Windsurf, and JupyterLab. The goal is simple, keep the spirit of Jupyter, make the medium ready for teams and agents. Check out the deep dive here, then tell me what you think: https://lnkd.in/eJHKFGX7

This is BIG. Deepnote, a company I’ve followed and used since its creation, just open-sourced their notebook framework. And honestly, this could be the final chapter for Jupyter. After 7 years of development, @DeepnoteHQ has built something that redefines the notebook experience: reactive, collaborative, AI-ready, and open by default. Core capabilities → Supports Python, SQL, and R → Interactive blocks, charts, and tables → Reactive execution: cells update automatically (no more "Run All") Integrations and compatibility → 60+ native data integrations → Fully .ipynb compatible with no lock-in → Runs locally or inside VS Code, Cursor, Windsurf, and JupyterLab Scalability → Move to Deepnote Cloud with a single command It’s everything we loved about Jupyter, rebuilt for 2025, a genuine open successor 🫶 Here's the repo they just open-sourced: ↳ https://lnkd.in/ejwuEtmJ Here are a few more goodies you can try: → VS Code Extension: https://lnkd.in/eGgS-qd5 → JupyterLab Viewer: https://lnkd.in/e8fSDhui → CLI: npx @deepnote/convert notebook.ipynb Heck I'm such a fan of Deepnote I've been on their homepage for years 😄 ↳ https://lnkd.in/eY_RWdSb! ___ ♻️ If this sparked an idea, hit repost so others can catch it too! Follow me here or on X → @datachaz for daily drops on LLMs, agents, and data workflows! 🦾

This could be the final nail in Jupyter's coffin. Deepnote is going open-source! Their kernel is way more powerful than Jupyter, but still backwards compatible. Notebooks are amazing: • They are perfect for data exploration • They are perfect for collaborating with AI agents • They are perfect for bringing technical and non‐technical users together But Jupyter doesn't cut it anymore in 2025. Here is what Deepnote gives you: • One workspace for technical and non‐technical users together • Native versioning, comments, and reviews • Human‐readable projects with clean diffs • Code with Python, SQL, or R • Use no‐code blocks with an AI agent • 100+ native connectors • Use in VS Code, Cursor, Windsurf, or JupyterLab • A modern UI and UX • Full Jupyter compatibility with round‐trip .ipynb Star this GitHub repository: https://lnkd.in/em3ABGjR

We've spent 7 years building the data notebook for AI era. Today, we're open sourcing it. Deepnote Open Source is successor to the Jupyter notebook. It acts as a drop-in replacement for Jupyter with an AI-first design, sleek UI, new blocks, and native data integrations. Use Python, R, and SQL locally in your favorite IDE, then scale to Deepnote cloud for real-time collaboration, Deepnote agent, and deployable data apps. Single-player notebooks were great in 2013. 2025 needs reactive, collaborative, AI-ready projects that integrate into your existing stack seamlessly. That's why we're making Deepnote open source - to offer the community an open standard for AI-native data notebooks and data apps. We're standing on the shoulders of Jupyter — it changed how the world explores data. But at team scale, the papercuts stack up: brittle reproducibility, no native data connectors, weak collaboration, and bolted-on AI features. In the enterprise context, this gets very tough to manage - and we're seeing an increasing demand from large companies to move away from Jupyter. What’s new: - Reactive execution (downstream block auto-update) - Powerful blocks beyond code: SQL, interactive inputs, charts, KPIs, buttons - 100+ data integrations - Code in your favorite IDE: Cursor, Windsurf, or VS Code - No lock‑in: open standard; export to `.ipynb` whenever you need Once you're ready to scale in your team, transfer to Deepnote Cloud with one command for beefier compute, powerful data apps from notebooks and agentic data science. Ty it now: Repo ➔ https://vist.ly/4ctcn Deepnote in VS Code ➔https://vist.ly/4ctcj Docs -> https://vist.ly/4ctc7 CLI ➔ `npx @deepnote/convert notebook.ipynb` P.S. This only works as an open standard. Tell us what’s missing, file issues, send PRs. Help define the data notebook for the AI era!

48K experts just called for a superintelligence pause. Claude topped the hardest coding benchmark. OpenAI’s $100B circular money machine is humming. Inside this issue: 🔵 SWE-Bench Pro: Claude Sonnet 4.5 leads a contamination-resistant, real-repo benchmark 🔵 AI safety goes mainstream: 48K signatories urge a superintelligence ban 🔵 MiniMax M2 goes open: #5 global at ~8% of Claude’s price 🔵 Claude lands in Microsoft 365: search across SharePoint, Outlook, Teams 🔵 Circular financing: NVIDIA ↔ OpenAI ↔ Microsoft ↔ Oracle, the loop powering AI’s spend Read the full rundown below.