Deepnote

Yesterday, we shipped one of our most requested updates! 🚀 🔵 Unlimited charting: Visualize datasets with millions of rows directly from Pandas DataFrames, without any preprocessing. 🔵 @ References in AI prompts: Boost precision by referencing specific integrations and modules in your AI prompts. Explore your data without limits, and make your AI insights more accurate than ever. Try them out today!

We burned weeks optimizing for accuracy. The users actually wanted usability. Spoiler: I learned it the hard way. No point in building the “perfect” AI if no one can use it. So, we build around three principles: - Human‑in‑the‑loop by default. Usability beats purity, because a 5‑second human check is faster than a 5‑month research project. - Ship, learn, iterate. We’d rather release v1 in a week and improve it 10× with real feedback than theorize our way to v∞. - Solve work, not wow. A feature isn’t impressive until it’s in production, saving someone time or money. Perfect accuracy isn’t the goal. Practical impact is. If you’re building human‑first AI and want to compare scars, let’s chat.

Who knew Jeff’s real secret weapon was psychology grad students? 👀 Data via tylervigen.com, recreated in Deepnote.

RAG recalls text. cognee remembers how it all fits together. See how graph-powered memory inside a Deepnote notebook boosts HotpotQA results from 0 to 50 % EM, slashes hallucinations, and lets you query data like a network graph.

Everyone’s busy launching agent stores and counting tokens. Here’s what actually mattered this week: - Models keep getting better: Grok 4 beats OpenAI's o3, while Mistral AI outperforms Whisper large-v3 on cost efficiency, with its new smaller Voxtral speech-to-text models. - Vibe-deleting: Replit agent deletes $1M worth of code, while Amazon Web Services (AWS) automates away most of its DevOps team with AI agents. - Deeper thinking still moves the needle: ThinkSound from Alibaba and HKUST first reasons, then generates, and it tops every objective and subjective metric, beating state-of-the-art benchmarks. - Zuck casually floated “psyops at scale.” Millions of AI agents to shape buyers and bully competitors. Because Cambridge Analytica clearly ended well for everyone. - Decart’s MirageLSD makes real-time video look easy. 16× faster than anyone else. Read 25 more news items in the newsletter.

Agent stores everywhere. Devs barely touching keyboards. One Replit agent just vibe-deleted a $1M prod DB. Inside this issue: 🔵 ChatGPT Agent goes fully autonomous (41.6% on Humanity’s Last Exam) 🔵 OpenAI’s $3B Windsurf deal nuked by Microsoft, Google swoops, Cognition cleans up 🔵 Amazon Web Services (AWS), Google, Microsoft race to be the App Store for agents 🔵 Vibe coding blows up: 75% of devs prompt, 0% read the fine print (see: $1M delete) 🔵 Decart’s MirageLSD: infinite video at <40 ms latency (16× faster than anyone) 🔵 Groq + Llama-3 top tool-use charts, Liquid AI puts GPT-3.5 speed on your phone Read the full scope.

Tomorrow’s best data engineers won’t be known for how much they code, but for how clearly they think. If Cursor can draft your entire pipeline before you finish coffee, why is your company still paying for your expensive coffee machine? Turns out, there’s quite a lot of cases where you’re more needed than ever before. - Problem framing. Knowing which question is actually worth solving. - Context distillation. Choosing which business logic, edge cases, and constraints to surface, and which to ignore. - Taste-driven editing. Spotting brittle logic or bloated prompts before they make it to prod. - Narrative thinking. Shaping raw data into arguments your org can act on. The job title won’t disappear; the skill stack will. Coding becomes the baseline. Agency & sense of taste are what will matter.

10× is giving every analyst an AI side-kick. 100× is letting agents pull your data for you. This is the extra mile to boost your productivity: - Serve your workspace over MCP, so the AI agent can fetch the tables, metrics, or notebooks it needs. - Wire it into something trivial-but-painful (like your CRM). - Watch the agent pull live numbers from Deepnote (and draft the reply). No extra integrations, no copy pasting, just data to action in one round-trip.

We don’t plan for more than 6 months in advance. Here’s why. In November 2022 we built a beautiful roadmap for 2023. Milestones, vision statements, OKRs. Rallied the team around it. 2 weeks later ChatGPT was released. KPIs were obsolete. Roadmap was obsolete. Job descriptions were obsolete. All the meticulous planning was a waste of time. It didn’t help with clarity, it slowed us down. Takeaway: It’s impossible to predict what’s going to happen after 6 months. Even if we get things right, we are wrong on timing. Features that don’t work today could work in 2 weeks.

We’re hosting some awesome events during the conference next week and you’re invited! 🍕 Python Rooftop After Party Join us for an evening of cold beers, Neapolitan pizza, and city views from our rooftop terrace. 📅 Wednesday | from 6:00 PM 🥐 Data Leaders Breakfast Start your day with fresh coffee, tasty breakfast, and great conversations with data leaders. 📅 Tuesday, Wednesday & Thursday | 8:00–9:30 AM RSVP: https://xmrrwallet.com/cmx.pvist.ly/3n949tc Spots are limited, so make sure to reserve yours soon. Hope to see you there!