Let's Data Science

The open-vs-closed LLM gap is now under 1 point. Kimi K2.5: 76.8% on SWE-bench. Claude Sonnet 4.5: 77.2%. 2026 deep-dive on the major open-weight families — benchmarks, VRAM, licensing, vLLM vs SGLang: letsdatascience.com/blog/open-sour… #DataScience

Anthropic limited who can use its Claude Mythos Preview — the model is good enough at finding and weaponizing software vulnerabilities. The UK's AI Security Institute clocked it finishing a 32-step simulated attack in 6 of 10 tries. letsdatascience.com/news/anthropic… #DataScience

df.dropna() is only safe when data is Missing Completely at Random. Otherwise you're not dropping noise — you're deleting a population. One loan dataset: no column >15% missing, yet deletion cut 20% of rows and skewed the model young. letsdatascience.com/blog/missing-d… #DataScience

Most SQL bugs come down to: I don't actually know what my JOIN returned. LDS ships a step-through SQL Visualizer on every one of its 1,585 problems. Watch each clause execute — row by row. letsdatascience.com/problems #DataScience

Anthropic is in early talks to raise at least $30B at a valuation above $900B, per Bloomberg. The round could close by end of month — though no term sheet has been signed yet. letsdatascience.com/news/anthropic… #DataScience

LLM interview Q: RAG vs fine-tuning — when which? RAG → fresh knowledge at query time. When facts change. Fine-tuning → bakes behavior into weights. For tone/format. Real systems use both: fine-tune for behavior, retrieve for facts. letsdatascience.com/quick-recall #DataScience

Fewer than 30% of data science candidates can write a correct Day-7 retention query from scratch. 13 worked SQL patterns — cohort tables, rolling retention, churn, L28, power user curve: letsdatascience.com/blog/sql-cohor… #DataScience

Google's Threat Intelligence Group has "high confidence" a criminal group used an AI model to weaponize a zero-day in an open-source admin tool. The exploit code showed textbook Python plus an apparent AI hallucination. letsdatascience.com/news/google-wa… #DataScience

Monday SQL — LinkedIn CX scenario. Pull every DM sent on or after Aug 1, 2025 that the user hasn't deleted. Return message_id, sender, recipient, sent_at, status — newest first. Mind the "deleted" status. letsdatascience.com/problems/sql/r… #DataScience

A doctor who forgets every patient would be dangerous. Most AI agents operate exactly this way. Agent memory isn't one vector DB. It mirrors human cognition — short-term, working, episodic, semantic, procedural. letsdatascience.com/blog/ai-agent-… #DataScience

53% of web traffic in 2025 was automated, per Thales — 40% of that classified as malicious. AI-driven bot attacks up 12.5x year over year. Human Security: AI agent traffic grew nearly 8,000%. letsdatascience.com/news/ai-bots-d… #DataScience

Sunday SQL challenge — Hard. Netflix Global Revenue Analyst: rank regions by captured subscription revenue, last 90 days. 3-table join + window functions + tie-breakers. RANK vs DENSE_RANK changes the answer. letsdatascience.com/problems/sql/r… #DataScience

This week's AI story has one through-line: security. → Anthropic's Mythos found thousands of zero-days → Google's gemini-cli got a CVSS 10 prompt-injection RCE → White House drafting pre-release model vetting Full recap: letsdatascience.com/news #DataScience

Anthropic's Claude Mythos found thousands of zero-days this week — 271 in a single Firefox run, plus decades-old bugs in OpenBSD and FreeBSD. The Fed convened bank CEOs over it. The White House is drafting AI vetting rules. letsdatascience.com/news/anthropic… #DataScience

Did you know? Running 10M tokens/day through GPT-5 costs ~$620,000/yr. Self-hosting Llama 3.3 70B on 2× H100s: ~$52,000/yr — an 88% reduction. The MMLU gap between open and closed models is now effectively zero. letsdatascience.com/blog/open-sour… #DataScience

@briansolis @nvidia Thanks @briansolis. Your "Return on Intelligence" framing, measuring AI by what it enables instead of what it eliminates, is exactly what the jobs debate needed. Full piece: briansolis.substack.com/p/the-new-roi-…

.@NVIDIA CEO Jensen Huang pushed back against alarmist claims that AI will destroy jobs, arguing instead that the technology creates opportunities. Full story by @letsdatascience 👉 letsdatascience.com/news/jensen-hu…

Most "Python Fundamentals" courses stop at lists and loops. LDS's free course goes further — 6 modules, ~6 hours, including a 12-section OOP module with magic methods, MRO, ABCs, and dataclasses. In-browser, interactive. letsdatascience.com/learn/python-f… #DataScience

Mozilla used AI-assisted, agentic pipelines (Anthropic Mythos + Opus 4.6) to find and fix 423 latent vulnerabilities in Firefox — including a 15-year-old HTML <legend> bug and a 20-year-old flaw. letsdatascience.com/news/firefox-f… #DataScience