ML coding rounds don't test leetcode recall — they test whether you can implement attention, k-means, and other ML primitives from scratch using only NumPy, while articulating the shape and complexity at every step. Today covers five high-frequency topics: vectorized thinking, softmax numerical stability, shape tracking and complexity analysis, padding/masking for batch inference, and how to verify correctness when hand-coding algorithms.
Callosum closes a $100M seed round led by Atomico, valuation undisclosed. The bet: the agent cost bottleneck is not the model itself but cramming every step into the same GPU. As inference spending eats over half of AI-native companies' revenue, the routing layer's value expands from model selection to chip selection.
Twin1 AI closed a $20M seed round co-led by Bessemer Venture Partners, Tribeca Venture Partners, and Aramco Ventures, with valuation undisclosed. The bet: the atomic unit of enterprise knowledge isn't the document — it's the person. While every Agent startup races to plug into document repositories, Twin1 goes after the context that lives in people's heads and was never written down.
The dividing line in growth interviews is whether you're talking about linear improvement or compound loops — adding an acquisition channel is marketing; making one user bring in two users is growth. Today we practice the Goal → Metric → Bottleneck → Hypothesis → Experiment → Measurement six-step diagnosis framework, with a question drawn from a real OpenAI Growth PM take-home.
DeepSeek open-sourced DeepSeek Harness, an MIT-licensed Agent execution framework, yet in the same week hiked peak-hour API prices by up to 1,100%. Alibaba open-sourced flagship weights Qwen3.8 Max (topping LongBench v2) and Qwen3.8-27B, a laptop-runnable model, plus context infrastructure MyContext. Zhipu's GLM-5.3 showed 'emergent' cybersecurity capabilities -- scoring higher than Anthropic's reference model on vulnerability discovery benchmark CyberGym -- prompting Zhipu to delay its planned open-source release. ByteDance and Tencent each received approval to import roughly 10,000 NVIDIA H200 chips, signaling marginal easing of chip export controls.
On 2026/8/19 Splunk published SVD-2026-0808, patching 17 vulnerabilities across the Cisco Talos add-on, AI Toolkit, Connect for Kafka, MCP Server app, and On-Call. The most severe, CVE-2026-76404 (CVSS 9.1), is in the Splunk MCP Server app's credential management component — unserialized stored data without type validation lets admin-role users execute arbitrary OS commands. CVE-2026-76395 (CVSS 8.8) in AI Toolkit triggers similar RCE when loading model files containing pickle payloads. No in-the-wild exploitation observed. Mitigation: upgrade MCP Server app to 1.2.1 and AI Toolkit to 6.0.1 immediately; disable the app if you cannot upgrade right away.
claude-scope is a Claude Code plugin that provides SQLite FTS5 full-text search over your session history. Install: claude plugin marketplace add waazy-w/claude-scope. It solves the dilemma of 'index-based tools go stale, grep-based tools rescan hundreds of MB every time' by using byte-offset incremental sync — each search only reads newly appended bytes, so even text you typed a minute ago is already searchable.
Three simultaneous acquisitions (SpaceX×Cursor $60B, Stripe×OpenRouter $7B, Anthropic×Decart $6B) prove what's being bought is complementary assets, not revenue; DeepSeek open-sourced a harness that hit 20K stars in one hour, as model companies race to claim the harness layer; a full week of memory papers plus GraphWake/CoSnitch attacks point to the same thing — memory is now both a complementary asset and an attack surface; agent framework security debt got priced in (Check Point: 11 vulns across 6 frameworks, CoreBreak dispatch-layer bypass, Splunk MCP CVSS 9.1).
D2ACCI introduces a dual-loop diagnostic protocol that localizes memory failures to specific pipeline stages, raising diagnostic success from 0% to 98–100%; Salesforce re-evaluates memory-based self-improving agents and finds that shuffling task order turns an expected +1.5% gain into a -4.5% drop; GraphWake shows that poisoning just 10% of agents' memories can drastically amplify group opinion polarization
GraphWake shows poisoning 10% of agent memory can sway group opinion; CoSnitch exploits the same idea against Copilot's persistent memory for real; CVE-2026-40369 lets AI agents inherit a browser sandbox escape; Grok 4.6 tops GDPVal-AA v2 but trails in hardcore coding; Taiwan is the only market among four Asian regions where AI usage intensity declined
Volcengine (ByteDance) open-sources OpenViking, replacing black-box vector search with a viking:// virtual filesystem for agent memory — benchmarks show 80%+ accuracy while saving 34-91% tokens. munder-difflin wraps multiple coding CLIs into a desktop office with shared memory; ai-memory solves cross-CLI amnesia with a Rust MCP server; mukul975's cybersecurity skill pack rockets to ~28K stars in a day. pydantic-ai v2.32.0 adds OpenRouter/xAI attachment search and instrumentation improvements.
The core of ML system design interviews isn't which model to pick — it's how to keep the model alive in production. Today covers four high-frequency topics: online/offline separation in feature stores, root causes and prevention of training-serving skew, deployment strategies (shadow/canary/blue-green), and ML-specific monitoring beyond HTTP error rates.
Prevalent AI closes a $22M first institutional round led by Integrity Growth Partners. The deal shows that 'prove the market first, raise later' still works in the agentic AI era — while most startups burn VC money searching for PMF, a company that bootstrapped for 9 years and already serves large enterprises chose to raise only when agentic AI needs it most.
Grok 4.6: 500K-token context window, $2 input / $6 output per 1M tokens (same as 4.5), AA Intelligence Index 61 (tied with GPT-5.6 Sol Max), GDPVal-AA v2 1753 Elo (highest overall), but DeepSWE and Terminal-Bench still trail GPT-5.6 Sol and Claude Fable 5
Strategy interviews don't test whether you can recite Porter's Five Forces — they test whether you can make well-reasoned trade-offs with incomplete information and convince others. Today we practice market positioning analysis, moat assessment, roadmap prioritization defense, and stakeholder alignment communication.
Varonis social-engineered Copilot into disclosing an undocumented ?autorun=1 parameter, then chained three exploits: auto-executing injected prompts, exfiltrating Gmail/Drive/Calendar data via OAuth connectors, and writing attacker instructions into persistent memory that survives password changes and session revocations. Microsoft patched on 2026/8/18, CVE-2026-24301, CVSS 8.8. Defenses: audit Copilot connector permissions, monitor AI assistants like privileged insiders, and treat links containing prompts with suspicion.
comfy-mcp is Comfy's official local MCP server that wraps the full comfy-cli feature set into 39 MCP tools. Install: pip install comfy-mcp "comfy-cli>=1.14.0". It solves the problem where agents trying to run image/video generation workflows for you still need you to manually open a terminal, type commands, and verify that the right nodes and models are installed.
QUMem uses episode segmentation plus a three-stage agent pipeline to infer user state, beating the strongest baseline by 4.6 pp overall success rate on KnowU-Bench; LENS retrieves without pre-built indexes, achieving 84.8% evidence recall vs ReAct's 50.4% with zero degradation when indexes go stale; Intent-Guided Decoding arbitrates between retrieved content and model memory at decode time, yielding up to 65.4 pp accuracy gains on factual-conflict benchmarks
DeepSeek Harness hit 20K stars in one hour — the fastest in GitHub history — as model companies race to own the harness layer. xAI completed its acquisition of Cursor, accelerating consolidation in the coding agent space. Anthropic's annualized revenue reached $65B ahead of IPO, while it accused DeepSeek/Moonshot/MiniMax of industrial-scale distillation of Claude. Chinese hackers deployed up to 8 coordinated AI agents to breach at least 85 Taiwanese government accounts in four days. Anthropic and EPFL disclosed 'mind virus' research showing self-propagating payloads can spread across agents via persistent memory files.
DeepSeek's open-source agent harness 'dsh' crossed 20K stars within an hour of its 8/13 launch and has since accumulated ~158K stars, with 2000+ plugin proposals flooding in within two days. Its core is a Cordis-powered 'everything is a plugin' architecture that can even call Claude Code and Codex as sub-agents. RightNow-AI reimagines agents at the OS level with Rust (openfang), NetEase Youdao ships a desktop Agent built on OpenClaw (LobsterAI), and PrimeIntellect's prime-agent features a self-improving reasoning loop. CrewAI 1.15.16 adds execution context tracking and flow error logging.