Mave raised $34M, including a $28M Series A led by Red Dot Capital Partners. The signal: the next battleground in SecOps isn't buying more security tools — it's letting an agent run the investigation across the tools you already have. Mave's ARR grew more than 4x in a year, proof that enterprises are already paying for agents to take over incident investigation, not just piloting a concept.
Raon-1 mini preview: KRAFTON AI (the studio behind PUBG) trained a 28.8B-total/4.3B-active MoE model from scratch, with 262,144-token context and Apache 2.0 open weights. Its hybrid attention stack (Raon Delta Attention + Gated MLA) shrinks KV cache to about 8 KiB per token, and a bundled DFlash2 draft model boosts throughput up to 2.18x on B300 GPUs. It scores 92.9 on an internal Korean chatbot benchmark — far ahead of the similarly-sized Qwen3.6-35B-A3B (74.5) and Gemma 4 26B-A4B (90.6) — but trails Qwen3.6-35B-A3B on MMLU-Pro and BFCL tool-calling. No official API pricing yet; it's open weights plus a time-limited free playground.
Claude Sonnet 4.5 stops responding on the Claude API, Claude Platform on AWS, and Microsoft Foundry on 2026-11-30, with no grace period. The recommended replacement, Claude Sonnet 5.5, is 33% cheaper per token (input $3→$2, output $15→$10 per 1M tokens), but its new tokenizer produces about 30% more tokens for the same text — so the real bill usually drops 10-15%, not the sticker-price 33%.
Today's skeleton: define growth with a North Star Metric plus a handful of input metrics, then use Eric Ries's three-engine framework (sticky, viral, paid) to figure out which engine the product is actually running on, instead of answering every growth question with 'add a sharing feature.' The practice question is a real diagnostic prompt — checkout funnel conversion dropped 20% in a week — where ruling out bad data and external causes before jumping to a UX hypothesis is itself the thing being graded. The case study is Clay, which went from a generic spreadsheet product with 20 paying customers to narrowing around cold-email agencies, then compounded through a loop where 'an agency delivers client work with Clay → publishes the workflow → looks like an expert → wins more clients → runs more work through Clay' — reaching a $7.1B valuation in September 2026. The most imitable part of this case isn't the LinkedIn posting, it's the design decision to make sharing the product economically self-interested for the people sharing it.
landmine is a CLI, pre-edit hook and MCP server rolled into one. It scans your lockfile for each direct dependency's exact pinned version, matches it against open GitHub issues filed against that version, and writes the known bugs and community workarounds into AGENTS.md/CLAUDE.md, or blocks the agent once when it's about to hit one. Install: `npx landmines scan && npx landmines install --rules`. It solves the problem of an agent writing code that trips a known bug in the exact version you have pinned.
CrowdStrike caught attackers who used Claude Code against South Korean banks, Apex Accelerator hijacked agent traffic through a malicious npm supply chain, and autonomous agents breached 1,200+ websites at $50 a site — the same week Google's Gemini 4 Argon, OpenAI's $1B "Patch the Planet" pledge, and open-sourced governance tooling from Microsoft and Cloudflare all shipped, scaling defense and offense in parallel; DeepSeek closed a $15B round two days after the US ordered federal agencies to strip its systems; four straight days of arXiv digests showed evaluation and monitoring infrastructure itself failing silently — nine LLM judges returned the identical score, and splitting an attack across multiple agents cut detector recall from 87% to 7%; Vertex AI, OpenAI, and Anthropic all retired legacy models this week, and the headline discount rarely matched the real savings; and in the same week Taiwan's financial regulator published its agentic-AI safety advisory, global capital poured eight large rounds into agent execution-layer infrastructure.
Once an agent chains together retrievers, sub-agents, and tool calls, checking whether the final answer is right isn't enough anymore. Today's practice is pulling evaluation apart layer by layer, worked through a multi-tool agent evaluation question.
Three things in Agno v3.1.8: (1) `PostgresCheckpointStore` lets sandbox checkpoints live in Postgres, so multi-replica deployments can restore a session without S3 or a shared filesystem; (2) a behavior change: sandboxes now keep only the 3 newest checkpoints per session by default, set `retain_checkpoints=None` to keep them all; (3) `Claude()` now defaults to `claude-sonnet-5-5`, replacing the soon-to-retire `claude-sonnet-4-5-20250929`.
Three things in Pydantic AI v2.59.0: (1) a high-severity security fix (GHSA-4whf-38hw-pxxh) — the `web_fetch` tool could stall unrelated work in the same process when it fetched attacker-crafted HTML with many elements; (2) `DaytonaSandbox`, which runs the whole harness toolset (Coder, Shell, FileSystem) inside a Daytona sandbox; (3) provider-neutral model error categories like `ModelRateLimitError` and `ModelContextWindowExceededError`, so retry logic no longer has to special-case each provider.
Legora raised $550M Series E led by Atreides Management, taking its valuation from $5.6B to $8.8B in just seven months. The signal: legal AI competition is no longer about whose model reads contracts faster — it's about whose agents can take over an entire case workflow. Two legal AI startups each raised $550M within six months, and capital is treating legal services as the next vertical due for an agent rebuild.
Pigment raised $250M Series E led by EQT's Scaleup Europe Fund at a $2.3B valuation — about 23x current ARR. The signal: the next battleground in enterprise planning software isn't 'better dashboards,' it's 'agents that turn decisions into action.' 85% of new metrics on the platform are already created by an agent.
Recursive Flint: a 40B-parameter MoE model (256 routed experts plus 1 shared expert, 6 routed experts active per token, ~2.5B active parameters), 262,144-token native context (official guidance caps practical use at 256K), Apache 2.0 license, released 2026-10-21 and covered by both TechCrunch and Wired. It scores 74.2% on SWE-bench Verified (edging out Qwen3.6-35B-A3B's own-reported 73.4%), 89.7% on SWE-bench Pro v2, 54.9% on Terminal-Bench 2.1, and 80.3% on LiveCodeBench v6 — the best self-reported scores among MoE models under 50B parameters. On a MacBook Pro M4 Pro it hits 131 tokens/s, nearly double the 58–69 tokens/s of comparable models like Gemma 4 and NVIDIA's Nemotron 3.5 Lightning, and tops 150 tokens/s on an RTX Spark. Quantizing to NVFP4 costs only 0.7 points versus the 16-bit checkpoint. There's no official paid API — Flint is a download-and-run open-weight model, though OpenRouter currently lists a free promotional hosted version. Recursive is led by Richard Socher, Salesforce's former chief scientist, and emerged from stealth in May 2026 with $650M raised at a $4.65B valuation. Flint is the company's first public model, and its headline claim is that almost the entire build pipeline — architecture search, quantization recipe, inference kernels — was handled by Recursive's own automated R&D system, the Eureka Machine, with a small team shipping it in under four months.
'Where should the human-in-the-loop go?' is one of the most common AI product design interview questions, and the easiest one to answer as a slogan. 'Important actions need a human to look at them' sounds right, but interviewers want to see whether you can turn 'important' into an operational rule. Today's drill uses Cloudsoft Solutions' three-tier matrix — error cost × reversibility → oversight level (high-cost irreversible actions need approval before they execute, medium-risk high-volume output can be sampled after the fact, low-risk output just needs exception handling and feedback) — paired with Tcules' eight human-AI collaboration UX patterns. The key point is that an approve button by itself doesn't prove real oversight: automation bias doesn't go away with training or reminders, and Anthropic's own data shows Claude Code users approve 93% of permission prompts — direct evidence of approval fatigue. The practice question is a made-up design prompt about an AI loan-approval system: instead of defaulting to 'review everything' or 'auto-approve everything,' the answer tiers review intensity by error cost and designs how to actually verify review is working, not just track the approval rate. The case study is Anthropic's Claude Code auto mode, published March 2026 — it redesigns 'approve every single action' into a three-tier architecture (a safe-tool allowlist, in-project file operations, and classifier-gated review). The two-stage classifier design cuts the false positive rate from 8.5% to 0.4%, but still misses 17% of real dangerous actions in testing — an honest disclosure that automated review isn't zero risk, it's trading approval fatigue for an acceptable amount of residual risk, which is exactly what's worth imitating in an interview answer.
ML system design differs from classic system design in one layer: what the system predicts, how you know the prediction works, and how it quietly decays. Today's practice uses a fraud-detection design question to work through that layer.
Three things in Pydantic AI v2.58.0: (1) `DelegationTasks.steer` lets a running delegated task be redirected mid-flight instead of being cancelled and restarted; (2) a run-level `timeout=` and `RunContext.deadline`, plus per-request `request_timeout`/`max_retries` on `ModelSettings`; (3) two medium-severity security fixes — provider-native MCP configuration could leak server URLs/tokens into OpenTelemetry traces, and a frontend-adapter misconfiguration could let a tool execute with client-chosen arguments without a model request. Also new: an `openai-chatgpt` provider and `ElevenLabsAgentsRealtimeModel`.
Arda raised a $160M Series A at a $1.5B valuation, founded by former OpenAI research chief Bob McGrew. The signal: the next frontier for AI agents isn't just code — it's compressing manufacturing's human bottleneck, where retooling a production line can take three to six months, down to the scale of days.
Daytona raised $60M Series B led by FirstMark, hitting a $1B valuation. The signal: once companies start running thousands of AI agents at once, the basic unit of cloud computing stops being a 'virtual machine for a person' and becomes a 'sandbox spun up in 60 milliseconds and thrown away.'
Ornith-1.5-122B: 122B total parameters (MoE, active parameter count undisclosed), 262,144-token native context (extendable to roughly 1M with YaRN scaling factor 4.0), MIT license, deployable on a single 80GB GPU; uploaded to HuggingFace on 2026-10-16 with no accompanying blog post, surfacing instead through community coverage and HF trending; scores 84.5% on SWE-bench Verified (non-hack) and 80.0% on CyberGym, both ahead of the similarly sized Qwen3.8-Flash-Next (125B); the training method carries over Ornith-1.0's self-improvement loop but extends it from optimizing scaffolds and rollouts alone to jointly optimizing task generation, scaffold construction, and rollouts; no official or third-party hosted API pricing exists yet.
OpenAI's deprecations page lists 12 legacy models shutting down on 2026-10-23. Migrating o1 to GPT-5.6 Sol cuts input from $15 to $4 (-73%) and output from $60 to $20 (-67%) — but o3-mini and o4-mini land on the same or a similar replacement at 1.8x to 3.6x their old price. Any code still calling these model IDs starts failing in three days.
What interviewers care about most in 2026 strategy questions isn't a well-structured answer anymore — a structured answer is the entry bar, not the differentiator. What separates candidates now is whether they bring a point of view five other well-prepared candidates couldn't have given. Today's drill uses Aced's (formerly Exponent) six-step framework: clarify the question, anchor to business goals, map the competitive landscape, set guiding principles, generate and filter options, argue and close — paired with Roger Martin's 'stupidity flip' test (if the opposite of your strategic choice sounds absurd when stated out loud, it wasn't a real choice) to judge whether a differentiator actually holds up. The practice question is a real Google interview prompt: 'You're Zoom's CPO facing heavy competition from Google Meet, what do you do?' The breakdown shows how to anchor first on what Zoom's revenue actually depends on (enterprise subscriptions and add-on services, not free consumer meeting volume), define the threat at the precise mechanism level (Google acquires distribution almost for free through Workspace bundling — this isn't a feature gap), then use a principle that can actually eliminate options to converge on one choice instead of a wish list. The case study is Netflix's 2000 counter-positioning move against Blockbuster — a classic from Hamilton Helmer's 7 Powers: Blockbuster collected nearly $800 million in late fees, 16% of its revenue, and Netflix's no-late-fee subscription model hit a pain point Blockbuster knew it had to match but couldn't, because matching it meant gutting its own primary cash flow.