Claude Design Ships, Figma Stock Drops, Vercel Hacked

Cerebras Systems filed to go public, reporting $510 million in revenue and an $87.9 million profit in 2025. The IPO filing disclosed a $1 billion loan from OpenAI tied to a $20 billion-plus inference infrastructure deal, with an option for OpenAI to add 1.25 gigawatts of capacity through 2030. AWS will also deploy Cerebras WSE-3 chips for the decode phase in its new disaggregated architecture. Separately, OpenAI secured a dedicated HBM4 production line with Samsung, supplementing its existing SK Hynix deal.

Microsoft acquired Fintool, a San Francisco startup building AI agents for financial research. Fintool's platform reads earnings transcripts, analyses filings, synthesises research, and builds financial models autonomously. Its V5 release earlier this year could build DCF models in Excel and prepare earnings presentations in PowerPoint without human input. The team will join the Office Product Group, initially focused on financial services before expanding to other verticals.

Cursor, the AI coding startup, is in advanced talks to raise $2 billion at a valuation north of $50 billion, according to Bloomberg. Andreessen Horowitz will co-lead the round, with Nvidia and Thrive Capital also participating. The raise underscores just how much venture capital is flooding into AI coding tools. Cursor was last covered in edition #81 for its multi-agent system that beat engineers on Blackwell kernel optimisations.

Anthropic shipped Claude Design, a prototyping tool that generates real HTML and JavaScript from text or image prompts. It works inside Claude's chat interface and outputs code directly, skipping proprietary design file formats entirely. Figma's stock fell on the announcement. The launch comes days after Anthropic released Opus 4.7, its latest frontier model focused on advanced software engineering and self-correction.

Vercel disclosed a security incident after threat actors claimed to have breached its systems and began selling stolen data. The company said the attack originated from a compromised "third-party AI tool," making it one of the first high-profile breaches attributed to an AI supply chain vector. Employee names, email addresses, and activity timestamps were exposed. ShinyHunters, the group behind the Rockstar Games hack, claimed responsibility.

OpenAI released a new model series built for life sciences research, starting with GPT-Rosalind, designed for reasoning in biochemistry and genomics. The company says drug development currently takes 10 to 15 years from target discovery to regulatory approval, with only one in ten clinical trial drugs getting approved. GPT-Rosalind is intended to help researchers move faster through the most analytically demanding phases, not replace expert judgement. Five competitor newsletters covered this.

A bipedal humanoid robot finished a half marathon in Beijing in under 51 minutes, faster than the human world record. The asterisk: the winner crashed and needed human help, and was telecontrolled rather than fully autonomous. Still, the race marked a major improvement from last year when only six of 21 humanoids finished, and the fastest took two hours and 40 minutes. China has made robotics a national priority, though analysts say the country's software still lags behind the US.

Bloomberg's Mark Gurman reports that Apple has already teased the new Siri interface coming in iOS 27 within the WWDC 2026 promotional graphic. The redesign reportedly replaces Siri's current full-screen takeover with a more integrated experience. WWDC's focus this year will centre on the Siri overhaul, including a standalone app, Dynamic Island integration, and an Ask Siri button.

A hardware developer connected Claude Code to a SPICE simulator and LeCroy oscilloscope via MCP servers, letting the AI validate circuits against real measurements instead of relying on natural language descriptions alone. The approach works well for SPICE circuit validation, embedded programming, and data analysis tasks like normalising time axes and aligning signals. Key lessons: always give Claude an explicit pinout map, save measurement data to files rather than dumping it into context, and use Makefiles for reproducible builds.

A Towards Data Science article demonstrates a failure mode where RAG systems retrieve correct documents but generate wrong answers because the context contains conflicting information. In three reproducible experiments, an extractive QA model picked preliminary earnings over audited revisions, outdated HR policies over current ones, and old API limits over updated ones. Each time, confidence exceeded 78%. The model has no mechanism to detect contradictions. The fix is a dedicated conflict detection layer between retrieval and generation.

A Shopify engineer took Karpathy's Autoresearch concept, originally designed for model training, and adapted it to optimise build times, test performance, and component rendering speeds across Shopify's codebase. The approach runs an agent in an infinite loop against a single metric, measuring each iteration and keeping only the improvements. CEO Tobi Lutke jumped in with a 32-commit PR to add multi-metric support, then they open-sourced it. The project has 3,600+ GitHub stars. Results include 65% faster builds and unit tests running 300x faster.

A designer argues that Figma accidentally excluded itself from the agentic era by storing designs in a proprietary format that LLMs were never trained on. Claude Design, rough as it is, outputs real HTML and JS. That means designs live in the same medium they ship in, and Claude Design connects directly to Claude Code. Figma's variables panel now has 946 colour variables and components with 16 variants. The author's prediction: design tooling splits into two shapes, and Figma's Sketch moment is coming.

Addy Osmani lays out why most production agents are held together by custom plumbing, fragile session logic, shared service accounts, and a security model held together by hope. He identifies four gaps: agents need real identities instead of borrowed OAuth tokens, universal context across systems instead of scraped windows, durable execution that survives laptop lids closing, and platform primitives so teams stop rebuilding orchestration from scratch. The underlying argument is that governance debt is accumulating silently and the first incident that reaches a CISO will force a full rewrite.

Security researcher Daniel Miessler published his weekly cancellation list: Zapier, Resend, Figma, Canva, Browserbase, and Supabase, all replaced by his own agent harness, repos, and Claude Code skills. His question to every SaaS company: what are you, exactly? An interface? A database? Data? What makes you hard to copy? He frames it as a massive evolutionary selection event where the ease of creating and running software has gone up so dramatically that mediocre, copyable companies can no longer survive on inertia alone.

Colossus profiles Notion's AI transformation from the inside. Co-founders Ivan Zhao and Simon Last locked themselves in a room in Mexico in October 2022, a month before ChatGPT launched, and declared Notion an AI company. Since then, the engineering team has rebuilt its agent harness twice (soon three times), ships 15 releases a month, and runs Project Applecart, a full product rewrite from scratch for agents. Custom Agents turned the decade of workspace context into a competitive moat, and revenue is accelerating faster than ever.

video-use is an open-source tool that uses Claude Code to automate video editing workflows. Drop raw footage into a folder and the AI cuts filler words, colour grades, adds subtitles, and generates animations. It processes audio transcripts with word-level timestamps for precise edits. The project pairs well with Hyperframes for a full agent-driven video pipeline from raw footage to published content.

Nvidia's Lyra 2.0, released last week, generates persistent, explorable 3D worlds from a single input image. The 14B-parameter model synthesises long-range video with geometric consistency, then reconstructs it into an explicit 3D representation that supports real-time rendering. It solves spatial forgetting through per-frame 3D geometry routing and temporal drifting through self-augmented training. Research use only under Nvidia's internal licence.

MZLA Technologies, a Mozilla subsidiary, launched Thunderbolt, an open-source AI workspace designed for organisations that want full control over their AI stack. It supports any model provider (commercial or local), connects to Haystack pipelines, MCP servers, and agents via the Agent Client Protocol, and runs on Windows, macOS, Linux, iOS, and Android. Trending on GitHub with 1,486 stars. The pitch is simple: run AI on your infrastructure, with your data, without vendor lock-in.

PgQue rebuilds PgQ, the queue architecture that ran messaging for hundreds of millions of Skype users, in pure PL/pgSQL. One SQL file to install, works on any managed Postgres 14+ including RDS, Aurora, Cloud SQL, and Neon. Unlike SKIP LOCKED queues that create dead tuples and eventually hit VACUUM pressure under sustained load, PgQue uses snapshot-based batching and TRUNCATE rotation for zero bloat by design. Benchmarks show 86k events per second on a laptop.

HeyGen released Hyperframes, an open-source framework that lets AI agents create and render video from HTML compositions. Write an HTML file with data attributes for timing, layering, and animation, then render to MP4. It ships with skills for Claude Code, Cursor, Gemini CLI, and Codex, so you can describe a video in natural language and the agent handles scaffolding, animation, and rendering. Includes 50+ ready-to-use blocks for social overlays, shader transitions, and data visualisations. Apache 2.0 licensed.