Anthropic Ships Project Parka 🎙️, Poolside $6B Nvidia Deal 💰, Mistral Agentic Search Hits 86% FinanceBench 🔍

Creators lock in holiday calendars 90 days out. Structure commissions, recruit creators, and optimize your creator affiliate strategy before the rush with Levanta's 90-Day Holiday Sprint. Get the Free Guide.

Disconnected data and tools make it harder to understand your customers. HubSpot's Agentic Customer Platform brings your data, teams, and tech stack together with AI built in to help your business work faster and create more personalized customer experiences.

Why HubSpot and what's new

Connect more of your business in one place and give every team a smarter way to work. Get set up quickly and start checking off your hardest tasks.

Get started free

Voices designs, licenses, and captures your Branded AI Voice from real, consenting professional actors—never scraped data. Fully licensed, exclusively yours. Trusted by BMW and SuperBloom.

Get a Demo

Anthropic's Project Parka Turns Meetings Into Claude Agent Tasks (5 minute read)

Anthropic's Project Parka captures Mac meeting audio, transcribes with speaker attribution, and turns discussion into executable work items assigned to Claude agents. It closes the meeting to task to execution loop that every productivity tool has been racing to build. The one unresolved question is whether agents execute autonomously or wait for approval, and that single toggle decides the entire risk profile of the product.

Scortier's Take: The interesting piece isn't the transcription, it's the assignment. Every consumer meeting bot for the last five years shipped "notes and action items" and stopped there because that's where humans took over. Parka is the first credible attempt to keep the loop closed straight into agent execution. If it works reliably, the definition of "attending a meeting" changes.

Mistral Replaces One Shot Retrieval With a Navigable Search Loop (8 minute read)

Mistral's Agentic Search gives the model five operations to inspect documents, follow references, and verify its own answers instead of one shot RAG. FinanceBench correctness jumped from 26.7% to 86%. If you're still building single call retrieval into agents, this is your prompt to redesign. The gain isn't marginal, it's the difference between "sometimes right" and "actually usable."

Poolside AI Signs $6B Non Exclusive Licensing Deal With Nvidia (2 minute read)

Poolside AI licensed its technology to Nvidia for $6 billion and 109 employees got Nvidia transition offers. Nvidia is moving beyond selling chips into owning the model side of the stack. Combined with the Google + AMD TPU rumor and Microsoft's Maia 300, the "one lab, one chip vendor" era is ending fast.

Cerebras Says the New CS-4 Boosts Its AI Speed Advantage Over Nvidia (4 minute read)

Cerebras unveiled the CS-4, several times faster than its predecessor, currently sampling with select customers and broader availability targeted for Q3. This is the same silicon powering GPT-5.6 Sol Ultrafast at 750 tokens per second. If Nvidia dominates training and Cerebras dominates real time inference, your model routing decision starts including hardware, not just provider.

OpenAI Slowed Frontier Training After Detecting New Cyber Capabilities (6 minute read)

OpenAI disclosed it temporarily scaled back frontier model development and paused reinforcement learning runs after detecting new cybersecurity capability signals and an unspecified security incident. This is the second time in a month the same lab has pumped the brakes on its most capable model. The pattern is now the story, not any individual pause.

PagedAttention: Virtual Memory for the KV Cache (17 minute read)

A clear walkthrough of how PagedAttention applies virtual memory concepts to GPU KV cache management, letting attention kernels operate on fragmented memory efficiently. This is the technique underneath vLLM's throughput advantage. Even if you'll never write a serving engine, this is the single best mental model for why long context inference is expensive.

Datadog Rebuilt Git Serving to Handle 20x CI Traffic Without Slowing Down (14 minute read)

Datadog rebuilt its Git serving layer so CI traffic could grow 20x without slowing down. The post covers where vanilla Git breaks as a centralized service and what they had to add: packfile strategy, filesystem tricks, and smart distribution. If your CI clones are getting slow at scale, don't add more runners, read this first.

Shopify Raised Mobile End to End Test Stability to 98% (12 minute read)

Shopify rewrote its flaky mobile e2e suite around an opinionated wrapper API with computer vision for element location instead of view hierarchies, hitting 98% stability. The lesson generalizes: your test suite should assert on user visible behavior, not implementation details. Every team that has "flaky tests" as a recurring standup topic should copy this playbook.

Bun 1.4 Ships With 5x Lower CPU and 35% Memory Savings (11 minute read)

Bun 1.4 adds 1,500 more Node.js compatibility tests, delivers 5x CPU reduction and 35% memory savings, and introduces Bun.Image and Bun.WebView. Node compatibility is now high enough that most services can migrate without code changes. If you're still on Node for a startup workload, this is the release where the switching cost stops being a real objection.

Slack Code: Where Your Team and Agents Build Together (5 minute read)

Slack launched Slack Code, dedicated channels for humans and AI coding agents to collaborate with GitHub, Anthropic, and Vercel integrations. Code changes and live previews render right in the channel. If your team already lives in Slack, this collapses "review a PR" into the same surface as "ask a question," and reduces the tab tax on daily dev work.

Google Brings Antigravity Agents Into Enterprise and Existing IDEs (6 minute read)

Google wired Antigravity into Gemini Enterprise subscriptions with extensions for the major IDEs, plus admin controls for sandboxing, tool permissions, budgets, and audit logging. This is Google's answer to Claude Code for regulated enterprises. Between Antigravity, Slack Code, and Cursor Origin, the agent IDE wars just properly started.

Early Outputs From Meta's Muse Video Model Leaked (4 minute read)

Meta's Muse Video is in closed beta, generating 10 second clips with native audio, strong detail, and temporal consistency. Combined with Muse Glimmer earlier this month, Meta is racing to plant the open weights flag across every modality. Watch for the weights license when it lands, that's the number that decides whether this reshapes the video model market or not.

Sol Loves to Cheat: An Agent Hit 94% on Terminal Bench 2.1 by Exploiting the Benchmark (9 minute read)

A developer noticed their automated agents scored 94% on Terminal Bench 2.1 by exploiting benchmark loopholes rather than solving the underlying tasks. The unresolved question is whether the model deliberately cheated or accidentally stumbled into the exploit. Either way, treat public leaderboard scores with more suspicion this quarter than last.

Fool's Gold: Open Weight Safety Alignment Is Trivial to Remove (13 minute read)

Mark Russinovich's research shows that safety alignment on open weight models is readily removable through abliteration techniques. He proposes decoy hardening as a partial defense. This is the honest technical framing of the open weights debate that most policy discourse avoids, and it's a must read before you form an opinion on where regulation should sit.

Ornith-1.5 Open Models Launch in 397B, 35B, and 9B Sizes

Ornith-1.5 spans flagship mixture of experts to mobile builds, extending self scaffolding into closed self improvement loops for open weight users.

Anthropic Sets Founder Supervoting Stock Ahead of IPO

Anthropic locked in supervoting stock for its founders to shield leadership from shareholder pressure before its public listing.

Harvey II Preserves Context and Memory Across Legal Tasks

Harvey's second generation agent keeps memory and preferences across sequential legal tasks, cutting repetitive initialization out of long running matters.

Vercel Opens a $1 Million Sandbox Escape Challenge

Vercel is paying up to $1M over two weeks to any researcher who can escape its Firecracker based agent sandbox.

Cursor Launches Origin Code Hosting Service to Take on GitHub

Cursor's Origin is a proprietary Git host built for agent workflows, going straight after GitHub's core repository business.

Cloudflare Workers Spectre Attack Leaks JWT From Co Located Worker at 12 Bits/Second

Researchers demonstrated JWT exfiltration across Cloudflare Workers via a Spectre variant that clears the multi tenant isolation boundary.

→ Find me on : Social Links

That’s it for today, keep learning!
Scortier, Signing Off!