Performance Marketing Manager
About the job
- At least 5 years of experience running and scaling paid performance campaigns at high-growth B2B SaaS, developer tools, or consumer technology companies.
- Demonstrated AI-native work habits, including building agents, automations, or workflows that materially change operating practices.
- Hands-on expertise with Meta Ads, Google Search and YouTube, and LinkedIn Ads.
- Track record managing $1 million to $5 million or more in annual paid media budgets and exceeding return on ad spend and customer acquisition cost goals.
- Fluency in data and analytics, campaign data storytelling, and testing and evaluating audience performance.
- Strong copywriting instincts and performance-creative judgment, with the ability to scale that judgment using AI.
- Clear written and verbal communication for technical developers and enterprise buyers.
- Own paid acquisition end to end across Google Search and YouTube, Meta, LinkedIn, and emerging channels, driving signups, activations, and pipeline.
- Manage multi-million-dollar annual paid media spend, exceed return on ad spend and customer acquisition cost targets, and lower customer acquisition cost over time.
- Build agents, automations, and AI-assisted workflows to accelerate campaign turnaround and scale creative production.
- Systematically test creative and audiences, identify winning creative and audience segments, and use AI to generate and iterate variants at high velocity.
- Build funnel reports and cohort analyses with Analytics, and own event-based tracking, pixels, and conversion APIs to maintain clean data.
- Provide leadership with actionable insights and clear performance recommendations on a regular cadence.
- Align with Product Marketing, Demand Generation, Lifecycle, Creative, and Marketing Operations on targeting, messaging, and down-funnel conversion.
- TikTok Ads or programmatic advertising experience.
About Vercel:
Vercel is the agentic infrastructure company, freeing people and agents to ship what's next. For more than a decade we've helped builders move from idea to production with speed, security, and exceptional developer experience.
Now we're scaling our products for both agents and people to ship and run software, built in the open and trusted by OpenAI, PayPal, Ramp, Supreme, and millions of developers worldwide.
About the role
You'll own paid acquisition at Vercel across Google, Meta, LinkedIn, YouTube, and emerging channels, with one goal driving the work: scalable, profitable customer acquisition for our self-serve and enterprise motions.
This role is built for an analytical, AI-native growth marketer who cares about results, testing velocity, and scale. We want someone who has already wired AI into how they work, building agents, automations, and workflows that let them operate like a much larger team. You'll bring operational rigor to how we launch and scale campaigns, own the systems that make us faster, and turn our acquisition engine into a compounding growth loop. The work sits between strategy and execution, in close partnership with Product Marketing, Lifecycle, Creative, Data, and Marketing Operations, so the right message reaches the right developer or buyer in the right channel.
What you'll do
- Paid acquisition end to end across Google Search and YouTube, Meta, LinkedIn, and emerging channels, driving signups, activations, and pipeline
- Multi-million-dollar annual paid media spend, where you beat ROAS and CAC targets and lower CAC over time
- AI-driven operating leverage, building agents, automations, and AI-assisted workflows that compress campaign turnaround and scale creative production
- Creative and audience testing, where you systematically find (and sometimes design) winning creative and audience segments, using AI to generate and iterate variants at high velocity
- Measurement and attribution, building funnel reports and cohort analyses with Analytics, and owning event-based tracking, pixels, and conversion APIs so the data stays clean
- Performance reporting, where you surface actionable insights and clear recommendations to leadership on a regular cadence
- Cross-functional alignment with Product Marketing, Demand Generation, Lifecycle, Creative, and Marketing Operations on targeting, messaging, and down-funnel conversion
What you need
- 5+ years running and scaling paid performance campaigns at high-growth B2B SaaS, developer tools, or consumer tech companies
- Demonstrated AI-native habits, where you reach for AI by default and have built agents, automations, or workflows that changed how you operate (if you've automated parts of your own job, that's a strong signal)
- Hands-on expertise with Meta Ads, Google Search and YouTube, and LinkedIn Ads, with TikTok Ads or programmatic as a bonus
- A track record managing $1M to $5M+ in annual paid media budgets and beating ROAS and CAC goals
- Fluency in data and analytics, storytelling with campaign data, and testing and evaluating audience performance
- Strong copywriting instincts and taste in performance creative, and ability to scale taste with AI
- Clear written and verbal communication that lands with both technical developers and enterprise buyers
- A bias for action - you can just ship things.
Benefits & Compensation:
- Competitive compensation package, including equity.
- Inclusive Healthcare Package.
- Learn and Grow - we provide mentorship and send you to events that help you build your network and skills.
- Flexible Time Off.
- We will provide you the gear you need to do your role, and a WFH budget for you to outfit your space as needed.
The San Francisco, CA base pay range for this role is $150,000 - $188,000. Actual salary will be based on job-related skills, experience, and location. Compensation outside of San Francisco may be adjusted based on employee location. The total compensation package may include benefits, equity-based compensation, and eligibility for a company bonus or variable pay program depending on the role. Your recruiter can share more details during the hiring process.
Disclosures:
- Privacy: Please review our Job Applicant Privacy Policy for more information on how we handle your data.
- Equal Opportunity: Vercel is committed to fostering and empowering an inclusive community within our organization. We do not discriminate on the basis of race, religion, color, gender expression or identity, sexual orientation, national origin, citizenship, age, marital status, veteran status, disability status, or any other characteristic protected by law. Vercel encourages everyone to apply for our available positions, even if they don't necessarily check every box on the job description.
About the company
Vercel runs a cloud platform for software teams to build, deploy and operate web applications and AI agents without managing their own servers. Its tools connect code repositories to automated builds and give teams separate environments for testing changes before publishing them. The platform combines application hosting, global content delivery and security controls, while its AI Gateway connects applications to models from several providers. Developers start on a free plan, and professional teams pay subscriptions, metered infrastructure charges or negotiated enterprise contracts. Vercel was founded in 2015 and is headquartered in San Francisco, California, United States.
Company Size
501-1,000
Company Stage
Series F
Total Funding
$863M
Headquarters
San Francisco, California
Founded
2015
Get referred to Vercel
See people who can refer or advise you
Simplify's Take
What believers are saying
- Developer reach is broadening: Vercel reported nearly 280 million skills.sh installs in September 2026.
- Enterprise purchasing gets easier: Vercel's AI Gateway sells through AWS Marketplace using customers' existing cloud commitments.
- AI engineers build across layers: Vercel's Gateway role spans provider integrations, distributed systems, and open-source contributions.
What critics are saying
- Customer trust faces repair: Vercel's April 2026 breach exposed variables not marked sensitive for a limited customer subset.
- Framework adoption does not ensure hosting revenue: Next.js's MIT license permits free use, distribution, and competing deployments.
- Nearby CDN engineers face commutes: Vercel requires Monday, Tuesday, and Friday office days for this role.
What makes Vercel unique
- Framework authors also run the hosting: Vercel builds Next.js and optimizes deployment infrastructure around its features.
- Waiting costs less: Vercel's Fluid Compute shares instances across requests and bills active CPU instead of idle time.
- Model outages have a fallback: Vercel's AI Gateway automatically routes the same model to another provider.
Help us improve and share your feedback! Did you find this helpful?
Benefits
- Health Insurance
- Stock Options
- Company Equity
- Professional Development Budget
- Unlimited Paid Time Off
- Remote Work Options
- Home Office Stipend
Growth & Insights and Company News
6 month growth
2%1 year growth
3%2 year growth
2%Israeli startup NanoCo, which builds security and governance controls for AI agents, has raised a $12 million seed round led by Valley Capital Partners. Docker, Vercel, and monday.com also took part. Founded by brothers Lazar and Gavriel Cohen, the company grew out of their frustration with OpenClaw, an agent platform they doubted could be trusted with company data. They built NanoClaw, an open-source alternative with about 4,000 lines of code, against OpenClaw's half a million. It runs agents in isolated sandboxes, and a gateway lets a human approve or deny requests. NanoClaw has gained more than 30,000 stars on GitHub. NanoCo's enterprise product deploys in a customer's own cloud, so the customer keeps control of its data. The company has about 20 employees across three countries.
Vercel's AI SDK ships Web Fetch Tool that works across any AI model. Vercel's AI SDK adds a first-party web fetch tool that returns pages as markdown, works with any function-calling model, and ships with domain allowlists. · 1 day ago Read 6 min * Vercel AI SDK adds a Web Fetch Tool that works with any function-calling model. * Pages are returned as markdown using defuddle, cutting token usage versus raw HTML. * Install via pnpm add @ai-sdk/tool-web-fetch, requires Node.js 22.13 or newer. * Security controls: allowedDomains, blockedDomains, requireUrlInUserMessage, and DNS rebinding protection on Node. * Supports HTML, text, JSON, XML, CSV; no PDFs, binaries, or JavaScript-rendered pages. * Built-in pagination via offset with short-lived per-instance cache; default 50k char limit, 10 MiB cap. Vercel's AI SDK team has released a first-party tool that lets agents fetch publicly reachable web pages with any AI SDK model that supports tool calling. The Web Fetch docs describe an integration that accepts a URL, retrieves the page, removes common boilerplate, and returns content the model can process. For server-rendered, text-based pages, the package can replace a custom scraping service, headless browser, or HTML-to-text pipeline. It runs as a standard AI SDK tool, so the same integration continues to work when an application changes model providers. A small integration surface. Install the separate package, register the tool with generateText, streamText, or ToolLoopAgent, and set a step limit to prevent an unbounded tool loop: npm install @ai-sdk/tool-web-fetch import {experimental_createWebFetchTool as createWebFetchTool,} from '@ai-sdk/tool-web-fetch'; import {generateText, isStepCount} from 'ai'; const result = await generateText({ model, tools: {webFetch: createWebFetchTool,}, stopWhen: isStepCount(5), prompt: 'Summarize https://ai-sdk.dev/docs/introduction.',}); console.log(result.text); The exported factory currently carries an experimental_ prefix, so teams should treat its API as subject to change. The isStepCount(5) condition caps the number of model and tool steps during this run. Cleaner input, fewer tokens. The tool requests Markdown by sending Accept: text/markdown. When a server returns HTML, the package uses Defuddle to extract the main content and remove navigation, footers, scripts, and other page furniture. | Response type | Tool behavior | | Markdown | Returns the server-provided content directly. | | HTML | Extracts the main content and converts it to Markdown. | | Plain text, JSON, XML, or CSV | Returns the text without conversion. | | PDF, image, audio, video, or another binary format | Rejects the response as unsupported. | Removing HTML tags, inline styles, scripts, and tracking markup reduces the tokens consumed by each page. The savings grow when an agent fetches several URLs within one tool loop. The model calls the tool with {url, offset?}. A successful response includes the requested URL, final URL after redirects, content type, title, extracted content, and pagination metadata. Failures return structured error information that the model can use to revise the URL or choose another source. Long pages stay bounded. Each call returns up to the configured maxLength, which defaults to 50,000 characters, and applies a 10 MiB response ceiling. When a document exceeds the character limit, the response provides an offset that the model can pass into a subsequent call. Processed documents are cached for a few minutes within each tool instance, allowing offset requests to continue reading without downloading and converting the page again. The cache belongs to that instance rather than serving as a shared application cache. Guardrails for hostile content. Fetched text must be treated as untrusted input because a page can contain instructions intended to override the agent's task. A compromised model could also place sensitive data in a query string and request an attacker-controlled URL. * allowedDomains limits requests to specified domains and their subdomains. * blockedDomains denies requests to listed domains. * requireUrlInUserMessage permits only URLs found in user messages. URLs introduced by system prompts, assistant messages, fetched pages, or tool results remain ineligible. * AI SDK tool approval can require a person to authorize each fetch. Domain restrictions are checked again across redirects. The transport also sends GET requests without cookies or credentials, rejects URLs containing embedded credentials, and blocks private or reserved addresses such as localhost and cloud metadata endpoints. On Node.js, the tool validates resolved IP addresses when the connection opens. This check reduces DNS rebinding risk, where a hostname initially resolves to a public address and later resolves to an internal service after validation. Choose controls by workflow. * User-submitted URL summaries: Enable requireUrlInUserMessage, add an allowlist when the expected domains are known, and consider approval for sensitive environments. * Citation-following research agents: Disable requireUrlInUserMessage only when the agent must follow links discovered in search results or fetched pages. Constrain egress with domain rules, a proxy, or network policy. * Internal applications: Keep private address blocking in place and avoid passing session cookies or service credentials through a custom transport. Where fetching stops. * No web discovery: The tool retrieves a known URL. Agents that need to find sources require a separate search tool. * No JavaScript execution: Client-rendered applications may return an empty shell or incomplete content because the package does not run a browser. * No authenticated browsing: Requests omit cookies and credentials, so pages behind a sign-in flow are generally unavailable. * No binary extraction: PDF-heavy workflows still need a document parser or dedicated ingestion pipeline. * Imperfect content detection: Automatic extraction can select the wrong section on unusual layouts. Known page structures can use an option such as defuddleOptions.contentSelector: 'article'. Portability without provider lock-in. Provider-native browsing features vary in their search, retrieval, citation, and execution behavior. The AI SDK implementation focuses on fetching known URLs inside the application's process, giving developers direct control over domain policy, logging, caching, and network routing. | Concern | AI SDK Web Fetch | | Model support | Any compatible model with tool calling | | Execution | The application's runtime | | URL discovery | Requires a separate search source | | Network control | Domain rules and a customizable fetch implementation | | Content handling | Text responses and extracted HTML | A custom fetch implementation can route traffic through a proxy or existing egress layer. Normal model, bandwidth, and compute costs still apply, while the application avoids depending on a provider-specific browsing interface. The package also complements the Bash Tool docs, which cover command execution in environments such as Vercel Sandbox or the in-memory just-bash runtime. One tool retrieves web content; the other runs code against files and command-line programs. A production checklist. * Pin package versions while the factory remains experimental. * Set stopWhen and choose a suitable maxLength. * Enable domain restrictions and requireUrlInUserMessage where the workflow allows them. * Log requested URLs, final redirect URLs, content types, tool errors, and approval decisions. * Test redirects, oversized documents, client-rendered pages, malformed content, and prompt injection attempts. * Provide separate tools for search, PDFs, authenticated pages, or browser-rendered applications. For existing AI SDK applications, integration requires one package and one tool registration. Production deployment also requires explicit egress rules, bounded tool loops, and testing against the page types the agent will encounter.
Nightview Capital LLC makes new Investment in Shopify Inc. $SHOP. October 21, 2026 Key points. * Nightview Capital LLC acquired 33,180 Shopify shares worth approximately $4.92 million in the third quarter, making Shopify about 3.7% of its portfolio. Institutional investors and hedge funds collectively own 69.27% of the company. * Shopify reported quarterly revenue of $3.58 billion, up 33.7% year over year, and earnings of $0.42 per share, beating estimates. Analysts remain broadly positive, with a "Moderate Buy" consensus rating and an average price target of $175.89. * The stock has risen sharply and trades at a high valuation, with a P/E ratio of 117.22 and a beta of 2.62. While AI initiatives, platform expansion and improving earnings expectations support growth, the recent rally increases the risk of volatility or profit-taking. * Want to know the moment analysts change their view on Shopify? Get MarketBeat All Access for $149 a year and keep that rate for life. Nightview Capital LLC bought a new position in Shopify Inc. (NASDAQ:SHOP - Free Report) TSE: SHOP during the 3rd quarter, according to the company in its most recent filing with the Securities and Exchange Commission. The firm bought 33,180 shares of the software maker's stock, valued at approximately $4,921,000. Shopify accounts for about 3.7% of Nightview Capital LLC's portfolio, making the stock its 12th biggest position. A number of other institutional investors and hedge funds have also recently bought and sold shares of the company. Norris Financial Group LLC bought a new position in Shopify in the 1st quarter worth approximately $1,586,000. Infusive Asset Management Inc. lifted its holdings in shares of Shopify by 27.4% during the first quarter. Infusive Asset Management Inc. now owns 43,611 shares of the software maker's stock worth $5,173,000 after buying an additional 9,374 shares during the last quarter. OMERS ADMINISTRATION Corp boosted its position in shares of Shopify by 18.6% in the first quarter. OMERS ADMINISTRATION Corp now owns 58,044 shares of the software maker's stock valued at $6,876,000 after acquiring an additional 9,096 shares during the period. Ninepoint Partners LP bought a new stake in shares of Shopify in the second quarter valued at approximately $29,027,000. Finally, The Manufacturers Life Insurance Company boosted its position in shares of Shopify by 8.5% in the first quarter. The Manufacturers Life Insurance Company now owns 8,418,903 shares of the software maker's stock valued at $998,137,000 after acquiring an additional 660,198 shares during the period. Institutional investors and hedge funds own 69.27% of the company's stock. Shopify news summary. Here are the key news stories impacting Shopify this week: * Analyst optimism and earnings outlook: Zacks upgraded Shopify to a #1 (Strong Buy) rating, citing improving earnings prospects. Other coverage points to expected revenue growth of approximately 31.5% this year, rising earnings estimates and expanding enterprise relationships with brands such as Dior and Clinique. * AI and platform expansion: Shopify is supported by Vercel Shop's open-source Next.js integration, enabling merchants to build AI-focused storefronts and use shopping agents for product discovery and checkout. Shopify's AI tools and Platinum membership in the PyTorch Foundation reinforce its strategy to improve merchant productivity. * Institutional support: Carnegie Investment Counsel increased its Shopify position by 3.2% to 115,804 shares, a modest signal of continued institutional confidence. * Product development: Shopify POS reportedly added a Smart Grid feature that recommends and optimizes tiles using sales data and usage frequency, potentially improving merchant workflow and platform engagement. * Valuation and profit-taking risk: Shopify has rallied about 35% in a month and trades at a high earnings multiple. Commentary increasingly asks whether investors should take profits rather than add exposure, raising the likelihood of volatility or a pullback despite strong long-term growth expectations. Shopify stock performance. Shopify stock opened at $173.48 on Wednesday. The company has a debt-to-equity ratio of 0.01, a current ratio of 5.35 and a quick ratio of 5.35. The stock has a fifty day moving average price of $149.95 and a two-hundred day moving average price of $129.12. Shopify Inc. has a 52-week low of $94.00 and a 52-week high of $182.19. The firm has a market capitalization of $223.21 billion, a P/E ratio of 117.22, a P/E/G ratio of 3.53 and a beta of 2.62. Discover more Build a Website Launch a Store Shopify (NASDAQ:SHOP - Get Free Report) TSE: SHOP last issued its earnings results on Wednesday, August 5th. The software maker reported $0.42 EPS for the quarter, beating the consensus estimate of $0.40 by $0.02. Shopify had a net margin of 14.53% and a return on equity of 12.58%. The company had revenue of $3.58 billion for the quarter. During the same quarter in the prior year, the firm posted $0.69 EPS. The firm's revenue for the quarter was up 33.7% on a year-over-year basis. As a group, analysts anticipate that Shopify Inc. will post 1.46 EPS for the current fiscal year. Analyst upgrades and downgrades. SHOP has been the topic of several research analyst reports. Piper Sandler restated an "overweight" rating on shares of Shopify in a research note on Monday, October 5th. Cantor Fitzgerald raised their price target on shares of Shopify from $145.00 to $158.00 and gave the stock a "neutral" rating in a research report on Monday, October 12th. Robert W. Baird boosted their price objective on shares of Shopify from $160.00 to $190.00 and gave the stock an "outperform" rating in a report on Wednesday, October 14th. Zacks Research upgraded Shopify from a "hold" rating to a "strong-buy" rating in a research report on Friday, October 16th. Finally, Rosenblatt Securities initiated coverage on shares of Shopify in a research report on Thursday, August 20th. They issued a "buy" rating and a $175.00 price target for the company. Three equities research analysts have rated the stock with a Strong Buy rating, thirty-six have given a Buy rating and nine have given a Hold rating to the company. According to MarketBeat, Shopify has an average rating of "Moderate Buy" and an average target price of $175.89. Shopify profile. Shopify Inc is a commerce technology company that provides tools for businesses of various sizes to start, operate and grow online and in physical retail environments. Its platform supports online storefronts, product and inventory management, order processing, marketing, analytics and customer engagement. The company offers subscription-based plans and merchant solutions, including Shopify Payments, Shopify POS, Shopify Capital and tools for shipping, checkout and fulfillment. Shopify also supports sales through multiple channels, such as merchants' own websites, mobile devices, social media and third-party marketplaces. Shopify was founded in 2006 by Tobias Lütke, Daniel Weinand and Scott Lake after the founders developed an online store for snowboarding equipment. This instant news alert was generated by narrative science technology and financial data from MarketBeat in order to provide readers with the fastest reporting and unbiased coverage. Please send any questions or comments about this story to [email protected]. Before you consider Shopify, you'll want to hear this. MarketBeat keeps track of Wall Street's top-rated and best performing research analysts and the stocks they recommend to their clients on a daily basis. MarketBeat has identified the five stocks that top analysts are quietly whispering to their clients to buy now before the broader market catches on... and Shopify wasn't on the list. While Shopify currently has a Moderate Buy rating among analysts, top-rated analysts believe these five stocks are better buys. Tesla, Nvidia, and Google helped shape the last era of market growth, but the next wave could come from a new group of companies. Inside this report, you'll find 7 stocks that could play a major role in the next tech-driven market boom. Continue following MarketBeat
AI daily|october 18, 2026. 12stories5sections10sourcesAbout 4 min DeepSeek has launched version V4.1-Flash-1017, delivering substantial performance gains in agentic and coding benchmarks alongside faster execution and lower token usage. This brief is compiled automatically by AI and may contain errors. Check the linked sources. * Bottom line:DeepSeek has launched version V4.1-Flash-1017, delivering substantial performance gains in agentic and coding benchmarks alongside faster execution and lower token usage. * Benchmark Performance:Achieved significant improvements on complex agent evaluations, raising Terminal-Bench 4.0 scores from 31.2 to 42.3. * Pareto Efficiency:Delivers faster task completion with shorter outputs, reducing token consumption and task costs without raising API pricing. * Open-Weights Availability:Released under an MIT license with checkpoints available on Hugging Face and ModelScope for self-hosting. * Source: 在 X 上查看 @OpenRouter 的貼文 * Bottom line:JetBrains has open-sourced Mellum 2.1, a 12 billion parameter model optimized specifically for low-latency coding agent workflows. * Parameter Size:Features 12B parameters tailored for developer tooling and autonomous programming tasks. * Open-Source Release:Made openly available by JetBrains to support community-driven coding agent ecosystems. * Source:JetBrains AI Blog * Bottom line:Kyutai and academic partners have released REKI, a 2.4B VLA model designed to provide efficient memory and control for robotic manipulation. * Architecture:Built with a compact 2.4 billion parameter Vision-Language-Action architecture. * Pre-training Data:Pre-trained exclusively on public data and released on Hugging Face under an open license. * Source:MarkTechPost Coverage * Bottom line:Vercel has introduced session affinity in AI Gateway logs to automatically route sequential chat turns to the same provider for maximum prompt caching benefits. * Routing Logic:Directs consecutive turns to the provider that served the previous request, retaining cached prompt states. * Session Analytics:Aggregates requests into session views displaying total costs, models, provider pathways, and cache hit rates. * Source:Vercel Changelog * Bottom line:DeepSeek has opened its proprietary enterprise search and web fetching APIs to developers, enabling real-time grounding for agent harnesses. * Core Capabilities:Combines multi-language web search and URL text extraction to supply verified context to LLM workflows. * Pricing Structure:Priced per successful request starting November 1, with unsuccessful or timed-out queries incurring no charge. * Source:DeepSeek API Documentation * Bottom line:OpenCode disclosed a security breach affecting 1.4% of its Console users linked to a large-scale distillation attack, prompting urgent API key revocations. * Impact Scope:Impacted 1.4% of OpenCode Console users who had third-party provider keys (OpenAI, Anthropic, Google) stored in the platform. * Security Remediation:Affected users were notified via email and advised to immediately revoke compromised keys and update dependent systems. * Source: 在 X 上查看 @opencode 的貼文 * Bottom line:Meta updated its commercial terms to remove phrasing that legal experts previously interpreted as allowing publisher domain scraping for AI model training. * Policy Adjustment:Removed specific language suggesting commercial account data could be used to train Meta products and AI models. * Official Stance:Meta clarified that its web crawler practices remain unchanged and scraping terms apply strictly to ads and commercial products. * Source:The Verge Report * Bottom line:An internal Anthropic report showed that junior technical hiring dropped significantly as internal agent work hours reached six times human operating levels. * Hiring Changes:Junior technical employee representation fell from 21% to 14%, and junior hiring share decreased from 27% to 9%. * Agent Workload:Telemetry data indicated that interactive AI agent work hours scaled to six times human operator levels by August 2026. * Source:TechCrunch Coverage * Bottom line:Researchers demonstrated an 8B reasoning model trained to propose diverse and novel research ideas in a single generation by weighting conceptual distance. * Methodology:Trains a small reasoning model to generate sets of research ideas while penalizing similarity to previously generated concepts. * Evaluation:Outperformed larger frontier models like Claude Opus 4.8 and GPT-6 Astra on creative utility and diversity metrics. * Source:ArXiv Paper * Bottom line:Tencent researchers proposed a recursive self-improvement framework using verifier-guided fine-tuning and PPO to dramatically elevate terminal coding agents. * Performance Gains:Elevated the four-suite Terminal-Bench average from 20.68% to 50.58%, scoring 86.52% on Terminal-Bench 2.1. * Training Architecture:Alternated between fine-tuning on verified successful trajectories and running PPO with an inherited critic across five stages. * Source:ArXiv Paper * Bottom line:Anthropic's Claude successfully reverse-engineered and cracked a 25-year-old NIST elliptic curve cryptography seed originating from the NSA in under 30 minutes. * Background:Addressed the NIST standardized ECDSA curve seed from 2000 whose unspecified NSA origins long prompted backdoor speculation. * Resolution Speed:Solved the complex cryptographic mystery in less than 30 minutes using advanced chain-of-thought reasoning. * Source:The Decoder Coverage * Bottom line:The Claude Code team published an engineering breakdown detailing how they overhauled code reviews as AI agents generate roughly 90% of their codebase. * Efficiency Metrics:Despite merged pull requests doubling between June and September, code revert and incident rates dropped by over 40%. * Multi-Layered Checks:Employs a Swiss-cheese verification model featuring self-checks, independent reviewer agent pipelines, and automated CI tests. * Source:Claude Engineering Blog Thanks for reading to the end. Stay curious. See you tomorrow. Ad · Scribis Experience Scribis: ultimate AI audio & video workflow. Scribis is an all-in-one AI audio tool designed for creators, researchers, and professionals. It combines local privacy computation with cloud API integration, providing millisecond-level speech-to-text, speech synthesis, and dynamic timeline editing.
How free low-latency embedding engines can unlock real-time document AI at scale. *Today's briefing looks at Vercel's free Beam model, a new DeepSeek update, a massive Indic pre-training effort, and improved observability for AI Gateway - all of which reshape the data platforms that power embeddings and document-level intelligence.* Lead story - Vercel ships Beam on AI Gateway for free. Vercel has added Beam, a low-latency, open-weight model from Reflection AI, to its AI Gateway and made it free to use during the launch window [Vercel AI](https://vercel.com/changelog/reflection-ai-beam-now-available-on-ai-gateway). Beam is hosted in the US, offers sub-second response times, and is positioned as an "open-w" model for embedding generation and lightweight reasoning. Why it matters: Most document-AI pipelines - RAG, classification, summarisation - rely on an embedding service that can keep up with high-throughput ingestion. Traditional hosted embeddings (e.g., OpenAI's text-embedding-ada-002) charge per-token and can become a bottleneck when scaling to millions of pages. Beam's free tier eliminates per-token fees and, thanks to its low latency, allows near-real-time indexing of newly created or updated documents. For a retail catalog that adds hundreds of product descriptions hourly, or a clinic that continuously receives scanned records, the ability to embed on the fly without queuing dramatically shortens the "data-to-insight" lag. Practical implication for businesses: 1. Plug-and-play integration - AI Gateway provides a standard HTTP endpoint, so existing ingestion services can swap in Beam with a single URL change. No SDK overhaul is required. 2. Cost-first experimentation - Since Beam is free for the launch period, product teams can prototype RAG-style support bots, compliance checkers, or personalised recommendation engines without worrying about embedding-related OPEX. 3. Latency-driven architecture - Sub-second embeddings make it feasible to run "on-the-edge" workflows (e.g., a browser-based document uploader that instantly surfaces relevant policies) while still centralising the model in a secure Vercel data centre. The caveat is that Beam's free window is limited; businesses should treat it as a pilot and plan migration paths (e.g., to a self-hosted open-weight model or a paid Vercel tier) before the promotion ends. Nevertheless, the announcement signals a shift toward free, low-latency embedding services that can be the backbone of real-time document AI stacks. DeepSeek V4.1 Flash 1017 raises the bar for multimodal embeddings. Vercel also rolled out DeepSeek V4.1 Flash 1017 on AI Gateway, a minor yet impactful upgrade that improves reasoning, coding, agentic tasks, and image understanding [Vercel AI](https://vercel.com/changelog/deepseek-v4-1-flash-1017-now-available-on-ai-gateway). The model's enhanced image-to-text capabilities mean it can generate visual embeddings alongside text, opening the door to richer document AI that handles scanned forms, receipts, or medical imaging alongside OCR text. *Implication:* Retailers can embed product photos together with descriptions, improving visual search relevance. Clinics can index radiology images with accompanying notes, enabling multimodal retrieval for diagnostic support. The upgrade is available through the same Gateway interface, so teams using Beam for pure text can add DeepSeek for mixed-media pipelines without additional infrastructure. Hugging Face reveals a 2.8-trillion-token Indic pre-training pipeline. A joint blog from Hugging Face and NVIDIA details an open data generation pipeline that pre-trained large language models on 2.8 trillion Indic tokens [Hugging Face](https://huggingface.co/blog/nvidia/scaling-indic-pretraining). The effort showcases how to build a scalable, reproducible data platform that harvests raw text, cleans it, and feeds it into distributed training jobs. *Why it matters for embeddings:* The pipeline's emphasis on open, version-controlled data assets and automated tokenisation provides a template for organisations that need domain-specific multilingual embeddings (e.g., a pan-Indian retail chain handling Hindi, Tamil, and Bengali product catalogs). By re-using the same pipeline, businesses can generate high-quality embeddings that respect local linguistic nuances, improving downstream retrieval accuracy. AI Gateway logs now expose session affinity - better observability for embedding pipelines. Vercel's latest logging feature groups requests into sessions and shows how "session affinity" routes each one [Vercel AI](https://vercel.com/changelog/ai-gateway-logs-now-show-session-affinity). For teams that stitch together multiple calls - file upload | OCR | embedding | RAG - this visibility is crucial. It lets engineers spot latency spikes, identify which model (e.g., Beam vs. DeepSeek) is handling a given session, and attribute costs accurately. *Practical tip:* Enable session logging on your embedding endpoint and correlate the session IDs with downstream RAG request logs. This end-to-end trace helps you fine-tune batch sizes, cache strategies, and fallback models, ensuring that real-time document AI remains both fast and cost-predictable. Enterprise workflow platforms are catching up. A survey of 12 enterprise workflow automation platforms notes that many now include native connectors for AI services, including Vercel's AI Gateway [n8n](https://blog.n8n.io/enterprise-workflow-automation/). While the article is a broader market overview, the key takeaway is that modern orchestration tools are ready to call out to low-latency embedding APIs as part of automated data-ingestion flows. This convergence reduces the need for custom glue code and accelerates time-to-value for AI-driven processes. IOTA's take. For clinics, retailers, and SaaS teams, the emerging ecosystem of free, low-latency embedding models (Beam, DeepSeek) combined with robust data-pipeline tooling (Hugging Face's Indic pipeline) and observable routing (AI Gateway session logs) means you can now build truly real-time document-AI services without ballooning budgets. Start by routing new documents through Vercel's Gateway, prototype with Beam, and layer in multimodal embeddings from DeepSeek as your use cases mature. The payoff is faster knowledge retrieval, higher user satisfaction, and a clearer line of sight on AI spend.