Pentagon RFP Overhaul, Perplexity's Open Embedding Model, and Vercel Connect Lead This Week's Developer Brief

Key Takeaways

1. Emil Michael is pushing to replace 300-page Pentagon RFPs with one-page problem statements and 90-day feedback loops, while the Drone Dominance program uses a fly-off format that ends in a $500M order split across 2-3 winners.
2. Perplexity opened email-based Computer tasks to anyone, open-sourced a context embedding model it says tops turbopuffer's context-bench, and rolled out inline charts and TradingView financial visualizations.
3. Vercel launched Connect to replace static agent credentials, and its AI Gateway is being positioned as a trusted token-flow data source with Microsoft AI models arriving day zero.
4. Grok Bot shipped team-shared bots, Plaid financial search, Cursor handoff, and GitHub/Origin PR management, with SpaceXAI reportedly using it to build itself.
5. Gemini 4 Argon's low hallucination rate comes with a lower max-mode accuracy, while GPT-6 Astra decoded a 217-year-old Napoleonic cipher from a single image in about six hours.
6. Jack's Loupe scanned 43 open-source Bitcoin repos producing 26,821 findings, and Cash App shipped a Spanish-language version with AI translation plus human review.
7. Alex Hormozi capped internal AI spend after finding per-employee token usage at 16x the tech-frontier average, tying token approval to demonstrable revenue or cost savings.

1. Defense Procurement and Drone Funding

1. Emil Michael, the U.S. Department of Defense's Under Secretary for Research and Engineering, criticized the Pentagon's 300-page RFP format at the DC Startup Industrial Base event, arguing it is designed for lawyers, lobbyists, and prime contractors: a 10-person startup that misreads one page is eliminated, or endures a 3-to-12-month cycle that stretches to two years, while primes have dedicated teams filling out every page. His proposed fix is a one-page statement such as "fly a missile 500 miles in this environment," with feedback targeted within 90 days. 1
2. Michael said the Drone Dominance program uses an application process, a top-20 fly-off, and objective metrics with round-by-round elimination, ultimately producing a $500M order with 2-3 winners rather than one. He framed the key metric as cost per intercept, arguing that using a Patriot missile against a $50,000 drone does not pencil out, which is why the program funds directed energy and high-power microwave at costs in the thousands rather than millions. 1

2. AI Models, Agents, and Developer Tooling

1. Perplexity opened an email entry point for Computer tasks: users can send, forward, or CC a task to [email protected] without a Perplexity account, free for a limited time, with each email task running as a regular Computer session viewable on web and mobile and retaining the same audit trail as in-app tasks. The company also open-sourced its context embedding model, which it says performs best on turbopuffer's context-bench; the pplx-embed-v2-context-9b-preview model is described as encoding each document chunk with the full document as context, and a third-party evaluator said Perplexity iterated over several weeks and at least nine benchmark revisions without privileged access. 1 2 3 4
2. Perplexity is rolling out inline charts, diagrams, and visualizations in Computer, including contextual financial charts from TradingView; the integration uses TradingView Lightweight Charts to render candlesticks, volume, and moving averages directly in the conversation thread. The company also launched an Amex-curated ready-to-use skill library for eligible U.S. Amex Business small-business card members, covering workflows such as cash flow forecasting and marketing campaign generation, and highlighted Seedance 2.5 inside Computer for generating commercial or brand videos. 1 2 3
3. Grok Bot shipped several updates: creating bots shared with team members, searching and managing finances through Plaid, searching chat history during calls and handling interruptions, handing coding tasks to Cursor, managing PRs via GitHub and Origin plugins, and sharing video demos of what was built. The account said Grok Bot combined with cloud agents is excellent and lets users create a dedicated computer for a bot, making it more like a manager than a direct coder; a repost claimed SpaceXAI is using Grok Bot to build Grok Bot, with agents participating in design, coding, testing, and release, and some PRs merged before a human reads them. 1 2 3 4
4. Paul Graham highlighted that Gemini 4 Argon has a 15% hallucination rate on Artificial Analysis, lower than Grok 4.7 (29%), GPT-6 Astra (45%), Opus 5.5 (59%), and Fable 5.1 (69%), but its max-mode accuracy is 50% versus Opus 5.5's 66%, with its advantage being that it says when it does not know rather than fabricating. Separately, GPT-6 Astra was used to decipher an unsolved cipher letter from a general under Napoleon that had gone unread for 217 years; a repost argued the notable part was not the deciphering itself but that Astra completed the entire multimodal workflow from a single image and a goal in about six hours. 1 2

3. Infrastructure, Security, and Platform Launches

1. Vercel launched Connect, opening it to service providers and claiming reach to more than 20 million developers and the billions of agents they will ship; the author argued code is now "free" and the final difficulty in building is connections, with Connect simpler and safer for both agents and apps. He said agent building is broadly stuck on static credentials, requiring many keys per agent or app, which is both a developer-experience disaster and a major liability risk, and Connect is meant to solve that. 1
2. Vercel's AI Gateway is being positioned as the largest trusted data source for understanding global AI token flows, based on real customer usage, a platform of more than 400,000 paying customers and thousands of enterprises, and zero markup; the author said most token aggregators have severe noise from vendors promoting tokens from training on user data and from vendors merely claiming zero data retention. He said Vercel rejects weekly "free token" promotions from companies with questionable ZDR claims or reputations, does not chase the most vendors because most are low quality, and invests equally in gateway engineering and legal, compliance, privacy, and back-office operations to serve trillions of tokens per day responsibly. 1
3. Vercel and Microsoft AI are partnering to bring the latter's models into AI Gateway, available day zero with MAI-Voice-2.1 for long-form and fast-reply speech and MAI-Transcribe-2-Streaming for transcription; the author said Microsoft's AI team is training good models and he looks forward to day-zero availability on Vercel. 1
4. Jack reported that Loupe produced 26,821 findings across 43 open-source Bitcoin repositories, with an average successful scan taking 2 hours 52 minutes and returning 53.6 findings, and the longest scan running 36 hours 58 minutes; 26,809 findings included a proof of concept and 12,774 included a proposed patch. Cash App launched a Spanish-language version with extensive app rewrites, AI translation plus human review, and customer support, with users whose phone language is Spanish set to auto-update within the coming week. 1 2 3

4. Business Signals and Operating Notes

1. Alex Hormozi issued an internal memo capping AI usage: the company's current token spend per employee is 16x the average of frontier tech companies based on Ramp data covering 70,000 businesses, so he will not approve additional AI budget because unlimited budget produces unlimited waste. The memo requires tokens to be used only for projects with clear revenue generation or cost savings, and promises training on judging what is worth doing, getting the same output with fewer tokens, and using AI-saved time more effectively; he noted revenue growth is roughly flat relative to AI usage, suggesting a skills gap where teams either apply AI to unimportant projects or spend saved time on unimportant things. 1
2. Jason Fried said he had wanted for years, possibly nearly a decade, to build a writing tool with only a few unique features that mirrors how he thinks while writing, offering clear functions and locations so he can adjust wording at any time. He said it did not fit as a 37signals product because it is not commercially viable and not worth pulling people off other work, and the old way would take months, so he built it himself over a weekend with Claude, named Write_On, with a short demo video; its core is "alternative control" at the word, sentence, and paragraph level, plus dimming and stashing content nearby, which he stressed is not version control but alternative control. 1
3. Brian Chesky said Airbnb made updates based on user feedback and will keep iterating with more coming soon: connecting with friends for travel inspiration, natural-language AI search that he called a small step toward agentic AI with related UI coming soon, and new services including laundry and baby-gear rentals. 1
4. Garry Tan relayed YC advice that the most valuable guidance is to explain a company with verbs rather than nouns: "we're reshaping the cloud for AI" is unintelligible, while "developers hand us their code, we run it on cloud servers, and they don't manage infrastructure themselves" is immediately understood, because people understand verbs rather than labels and no one buys or invests in what they do not understand. 1