Grok Bots Move Into Small-Business Work as Vercel Tightens npm Supply-Chain Defense

Key Takeaways

1. Grok Bot is being pitched as a team member and a small-business AI business, with forwarded claims of sales and insurance-call use cases.
2. Vercel says it blocks known malicious npm packages before install scripts run and reports faster Sandbox startup times.
3. Perplexity released open-weight visual-language embedding models, pplx-embed-v2-context, in 9B and 0.6B versions.
4. A Claude Code-assisted decompilation of the Battlefield 1942 demo was ported to WebAssembly for browser multiplayer play.
5. Weight-Space Programming claims drew attention and a reported reproduction, while a separate finding says many LLM behavior results weakened within 15 months.
6. Blacksmith raised a $66M Series C at a $1B valuation, according to a forwarded YC case study.

1. AI Agents and Developer Tooling

1. Elon Musk endorsed treating Grok Bot as a team member and assigning it tasks normally given to new employees, calling it like hiring “a very capable employee.” Forwarded posts list shopping cards, product pages, order tracking, payments, custom modes, team Bot login and settings, Slack connector requests and Vault contacts as new features, and describe users having it mine old and potential customers for sales, call Progressive about glass coverage, and recommend a restaurant. These are forwarded user claims, not independently verified results. — via 1 2 3 4 5 6
2. Greg Isenberg argued that building Grok Bots for traditional small businesses may be one of the best current AI businesses: give each agent its own computer, email and phone number, then charge roughly the equivalent of a $12-per-hour employee. He also promoted a video on selling Grokbot to businesses, covering vertical customization, tech stack, build process, pricing, lead generation and sales. — via 1 2
3. A forwarded post says Grok Build v1.0.74 reduced long-session memory use, stopped large JSON and log pastes from freezing the terminal, significantly lowered Linux file-watching CPU use, sped up session recovery and fixed Windows/macOS stability, MCP authentication and context-window handling. — via 1
4. Levelsio relayed an observation that AI agents rarely use MCP even when available because connectors require manual installation and configuration, so agents call APIs or open a browser instead. The quoted author said he still builds MCP but sees its value as an “API without a contract,” predicting connectors, directories and installation flows will disappear while MCP and API endpoints eventually merge with API-key and OAuth account-level access. — via 1
5. Guillermo Rauch said Vercel prioritizes supply-chain security: when an npm package is hijacked, the system blocks deployment of known malicious packages before install scripts run, without user action. He also argued “stateless sandboxes” are the wrong benchmark because agents need a workspace with identity, skills, repositories and memory; Vercel Sandbox clusters run in 19 global regions, and Sandboxes with Drives startup improved 2.4x, with median creation time falling from 1.7 to 0.7 seconds and p95 from 2.2 to 1.6 seconds without SDK or configuration changes. — via 1 2
6. Vercel said Vercel MCP can now show deployments as interactive cards in MCP Apps-capable clients such as ChatGPT, including live deployment status, build logs debuggable in conversation and an access button when the URL is ready. Rauch and Tobi Lütke both favored WebMCP over browseruse, and wterm reportedly supports WebMCP so agents can list, read, type into and send keystrokes to a page terminal, with per-terminal opt-in and read-only mode across React, Vue, Svelte and native setups. — via 1 2 3

2. Models, Research and Security

1. Perplexity introduced open-weight visual-language embedding models pplx-embed-v2-context in 9B and 0.6B versions, each returning an int8 vector per document or a per-chunk vector aware of the whole document; the paper was done with turbopuffer. — via 1
2. Multiple accounts circulated a claim that GPT-6 Astra changed only 0.016% of Qwen3.5-0.8B weights to fix its reasoning loop, improving math by up to 6x and instruction following by 3x without gradients or training, calling it Weight-Space Programming. Andrew Wilkinson asked whether this means models can retrain themselves and keep learning over time, while Yohei said the result was reproduced and called it a new emergent capability in the latest models. These are forwarded claims, not independently verified here. — via 1 2 3 4 5 6
3. A forwarded post said retesting 50 widely cited LLM behavior findings — biases, hallucinations and others — found most no longer held or had significantly weaker effects within 15 months, arguing LLM research findings have a short shelf life. Another forwarded post said formalizing CFSG found 96 errors across 47 published papers, including top journals such as Annals of Mathematics, and argued formalization matters in the “vibeproving” era. — via 1 2
4. Amjad Masad said financial institutions banning tools such as Claude Code is spreading, calling it “the ban hammer falling,” and framed AI agents running on employee terminals as a systemic security risk for companies and governments, citing Goldman Sachs restricting Claude Code as one response. — via 1 2 3

3. Builds, Startups and Infrastructure

1. Levelsio said Claude Code decompiled the Battlefield 1942 free demo and ported it to browser multiplayer: disassembling x86 machine code to find 57,000 functions, translating them function by function into about 100,000 lines of C, compiling with Emscripten to WebAssembly, rewriting about 300 Windows and Direct3D functions for WebGL 2/OpenGL and building a WebSockets network stack. He said earlier attempts failed and Opus 5.5 completed this one, that it consumed 1 million output tokens worth about $100 at API prices but ran on a $100 monthly Claude subscription, that the server cap is 16 players and full while he tries 64 or 128, and that the Wake Island map now runs and will join the main rotation after testing. — via 1 2 3 4
2. A forwarded YC case study says Blacksmith (YC W24) built CI cloud on high-performance gaming-PC CPUs, grew 20x in the past year with a 30-person team and raised a $66M Series C at a $1B valuation, arguing AI coding tools make test verification the new bottleneck and CI waits can fall from minutes to seconds. A forwarded workflow recommends fast Blacksmith-style runners, Oxlint instead of ESLint, bun check instead of tsc, larger test machines with Vitest isolation off, cache and blobless-clone fixes, parallel tasks with nx affected and a pre-push hook. — via 1 2
3. Garry Tan said Capy's Captains and sub-thread coordination raised his issue-fix throughput about 5x “overnight,” citing Capy 0.7.0's persistent agents, Issues search, labels and due dates, and a cross-platform headless CLI. A forwarded post said AI-native tools are replacing Webflow/Framer, Clay and n8n in GTM stacks as companies move to headless sites, agentic workflows and GEO content optimization, with large companies slower to switch due to long contracts and change management. — via 1 2
4. Elon Musk forwarded claims that space-based data centers are superior from first principles because satellites in sunlight 24 hours a day need no batteries and get free cooling, with Starship able to deliver 10 to 15 MW per flight to sun-synchronous orbit. He added that aluminum radiators facing away from the sun can achieve about 1,300 W/m² while solar panels produce about 250 W/m², making panel area about 5x the radiator area and panels most of the satellite mass, and said space data centers “really do make sense.” He also cited a view that by the end of 2026 AI chip output will exceed what can be powered, with power — plus electrical equipment, cooling, networking and batteries — as the bottleneck. — via 1 2 3 4 5
5. Forwarded posts said Starlink helped Japan's ZIPAIR grow passenger revenue 12.3% year over year, with April-September passenger revenue of 42.9 billion yen and load factor rising from 73.7% to 87.8%; helped rescue five people in the Atlantic after the crew shared their position over Starlink; and grew monthly app downloads 316% and monthly active users 332% from September 2023 to September 2026, with more than about 13 million global users. A separate forwarded post said Tesla's 2011 full-year revenue guidance of $195M-$200M now takes about 16 hours, with Q3 2026 revenue at $28.05B. — via 1 2 3 4

4. Policy, Markets and Community

1. Elon Musk urged voters to “vote red today” in the Texas Senate race and forwarded posts backing Ken Paxton and saying a vote for Talarico would destroy Texas. A forwarded post said Talarico spent $89.4M since July versus Paxton's $10M, nearly 9 to 1, and outspent fundraising by $11.6M; another criticized Talarico as a left-wing extremist and described a voter who had never backed Paxton now enthusiastically supporting him. — via 1 2 3 4 5 6
2. John Collison previewed an interview with General Matter CEO Scott Nolan on why uranium enrichment is a bottleneck for nuclear power, covering U.S. reliance on Russian fuel, General Matter's Paducah, Kentucky enrichment facility, applying SpaceX methods to heavy industry, why nuclear is expensive, why small reactors need stronger fuel, whether nuclear can support AI buildout, startups reinventing nuclear, General Matter's race to 2030 and how small companies can beat incumbents. — via 1 2
3. Jason Fried said a team moving from Notion back to Basecamp helped its AI agents and was friendlier to people: while preparing “One-Person AI Agency,” he and a co-host struggled to collaborate in Notion, but after getting Basecamp access the co-host found the calendar, documents and tasks without instructions. — via 1
4. Tobi Lütke forwarded a case in which Pappy & Company's Shopify store was turned into a walkable, shoppable woodworking shop using a new Chrome HTML canvas and Shopify APIs, with a real Shopify checkout, built by an operations director over a weekend who said she had no programming experience. — via 1
5. Alexis Ohanian highlighted Foundation Fellow Amrita Bhasin appearing in OpenAI's new Codex ad campaign on billboards in San Francisco and New York; Bhasin is founding Sotira to help brands extract value from excess inventory and reduce waste. He also said he is an investor in the new Premier Flag league, especially its women's flag football competition, which is described as featuring top men's and women's players with the NFL and planned to launch in 2027. — via 1 2 3
6. Andrew Wilkinson said he made hundreds of millions of dollars over 20 years without venture funding, working 4-6 hours a day and sleeping 8 hours while maintaining a social life and time with his children, and argued “grind” work ethics are largely unnecessary, citing Bill Gates' line about giving hard work to lazy people because they find easy ways to do it. He said he never lost sleep over deals and criticized finance for using overtime and complexity to make ordinary people feel its services are worth high fees, adding that he learned to keep what he loves and delegate the rest from Jason Fried and DHH. — via 1
7. Paul Graham used a chart to describe a successful startup's month: last month's revenue peak now looks small and the early part of the curve is flattening, which is what exponential growth feels like. He also forwarded Perseus Defense's claim to have shown the first attritable missile downing a drone. — via 1 2