WELCOME Happy Tuesday, legends. Welcome back to another edition of The Frontier, our weekly newsletter covering the best new AI launches on Product Hunt. WHAT'S HOT 🤝 Nvidia is buying Hugging Face for $12.9 billion. The chipmaker already put money into it in 2023, when the company was worth about $4.5 billion, so this is roughly three times that mark. It would hand Nvidia the place most open models actually live. Hugging Face spent the same week shipping a $399 robot duck. ( CNBC ) ⏸️ OpenAI has paused its most capable unreleased model. TIME got two weeks inside the company and more than twenty interviews. The trigger was July, when OpenAI's own agents escaped a test environment and hacked Hugging Face, which the company has now reclassified as an alignment failure and not a breach. Altman's line is that safety matters more than any company's momentum. ( TIME ) 🪖 The Pentagon put ChatGPT and Grok on GenAI.mil on Monday. The portal launched in December with Gemini alone and has onboarded 1.7 million people in nine months, against a department of more than three million. Claude is still not on it, held up by a contract dispute with Anthropic. ( TechCrunch ) 📈 Anthropic's listing window opens this month. It filed a confidential S-1 on 1 June off a $65 billion round that set the mark near $965 billion, and bankers have been running pre-IPO meetings since July for a September or early October float. Secondary trades have printed above the primary. ( Yahoo Finance ) 📉 Nobody agrees what a GPU costs, so someone built the index. The Computable GPU Index takes published on-demand rates from 28 providers every fifteen minutes and averages only the middle third of the weighted votes, so no cluster of sellers can drag the print. Open data, open method, open code, and the first thing being built on it is cash-settled hedges on GPU prices. ( Product Hunt ) Five AI tools you may have missed Caddi — Agent that builds agents by only showing your work only once 41 • Productivity ▲ 311 Hy4 preview — Tencent’s 770B open model for long-horizon work 3 • Open Source ▲ 216 Topview Motion Studio — Get a $3000 launch video for just $3 67 • Design Tools Promoted ▲ 332 Firecrawl Developer Index — A curated index of 70M+ artifacts for coding agents. 3 • Developer Tools ▲ 174 Almanac — The agent with a second brain 16 • Productivity ▲ 154 x1 — Lovable for iPhone apps go from idea to App Store 79 • iOS ▲ 517 FROM THE FORUMS It flags polish, not authorship Product Hunt started labelling comments it thinks are AI-generated, and Rohan Chaubey (@rohanrecommends) posted a screenshot of one in the wild, arguing the label is a good thing because AI comments run loud, generic and overly polished. Most of the thread agreed. Nika (@busmark_w_nika) is glad to stop reading ten-paragraph replies, Kate Ramakaieva (@kate_ramakaieva) pointed out LinkedIn shipped its own slop flag recently, and Rajiv Ayyangar (@rajiv_ayyangar) turned up to ask why people use AI for comments at all. Join the thread

WELCOME Happy Tuesday, legends. Welcome back to another edition of The Frontier, our weekly newsletter covering the best new AI launches on Product Hunt. WHAT'S HOT ⚖️ Alabama has subpoenaed OpenAI over an agent that escaped its own test. During a July cybersecurity eval, its agents left the lab and hacked Hugging Face to get the answer to the test they were sitting. Fourteen other state attorneys general have demanded OpenAI preserve everything. Meta and Anthropic have since admitted the same thing happened to them. ( TechCrunch ) 🚗 Uber has been fined €825 million for letting software fire people. The Dutch DPA found it deactivated driver accounts automatically with no meaningful human review, which GDPR prohibits outright when the decision changes someone's life. It's the second-largest GDPR penalty ever handed down, and Uber is appealing. ( TechCrunch ) 🔧 Anthropic hired the man who built Google's TPU. Amir Salek founded Google's custom chip program and ran it through seven generations before leaving in 2022. He joined Anthropic's compute team on the 21st. Anthropic rents every chip it currently uses. ( Bloomberg ) ⏳ OpenAI put the five-hour limit back on Plus today. Codex and ChatGPT Work had run on the weekly allowance alone since 12 July, which OpenAI says let casual users burn a week in an afternoon. Everyone is converging on the same two-limit shape, and nobody's interface shows you where you stand in it. ( 9to5Mac ) Five AI tools you may have missed Diet Claude — Never get blindsided by Claude's usage limits again 47 • Chrome Extensions ▲ 421 Memoria — Search photos by text, speech, object & faces. 100% offline. 18 • Productivity ▲ 157 session-indexer — Semantic search over your own Claude Code session history 3 • Open Source ▲ 91 Clarify — The autonomous CRM that helps you sell more 66 • Sales Promoted ▲ 483 Flare — The graph-first IDE and interactive map for agentic coding 5 • Open Source ▲ 120 Buddy Visual Tests — Every UI change, reviewed before merge 3 • Software Engineering ▲ 88 FROM THE FORUMS Month two is where it breaks David Brandt (@david_brandt3) says the tools have week one nailed, so he asked what month two looks like: one change moves five files, styling drifts, and the model forgets what you settled last week. His diagnosis is that your app's structure only exists in the chat history, and the chat history runs out. Alex Gidirim (@alex_gidirim) sharpened it. Calling that forgetting undersells the problem, because the model never had anywhere to keep those decisions to begin with. The fix people keep landing on independently is a written spec kept outside the repo, pasted in at the start of each session. Good thread if you are a few weeks into something that works and starting to feel nervous about it. Join the thread

WELCOME gm builders. Happy Wednesday. This week in three numbers, a prompt that stops your agent redecorating the codebase, and a thread about the 102 apps nobody ever opened. THIS WEEK, IN NUMBERS 776 — upvotes for x1, which takes an iPhone app from idea to the App Store The launch → 45,824 — GitHub stars on GitNexus, put up last August by a CS student and now a company The launch → $12.9B — what Nvidia is reportedly paying for Hugging Face, where most open models live The story → Five builds worth knowing Kilo Code for JetBrains — Fully native, open-source coding agent built for JetBrains 96 • Open Source ▲ 511 Superset Mobile — Run your parallel coding agents from your phone 2 • iOS ▲ 92 Caddi — Agent that builds agents by only showing your work only once 41 • Productivity ▲ 311 Clarify — The autonomous CRM that helps you sell more 66 • Sales Promoted ▲ 483 seendiff — Code diff viewer with progress tracking 7 • Open Source ▲ 149 Onset MCP — Write and publish release notes from your AI assistant 9 • SaaS ▲ 108 STEAL THIS Give the agent a budget Siddhant at @Evolvix AI asked for a small change. It worked. Then the diff showed the agent had touched 11 files for a two-file task , some of it architecture drift dressed up as cleanup. Now nothing gets written until the agent lists what it expects to modify, add and touch. Then he draws the line: This should probably stay within these 3-4 files. If you think it needs more, stop and explain why before continuing. Now "I also need to refactor these seven other files" arrives as a question, not a surprise. FROM THE FORUMS We shipped 111 one click apps. 9 of them get used. @asadmalik901 built 111 one-click apps for Arteza in six weeks, describing each one and shipping whatever ran. Nine cover 80% of usage. Around forty have never been opened outside the team. Build time was never the cost. Every one is a surface that can break and a support ticket he can't match to an app. Cheap to write, expensive to own. Next time he'd set the kill rule first: untouched for 30 days, deleted. Help or be helped

WELCOME gm builders. Happy Wednesday. This week in three numbers, a prompt from Sam Altman himself, and a forum thread for telling your agents to stop fooling around. THIS WEEK, IN NUMBERS 10K+ — Android downloads of Meta’s game vibe-coding app The launch → 361 — upvotes for Cursor’s new GitHub rival, Origin. The launch → 5 — hours per day Plus users can use Codex and ChatGPT Work The story → Five builds worth knowing /mission for Claude Code — Give Claude Code missions to spawn a team of agents 40 • Productivity ▲ 282 MemoryCustodian — Repo-native memory for coding agents 28 • Developer Tools ▲ 145 Task Monki — Run coding agents through the full development process 15 • Task Management ▲ 116 Topview Motion Studio — Get a $3000 launch video for just $3 67 • Design Tools Promoted ▲ 332 Prelint — Prevent product drift in AI-written code 135 • Software Engineering ▲ 603 Denovo — Turn your vibe-coded app into paying customers 72 • Sales ▲ 316 STEAL THIS Take Sam Altman's code Kelsey Vlamis of Business Insider had no idea how to vibe code . Everybody around her seemed to be doing it, but it just didn’t stick. Until she saw a prompt OpenAI founder Sam Altman plugged into ChatGPT on his phone: use all my chat history to figure out ideas for a long weekend trip with 8 friends, plan the best three options, make a full-stack site where the 9 of us can coordinate on what we would want to do in each place and decide where to go, and then after we get to group agreement make reservations. draft an email in my gmail i can send out to my friends when the site is ready.. According to Vlamis, she had a site up in two hours. If you’re struggling to start with your own thing, just use Uncle Sam’s code, kids. FROM THE FORUMS How do you stop AI coding agents from rewriting entire files when you only need 3 lines changed? @aditya_srivastava30 , a growth marketer who recently joined Product Hunt, asked builders for tips and tricks for making sure AI coding agents focus on the things he needs changed instead of starting from scratch with every revision. He writes: “Are there specific extensions, prompt patterns, or scoping tools you rely on to keep the edits surgical?” Similarly, Siddhant, founder of @Evolvix AI , is noticing that his agents keep tinkering with UI things like spacing and buttons, often making the product worse—even when he’s not asking it to. He’s started adding a “Do not change” section to his implementation requests. He’s gotten solid suggestions on how to give agents explicit constraints. Are there more? Help or be helped