Skip to content

#其他

0 today

Aug 20Thursday

TechCrunch · AI

AI was supposed to win people over by now — it hasn't

TechCrunch argues that AI's public reputation is worsening despite technical progress. Axios reported the NRSC sent a memo warning AI data centers are hurting GOP chances in a key Ohio race. Pew Research found most Americans now view AI negatively. Protests against data centers have also appeared in Vancouver, Canada. The post doesn't disclose Pew's exact numbers or the full memo text, but the direction is clear: widespread use isn't turning into acceptance.

Why it matters: A GOP Senate committee memo obtained by Axios ties AI data centers to falling support in a key Ohio race, as Pew finds most Americans view AI negatively.

Financial Times · Technology

Stripe to acquire model router OpenRouter in $8bn deal

Stripe is buying OpenRouter for $8bn—its largest acquisition ever. OpenRouter gives developers a single API to access 300+ models, earning roughly $300M in revenue last year by taking a cut of usage. The article is paywalled, so deal terms, integration plans, and team details aren't disclosed. I'd flag the valuation: $8bn for a $300M-revenue API reseller is steep unless Stripe can bundle model access tightly with payments.

Why it matters: Stripe is buying model router OpenRouter for $8bn, a price that sits against roughly $300m in annual revenue and shows where the model-call middle layer sits.

OpenAI News

OpenAI previews Private Safety Processing to keep Zero Data Retention for frontier models

On Aug 19, OpenAI previewed Private Safety Processing, which lets Zero Data Retention customers get cross-interaction safety monitoring without exposing raw content to OpenAI staff. Automated systems detect misuse patterns across related requests; customer data stays on customer-controlled infra or is encrypted with customer-held keys on OpenAI storage. When a risk fires, OpenAI receives only an activity-type signal and severity—no content. The feature is in early-customer testing, with Glean, Databricks, and Microsoft voicing support.

Why it matters: OpenAI previews privacy-safe processing that gives zero data retention customers cross-session abuse monitoring while data stays on their servers or under their keys.

Hacker News front page

LLMs turn static software into something users can extend by asking

Jeremy Morrell argues that most web apps only serve the top of the demand curve, leaving a long tail of unmet needs. LLMs now let users “speak code into existence,” so a product can keep a stable core while AI-generated extensions cover the long tail. He points to Pi and DeepSeek Harness as early examples: users ask for a feature, the system writes a TypeScript extension and hot-reloads it. The post then focuses on how to bring this to the web—traditional webhooks set the bar too high; sandboxed runtimes like Cloudflare Dynamic Workers could let non-developers safely run custom logic. No timeline or performance numbers are disclosed; the piece is a product-direction argument.

Why it matters: Jeremy Morrell argues web apps can keep a stable core and have an LLM generate and hot-load TypeScript extensions on demand, with Pi and DeepSeek Harness as examples.

Financial Times · Technology

Google signs $12bn AI chip deal with Marvell

Google awarded Marvell a $12bn multi-year contract to design and package its custom AI chips, primarily next-gen TPUs. TSMC will handle manufacturing. It's Marvell's largest deal since pivoting from networking silicon to data center compute. Google didn't disclose delivery timelines, but the deal size signals a long-term bet on in-house chips to reduce reliance on Nvidia.

Why it matters: Marvell won a $12bn multiyear deal to design and package Google's next TPU, with TSMC manufacturing, a scale and split that show Google's long chip commitment.

Aug 19Wednesday

Hacker News front page

Ornith-1.5 uses self-generated tasks for RL, with three model sizes beating comparable open-source models on coding and agent benchmarks

Ornith-1.5 extends the self-scaffolding idea from Ornith-1.0 into a full self-improvement loop: the model proposes tasks, builds scaffolds, generates solution rollouts, and improves via RL. The 397B MoE flagship scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, matching Claude Opus 4.8 (85.0, 59.0) and beating GLM-5.2 and DeepSeek-V4-Flash-0731. The 35B MoE activates only 3B parameters per token yet outperforms Gemma 4-31B and Meta Muse Glimmer-30B on agentic coding. The 9B dense model has a quantized mobile version that runs on phones and scores 47.0 on Terminal-Bench 2.1 and 70.6 on SWE-Bench Verified, beating many larger models. Task reward multiplies validity, frontier difficulty (targeting a 20% success rate), and novelty. The post does not disclose training compute, data scale, or a release timeline.

Why it matters: Ornith-1.5 has the model generate its own tasks, build solving scaffolds and produce multi-turn answers, then iterates on that experience with RL across three sizes.

MIT Technology Review · AI

AI's recursive self-improvement might not come so quickly after all

A new study finds AI agents still can't do open-ended AI research—the kind without clear-cut answers that requires judgment and creativity. This tempers claims that recursive self-improvement is near. The open question is whether AI can grind its way there through narrow tasks alone, or if open-ended research is essential. The post doesn't settle that, but the results suggest current timelines may be too optimistic.

Why it matters: A new study finds today's AI agents still can't do open-ended AI research that has no standard answer and needs judgment, giving an empirical reference point for timelines of recursive self-improvement.

Hacker News front page

Mojo 1.0 is now open source under Apache 2.0, alongside Modular Cloud and new silicon support

At ModCon 2026, Modular open-sourced Mojo 1.0 under Apache 2.0 and made Modular Cloud publicly available, already serving customers like MiniMax. The platform now supports AWS Trainium, Google TPUs, and Qualcomm Cloud AI 100 and Dragonfly accelerators. Mojo is getting native Windows support via a Microsoft collaboration. MAX drops device usage restrictions and moves to source-available with an open alliance program.

Why it matters: Mojo 1.0 is fully open source under Apache 2.0, free to use, modify and ship commercially, with the Modular Cloud inference platform opening to the public at the same time.

OpenAI News

Replit launches Free Mode powered by GPT-5.6 Luna, removing token costs for software creation

Replit introduced Free Mode running on GPT-5.6 Luna, so users can plan, ideate, and explore projects without tracking token spend. CEO Amjad Masad credits recent OpenAI price cuts for making the free tier viable at millions-of-users scale. Complex reasoning tasks get routed to GPT-5.6 Sol, then return to Luna while preserving project context. Sam Altman frames it as a step toward anyone with internet building a product or startup. The post does not disclose Free Mode quotas, concurrency limits, or the exact launch date.

Why it matters: Replit launched a free tier that runs everyday planning and exploration on GPT-5.6 Luna and switches complex tasks to Sol, made possible by OpenAI's price cuts.

Hacker News front page

Bun 1.4 Rust rewrite is three months late and the community is losing trust

Bun has gone three months without a stable release for the first time since 2022. Founder Jarred Sumner has been promising v1.4 since June, but dates keep slipping and community replies now openly mock the repeated 'tomorrow' promises. The rewrite moves the codebase from Zig to Rust. In the past month, 15.8k commits came from robobun, 1.6k from autofix-ci[bot], and only 790 from Jarred; he himself noted most PRs are now Claude prompting Claude. Zig creator Andrew Kelley called the original Bun code 'hacks on top of hacks' and said Jarred was writing slop before LLMs. The author argues the rewrite looks more like an Anthropic ad than a genuine memory-safety fix, pointing to the number of unsafe blocks in the new Rust code. The project now has over 5,000 open PRs, far exceeding GitHub's recommended 1,000 limit.

Why it matters: Bun has gone three months without a stable release, its longest gap since 2022, after the founder teased v1.4 in June and kept pushing it back, drawing open mockery in the comments.

New York Times Chinese

Unitree Robotics shares surge nearly 500% on Shanghai trading debut, hitting ~$53B market cap

Unitree Robotics, the Chinese humanoid robot maker, surged nearly 500% on its first day of trading in Shanghai, briefly hitting a 629% intraday gain and a market cap of about 360 billion yuan (~$53B). The IPO raised roughly 6.1 billion yuan at 150.80 yuan per share. An Omdia analyst flagged heavy speculative froth, noting Chinese retail investors were willing to buy at almost any price. Unitree’s 2025 revenue was about $250M—up over 4x and profitable—but large-scale humanoid deployment hasn’t arrived; most factory use is still in pilots. The U.S. FCC recently proposed banning imports of foreign-made humanoid and quadrupedal robots on national-security grounds, and Unitree’s prospectus lists U.S. market access as a risk; U.S. buyers accounted for 13% of sales last year.

Why it matters: Unitree's Shanghai debut closed up nearly 500% on about $250M of 2024 revenue, a valuation analysts call heavily speculative while humanoid robots stay in pilots.

AI HOT (Curated Pool)

Grok Build is now available on web and mobile for all plans

xAI moved Grok Build from Early Beta to general availability. Any plan user on web, iOS, or Android can now describe an app, game, or dashboard in natural language and get a working version live in chat. Published apps get a grok.me link, support custom domains, can export to GitHub, and can call Grok's APIs for chat, images, and voice. Shared apps render as inline cards on X. The post shows four real published examples: a forest driving game, an isometric city builder, a 3D physics playground, and a browser beat machine.

Why it matters: xAI moves Grok Build from beta to all plan tiers, letting web and mobile users generate runnable apps in plain language, with domain binding and GitHub export.

Computing Life · Share · Yage

NVIDIA guarantees up to $105B for OpenAI's data center lease—using a year's cash flow as collateral

NVIDIA signed residual value guarantees for OpenAI's Ohio data center lease, capping its exposure at $105 billion—roughly its entire FY2026 operating cash flow. OpenAI lacks a credit rating, so the guarantee lets SB Energy borrow to build the campus. In return, the site must exclusively use NVIDIA's full-stack hardware, and NVIDIA also invested $1.5 billion in SB Energy. The deal makes NVIDIA supplier, landlord shareholder, and tenant guarantor all at once, with chip payments ultimately flowing back to it. Payouts trigger only if OpenAI defaults, and only cover the shortfall after the facility is re-leased or sold. The article argues this is closer to vendor credit enhancement than a subprime rerun: no margin calls, and the debt sits mostly in private credit. If the AI cycle turns, the most exposed are GPU-collateralized neocloud lenders, OpenAI's cash burn, and SoftBank's bridge loan—not NVIDIA's balance sheet.

Why it matters: NVIDIA backstops OpenAI's Ohio data center lease up to $105 billion, close to its full fiscal 2026 operating cash flow, letting an unrated OpenAI pull bank loans through.

Computing Life · Share · Yage

$60B for a Data Flywheel, $10M for a Working Memory

The AI race in 2026 is shifting from compute to real-world behavioral data. SpaceX acquired Cursor for $60B in stock, gaining access to coding interaction traces from over 50,000 enterprise clients—data fed directly into Grok 4.5 and 4.6 training. Google bid $10M in Spirit Airlines' bankruptcy auction for roughly 100M emails, 500M Teams messages, and 30M lines of internal code. Atlassian updated its terms to turn Jira and Confluence usage data into a continuous model-training input. DeepSeek is building its own front-end harness to capture local developer workflows that API-only access misses. The article maps these four paths—acquisition, bankruptcy purchase, SaaS terms, and self-built harness—and notes that the actual model gains from data flywheels, signal retention after de-identification, and real-world compliance friction remain unverified.

Why it matters: SpaceX buys Cursor for $60 billion in stock, taking over coding interaction records from 50,000+ enterprise customers, while Google pays $10 million for Spirit Airlines' email, messages and code, pricing work-scenario behavior data.

AI HOT (Curated Pool)

OpenRouter is joining Stripe to scale its neutral model gateway

OpenRouter announced it is joining Stripe. The product, name, and roadmap stay the same. It now processes 10+ trillion tokens daily across 400+ models for over 10 million developers. Routing decisions remain driven by what's best for users, with no model or provider bias. The deal is expected to close in weeks; the post does not disclose the acquisition price.

Why it matters: OpenRouter handles over 10 trillion tokens a day across 400-plus models, and its pledge to keep routing neutral is the part developers should check.

Hacker News front page

modelmap: paste a HuggingFace model ID, get an interactive architecture map

modelmap.cc turns any HuggingFace model ID into an interactive architecture diagram—no weights downloaded. It instantiates the model on a meta device to get the structure, then runs a traced fake forward pass to infer tensor shapes. The homepage shows trending models like Qwen3.8-27B, DeepSeek-V4-Pro-0813, and Kimi-K3, plus classic reference architectures such as GPT-2, BERT, and DeepSeek-V3.1. Public repos work out of the box; gated ones work after adding a token. The post doesn't disclose whether it's open source, backend costs, or concurrency limits.

Why it matters: modelmap.cc turns any HuggingFace model ID into an interactive architecture diagram, instantiating it on meta devices and running a fake forward pass to infer tensor shapes without downloading weights.

Hacker News front page

Purely AI-generated code has no author and no copyright under US law

This site walks founders and engineering leads through a hard legal reality: under current US copyright law, code generated entirely by AI has no human author, so it can't be copyrighted or defended as an owned asset. It cites four settled anchors—including the Supreme Court's March 2026 denial of cert in Thaler and the first rejection of a fair-use defense for AI training in Thomson Reuters v. Ross—to show the rule is already locked in. It also breaks down four common blind spots: pure AI output isn't yours, vibe coding where the AI makes creative choices leaves code unprotected, mixed codebases only protect the human-authored parts, and open-source licenses are unenforceable on code no one owns. The post doesn't offer fixes; it's a risk primer with a self-assessment quiz.

Why it matters: Under US copyright law, fully AI-generated code has no human author, so it cannot be copyrighted or protected as a company asset, which directly affects how teams using AI to write code handle ownership.

Hacker News front page

Linear publishes 2026 report on AI usage patterns in software teams

Linear analyzed 127,000 paid users and found AI feature adoption more than doubled across every function in H1 2026. Product roles jumped from 12% to 34%, the steepest climb. CEOs at companies with 201+ employees went from 9% to 36%, the single largest increase in the report. Company size barely mattered—adoption roughly tripled everywhere. Meanwhile, engineers spent about 17% more time creating, triaging, and commenting, suggesting AI is driving more coordination and context into the system rather than eliminating it.

Why it matters: Drawing on 127,000 paying users, Linear maps how much AI feature usage rose by role in the first half of 2026, with product roles and CEOs at large companies showing the biggest adoption gains.

Latent Space

Glean CEO on model routing: frontier model costs and open-weight popularity push enterprises toward automatic model selection

Glean CEO Arvind Jain says model routing is gaining enterprise traction mostly because of cost. Per-token pricing for Opus or the latest GPT models is 2-4x higher than previous generations, and users run longer tasks on them, pushing per-user spend up 10-20x year-over-year. Glean's automatic mode picks a model per task—or skips an LLM entirely. Engineering lead Tony Gentilcore claims Glean averages $0.45 per task vs. $1.84 for Claude Code, a 4x gap. Glean hit $300M ARR, tripling in 15 months. The post does not detail Waldo's routing mechanics.

Why it matters: Glean CEO Arvind Jain traces model routing demand to cost, with frontier models 2 to 4 times pricier per token than the previous generation and per-person spend up 10 to 20 times in a year.

Hacker News front page

Claude Code reverse-engineered a Mac driver for an unsupported HP printer in 4 hours

The author spent ~4 hours with Claude Code getting an HP Laser 1008a—a rebadged Samsung with no macOS driver and no AirPrint—to print natively from Cmd-P. The session went from a generic driver hanging at 'connecting-to-device', to reading the printer's own error pages to figure out its proprietary SPL3 raster language, bypassing the macOS USB sandbox, running HP's real rastertospl codec inside a Linux container, and finally shipping a reboot-safe daemon and an MIT-licensed one-command installer. The post is the full lightly-redacted transcript.

Why it matters: The author used Claude Code to write a native macOS driver for an HP Laser 1008a that has none, working from the private SPL3 raster language through to system integration in four hours.

AI HOT (Curated Pool)

Claude can now send Gmail and manage Google Drive files

Claude adds Gmail and Google Drive connectors for all paid plans. It can draft and send replies to email threads, with an optional approval step before sending. It can also manage files in Drive. The post doesn't detail permission scopes or specific file operations.

Why it matters: Anthropic added Gmail and Google Drive connectors to paid Claude plans, letting it draft and send replies inside email threads with an optional human confirmation before sending.

TechCrunch · AI

Etched’s valuation doubles to $21B in a month

AI chip startup Etched raised another $700M at a $21B valuation, led by Jane Street. That's nearly double the $10.3B valuation from its Series C just one month ago. Jane Street first installed Etched's first shipped inference cluster and was impressed enough to lead the new round. Etched says it designed two new components from scratch to speed up inference's prefill and decode stages, though the post doesn't detail the tech. I'd discount the valuation jump a bit — it's largely a signal boost from Jane Street's name, with no public deployment scale yet.

Why it matters: Etched doubled its valuation from $10.3B to $21B in a month, led by Jane Street, which had already bought its first shipped inference cluster, an unusual buy-then-invest pattern for a chip company.

Hacker News front page

Superpowers, Not Superintelligence

Bond responds to Zuckerberg's 'AI for everyone' essay, arguing he ignores data concentration. Meta's glasses and agents collect ambient data through constant observation, making users the object, not the owner. Real AI tools should require active input and give people superpowers—like phones, cameras, search engines—not build machines you feed. The post cites Meta's December 2025 privacy update: private chats with Meta AI now personalize ads, backed by ~$200B in ad revenue. The article does not detail how Bond's own product implements active input.

Why it matters: Bond cites Meta's December 2025 privacy update to argue its glasses and AI assistant collect data through all-day observation, with private chats feeding ad targeting backed by roughly $200B in ad revenue.

Aug 18Tuesday

Hacker News front page

Nova3D generates 3D assets as executable Blender code, not opaque meshes

Nova3D outputs Blender source code instead of a final mesh; the compiled glTF is just an artifact. All 54 benchmark items produce a valid executable program and model, each exposing named parts and a parent-child assembly tree. It satisfies 51 of 52 numeric and count constraints (best baseline: 11), defines 59 joints across 12 assets at 98.3% geometric validity, and passes 14 of 18 blinded local edits with locality preserved in all 18. Texture realism trails baked-PBR systems, but shape quality ranks second in structured domains. The key result is representational: code-native generation bakes semantic handles in at creation time, so downstream systems can inspect, measure, edit, and animate without post-hoc segmentation or rigging.

Why it matters: Nova3D has 3D generation models output runnable Blender code, so the model file is just the result and assets carry part names, assembly trees and joint constraints, skipping later segmentation and rigging.

Hacker News front page

Shoehorn: Quantize any model to fit your exact memory budget, down to the byte

Shoehorn is an open-source quantizer that starts from your available memory, subtracts inference overhead, then solves a per-tensor mixed-precision assignment that routinely uses over 99.99% of the budget. It avoids preset quantization tiers that either waste hundreds of megabytes or fail at load time. The local web UI measures your machine, streams the fit, shows perplexity cost, and launches a chat. It requires llama.cpp on PATH, outputs standard GGUF v3, and runs on macOS Apple Silicon, Linux, and Windows. The quantizer is written from scratch in Rust.

Why it matters: The open-source quantization tool Shoehorn measures available machine memory first, then assigns mixed precision layer by layer, leaving 13 KB of headroom on a 519.2 MiB budget in the official example.

Hacker News front page

Muse Glimmer fits an agent on-device with a memory hierarchy disguised as a 30B Transformer

Meta's Muse Glimmer is a ~30B multimodal model built to run agentic tasks offline on consumer hardware. The BF16 checkpoint is 55 GiB; Meta ships ~4-bit quantized versions that bring the language model under 20 GB. Architecturally, only every fourth of the 52 layers uses full-context attention—the other 39 use a 2,048-token sliding window. Global layers drop RoPE and retrieve by content. The KV cache stores just two key/value heads while 32 query heads provide diverse retrieval behaviors. A large ViT handles perception once, compresses neighboring patches 4:1, and feeds them as tokens. The result is a memory hierarchy: local layers build ordered representations, global layers search across the full sequence, and the tiny KV cache means quantization savings translate directly into longer context or larger batches.

Why it matters: Meta quantized its 55 GiB BF16 Muse Glimmer model to 4-bit so the roughly 30B-parameter model runs offline on a phone or laptop, with the language part under 20 GB.

Hacker News front page

NeoBrowser: An MCP server that drives real Chrome with your logged-in sessions

NeoBrowser is an MCP server that lets AI control your real, logged-in Chrome instead of a headless browser or API simulation. It's a single static Rust binary with 43 tools, uses human-like input patterns, and passes bot detection checks like bot.sannysoft. Built for automation that needs persistent sessions and bot-wall evasion.

Why it matters: NeoBrowser drives the user's already logged-in Chrome instead of a fresh headless browser, shipping 43 tools in one static Rust binary with more human-like input.

Hacker News front page

Google bought bankrupt airline Spirit's data at auction for $10M, because AI

Google paid $10 million at a bankruptcy auction for all of Spirit Airlines' data. The haul includes over 100 million emails, 30 million recorded phone calls, and reams of internal records from Teams, Oracle, and SAP. Spirit collapsed in May 2026 after years of post-COVID losses. Google wants the data to train AI models—real enterprise communication and customer service logs are expensive feedstock. The article doesn't say how Google plans to handle the personal data inside, or whether regulators have weighed in.

Why it matters: Google paid $10 million at Spirit Airlines' bankruptcy auction for 100 million emails, 30 million call recordings and internal system records to train AI, raising whether user data can be transferred this way in liquidation.

OpenAI News

Asana cleared 5 years of engineering work in 2 weeks with Codex

Asana used OpenAI Codex to fully remove Enzyme, an outdated testing framework, from its codebase. The work was originally estimated at five years and roughly $6M; it took two calendar weeks and $12K in model and infrastructure costs. Engineers wrote a five-sentence prompt, ran up to four coding agents in parallel, and reviewed every proposed change twice a day. Asana's CTO noted that not every multi-year project will collapse into weeks, but agents make once-impossible engineering work worth attempting.

Why it matters: Asana engineers used Codex to remove the outdated Enzyme test framework, work internally estimated at five years and about $6 million, finishing in two weeks for roughly $12,000 in model and infrastructure cost.

Bloomberg Technology

DeepSeek, Qwen, and Moonshot worry US AI rivals not by leading in tech, but by being dramatically cheaper

Bloomberg argues the real threat from Chinese AI firms isn't superior model capability—it's cost. DeepSeek, Alibaba's Qwen, and Moonshot are delivering comparable performance at a fraction of the price, forcing US rivals to rethink their economics. The article credits more efficient training methods and hardware utilization, but doesn't provide specific pricing comparisons or recent benchmark figures. Treat this as an industry trend piece rather than a technical deep dive.

Why it matters: Bloomberg argues DeepSeek, Qwen and Moonshot hold costs down through more efficient training methods and hardware utilization, pushing US rivals to recalculate their spending.

Computing Life · Share · Yage

The Company Selling You AI Product Managers Doesn't Give Its Own Agents Job Titles: Roles, Isolation, and Code in Multi-Agent Systems

Grok Bot markets agents as named coworkers like sales or finance, but its engineering core Grok Build uses only functional names such as researcher-0 and verifier—zero personas. The article breaks multi-agent design into three layers: the mechanism layer relies on context-window isolation for output quality; the orchestration layer decides who controls the next step (teammate persona, main agent, or script); the interface layer uses job hats to lower the human adoption barrier. Hats solve three human problems—delegation intuition, approval anchors, and memory partitioning—but do nothing for model reasoning. All coworkers under one account share the same cloud computer and credentials; security boundaries depend solely on manual approval gates. When building your own system, nail window isolation first, then pick an orchestration style based on task reliability needs, and save persona packaging for last.

Why it matters: Grok Bot presents agents externally as named colleagues while the engineering core uses only functional IDs like researcher-0, and the article uses that gap to examine context isolation and role design in multi-agent systems.

TechCrunch · AI

Anthropic's annualized revenue hits $65B, up $18B in two months

Anthropic's annualized revenue run rate passed $65B by end of July, up from $47B in May and $9B at end of 2025. Investors expect $100B–$120B for full-year 2026. OpenAI's run rate doubled to $40B in the same window. Both have filed confidential IPO paperwork; Anthropic may go public this fall targeting a $2T+ valuation. The post doesn't spell out how each company calculates revenue, so direct comparisons need a grain of salt.

Why it matters: Anthropic's annualized revenue went from $9bn late last year to over $65bn by end of July, while OpenAI doubled to $40bn, and that growth gap is the direct read here.

AI HOT (Curated Pool)

Google shows how to build zero-trust AI agents with ADK, using three hard security layers against prompt injection

Google's developer blog open-sourced a customer support refund agent to show why system prompts aren't security boundaries. A single prompt injection can bypass refund caps or leak environment variables. The fix is three hard layers: every database write is signed with a Cloud KMS hardware-backed key, dynamically generated code runs inside a gVisor sandbox with no network egress, and all I/O passes through deterministic semantic gateways. Full code and a local demo using HMAC to simulate KMS are on GitHub.

Why it matters: Google built a refund agent with ADK, showed one prompt injection bypassing the refund cap, and laid out three isolation layers including Cloud KMS hardware-key signing.

Latent Space

Stripe acquires OpenRouter for $7B, repricing the model routing layer

Stripe is acquiring model router OpenRouter for $7B, just 90 days after its $1.3B Series B. OpenRouter had $140M annualized revenue, ~$100M gross profit at 70% margin, and 250T tokens/month volume. The 50x multiple is standard for top-tier AI, but routing margins are under pressure—both OpenRouter and Vercel cut GPT-5.6 Sol pricing. The post also covers OpenAI's 8 GW Ohio campus plan, Cursor's Origin launch aiming to own the full dev loop, multi-agent systems moving from demos to operating patterns, and Vanta/LangChain productizing sandboxed agent execution.

Why it matters: Stripe bought model router OpenRouter for $7bn, and OpenRouter's $140m annualized revenue at 70% gross margin gives a pricing reference for the AI middle layer.

AI HOT (Curated Pool)

Cursor launches Origin code hosting as a GitHub alternative with agent-native repos

Cursor is rolling out Origin, its own code hosting service, in early beta for paid users. You can create repos, open pull requests, browse code, and sync existing GitHub repos with real-time two-way PR comments. Agents live inside every repo—ask questions, make changes, or push branches. First app integrations include Vercel for preview deploys, plus Depot and Buildkite for CI. The post doesn't say when free-tier access will arrive.

Why it matters: Cursor opened early beta of Origin to paying users, putting repos, pull requests and an AI assistant inside the editor with two-way GitHub sync, a first-hand product reference for anyone watching code hosting.

Hacker News front page

Rachel Thomas returns to AI, joins Answer.AI while her friends all hate AI

Rachel Thomas, co-founder of fast.ai, is returning to AI after a two-year break, joining Answer.AI to work on AI in education. She spent a decade criticizing Big AI's concentration of power, hollow ethics, and overhyped claims. Now she wants to show alternatives exist—SolveIt, built by Answer.AI, forces users to work through Polya's four problem-solving stages instead of getting instant chatbot answers. The post doesn't disclose her exact role or start date.

Why it matters: Rachel Thomas joined Answer.AI after two years away from the field, following a decade of criticizing big-company power concentration and ethics, offering a career path unlike the mainstream narrative.

Hacker News front page

OpenAI cuts GPT-5.6 Sol API pricing by 50%

GPT-5.6 Sol's listed price on OpenRouter just got slashed by 50% — $2.50/M input and $15/M output. It's the flagship of OpenAI's GPT-5.6 series, built for complex reasoning, coding, and multi-step agent workflows with a 1M-token context window. The actual weighted average is even lower: $0.81/M input via OpenAI's own channel thanks to an 86% cache hit rate. Direct latency sits at 2.78s P50. The post doesn't say whether the cut is permanent or a limited promo, nor whether it's tied to the Gemini 3.7 Flash discount.

Why it matters: OpenAI halved GPT-5.6 Sol's API price on OpenRouter to $2.50 per million input tokens and $15 output, so high-volume teams can recompute their call costs.

Hacker News front page

Israel created a fake think tank to feed pro-Israel narratives into AI chatbots

Israel's Government Advertising Agency commissioned a firm to build a fake think tank, the Hanover Institute for Public Policy, which published over 100 reports on Israel/Palestine in just over a week. The reports have no bylines, use a neutral tone, and include footnotes and tables of contents—designed to appeal to chatbots like Claude and Gemini. The contractor, Piro, calls this 'AI Story Optimization'; others call it LLM poisoning. The article does not confirm whether these reports have actually been ingested by any major model or what the measurable impact is.

Why it matters: A government paid for a fake think tank whose neutral, footnoted reports were built to be scraped and trusted by chatbots.

Bloomberg Technology

Anthropic's annualized revenue tops $6.5 billion ahead of IPO

Anthropic's annualized revenue run rate has passed $6.5 billion as it prepares for an IPO, Bloomberg reports. The figure is a single month's revenue annualized, not actual full-year cash. The post doesn't specify which month or disclose profit. Run rate can overstate seasonal bumps, but a $6.5B number signals strong enterprise adoption and will anchor IPO pricing.

Why it matters: Anthropic's annualized revenue passed $6.5 billion, a figure reached by multiplying one month's revenue by 12, and that scale feeds straight into its IPO pricing.

Hacker News front page

Hidden AirTag reveals Amazon is trashing rare books to train AI

404 Media planted an AirTag in a rare book and tracked it to an Amazon AI training facility in Las Vegas. A team there tears books from their spines and scans the pages; their door logo shows a T. rex devouring a book. Worker forum posts said Amazon ran out of books to scan earlier this year and feared the warehouse would shut down. Amazon's statement says it buys books through commercial channels to improve products and services, without mentioning AI training. Anthropic and xAI have publicly said they don't train on rare books, so Amazon's practice gives it a data advantage. Booksellers suspect AI firms are working through ISBN lists to scan every printed book, and this investigation adds hard evidence to that theory.

Why it matters: 404 Media hid an AirTag in a bulk order of rare books and traced it to the Amazon warehouse used for AI training, recording workers tearing off spines and scanning page by page.