# StartupXO: Front page (en), page 2

> Picked tech news and community posts from the past 30 days, newest first. Job posts are listed separately. HTML: https://startupxo.com/en/?p=2

31. [How many people can use an internal document search chatbot on L40S GPU 1 unit, measured by model and setting](https://zenn.dev/epicai_techblog/articles/29b509117755af) (zenn.dev)
   0 points by startupxo, 2026-11-18T00:43:01.943Z | 0 comments | https://startupxo.com/en/1413/ | https://startupxo.com/en/1413.md
   - The author measured how many people can use a chatbot that finds and answers from internal documents at the same time when it runs on one GPU (NVIDIA L40S 48GB). Inputs averaged about 7,050 tokens, and the criteria were a first token within 10 seconds and at least 20 tokens per second.
   - The number of people who met the criteria ranged from 1 to 60 depending on the configuration. The highest was 60 people with Gemma 4 26B-A4B compressed to 4 bits and run on vLLM, and the lowest was 1 people with Qwen3.8-27B run on llama.cpp.
   - Compressing to 4 bits on vLLM increased the number of people while the average accuracy dropped slightly. The task of copying unit prices from long documents did not get worse, but for the tax-inclusive total calculation, every configuration got only 0 of 50 questions right.
   - Placing the retrieval embedding model on the same GPU met the criteria when only question search was done. When document embedding ran continuously, output speed dropped by 41 to 46%, and even 1 people fell short of the criteria. Limitations: 2 models, 1 GPU, and accuracy measured 1 times per configuration.
32. [2026년 1~10월 패션 카드 결제추정금액 1위 무신사/29CM, 2위 유니클로](https://wowtale.net/2026/11/18/268264/?utm_source=rss&utm_medium=rss&utm_campaign=%25ed%258c%25a8%25ec%2585%2598-%25ea%25b2%25b0%25ec%25a0%259c%25ec%25b6%2594%25ec%25a0%2595%25ea%25b8%2588%25ec%2595%25a1-1%25ec%259c%2584-%25eb%25ac%25b4%25ec%258b%25a0%25ec%2582%25ac-2%25ec%259c%2584-%25ec%259c%25a0%25eb%258b%2588%25ed%2581%25b4%25eb%25a1%259c) (wowtale.net)
   0 points by startupxo, 2026-11-18T00:42:48.166Z | 0 comments | https://startupxo.com/en/1412/ | https://startupxo.com/en/1412.md
   - 와이즈앱, 리테일이 한국인의 신용카드, 체크카드 결제 내역을 표본 조사해 브랜드별 결제금액을 추정했다. 2026년 1~10월 누적 패션 전문 브랜드 중 무신사/29CM이 가장 높았다.
   - 무신사/29CM을 100으로 환산하면 유니클로 87.9, 크림 50.1, 지그재그/Z결제 32.9, 에이블리 27.2, LF 26.4, 신성통상 패션부문 22.8 순이다.
   - 2026년 10월 결제추정금액 100억 원 이상 브랜드 기준으로 30대 이하는 무신사/29CM이, 40대 이상은 유니클로가 1위다. 유니클로는 30대 이하에서도 2위다.
   - 계좌이체, 현금, 상품권 결제는 포함되지 않았다. 와이즈앱, 리테일은 이 수치가 개별 기업의 실제 매출액과 다르다고 밝혔다.
33. [Use of AI "decision models" that quickly pick one of a set of options instead of writing text surges, with custom training becoming the point of competition](https://techcrunch.com/2026/11/17/as-decision-model-use-skyrockets-is-customization-is-the-next-frontier/) (techcrunch.com)
   0 points by startupxo, 2026-11-17T18:25:40.706Z | 0 comments | https://startupxo.com/en/1411/ | https://startupxo.com/en/1411.md
   - A decision model is an AI that, instead of generating long text, quickly and cheaply picks one of several preset options and also reports how confident it is in that choice. It is used in service environments where a fast judgment matters more than the slow reasoning of a large model.
   - According to TechCrunch, in October more than 4 in 1 requests on Vercel's AI model router (a service that splits requests among multiple models) went to decision models. On OpenRouter, the TypeSafe model received more than 1.5 billion requests last week.
   - Microsoft said it will make Microsoft-Decision-1.1 widely available directly on Azure. The UK startup Sqwish received $3 million in pre-seed funding for its kettei fat model, and highlights as its strength a tool that keeps improving the model with customers' real operating data.
   - Ranking comparisons are difficult, and many say developers can simply swap models in and test them directly on their own work. A Microsoft executive said by email that models tuned only to benchmarks quickly show their limits in real enterprise environments.
34. [Google September AI roundup: external apps in Gemini, a Windows app, and new voice model](https://blog.google/innovation-and-ai/technology/ai/google-ai-updates-september-2026/?ref=welcome.ai) (blog.google)
   0 points by startupxo, 2026-11-17T18:25:22.105Z | 0 comments | https://startupxo.com/en/1410/ | https://startupxo.com/en/1410.md
   - Google said that by connecting external apps such as Airtable, Linear, Adobe, Webflow and Peloton in Gemini app settings, you can handle tasks in those apps inside Gemini without switching tabs.
   - The Gemini app for Windows opens instantly while you work with the Alt+Space shortcut. Google described it as a single place to delegate multistep tasks, synthesize Gmail and Drive files, and generate images and videos.
   - Gemini 4 Argon, the latest model, is strong in coding and cybersecurity defense, according to Google. It is currently available only through the Fairwind program for trusted security defenders. Expansion to developers and general users will come after safeguards are improved based on early tester feedback.
   - Google said the voice model Gemini 3.8 Live is available in Google AI Studio and the Gemini API for developers, as well as in Search Live, Gemini Live, Gmail, Docs and Keep. The weather forecasting model WeatherNext 3 has also been integrated into Search, Maps and more.
35. [SBVA supports 15 Korean founders attending YC Startup School in Taiwan](https://wowtale.net/2026/11/17/268214/?utm_source=rss&utm_medium=rss&utm_campaign=sbva-yc-%25ec%258a%25a4%25ed%2583%2580%25ed%258a%25b8%25ec%2597%2585-%25ec%258a%25a4%25ec%25bf%25a8-%25ed%2583%2580%25ec%259d%25b4%25eb%25b2%25a0%25ec%259d%25b4-%25e9%259f%2593-%25ec%25b0%25bd%25ec%2597%2585%25ea%25b0%2580-15%25eb%25aa%2585-%25ec%25b0%25b8%25ea%25b0%2580-%25ec%25a7%2580%25ec%259b%2590) (wowtale.net)
   0 points by startupxo, 2026-11-17T12:28:56.725Z | 0 comments | https://startupxo.com/en/1409/ | https://startupxo.com/en/1409.md
   - YC Startup School is a program where founders and aspiring founders receive practical advice from Y Combinator (YC) partners and experienced founders and network with them. This Taipei event was the first held in Taiwan.
   - Venture capital firm SBVA said on the 17 that it formed a delegation of 15 Korean founders in fields such as AI applications, AI infrastructure, robotics and semiconductors, and supported their attendance at the event held on the 12.
   - The event shared experiences building AI developer tools, a case of building payment infrastructure in Southeast Asia, and DoorDash's growth story. It also covered how to release an early product quickly and validate business hypotheses through user reactions.
   - Networking sessions discussed the performance and reliability of AI agents, the possibility of connecting with sovereign AI models, and changes in ways of working due to AI adoption. Local exchanges were supported in cooperation with Silicon Corridor.
36. [How to run AI models directly on your PC and browser: generation speed is governed by memory bandwidth](https://zenn.dev/port_inc/articles/ff974c7456f4db) (zenn.dev)
   0 points by startupxo, 2026-11-17T12:28:20.676Z | 0 comments | https://startupxo.com/en/1408/ | https://startupxo.com/en/1408.md
   - An AI model is essentially a file containing a huge number of values called weights. You can choose to run it in the cloud, on an in-house server, or on a personal device, and if you run it on your device, your input never leaves the device. After the first download, it can be used without a connection.
   - The speed of generating text one token at a time is governed by memory bandwidth rather than compute performance. By the author's calculation, llama.cpp benchmark measurements reached 6 to 8 tenths of the theoretical upper limit, which is bandwidth divided by model size. When you feed in a long document, the KV cache (storage for earlier computation results) grows and speed drops.
   - With 4-bit quantization (a method of coarsely reducing numbers to cut size), examples were given such as Qwen3.5-4B for GPU memory of the 8GB class and Qwen3.6-27B for the 24GB class. Epoch AI compared that the best model running on 1 RTX 5090 cards lags the top-tier models by an average of 6 to 12 months.
   - In the browser, you can run models with WebLLM, which is based on WebGPU, and wllama, the browser version of llama.cpp. The author explains that small models can be used for classification, reading only the probabilities of options without writing text, or can be tailored to a purpose with distillation and LoRA (a method of additionally training only some of the weights).
37. [Corma, which builds AI models and agents for cybersecurity defense, reaches a valuation of $1 billion 3 months after launching its product](https://techcrunch.com/2026/11/16/cybersecurity-ai-startup-corma-reaches-1b-valuation-in-about-3-months/) (techcrunch.com)
   0 points by startupxo, 2026-11-17T06:32:20.361Z | 0 comments | https://startupxo.com/en/1407/ | https://startupxo.com/en/1407.md
   - Corma provides AI models and AI agents (programs that carry out tasks continuously on behalf of people) trained for cybersecurity defense. It came out of stealth and launched its product in August.
   - According to the company, it raised a Series A of $125 million led by NEA, and its valuation is $1 billion. Total funding is $185 million, with participation from Sequoia, Khosla Ventures, Cowtu and others.
   - On October 21, a critical PHP flaw, a Node.js flaw and a LiteLLM security incident were disclosed on the same day, leaving security leaders scrambling to respond, and interest in Corma's product reportedly grew significantly afterward, the article says.
   - Competitors mentioned include Kevin Mandia's Armadin, Anthropic, OpenAI and Microsoft. Corma claims its differentiator is that it is a frontier AI lab focused on defense.
38. [Carnegie Mellon computer science courses offered free on YouTube, with self-study programs in preparation](https://www.cs.cmu.edu/learn/) (cs.cmu.edu)
   0 points by startupxo, 2026-11-17T06:32:02.590Z | 0 comments | https://startupxo.com/en/1406/ | https://startupxo.com/en/1406.md
   - Carnegie Mellon University has launched a free online learning site called CS Learn. Anyone can watch computer science courses taught by its faculty on YouTube at their own pace.
   - Starting in November, courses will be released in turn, including machine learning, software design and AI agents. In January 2027, probability and computing, multimodal machine learning, and product management basics, among others, will be added.
   - As a next step after the lecture videos, the site says self-paced course materials based on actual CMU classes will be released. They will offer instant feedback and badges, and AI tutoring is planned to be added later.
   - For those who need course credit or official certification, the site points to CMU Online's assessed courses and official certificates. The site was built by students with support from ScottyLabs.
39. [How to keep API logs searchable later: attach an ID to each request to group records of the same request](https://qiita.com/masashige0904/items/22f2836485ecc25ee527?utm_campaign=popular_items&utm_medium=feed&utm_source=popular_items) (qiita.com)
   0 points by startupxo, 2026-11-16T18:26:17.203Z | 0 comments | https://startupxo.com/en/1405/ | https://startupxo.com/en/1405.md
   - If you attach a unique number (requestId) to each request in the records a server writes, you can avoid the problem of being unable to tell, from timestamps alone, which of the concurrently handled requests failed. This is a design example written on Qiita by the Japanese developer masashige0904.
   - Each log entry is written as one line of JSON, with searchable values such as the event name, request ID and failure category kept as separate fields. Multi-line error records may be split into several events depending on how they are collected, it is said.
   - The article proposes a policy of not logging passwords, authentication tokens, emails, request bodies, and raw exception text or SQL whose safety has not been verified, and creating a common logging gateway so this rule can be checked in one place.
   - In AWS CloudWatch Logs Insights, filtering by requestId and sorting in chronological order lets you follow the start and failure of a single request. The article states that its search result table is a hypothetical example, not run on AWS, and that retention periods and fees must be checked separately.
40. [A Qiita post argues for dropping content you can ask AI about from internal tech study sessions](https://qiita.com/kotowazaman/items/9ee7c475a475e7527b0e) (qiita.com)
   0 points by startupxo, 2026-11-16T18:26:02.385Z | 0 comments | https://startupxo.com/en/1404/ | https://startupxo.com/en/1404.md
   - This is a personal opinion post on Qiita, a Japanese developer knowledge-sharing site. It argues that internal tech study sessions, where one person compiles and presents general knowledge that can be found through search or AI, should be stopped.
   - As evidence, it cites the Stack Overflow 2026 developer survey. The share of people who named online search when looking for answers is 83%, and the share who ask AI agents is 70%. The author states that the survey did not measure the value of such sessions.
   - According to the author's hypothetical calculation, a session with 10 people gathering for 1 hours, with 3 hours of presentation preparation, held 4 times a month, costs 624 person-hours a year, about 3.9 person-months.
   - The author names sessions worth holding: ones where people review the results of applying a new technology to their own projects, and ones where they share work knowledge not found in documents and experience handling incidents. Before holding one, the author suggests asking whether AI can answer it, whether a wiki is enough, and whether you can write in one sentence what to do afterward.
41. [Amazon-owned Zoox's steering-wheel-free robotaxis can be hailed through the Uber app in Las Vegas](https://www.theverge.com/transportation/1032984/uber-zoox-robotaxi-las-vegas-launch-mgm-resorts-los-angeles) (theverge.com)
   0 points by startupxo, 2026-11-16T12:37:14.161Z | 0 comments | https://startupxo.com/en/1403/ | https://startupxo.com/en/1403.md
   - When you request a ride in the Uber app in Las Vegas, you may be matched with a driverless robotaxi without a steering wheel from Zoox, which is owned by Amazon. It is the first time Uber has added this kind of purpose-built autonomous vehicle to its network, and the first time Zoox has opened its vehicles to an outside app.
   - If you request UberX, Uber Comfort, Uber Electric, Uber Comfort Electric or Wait & Save, you may be matched with a Zoox vehicle at no extra charge, according to Uber. Choosing autonomous vehicles in the ride preferences in the app raises the chance of being matched.
   - Zoox's service area is limited to places popular with tourists, such as the Strip, Downtown and the Arts District. Zoox's own app will continue to operate separately from the Uber integration.
   - Under a multi-year agreement, Uber and Zoox have become the official ride-hailing and robotaxi partners of MGM Resorts International in Las Vegas. The three companies said they are improving pickup and drop-off guidance inside resorts and guidance in the app. In Los Angeles, the goal is to operate on Zoox's own app before the end of 2026, and to integrate with Uber in 2027.
42. [Dream team of former Freshworks executives raises $53 million for an AI customer management tool that automatically records emails and conversations](https://techcrunch.com/2026/11/16/freshworks-alums-raise-53m-to-take-on-salesforce-hubspot-with-ai-native-crm/) (techcrunch.com)
   0 points by startupxo, 2026-11-16T12:37:00.221Z | 0 comments | https://startupxo.com/en/1402/ | https://startupxo.com/en/1402.md
   - It reduces the work sales reps used to enter by hand into customer relationship management (CRM) software. It automatically gathers content from emails and customer conversations, and AI agents handle follow-up emails, meeting preparation, and sales pipeline checks.
   - At its core is a shared memory store the company calls the "Company Brain." Based on it, 4 built-in AI agents research prospects, draft follow-up emails, flag deal risks, and set up CRM workflows, while the sales team controls what is sent to customers.
   - According to CEO Anand Venkatraman, within 3 months of full-scale sales it has won more than 100 customers in 13 countries, most of them paying. Some switched from Salesforce or HubSpot. All of this is the company's own account.
   - The $53 million is the combined amount of the seed and Series A rounds, and the size of each round and the valuation are undisclosed. The venture capital firm a16z expects existing CRMs to remain as systems of record while AI tools capture the value on top of them.
43. [5 technologies reaching everyday life, from live interpreting earbuds to satellite texting and driverless taxis](https://www.artiverse.ca/five-technologies-moving-from-future-promise-into-daily-life/) (artiverse.ca)
   0 points by startupxo, 2026-11-16T06:47:49.046Z | 0 comments | https://startupxo.com/en/1401/ | https://startupxo.com/en/1401.md
   - In 2025 September, Apple added Live Translation to its latest AirPods, which translates what the other person says. You can hold a conversation with just the earbuds, without an interpreter or a translation app.
   - T-Mobile's T-Satellite uses SpaceX's satellite network to send texts, photos and emergency reports even where there is no cell coverage. Voice calls work only on WhatsApp, and apps that are not optimized do not work.
   - In 2026 March, Google researchers estimated that the elliptic curve cryptography used by the web and Bitcoin could be broken with fewer than 500,000 qubits. Chrome has enabled quantum-resistant key exchange by default since the end of 2024.
   - Waymo's driverless ride-hailing taxis have expanded to 15 cities in the US. The original says smart glasses promote themselves as monitor replacements, but because of heat, resolution and privacy issues they may be better suited to video and games for now.
44. [Can malicious instructions in external documents be stored in an AI agent's long-term memory? An AWS AgentCore Memory experiment](https://zenn.dev/takuyousou/articles/agentcore-memory-poisoning-guardrails) (zenn.dev)
   0 points by startupxo, 2026-11-16T06:47:26.779Z | 0 comments | https://startupxo.com/en/1400/ | https://startupxo.com/en/1400.md
   - A Japanese Zenn author tested with AWS AgentCore Memory whether bad instructions hidden in an external document can get into the feature where an AI agent picks out important points from conversations and remembers them. The author used a fictional customer and mock materials.
   - An English sentence telling the agent to also read Customer B's materials when writing Customer A's report was stored as a "user preference." A request beyond the agent's authorization was organized as if it were a report-writing habit.
   - In the pre-storage check, the Prompt Attack filter did not catch this sentence. After adding Denied Topics that prohibit referencing materials across customers, it was blocked. The author implemented the step that stops storage on the app side.
   - The contaminated memory was passed to new sessions and run 5 times each, 10 times in total, but there was no request for Customer B's materials. The reason could not be determined. The author tested only one document and said false positives and bypasses are possible.
45. [신용보증기금, 서울 프론트원에 비수도권 스타트업용 무료 업무, 행사 공간 NODE 열어](https://wowtale.net/2026/11/16/268107/?utm_source=rss&utm_medium=rss&utm_campaign=%25ec%258b%25a0%25eb%25b3%25b4-%25eb%25b9%2584%25ec%2588%2598%25eb%258f%2584%25ea%25b6%258c-%25ec%258a%25a4%25ed%2583%2580%25ed%258a%25b8%25ec%2597%2585-%25eb%25b3%25b5%25ed%2595%25a9%25ea%25b3%25b5%25ea%25b0%2584-node-%25ea%25b0%259c%25ec%2586%258c) (wowtale.net)
   0 points by startupxo, 2026-11-16T00:48:24.503Z | 0 comments | https://startupxo.com/en/1399/ | https://startupxo.com/en/1399.md
   - 지방 스타트업이 서울에서 투자자를 만나거나 사업을 할 때 쓸 수 있는 업무 공간과 행사 공간을 신용보증기금이 서울 마포구 프론트원에 열었다. 이름은 NODE이고 이용은 무료다.
   - 지방 소재 스타트업과 지방 소재 대학 예비 창업자에게 우선 이용권을 준다. 기존에 사무실과 서고로 쓰던 유휴공간을 고쳐 총 553㎡로 꾸몄다.
   - 약 60명이 들어가는 IR라운지는 대형 LED 화면과 음향시설을 갖췄고 기업설명회, 강연, 투자자 미팅에 쓴다.
   - 최대 100여 명이 쓰는 Flex오피스에는 자유석, 휴게라운지, 6인실 2개와 12인실 1개가 있다. 신보는 임차 비용 부담이 줄 것으로 내다봤다.
46. [A free service that, once you choose the IT products you use, shows only the related security vulnerabilities with Japanese action guidance and deadlines](https://qiita.com/songchong/items/fa50cc82e49b24925fe3?utm_campaign=popular_items&utm_medium=feed&utm_source=popular_items) (qiita.com)
   0 points by startupxo, 2026-11-16T00:48:13.629Z | 0 comments | https://startupxo.com/en/1398/ | https://startupxo.com/en/1398.md
   - The service was created by the editorial team of IT Renkei Map in Japan. When you select the products and versions you use, it gathers only the relevant vulnerabilities and security incidents and adds tasks and deadlines in Japanese for each role: users, IT staff and developers.
   - According to the editorial team's measurements, the median time for a notice from the publisher to appear on this service was 46 seconds (37 cases). Japanese task guidance took 17 minutes 56 seconds, and email took about 37 minutes. The Japanese vulnerability database (JVN iPedia) has a median of 4.6 days after a CVE is published.
   - If there is no official deadline, the deadline is calculated based on the US CISA. When the editorial team checked 162 cases designated by CISA, 144 cases (88.9%) had the same number of days. It is free and requires no registration, and only email notifications require login.
   - This is information compiled from official announcements, not an official notification. 15 minutes is a target, and 7 of 37 measured cases exceeded it. It cannot replace your own company's configuration list (SBOM), so you need to check the original source before taking action.
47. [Developer Jakub Jankiewicz's blog argues that AI-generated code should be released as open source](https://jakub.jankiewicz.org/blog/ai-code-open-source/) (jakub.jankiewicz.org)
   0 points by startupxo, 2026-11-15T18:23:01.380Z | 0 comments | https://startupxo.com/en/1397/ | https://startupxo.com/en/1397.md
   - The opinion is that people who produce code with AI should not keep it to themselves but release it for anyone to use. Developer Jakub Jankiewicz argued on his personal blog that it should be released under an open source license or CC0 (a copyright waiver declaration).
   - The first argument is that AI models were trained on open source code and posts from sites like Stack Overflow without permission or attribution. He sees turning that into closed code as something like license laundering.
   - The second argument is that purely AI-generated output not written by a person should not be subject to copyright. The author cites the monkey selfie dispute and the U.S. Copyright Office's position on AI.
   - He acknowledges that some countries, such as the UK, grant limited copyright to computer-generated works. His position is that code a person barely reviewed or edited beyond writing the prompt is especially one that should be released.
48. [HGhub, a universal player that adds video sources through plugins, releases its source with no built-in content](https://github.com/Minireel/HGhub) (github.com)
   0 points by startupxo, 2026-11-15T18:22:44.501Z | 0 comments | https://startupxo.com/en/1396/ | https://startupxo.com/en/1396.md
   - HGhub is a player that brings video browsing, search, playback, and a personal library into one app. When users import JavaScript plugins, they can watch and manage videos from multiple sources on the same screen.
   - The development team stated that the app has no built-in content sources and does not provide any 3rd-party plugins or an official source list. No sign-up is required, and if a source requires login, the corresponding plugin provides a login window.
   - It plays HLS (a streaming format) and regular video, and lets you change the playback route, quality, and resolver. It supports favorites, watch history, resume playback, and tracked viewing, and detailed features vary depending on the installed plugins.
   - The plugin version is in early development, so for now you must run it from source or build it yourself. The team is polishing Windows and Android first. It is maintained by an individual, and the license prohibits commercial use.
49. [Weekly roundup: Nvidia partners with 5 Korean companies at once, the SMEs Ministry's Beyond Unicorn strategy, and Anthropic's IPO filing](https://wowtale.net/2026/11/15/268082/?utm_source=rss&utm_medium=rss&utm_campaign=%25ec%259c%2584%25ed%2581%25b4%25eb%25a6%25ac-%25ec%259d%25b8%25ed%2585%2594%25eb%25a6%25ac%25ec%25a0%2584%25ec%258a%25a4-29-%25ec%2597%2594%25eb%25b9%2584%25eb%2594%2594%25ec%2595%2584-%25ec%25a0%2584%25eb%25b0%25a9%25ec%259c%2584-%25ec%2583%2581%25eb%25a5%2599%25ec%2597%2590%25eb%258f%2584-%25eb%25a6%25ac%25eb%258d%2594) (wowtale.net)
   0 points by startupxo, 2026-11-15T06:29:22.828Z | 0 comments | https://startupxo.com/en/1395/ | https://startupxo.com/en/1395.md
   - A weekly roundup of startup, investment and tech news in Korea and abroad from November 9, 2026 to 15. Nvidia announced investment and technology partnerships with Rebellions, SKT, Motif, Upstage, Realworld, Kakao Mobility, Aim Intelligence and others.
   - The Ministry of SMEs and Startups and the National AI Strategy Committee announced the Beyond Unicorn strategy to nurture 1 AI decacorns and 20 unicorns by 2030. The plan is to concentrate growth funding of at least 10 billion won per company, along with market and regulatory support.
   - According to the original text, Anthropic filed amendment no. 2 to its IPO registration statement (S-1). The offer price is 70 to 78 dollars, and the valuation is up to $1.405 trillion on a fully diluted basis. Revenue for 1 to September is $33.2 billion, and the net loss is $104 billion.
   - TSMC said it will invest up to $1 billion in Anthropic and OpenAI. Spend management platform Menlo Ventures said enterprise AI spending is $176.8 billion per year, of which 69% goes to frontier labs.
50. [Claude API lets you add instructions and tools mid-conversation and change the thinking level](https://qiita.com/moritalous/items/9e7c02b749037238b2bc?utm_campaign=popular_items&utm_medium=feed&utm_source=popular_items) (qiita.com)
   0 points by startupxo, 2026-11-15T06:29:11.969Z | 0 comments | https://startupxo.com/en/1394/ | https://startupxo.com/en/1394.md
   - You can append an instruction to the AI (system prompt) in the middle of a conversation to change behavior such as tone or language. The author explains that because the earlier part is not changed, the prompt cache (reuse of previously processed content) is preserved.
   - Tool definitions can also be added or removed during a conversation. In the author's example, when a time lookup tool was added to a conversation that had only a weather tool, both tools were called together. The author states this was a beta feature at the time of posting.
   - The thinking level (effort) can also be changed during a conversation. In the author's run, the thinking tokens were high 465, low 0, and high again 169. This feature also requires specifying a beta header.
   - You can also insert a temporary instruction that applies only until the next user turn. In the author's example, only response number 5 was fixed to French, and response number 6 returned to normal. This feature is also in beta.
51. [a16z argues that if AI improves the success rate and timeline of drug development, the pharmaceutical industry will grow larger](https://wowtale.net/2026/11/15/268066/?utm_source=rss&utm_medium=rss&utm_campaign=ai%25ea%25b0%2580-%25ec%25a0%259c%25ec%2595%25bd%25ec%2597%2585%25ec%259d%2584-%25eb%258d%2594-%25ed%2582%25a4%25ec%259a%25b4%25eb%258b%25a4%25eb%258a%2594-a16z%25ec%259d%2598-%25eb%2585%25bc%25eb%25a6%25ac-%25ec%2588%25ab%25ec%259e%2590%25eb%25a1%259c-%25eb%2594%25b0%25ec%25a0%25b8%25eb%25b3%25b4%25eb%258b%2588) (wowtale.net)
   0 points by startupxo, 2026-11-15T00:45:13.202Z | 0 comments | https://startupxo.com/en/1393/ | https://startupxo.com/en/1393.md
   - In a newsletter, the venture capital firm a16z argued that if AI helps pick drug candidates and design clinical trials, raising the probability of success and shortening the time to approval, the profit structure of the pharmaceutical industry will change.
   - By a16z's estimate, bringing one new drug to market costs about $2 billion and takes an average of 13 years. It calculated that if development costs fall from $1.6 billion to $800 million, the 10-year internal rate of return for a drug with peak sales of $500 million rises from 8% to 22%.
   - a16z stated that this calculation is an illustration, not a forecast of AI cost savings. It does not reflect the probability of clinical and regulatory failure, and it acknowledged that stages whose schedules are set by the FDA, such as long-term toxicity studies, are hard to shorten.
   - The author has invested in life sciences since 2015, and many of the example companies are investments. The comparison scope is also broad: the 16 software companies with a market capitalization above $200 billion include Apple, Samsung Electronics and others.
52. [The 5 components for building an always-on AI agent, a Claude official article summarized as slides](https://speakerdeck.com/minorun365/jouji-kadou-ejento-no-tsukurikata) (speakerdeck.com)
   0 points by startupxo, 2026-11-15T00:45:00.226Z | 0 comments | https://startupxo.com/en/1392/ | https://startupxo.com/en/1392.md
   - An always-on agent is an AI that carries context across changing conversations and starts work on its own without a person watching. Its core task loop is the same as a regular agent's; the surrounding machinery is different.
   - In practice it sleeps most of the time. When a notification arrives it wakes up, works, leaves a record, and goes back to sleep. The article explains that it does not use tokens (the unit of AI usage) while asleep.
   - The surrounding components are an event inbox, a timer, memory, permissions, and an outbox. The inbox gathers requests in one place, preventing conflicts between requests that arrive at the same time and loss of requests if work is interrupted.
   - Permissions are named as the riskiest part. The approach introduced is that an external proxy injects credentials, a classifier blocks dangerous actions, and important decisions proceed only after user confirmation.
   - This is an experimental slide deck made by a Japanese presenter, Minorun, who put the URL of the official Claude blog into their own Claude Code. The Opus 5.5 cost reduction figure is also a claim from the original article.
53. [PyPI project page redesign: Security tab shows package provenance and flags anomalies](https://blog.pypi.org/posts/2026-11-06-ui-updates-complete/) (blog.pypi.org)
   0 points by startupxo, 2026-11-14T18:22:38.555Z | 0 comments | https://startupxo.com/en/1391/ | https://startupxo.com/en/1391.md
   - PyPI, the Python package repository, added a new Security tab to project pages. For each file, you can view records (provenance attestations) proving who built the file you are about to download and where.
   - If every file has an attestation, the page shows full provenance. If none do, it shows no provenance. PyPI explains that none does not mean malicious, only that the origin cannot be cryptographically verified.
   - If only some files have attestations, the source repository differs from the previous release, or attestations disappear, a warning appears at the top of the page. Since this could be an account takeover, PyPI recommends checking before installing.
   - The release files tab gained search and filters, and the release history shows the number of files and the provenance status. This work was implemented by Kabu Creative with OpenSSF funding.
54. [PerlOS built in 1 hours: a hands-on guide to a homemade OS that boots into a Perl shell](https://zenn.dev/takesako/articles/e9b10360d889ca) (zenn.dev)
   0 points by startupxo, 2026-11-14T18:22:21.531Z | 0 comments | https://startupxo.com/en/1390/ | https://startupxo.com/en/1390.md
   - This is a hands-on article that teaches what happens between powering on a computer and text appearing on the screen, by building a small operating system yourself. The author is Zenn writer takesako.
   - The exercise starts in QEMU (an emulator that imitates a CPU and peripherals in software). You control the serial communication device (UART) directly to print characters, and write printf and malloc by hand.
   - The final goal is to load microperl, a Perl interpreter for embedded use, and build PerlOS, which boots into a Perl shell. After that, you build it for real hardware on the Raspberry Pi Pico 2.
   - The real hardware version has little RAM, so reportedly Perl's numeric type is changed from double to float and malloc is improved. The provided text ends at the introduction of exercise 7, so the later steps are not confirmed.
55. [10 deep tech startups recommended by 5 Yeongnam Creative Economy Innovation Centers hold joint IR for investors](https://wowtale.net/2026/11/14/268056/?utm_source=rss&utm_medium=rss&utm_campaign=%25ec%2598%2581%25eb%2582%25a8%25ea%25b6%258c-5%25ea%25b0%259c-%25ec%25b0%25bd%25ea%25b2%25bd%25ec%2584%25bc%25ed%2584%25b0-%25ec%258a%25a4%25ed%2583%2580%25ed%258a%25b8%25ec%2597%2585-10%25ea%25b0%259c%25ec%2582%25ac-%25ec%25b0%25b8%25ec%2597%25ac-%25ec%25b0%25bd%25ec%2597%2585-bus-%25ec%2597%25b0) (wowtale.net)
   0 points by startupxo, 2026-11-14T12:25:28.679Z | 0 comments | https://startupxo.com/en/1389/ | https://startupxo.com/en/1389.md
   - This is an event where regional startups present their businesses to investors and hold consultations and networking. It was the 2th Yeongnam Joint IR hosted by the Ministry of SMEs and Startups, held on November 12 at the Grand Josun Hotel in Busan.
   - 10 companies recommended by 5 Creative Economy Innovation Centers in Daegu, Gyeongbuk, Busan, Ulsan and Gyeongnam took part. The Daegu center recommended RealTrial, an AI clinical trial design platform, and YesNet, an infrastructure connecting corporate operating data.
   - Startup-BuS is a program that supports regional deep tech startups from early-stage investment through follow-on investment linkage. The source says it provides stage-by-stage support, including direct investment and TIPS recommendations.
   - The Daegu center runs Startup-BuS Station IR every Thursday. Companies selected through IR pitches receive support including direct investment review, TIPS recommendation and follow-up acceleration.
56. [Hands-on with Amazon's new Kindle Colorsoft: good for comics and cookbooks, but the price starting at 290 dollars is a hurdle](https://techcrunch.com/2026/11/13/amazons-new-kindle-colorsoft-is-a-premium-e-reader-with-a-premium-price/) (techcrunch.com)
   0 points by startupxo, 2026-11-14T12:25:00.142Z | 0 comments | https://startupxo.com/en/1388/ | https://startupxo.com/en/1388.md
   - The Kindle Colorsoft is a 7-inch color e-reader that shows book covers and illustrations in their original colors. The reviewer said it is also useful for readers who use different highlight colors.
   - The base model costs 290 dollars with 16GB of storage. The Signature Edition costs 320 dollars with 32GB, and comes with an aluminum back in graphite or a reddish-brown fig color.
   - After using it for several weeks, the TechCrunch reviewer judged the colors more muted than an iPad's but clearly better than a black-and-white Kindle. The screen refreshes rarely, and the reviewer said they noticed no ghosting.
   - The power button has moved to the top right, and the device is waterproof. Amazon says the battery lasts up to 8 weeks. The reviewer thought first-generation Colorsoft owners have little reason to upgrade, and that if you only read, the 169.99 dollar base model or the 199.99 dollar Paperwhite is better.
57. [How to answer development estimate questions without guessing: give a range based on screen count and past results](https://qiita.com/tsutorm/items/c9ab1e1c64474c341d77?utm_campaign=popular_items&utm_medium=feed&utm_source=popular_items) (qiita.com)
   0 points by startupxo, 2026-11-14T00:40:00.587Z | 0 comments | https://startupxo.com/en/1387/ | https://startupxo.com/en/1387.md
   - A Japanese developer (a Qiita author) proposes that when asked for a development estimate, you should not give a single gut-feeling number, but calculate it as number of screens × effort per screen. The author believes a number once spoken easily becomes the baseline against which the formal estimate is compared later.
   - Effort per screen is not a guess. It is the average obtained by dividing the total effort of past projects by their number of screens. Design, testing, batch and API effort are also included in the total effort and absorbed into the average, and the team fixes the rules for counting screens.
   - Because an estimate is a prediction, the author says to convey it as a range with three values: optimistic, most likely and pessimistic. The article's example (a hypothetical for explanation) is 8 to 18 person-months, an expected value of 11 person-months, and a line of about 12.7 person-months at which the work finishes with roughly 80% probability.
   - The author recommends placing commitments and budgets at the roughly 80% line, not at the expected value. The business side can reduce the amount and the range by cutting screens or finalizing specifications early. The author adds that for large projects effort per screen grows and productivity differs by team, so adjustment is needed.
58. [I let Fable 5.2 write an instrument sound simulator and compose music: piano and strings went well, woodwinds were hard](https://nowokay.hatenablog.com/entry/2026/11/13/101644) (nowokay.hatenablog.com)
   0 points by startupxo, 2026-11-14T00:39:46.518Z | 0 comments | https://startupxo.com/en/1386/ | https://startupxo.com/en/1386.md
   - Japanese developer Naoki Kishida had the AI model Fable 5.2 build, together with a piece of music, a program that creates sound by calculating how instruments resonate (a physical modeling synthesizer). The result is a demo that plays in a browser, and it attempted strings, piano, brass, woodwinds, guitar and vocals.
   - According to the author, piano and guitar were easy to reproduce, while woodwinds with reeds were hard. The flute sounded mechanical, the saxophone was too clean, and the vocals were barely intelligible.
   - The string 4 was not instructed to use sonata form or key changes, yet they appeared. For the piano, once it was fixed so that other strings resonate along with the pressed note, the harmony became softer. The author wrote that Opus 5.5, given the same instructions, produced notes that ran together without breaks and a simple musical structure.
   - There are limitations too. The onset of notes, such as the scraping sound of a bow, was hard to calculate and was replaced with noise. The classical guitar performance used fingerings that are hard for a person to play, and the electric guitar is assessed as sounding convincing thanks to an external amp model.
59. [Keychron wireless trackball mouse T1 HE, with a ball position adjustable to fit the thumb, up for preorder at 59.99 dollars](https://www.theverge.com/tech/1033159/keychron-t1-he-trackball-wireless-mouse-adjustable) (theverge.com)
   0 points by startupxo, 2026-11-13T18:22:58.347Z | 0 comments | https://startupxo.com/en/1385/ | https://startupxo.com/en/1385.md
   - This is a trackball mouse where you move the cursor by rolling the ball with your thumb instead of moving your wrist. Two dials on the bottom let you adjust the ball's position up, down, left and right, and you can also pull the ball forward or push it in.
   - The adjustment range is only a few millimeters in each direction. The original reporter thinks even such small adjustments can help with comfort during daily use. The ball diameter is 34mm.
   - Cursor speed (DPI) can be changed from 100 to 2,400 with a button or the Keychron Launcher software. There are 6 buttons, and the software allows remapping functions and setting macros.
   - The usage times stated in the original are up to 106 hours over Bluetooth on battery 600mAh, and 98 hours with the included 2.4GHz receiver. A USB-C cable connection is also supported. Colors are black, white and transparent, and shipping is expected in early December.
60. [ChatGPT launches wallet feature, and an AI agent picks products and even pays for them](https://techcrunch.com/2026/11/13/chatgpts-new-wallet-lets-ai-agents-shop-and-check-out-for-you/) (techcrunch.com)
   0 points by startupxo, 2026-11-13T18:22:45.033Z | 0 comments | https://startupxo.com/en/1384/ | https://startupxo.com/en/1384.md
   - OpenAI has added a wallet feature to ChatGPT. An AI agent called dots helps choose what to buy and completes the payment on the user's behalf.
   - At first, it is open only to some ChatGPT Pro subscribers in the US aged 18 or older. A Visa, Mastercard or American Express card issued in the US is required.
   - Before payment, users check the product, the total and the card to be used, then approve with Face ID or a fingerprint. OpenAI said it uses payment tokens, so the actual card details are not shared with dots or merchants.
   - OpenAI explains that benefits such as card rewards and the ability to file payment disputes are also kept. It announced plans for spending limit settings and more supported merchants, but some sites, such as Amazon, block AI agents.

Next page: https://startupxo.com/en/index.md?p=3
JSON: https://startupxo.com/api/v1/front.json?lang=en&page=2
Other languages: https://startupxo.com/ko/index.md https://startupxo.com/ja/index.md https://startupxo.com/es/index.md https://startupxo.com/ar/index.md
