Boss lens: platforms now hand merchants ready-made AI for selection, images and service; the sharper warning is that when buyers send AI agents to shop, stores with unclear specs, traceability or after-sales terms may simply be skipped (inference).
AI NEWS
Today’s AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
Latest day
November 12, 2026
11-11 07:50 – 11-12 07:50 (UTC+8)
Top 3
Boss lens: most firms spend about $16 a head on AI while top firms spend hundreds of times more—the gap is about embedding AI in real workflows, not tools; with token prices down 21% in a month, re-quote annual plans before renewal (inference).
Boss lens: “judgment” work—screening documents, tagging customer messages, checking product images for compliance—can now be batch-scored cheaply; have whoever runs your systems assess which repetitive review steps it could pre-screen (inference).
All
Singles Day becomes an AI testbed: over 10 million SMBs trade on Alibaba’s Accio with token use up 10x, and JD’s Jingmai ships 60+ AI tools from product selection to customer service
Boss lens: platforms now hand merchants ready-made AI for selection, images and service; the sharper warning is that when buyers send AI agents to shop, stores with unclear specs, traceability or after-sales terms may simply be skipped (inference).
CollapseSingles Day becomes an AI testbed: over 10 million SMBs trade on Alibaba’s Accio with token use up 10x, and JD’s Jingmai ships 60+ AI tools from product selection to customer service
Xinhua reported on Nov 11 that every major e-commerce platform put AI front and center this Singles Day: Alibaba wired Qwen into Taobao, JD rolled out its AI shopping assistant “Dongdong,” Douyin added Doubao, and tokens were listed as a standalone product on several platforms. Alibaba’s global B2B president Zhang Kuo said more than 10 million SMBs now run their business on its AI workbench Accio, with token usage up 10x year on year; one solo founder profiled runs a custom-cap export business with more than 80 agents handling research, design sketches and overseas inquiries, and had 20-plus brand orders in progress. JD said its Jingmai AI center launched with 60+ tools across product selection, creative, marketing, fulfillment and customer service, helping 80% of new merchants work markedly more efficiently. The report also quotes industry voices cautioning that the economics from model costs to actual orders need a longer period to judge.
Ramp AI Index: US business AI spend rose only 1.65% in October—OpenAI +2.1%, Anthropic +0.5%—and the median firm spends about $16 per employee per month
Boss lens: most firms spend about $16 a head on AI while top firms spend hundreds of times more—the gap is about embedding AI in real workflows, not tools; with token prices down 21% in a month, re-quote annual plans before renewal (inference).
CollapseRamp AI Index: US business AI spend rose only 1.65% in October—OpenAI +2.1%, Anthropic +0.5%—and the median firm spends about $16 per employee per month
Ramp’s November AI Index, built on payments from more than 85,000 US businesses, shows AI spend hit a new high in October but grew just 1.65% month over month—OpenAI up 2.1% and Anthropic up 0.5%, far below the double-digit growth seen for most of the year. Among firms paying for AI, the median spends $16.32 per employee per month, the top 10% about $801 and the top 1% about $10,200; 1% of firms drive roughly 80% of OpenAI and Anthropic revenue. Ramp says the average price per million tokens fell 21% over the past month, so usage is growing faster than spend, and engineering remains the biggest spending department. Ramp notes its customer base skews toward tech-savvy, growth-oriented companies.
Google launches Gemini Decision 1: it makes judgments instead of writing answers, takes up to 1,000 questions per call, and costs $0.10 per million input tokens with no output charge
Boss lens: “judgment” work—screening documents, tagging customer messages, checking product images for compliance—can now be batch-scored cheaply; have whoever runs your systems assess which repetitive review steps it could pre-screen (inference).
CollapseGoogle launches Gemini Decision 1: it makes judgments instead of writing answers, takes up to 1,000 questions per call, and costs $0.10 per million input tokens with no output charge
On Nov 11 Google added the Decisions API and the Gemini Decision 1 model to the Gemini API, built for bounded judgments—yes/no, pick-one, or scoring against a rubric—rather than generating text. One request can carry a single text, image, audio, video or PDF input of up to about 256K tokens plus up to 1,000 independent questions, returning probabilities and confidence in about 150–250 ms end to end, according to Google. Pricing is $0.10 per million input tokens with zero output tokens billed, the shared input is charged once regardless of question count, and there is a free tier in AI Studio. Artificial Analysis scored it 66 on its Decision Index.
Bloomberg: Anthropic’s IPO is multiple times oversubscribed as it seeks up to about $90.4 billion including a private placementUnconfirmed
Boss lens: investors rushing in means model vendors will keep cutting prices to win customers, but public companies answer to quarterly results, so expect more frequent plan changes—keep repricing clauses in long contracts (inference).
CollapseBloomberg: Anthropic’s IPO is multiple times oversubscribed as it seeks up to about $90.4 billion including a private placement
Bloomberg reported on Nov 11, citing people familiar, that orders for Anthropic’s IPO already exceed the shares available several times over, including large anchor-size indications of interest. The company and its co-founders aim to raise up to about $90.4 billion through the IPO and a concurrent private placement, with the range still $70 to $78 a share. CNBC earlier reported pricing is planned for Nov 18 with trading on Nov 19. During the roadshow, CEO Dario Amodei also predicted AI could make cancer, Alzheimer’s and other major diseases history in under a decade—his own forecast. The oversubscription comes from anonymous sources.
Bloomberg: ByteDance will launch a Muse-style proactive AI agent in the coming weeks, and Zhang Yiming says catching OpenAI and Anthropic will take until at least 2029Unconfirmed
Boss lens: pair Doubao and Douyin’s reach with an agent that acts on its own and “let the AI find and book it” arrives faster in China—merchants on Douyin should write product and service info an AI can parse (inference).
CollapseBloomberg: ByteDance will launch a Muse-style proactive AI agent in the coming weeks, and Zhang Yiming says catching OpenAI and Anthropic will take until at least 2029
Bloomberg reported on Nov 11, citing people familiar, that ByteDance plans to release an AI assistant similar to Meta’s Muse—a voice agent with an animated avatar launched in September that has over 10 million users—in the coming weeks. At an all-hands of the AI team, Zhang Yiming said ByteDance can compete with Anthropic and OpenAI at the frontier but not before 2029 or 2030; his top concern is data, saying Seed models are pre-trained on datasets orders of magnitude smaller than rivals’, and he reiterated the company should not rely on distillation. The report says ByteDance plans to spend up to $70 billion to make Seedance China’s leading video AI platform. The agent’s name, form and any hardware are unknown, and ByteDance did not comment.
Baidu releases ERNIE 5.5 Flash Preview: 280B total parameters and 1M context, live on Qianfan at half price for a limited time, with weights due as open source by Nov 25
Boss lens: another low-cost domestic model, with off-peak prices halved again—good for non-urgent batch work like product copy or document cleanup; wait for the final release and license before using it in production (inference).
CollapseBaidu releases ERNIE 5.5 Flash Preview: 280B total parameters and 1M context, live on Qianfan at half price for a limited time, with weights due as open source by Nov 25
Baidu on Nov 11 released ERNIE 5.5 Flash Preview, the first post-trained version of the Base model open-sourced on Nov 5: 280B total parameters with 14B active and a 1M-token context, pitched at coding, 3D modeling, knowledge work and content creation, with output speeds above 200 tokens per second. The API is live on Baidu AI Cloud Qianfan; per Techub, it costs RMB 2 per million input tokens at peak and RMB 1 off-peak, and RMB 8 / 4 for output, currently at half price; weights are promised under Apache 2.0 by Nov 25. Baidu says this is a mid-training preview that may have gaps, and reviewers say it still trails frontier models.
Alibaba opens Qwen4-Max Preview until Nov 25: Token Plan calls cost as little as 20% of standard, less at night, and it is also in the Qwen app and Qoder
Boss lens: a two-week window at 20% of list price is a cheap chance for staff to test it on real documents and spreadsheets against current tools—but it retires on Nov 25, so do not build live workflows on it (inference).
CollapseAlibaba opens Qwen4-Max Preview until Nov 25: Token Plan calls cost as little as 20% of standard, less at night, and it is also in the Qwen app and Qoder
From 10:00 on Nov 11 Alibaba opened a preview of its next flagship Qwen4-Max, which it says improves knowledge, coding, agentic tasks, office work and front-end design, with a 1M-token context, image and video understanding, web search and tool calling. Developers can only reach it through the subscription-based Token Plan, not pay-as-you-go; during the preview usage is billed at as little as 20% of the standard rate, with an extra discount from 22:00 to 08:00, and the Lite personal plan is $3 for new users’ first month. It is also live in the Qwen app, Qwen Office, Qoder and QwenStudio. The preview ends at midnight on Nov 25 when the model is retired; Alibaba has published no benchmarks yet, and community reviews of its coding are mixed.
Alibaba open-sources Qwen Code 1.0, merging Qoder’s agent harness so companies can embed it, with free Qwen4 credits for sign-ins through Nov 24
Boss lens: companies wanting a hands-on AI assistant inside their own systems get a free open-source base; it is a call for the tech team—the boss just needs to ask how much it saves versus buying, and who maintains it (inference).
CollapseAlibaba open-sources Qwen Code 1.0, merging Qoder’s agent harness so companies can embed it, with free Qwen4 credits for sign-ins through Nov 24
Alibaba on Nov 11 released Qwen Code 1.0 Preview, open-sourcing Qoder’s agent harness—the layer around a model that calls tools and manages context—and merging it into Qwen Code so developers can embed it in their own products. Per IT Home and others, signing in to Qoder inside Qwen Code between Nov 11 and Nov 24 grants free Qwen4 trial credits valid for 30 days. The code is on GitHub. NFT_Chen’s post drew about 230 likes and 170 bookmarks.
Altman told Samsung executives AI demand has been double-counted: expect short-term volatility, with only OpenAI and Anthropic truly driving demand nowUnconfirmed
Boss lens: when the biggest buyer says demand is inflated, compute, server and memory prices may swing over the next year—companies planning their own hardware should buy in stages (inference).
CollapseAltman told Samsung executives AI demand has been double-counted: expect short-term volatility, with only OpenAI and Anthropic truly driving demand now
Korea’s Financial News reported exclusively on Nov 11 that OpenAI CEO Sam Altman told about 50 Samsung executives in Korea on Oct 31 that AI demand for memory and data centers has been double-counted across supply-chain forecasts, which could bring short-term volatility or even oversupply; he said only OpenAI and Anthropic are really driving demand growth right now. He added that long-term demand will explode and chips and data centers remain the bottleneck. The account comes from the paper’s exclusive; OpenAI has not commented publicly.
SemiAnalysis says persistent agents went from 0% to 45% of its daily AI spend between February and early NovemberUnconfirmed
Boss lens: AI spend is shifting from staff asking a chatbot to agents working on their own—pick one daily task (say, a morning roundup of orders and bad reviews), run it for a month and read the bill before expanding (inference).
CollapseSemiAnalysis says persistent agents went from 0% to 45% of its daily AI spend between February and early November
According to FourWeekMBA’s write-up of a SemiAnalysis podcast, CEO Dylan Patel said persistent agents—AI that stays running and works on schedules or triggers—went from 0% of the firm’s daily AI spend when tracking began on Feb 1 to 45% in the week of Nov 8. He said the main products are Perplexity Computer and Claude-based tools, with Perplexity mostly calling Anthropic models for their use cases. This is a single research firm’s self-reported figure, not independently verified and not representative of typical businesses.
n8n Labs open-sources Nod, a personal agent with its own computer and browser that chases invoices, cancels subscriptions and files supplier bills, checking in once a task passes $5
Boss lens: chores like reconciling, chasing payments and downloading invoices now have a free agent to take them on—but it logs into your accounts and cards, so have a tech-savvy colleague trial it on a spare machine with one low-risk job first (inference).
Collapsen8n Labs open-sources Nod, a personal agent with its own computer and browser that chases invoices, cancels subscriptions and files supplier bills, checking in once a task passes $5
n8n Labs, an experimental team at the automation company n8n, released Nod as a free, open-source (Apache 2.0) research preview for Mac and Windows on Nov 10. It runs in a sandboxed local VM with its own browser, working in connected apps like Gmail, Calendar and Slack and on any website to finish jobs rather than hand back a to-do list—examples include cancelling a subscription and asking for a refund, logging into supplier portals monthly to file invoices for the accountant, and copying last week’s numbers from a dashboard with no export. You can assign jobs over Telegram or WhatsApp. A separate check confirms you asked before it sends, pays or deletes; text on web pages or emails cannot give it orders; and logins and cards are encrypted locally. Each task shows its cost and it checks in past $5. You bring your own model account, and it stops when the laptop sleeps.
a16z: AI agents are now most web traffic, and orders from AI search at Shopify stores grew nearly 13x—businesses must learn to sell to agents
Boss lens: it echoes today’s Singles Day item—more buyers ask an AI before purchasing; have marketing check whether your site, store pages and price lists are readable by AI and up to date (inference).
Collapsea16z: AI agents are now most web traffic, and orders from AI search at Shopify stores grew nearly 13x—businesses must learn to sell to agents
In a Nov 10 essay a16z described the “B2A economy”: agents now make up the majority of web traffic, write more than half of new online articles and generate most code shipped to production. Cited figures include Shopify saying AI-driven traffic to its stores grew 8x year over year in Q1 and orders from AI search nearly 13x; Human Security seeing agentic browser traffic grow 8x from August to October; and AI-referred shoppers converting better with lower bounce rates. The authors argue agents will increasingly decide what to buy and subscribe to, so businesses that once blocked bots will have to compete for them. The majority-traffic claim cites no specific source, and the Shopify figures are quoted secondhand.
IBM and Microsoft launch Imagine Enterprise, an agent platform with 6,000+ prebuilt agents starting with finance, insurance and retail
Boss lens: big vendors now sell agents on control—permissions, logs, traceability—rather than smarts; smaller firms will not need this platform, but can ask any AI vendor the same questions (inference).
CollapseIBM and Microsoft launch Imagine Enterprise, an agent platform with 6,000+ prebuilt agents starting with finance, insurance and retail
IBM and Microsoft on Nov 10 launched Imagine Enterprise, a platform for companies to build, run and govern AI agents on their own business data. IBM Consulting contributes more than 6,000 prebuilt agents and industry assets, and Microsoft provides AI, cloud, identity, governance and security; the first industry frameworks target financial services, insurance and retail. It keeps human oversight, permission controls, audit logs and task tracing, and can be adopted through IBM Consulting engagements or self-managed.
One Claude run cost $1.99 and 94% went to reading files: Spotify routes file reads to a cheaper model with a hook and cuts tokens about 90%
Boss lens: much AI spend goes to reading, not thinking—have whoever owns the bill check how much goes to reading files and long documents; routing that to a cheaper model often saves most of it (inference).
CollapseOne Claude run cost $1.99 and 94% went to reading files: Spotify routes file reads to a cheaper model with a hook and cuts tokens about 90%
Product writer Aakash Gupta posted on Nov 11 that one Claude Opus coding run cost them $1.99, of which only 13 cents went to writing the answer and 94% to the model reading files; they then tested Spotify’s approach on two codebases. Spotify’s open-source Shunt plugin uses Claude Code hooks to block full reads of files over 350 lines and hand them to a cheaper model that returns a summary, cutting tokens by about 90% in bulk-read scenarios in Spotify’s Java monorepo. The post drew 70 bookmarks versus 63 likes.
levelsio adds video to Interior AI: upload a living-room video, photos or a floor plan to empty the room, move furniture and repaint surfaces
Boss lens: for home-improvement and furniture shops, turning a customer’s phone video into an “after” view could speed up closing; similar tools exist in China—have a designer or salesperson try one on a real customer’s home (inference).
Collapselevelsio adds video to Interior AI: upload a living-room video, photos or a floor plan to empty the room, move furniture and repaint surfaces
Indie maker levelsio posted on Nov 11 that Interior AI, his four-year-old interior design app that until now only turned room photos into redesign images, can now take a video, a few photos or just a floor plan, then empty the room, move or remove furniture and recolor walls, floors, ceilings or any object. Next up are a top-down floor plan view, one-click style changes and folding it all into the product. The post drew about 98 likes and 73 bookmarks; the feature is still in development.
ChatGPT retires Advanced and Standard Voice on Dec 11, moving everyone to Live voice
Boss lens: if staff use ChatGPT voice for sales-script practice, language training or demos, test the new Live voice before Dec 11—behavior may differ (inference).
CollapseChatGPT retires Advanced and Standard Voice on Dec 11, moving everyone to Live voice
OpenAI’s help center says ChatGPT’s Advanced and Standard Voice modes will be retired on Dec 11, with users automatically moved to the new Live voice and the old modes removed from Settings. Existing voice conversations stay in history and can be continued by text or voice. Gadgets 360 reported the change follows a spike in GPT-Live usage.
Gergely Orosz: fresh data shows native mobile and frontend roles shrinking faster, with only senior people landing them—generalist engineers who use AI are safer
Boss lens: app and web outsourcing quotes and team shapes will change—one engineer with AI can cover what a small team did; when hiring or outsourcing, ask how they deliver with AI and how short the timeline can get (inference).
CollapseGergely Orosz: fresh data shows native mobile and frontend roles shrinking faster, with only senior people landing them—generalist engineers who use AI are safer
Gergely Orosz, author of The Pragmatic Engineer, posted on Nov 11 that fresh hiring data confirms what he said in April: native mobile and frontend roles are dying, and faster, with mostly very experienced people still landing them. His advice is to become a more general software engineer who works with AI for career stability. The post drew about 1,630 likes and 760 bookmarks, with many replies describing similar shifts at their companies; it reflects US tech hiring data.
Allie K. Miller shares five ways to get sharper AI critiques: three reviewer personas, cross-checking with other AIs, and telling it the work is someone else’s
Boss lens: once a proposal, contract, quote or poster is done, spend five minutes having AI play a picky customer, a competitor and a confused newcomer—far more useful than asking “is this good?”, and no tech skills needed (inference).
CollapseAllie K. Miller shares five ways to get sharper AI critiques: three reviewer personas, cross-checking with other AIs, and telling it the work is someone else’s
AI adviser Allie K. Miller shared five non-coding ways to make AI adversarial reviews more useful on Nov 11: have Claude spin up three reviewer personas—a superfan, someone who wants to destroy you, and a confused newcomer; make reviews multimodal with walkthrough videos and screenshots; cross-review with other AI systems (Codex reviews what Claude built and vice versa); bounce the review back and forth between main agent and personas; and tell the AI the work is someone else’s, which she finds reduces flattery. The post drew 85 bookmarks versus 56 likes.
England’s university regulator warns AI lets students finish unsupervised assessments in minutes and recommends closed-book exams and live presentations
Boss lens: résumés, take-home tests and online assessments may all be AI-made—add a live task or in-person walkthrough when hiring; it reveals real ability better than submitted work (inference).
CollapseEngland’s university regulator warns AI lets students finish unsupervised assessments in minutes and recommends closed-book exams and live presentations
England’s Office for Students and Advance HE published a report on Nov 11 drawing on 244 staff and 91 students at at least 127 institutions, finding that AI produces high-quality work quickly and has undermined unsupervised assessments. It recommends universities help students build AI skills while relying more on secure methods such as closed-book exams, presentations and practical assessments. The Telegraph reported universities that fail to prevent AI cheating could face fines.
ACE Music v1 tops Artificial Analysis’s vocal music leaderboard ahead of Suno v6, generating full songs from a prompt with 20 free credits a day
Boss lens: store background music, short-video soundtracks and brand jingles can be generated in-house—read the platform’s copyright and commercial-use terms before running them in ads (inference).
CollapseACE Music v1 tops Artificial Analysis’s vocal music leaderboard ahead of Suno v6, generating full songs from a prompt with 20 free credits a day
Artificial Analysis said on Nov 11 that ACE Studio’s ACE Music v1 ranks first on its vocal music leaderboard (AA-Music-Vocal v1.1), ahead of Suno v6. It generates full vocal or instrumental tracks from a description or your own lyrics, doing best on reggae and classical/orchestral prompts. ACE Studio, founded in 2018, co-develops the open-weight ACE-Step music models with StepFun. There is no API; it is available in the ACE Music app with 20 free credits a day and paid plans from $7.90 a month billed yearly for up to 500 songs a month. The post drew about 570 likes and 340 bookmarks.
Tw93 deleted nearly a third of his AI Skills and shared a 14-step prompt that has the agent list proposed cuts for approval before trimming rules
Boss lens: the “handbook” you give AI needs regular pruning, just like company policies that pile up until nobody follows them; the rule worth copying is the floor—never drop the rules on spending, deleting or publishing (inference).
CollapseTw93 deleted nearly a third of his AI Skills and shared a 14-step prompt that has the agent list proposed cuts for approval before trimming rules
Developer Tw93 posted on Nov 11 that he deleted nearly a third of his installed Skills and slimmed the context every project gives the model, arguing that as models get stronger, old rules start getting in the way. He packaged the method as a prompt to hand an agent: inventory what loads every turn and how long it is, propose fix/delete/move/rewrite/keep for each section as a checklist and change nothing until approved; write goals and “what done looks like” rather than steps; but never delete personal preferences, known pitfalls, or rules on irreversible actions like deleting data or spending money. A second agent that did not do the edit then checks nothing was over-cut. The post drew 545 bookmarks versus 446 likes.
Baoyu: Stanford’s CS329Z “Engineering AI Agents” goes fully public, and the first assignment builds a Claude Code-style agent from scratch for under $5
Boss lens: for whoever handles tech or digitization, this is a free, hands-on curriculum on how “AI employees” are built—bosses need not study it, but can pass it to whoever evaluates AI projects (inference).
CollapseBaoyu: Stanford’s CS329Z “Engineering AI Agents” goes fully public, and the first assignment builds a Claude Code-style agent from scratch for under $5
Baoyu wrote on Nov 11 that Stanford’s new course CS329Z, Engineering AI Agents, is releasing all materials publicly: one lecture a day on Stanford Online’s YouTube and both assignments on GitHub under MIT. The first bans agent frameworks and starts from a bare text-in, text-out interface: students build email triage, daily digests and retrieval Q&A over a real corporate email archive, then an internal assistant for a fictional energy company with tool calls, context compression, cross-session memory and user approval before actions; the instructor says it takes under $5 of API credit. The second teaches how to write benchmark tasks for agents. The post drew about 1,460 bookmarks versus about 930 likes.
Baoyu: don’t ask an AI what it can do—give it the goal and a high bar, the way a good manager directs staff
Boss lens: treat AI like a capable new hire—skip “can you do reports?” and say “I need a shareholder-ready monthly review by Monday; the data is in these sheets”; retry what failed last time after a model upgrade (inference).
CollapseBaoyu: don’t ask an AI what it can do—give it the goal and a high bar, the way a good manager directs staff
Baoyu posted on Nov 11 that a common mistake is asking an AI what it can do and then only assigning what it claims, because models do not really know their own abilities, especially when unaware of available tools and Skills. His advice: state the goal and set a high bar, like a strong manager who does not ask staff what they can do but explains the goal, gives direction and supplies context and tools. He adds that after every model upgrade, tasks that failed before are worth retrying. The post drew about 280 likes, 160 bookmarks and 130-plus replies.
A Hermes Desktop data plugin: drop in an Excel or CSV file and the AI suggests a few angles and three to five charts for you to pick
Boss lens: for sales, inventory or store ledgers, “AI suggests angles, a person picks” beats letting AI jump to conclusions; have whoever handles data try it on last month’s real sheet (inference).
CollapseA Hermes Desktop data plugin: drop in an Excel or CSV file and the AI suggests a few angles and three to five charts for you to pick
Blogger geekbb highlighted a data-analysis plugin for Hermes Desktop on Nov 11: drop in an Excel, CSV or JSON table and, instead of dumping conclusions, the AI first suggests a few angles plus a title and three to five chart ideas beside the page for you to select. The post drew about 17 likes and 39 bookmarks, more than twice as many bookmarks as likes; it is a community plugin, described by its author.