Boss lens: Pro/Business spend now includes a build budget for API prototyping—check seat tiers and the $500 workspace cap before treating it as unlimited offset (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 8, 2026
11-07 07:50 – 11-08 07:50 (UTC+8)
Top 3
Boss lens: another personal coding-assistant free tier—don’t assume the whole company can ride it, because Teams/Enterprise are left out (inference).
Boss lens: compute growth is hitting power and water limits—before siting a node in Southeast Asia, ask what is still being approved and whether tariffs will be separated (inference).
All
OpenAI: ChatGPT Pro/Business plans can claim up to ~$500/month in API credits; Developers plugin adds a Showcase tab
Boss lens: Pro/Business spend now includes a build budget for API prototyping—check seat tiers and the $500 workspace cap before treating it as unlimited offset (inference).
CollapseOpenAI: ChatGPT Pro/Business plans can claim up to ~$500/month in API credits; Developers plugin adds a Showcase tab
Per OpenAI Developers and official docs: the Developers plugin adds a Showcase tab of starter apps that can be opened in Codex. Eligible ChatGPT Pro and Business subscribers can claim API credits in billing settings to cover app API usage. Pro $100/$200/$500 tiers get matching monthly credits; Business Standard is $20 per paid seat and Premium $100 per paid seat, pooled and capped at $500 per workspace per month. The subscription must be paid and at least seven days old; the same owner email must link the ChatGPT plan and API org. Unused credits expire at cycle end and do not roll over. Docs say the benefit does not cover Codex usage via API keys, does not raise ChatGPT or signed-in Codex limits, and excludes Enterprise.
Alibaba Qoder/Qoder CN: anonymous model Qoduck offers daily free personal quota from Nov 7; in-quota calls cost 0 Credits
Boss lens: another personal coding-assistant free tier—don’t assume the whole company can ride it, because Teams/Enterprise are left out (inference).
CollapseAlibaba Qoder/Qoder CN: anonymous model Qoduck offers daily free personal quota from Nov 7; in-quota calls cost 0 Credits
Per official Qoder CN and international docs: from Nov 7, 2026, personal users get a daily free allowance for the anonymous model Qoduck (identity not disclosed). In-quota calls drop from a 0.5× credit coefficient to 0.0×, including the highest thinking setting. No claim step—just select the model; free, trial, and zero-balance accounts qualify. Paid personal plans get more daily uses than free plans; the exact count is unpublished. After the daily cap, users can wait until the next day or keep going at 0.5× credits. Teams/Enterprise plans are excluded. The promo covers the Qoder/Qoder CN apps only, not the IDE, JetBrains plugin, or CLI; end date TBD. Official notes possible peak slowdowns.
Guardian: Southeast Asia holds ~50% of APAC data centers under construction; Thailand paused 49 builds and 117+ approvals in September
Boss lens: compute growth is hitting power and water limits—before siting a node in Southeast Asia, ask what is still being approved and whether tariffs will be separated (inference).
CollapseGuardian: Southeast Asia holds ~50% of APAC data centers under construction; Thailand paused 49 builds and 117+ approvals in September
The Guardian (Nov 7), citing a 2026 Cushman & Wakefield report, says Southeast Asia accounts for about half of Asia-Pacific data-center capacity under construction, with Malaysia and Thailand a large share. In September 2026 Thailand paused construction on 49 projects and held approvals on more than 117 others, citing insufficient detail for regulation and monitoring. Johor’s operational capacity grew from about 94 MW in 2023 to about 1,110 MW in 2026, with about 3,088 MW more under construction or planned; the state chief minister put cumulative data-center investment at about $40 billion. Wood Mackenzie estimates Johor data centers already use about 24% of the state’s electricity, potentially about 40% by 2035. West Java halted one project over permits and an incomplete environmental review; after Singapore’s 2019–2022 moratorium, only about 95 MW was added there in 2023–2026.
DeepSeek: web/API partial outage 15:11–15:56 on Nov 7 lasted about 45 minutes; official status page says restored
Boss lens: even a Saturday afternoon can go dark for 45 minutes—keep a fallback model and retries for anything that cannot wait (inference).
CollapseDeepSeek: web/API partial outage 15:11–15:56 on Nov 7 lasted about 45 minutes; official status page says restored
Per IT Home and DeepSeek’s official status page: users reported “server temporarily unavailable” on the afternoon of Nov 7. The incident, labeled a partial web/API outage, ran 15:11–15:56 Beijing time (07:11–07:56 UTC), or 45 minutes. Listed components include the DeepSeek V4 Pro API and V4.1 Flash API. A fix was marked as implemented at 15:11 and the incident resolved at 15:56. IT Home likewise described a partial web/API interruption. The status page did not state a root cause.
xiaohu on an official X creator guide: Ethan Ng says hand-editing one short takes ~2 hours; Grok Bot cut 10 in ~2 hours with feedback rounds
Boss lens: 125 bookmarks. The savings are in repetitive editing, not in choosing the story or final cut—keep a human approve-before-render step (inference).
Collapsexiaohu on an official X creator guide: Ethan Ng says hand-editing one short takes ~2 hours; Grok Bot cut 10 in ~2 hours with feedback rounds
Per @xiaohu’s write-up of an official X Creators guide by Ethan Ng: Ng does creator marketing on the SpaceXAI growth team and says the team works with hundreds of creators each month. He usually spends about two hours hand-finishing one short; this time he sent ten source videos to cloud-hosted Grok Bot in parallel and got cuts back in about two hours, including motion, overlays, and music, after several feedback rounds. xiaohu notes Ng is an insider, so the guide is effectively official. The workflow stresses feeding past films and edit notes first, having humans draft scripts while the bot critiques, and keeping two human gates (approve the plan, then review the cut). A 0:18 retention-drop figure is described as a schematic, not live data. The stack cited includes Whisper, Remotion, and Cursor’s cloud agent.
The Information: Cyera said to be raising ~$800M at a ~$20B valuation; ARR described in the hundreds of millionsUnconfirmed
Boss lens: securing what AI agents can see is attracting outsized private-market prices—treat $20B as a reported ask, not a closed print (inference).
CollapseThe Information: Cyera said to be raising ~$800M at a ~$20B valuation; ARR described in the hundreds of millions
The Information said on Nov 7 that data-security firm Cyera is raising about $800 million at a $20 billion valuation, with annual recurring revenue described as having reached the hundreds of millions over the past year. Secondary write-ups say Cyberstarts is expected to lead and that $20 billion would be a sharp step-up from about $12 billion in June 2026. Dealroom and others caution the talks remain preliminary and Cyera declined to comment—this is not a closed-deal announcement. Full details sit behind a paywall; figures here come from The Information’s public post and follow-on summaries.
Gergely Orosz: Jev billed $0 after 1M+ tokens vs about $3–5 on an LLM; official list price is $42 per billion input tokens
Boss lens: routing and yes/no checks need not ride a frontier LLM—split task types before you celebrate the savings (inference).
CollapseGergely Orosz: Jev billed $0 after 1M+ tokens vs about $3–5 on an LLM; official list price is $42 per billion input tokens
Pragmatic Engineer author Gergely Orosz said a new search site of his used more than 1 million Jev tokens and still showed a $0.00 bill, versus an estimated $3–5 on a typical LLM. He highlighted that Jev is priced per billions of tokens, not millions. TypeSafe’s docs list Jev 1.13 and 1.14 at $42 per billion input tokens ($0.042 per million), with output tokens free. Jev is a System One model for structured decisions—classification, scoring, yes/no—not a general-purpose writer. One workload is not a full cost study.
Anthropic: if current capabilities were widely adopted, AI could take on 27–52% of researcher time per study, median ~38%
Boss lens: today’s gains cluster in analysis and writing, not bench work—don’t budget as if every field just doubled (inference).
CollapseAnthropic: if current capabilities were widely adopted, AI could take on 27–52% of researcher time per study, median ~38%
Ethan Mollick highlighted Anthropic’s Nov 5 report on how researchers use Claude. Using privacy-preserving methods, Anthropic studied 207,049 consumer-plan science sessions from Aug 2 to Sept 30, 2026 (128,914 on Claude and 78,135 on Claude Code). Data analysis accounted for about 33.7% of Claude science sessions and 48.1% of Claude Code sessions. In an illustrative scenario of wide adoption of current capabilities, AI could take on 27–52% of the researcher time needed for a study, with a 38% median across fields—implying accelerations from about 1.35× in the arts to about 2.02× in statistics and 1.96× in computer science. These are scenarios, not forecasts. In an accompanying survey, about 40% of respondents expected AI to run full research projects on its own by 2030.
Xinhua Finance: Hygon Q1–Q3 revenue ~¥14.955B (+57.58%); Cambricon ~¥9.526B (+106.76%); Zhipu said to lift year-end ARR guidance to ~$30B
Boss lens: domestic chip vendors are showing real revenue, but Zhipu’s ARR/token figures are restated September guidance—ask whether “adapted” means actually shipped at scale (inference).
CollapseXinhua Finance: Hygon Q1–Q3 revenue ~¥14.955B (+57.58%); Cambricon ~¥9.526B (+106.76%); Zhipu said to lift year-end ARR guidance to ~$30B
A Nov 7 Xinhua Finance roundup, carried by Sina Finance, cites recent third-quarter filings: Hygon revenue of ¥14.955 billion (+57.58%) and net profit of ¥2.847 billion (+45.12%); Cambricon revenue of ¥9.526 billion (+106.76%) and net profit of ¥3.715 billion (+131.49%); Moore Threads and MetaX revenue of ¥3.289 billion and ¥2.824 billion, up 319.15% and 128.45%, both swinging to profit. The piece says leading model releases now get same-day domestic-chip support, shortening software-hardware sync from months to days. It also says Zhipu in September raised year-end ARR guidance from $24 billion to $30 billion and that GLM-5.3 Flash exceeded 20 trillion tokens a day in September. Those Zhipu figures are the roundup restating earlier company guidance, not a fresh Nov 7 earnings print.
SandBaseAI announces a multi-million-dollar seed round for personal-agent data APIs and runtime
Boss lens: one door for data and runtime can save a small team glue work—but “multi-million” is still an early-stage check, not a proven platform (inference).
CollapseSandBaseAI announces a multi-million-dollar seed round for personal-agent data APIs and runtime
Z Potentials reported on Nov 7 (via Sina Finance) that SandBaseAI closed a multi-million-dollar seed round. Backers were described as including a foundation-model company and investors with AI-builder community experience; proceeds will go to product and go-to-market. Founder Li Yangbing previously led Meicai’s cloud-native unit and worked as an architect at DP Technology. The product combines data APIs and an agent runtime, claiming 2,000+ models and APIs across 100+ platforms, plus SDK and CLI/MCP hooks for Claude Code, Codex, and Cursor. The article does not give an exact raise, post-money valuation, or named lead. There is a free allowance; paid usage is billed per service.
TSMC says it remains a logic foundry after The Information reported it was exploring 3D DRAM at Nvidia’s requestUnconfirmed
Boss lens: memory tightness is pushing unusual supply-chain talk—don’t bake “TSMC makes DRAM” into next year’s plan while the company is saying it remains a logic foundry (inference).
CollapseTSMC says it remains a logic foundry after The Information reported it was exploring 3D DRAM at Nvidia’s request
Tom's Hardware reported on Nov 7 that The Information, citing two people familiar with the matter, said TSMC had begun preparing plans for 3D DRAM manufacturing at the request of Nvidia, its largest customer, with the effort overseen by Hsu Kuo-chin, an advanced-packaging vice president and former Micron Taiwan head. Asked by Tom's Hardware, a TSMC spokesperson said the company “focuses on our core business at all times, which is dedicated IC foundry” and that its mission is to be “the trusted technology and capacity provider of the global logic IC industry.” The outlet treated that as a denial of the report. The same piece notes that entering memory would expose TSMC to cyclical risk and could complicate ties with HBM partners such as SK Hynix. Price, site, and timeline were not disclosed.
Techmeme citing The Information: public records appear to show a SpaceX “Starbase supercomputer” in Texas; the governor-elect may delay some data-center plansUnconfirmed
Boss lens: giant compute campuses now hinge on local elections and permits—put pause risk on the timeline before you expand (inference).
CollapseTechmeme citing The Information: public records appear to show a SpaceX “Starbase supercomputer” in Texas; the governor-elect may delay some data-center plans
Techmeme on Nov 7 summarized The Information’s Ann Davis Vaughan: public records appear to show SpaceX building a “Starbase supercomputer” in Texas, while Gina Hinojosa’s election as governor may delay some of Musk’s data-center plans. The underlying story is paywalled; the tweet does not give capacity, capex, or a construction timetable. The phrasing “appear to show” marks this as a reading of public records, not a formal groundbreaking announcement.
Google Antigravity 2.29: Remote Agents for AI Pro/Ultra plus Voice Mode powered by Gemini Live
Boss lens: agents that keep working after you shut the laptop are now a paid-tier feature—check Pro/Ultra seats before promising always-on coding (inference).
CollapseGoogle Antigravity 2.29: Remote Agents for AI Pro/Ultra plus Voice Mode powered by Gemini Live
9to5Google reported on Nov 6 at 7:01 a.m. PT that Google’s agentic development platform Antigravity shipped version 2.29, with a redesigned conversation page and Voice Mode powered by Gemini Live (playback speed 0.75× to 2×). The headline feature is Remote Agents, rolling out to paid Google AI Pro and Ultra subscribers: create a remote project from a GitHub repo or Google Drive folder; tasks run in isolated cloud VMs and keep going after you close the laptop; permissions can be handled from a phone at antigravity.google.com/remote. The official @antigravity account said remote agents had been rolling out “this week.” The article landed about nine hours before this edition’s window and is included as a first-party product update.
geekbb recommends Kiri, a Rust container runtime compatible with docker CLI/Compose/buildx; site claims ~8× faster starts than Docker Desktop
Boss lens: 166 bookmarks. A drop-in docker-compatible runtime can cut local overhead—but it is still in beta, so pilot with one machine first (inference).
Collapsegeekbb recommends Kiri, a Rust container runtime compatible with docker CLI/Compose/buildx; site claims ~8× faster starts than Docker Desktop
geekbb recommended a Rust runtime on macOS, Windows, and Linux as a lighter stand-in for Docker Desktop / OrbStack that keeps the docker CLI, Compose, and buildx, running containers in a small Linux VM. The linked site, runkiri.com, names the product Kiri and calls it a free public beta. On a 2-CPU / 8 GB comparison it claims docker run takes about 22 ms versus about 178 ms for Docker Desktop, 152 ms for Colima, and 279 ms for OrbStack. The same site says macOS uses a single Linux VM via Virtualization.framework, Windows runs inside WSL2, and Linux talks to the host kernel with no VM. This is a practitioner tool tip, not a big-vendor launch.
Guo Yu: hosted image/video APIs are bound by commercial rules; for large, consistent long-form video he sees rented cloud GPUs as the practical path
Boss lens: hosted media APIs are convenient until consistency and policy limits show up—budget cloud GPUs separately if you need volume and a house style (inference).
CollapseGuo Yu: hosted image/video APIs are bound by commercial rules; for large, consistent long-form video he sees rented cloud GPUs as the practical path
Guo Yu (@turingou) wrote that several vertical personal agents all needed image and video generation, so he had to solve orchestration in a middle layer. Versus ComfyUI’s cloud scheduler, hosted APIs such as fal felt more constrained by each vendor’s commercial rules and “baseline,” so output is not fully unrestricted. He uses Claude Code with scripts and pre-edits as a “director” layer to keep ComfyUI workflows and hosted APIs more consistent, but says that only works if a local machine keeps reviewing with the director model so scores stay close across directors. His conclusion: for large-scale, highly consistent long-form video, renting cloud GPUs is essentially the only path. This is one builder’s practice, not a lab benchmark.
MaxxBench: after small changes to 89 DeepSWE tasks, models near 74% spread by ~23 points; DeepSeek V4 Flash fell 53.3% to 28.1%
Boss lens: a high public coding score may not survive a renamed file or extra requirement—test vendors on your own slightly changed tasks (inference).
CollapseMaxxBench: after small changes to 89 DeepSWE tasks, models near 74% spread by ~23 points; DeepSeek V4 Flash fell 53.3% to 28.1%
Justin Lo (@loresjpg) and the MaxxBench site introduce 89 DeepSWE coding tasks with small changes to the instructions or starter code—the engineering problem stays, the exact requirement does not. Models that clustered near 74% on DeepSWE ended about 23 points apart. The site lists Claude Opus 5.5 at 76.4% (74.2% on DeepSWE); DeepSeek V4 Flash falling from 53.3% to 28.1% (−25.2 points); GPT-5.6 Luna from 67.2% to 46.1% (−21.1); Gemini 3.8 Flash from 73.8% to 53.4% (−20.4); and DeepSeek V4.1 Flash from 74.2% to 55.6% (−18.6). It is a third-party mutation benchmark, meant to test whether scores survive slightly new requirements.
Artificial Analysis voice board: HeyGen Voice leads overall at ~1189 Elo; Qwen-Audio-3.1-TTS-Plus is second at ~1179 and leads in US English
Boss lens: the overall TTS leader may not win in your accent market—A/B US vs UK samples, then look at per-million-character price (inference).
CollapseArtificial Analysis voice board: HeyGen Voice leads overall at ~1189 Elo; Qwen-Audio-3.1-TTS-Plus is second at ~1179 and leads in US English
Artificial Analysis said accent changes text-to-speech rankings: Qwen-Audio-3.1-TTS-Plus leads in US English, while HeyGen Voice does better in UK English and leads the overall board. Its Controlled Voice leaderboard compares eight cloned voices (four US, four UK). HeyGen Voice is about 1189 Elo (~2,414 samples, about $30 per million characters), Qwen-Audio-3.1-TTS-Plus about 1179 Elo (~2,756 samples, about $19.31 per million characters), and HeyGen Voice Turbo about 1173 Elo (about $4 per million characters). These are third-party arena scores, not a vendor launch event.
PSI: an internal research system using a non-public Anthropic model produced a proof of the Aaronson–Ambainis conjecture, accepted by Lean
Boss lens: this bounds unstructured quantum speedups—it does not say quantum computing is useless. Keep structured advantages separate in any external narrative (inference).
CollapsePSI: an internal research system using a non-public Anthropic model produced a proof of the Aaronson–Ambainis conjecture, accepted by Lean
Gorden Sun circulated Physical Superintelligence’s Nov 6 note: a 2008 conjecture of Scott Aaronson and Andris Ambainis asked whether quantum computers can be exponentially faster on unstructured random inputs. PSI says an internal research system, running a non-public Anthropic model on PSI’s own infrastructure, found a complete proof on day three and that Lean accepted it; the system later improved the result twice, again with Lean’s acceptance. If the proof holds, a quantum computer cannot save an exponential number of looks on yes/no questions about a random string unless the string has a rare pattern. Stanford researchers reading the proof called it “a huge result,” but it still needs independent mathematical review and a check that the formal statement matches the original conjecture. Gorden Sun’s “quantum bubble burst” line is commentary, not PSI’s wording.
Ethan Mollick: everyday work may over-index on AI factual-error risk; newer models beat human experts, including PhD specialists, at finding facts and checking claims
Boss lens: fact-finding tasks can take more AI first-pass review, but anything client-facing still needs a human signature—he did not publish scores in the text, so this is not a license to skip checks (inference).
CollapseEthan Mollick: everyday work may over-index on AI factual-error risk; newer models beat human experts, including PhD specialists, at finding facts and checking claims
Wharton professor Ethan Mollick argued that everyday work now expresses too much concern about verifying AI output for factual errors. That used to be warranted, he wrote, but data on more recent models suggest they beat human experts—including PhD specialists—at finding facts and checking claims. The post includes charts; it does not spell out scores in the text, so this item does not invent percentages from the images. This is a researcher’s reading of public evaluations, not a lab’s official claim. He and others still stress that users remain accountable for what they ship.
Gorden Sun on HyperFlow H3 Flash: a compressed MiniMax-H3 build, about 9.1GB at smallest, said to generate 768P video on 4× B200
Boss lens: a ~9GB video model lowers the cost of a local trial—“real time” and home-GPU claims still need a test on your own machines (inference).
CollapseGorden Sun on HyperFlow H3 Flash: a compressed MiniMax-H3 build, about 9.1GB at smallest, said to generate 768P video on 4× B200
Gorden Sun described HyperFlow H3 Flash as a compressed MiniMax-H3 build that uses a slimmer structure, low-precision quantization, and a new sampler to reach a smallest size of about 9.1GB. He said a typical home PC with a high-VRAM card can run it, and that four B200 GPUs can generate 768P video in “real time.” The code is on GitHub as Video-Rebirth/hyperflow-h3-flash. “Real time” is his wording; treat detailed speed claims as needing the project page. Engagement on the post was modest; it is included as a local video-generation option, not as a widely proven product.