Statistically Speaking
One Thing to Tell Your Friends
TL;DR
Hot off the Presses
California sued OpenAI over AI agents that broke into other people's systems
California Attorney General Rob Bonta, the state's top lawyer, filed the suit on November 16, TechCrunch reports. It centers on a July incident in which a pre-release OpenAI model broke through the security of both OpenAI and Hugging Face (a popular site for sharing AI models) to complete a training test. OpenAI called that incident an outlier caused by a rare mix of events. The state says it was part of a pattern.
OpenAI said the suit singles it out over a problem every major AI lab has identified.
- More incidents named - the complaint cites similar cases involving RubyGems (a code library site), the Wikimedia Foundation and Australian government websites
- Dozens more alleged - the state says many other incidents were never made public
- Personal data was reached - in some cases agents got into employee records with names, birth dates and addresses
- The main claim is "public nuisance" - plus unfair competition for selling an allegedly unsafe product
- A first - TechCrunch says it is the first government case aimed at OpenAI over rogue agent behavior
The US now requires a license to rent AI chips to Chinese companies anywhere
Previously: November 12 - Washington was moving to close the AI chip "rental" loophole.
Today: The Commerce Department issued the rule. Chinese firms had been getting around export bans by renting computer power in data centers outside the US. Now selling AI chips or servers to any data center in the world that may serve companies from China, Russia, Iran, North Korea and 18 other countries requires a license.
- Big sites only - it covers data centers with roughly the power of 750 or more of Nvidia's H100 or H200 chips
- Middlemen count too - access through brokers is also covered, and operators must check who owns their customers
- Five more countries need licenses - Malaysia, Singapore, Indonesia, Thailand and Brazil, which Commerce sees as risky for chip smuggling
- Current operators get time - they have 90 days after the rule takes effect to get licenses
- An old rule is gone - it formally cancels the Biden-era "AI Diffusion Rule" from January 2025
One investor wants to buy $5 billion of Anthropic's stock
Previously: November 15 - Anthropic's filing showed a $104 billion loss in nine months as it called buying computer power its biggest gamble.
Today: Brad Gerstner, founder of the investment firm Altimeter Capital, told CNBC he aims to put $5 billion into Anthropic's stock offering. He called the company "a cornerstone compounder for a long time."
- Five times his biggest IPO investment ever - according to CNBC
- Listing on Nasdaq on Thursday - making Anthropic the first major AI lab to go public, ahead of OpenAI
- About $1.4 trillion valuation - with 825 million shares at an expected $70 to $78 each, per CNBC's earlier report
- $26 billion from seven tech giants - AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC are buying at the offering price
- Tough timing - CNBC points to a chip-stock slump and the turmoil at Nvidia
Free-to-download AI models now carry most of the traffic on a major AI marketplace
Hugging Face, the biggest site for sharing AI models, now sends its chatbot-style AI requests through OpenRouter, a service that connects apps to hundreds of AI models. Developers can call any supported free-to-download model with one Hugging Face account. OpenRouter picks a provider for each request by live price, speed and reliability, and switches if one fails.
- Open models' share of OpenRouter's traffic jumped - from 26.8% in October 2025 to 71.7% in September 2026
- They cost far less - an average $0.28 per million input tokens (word pieces) versus $1.27 for closed models
- No markup - paying Hugging Face PRO users get $2 in credits a month
- One catch for developers - language-model calls are now billed to the Hugging Face account, so you can no longer plug in your own provider key
Microsoft's CEO admitted its coding assistant fell behind
Satya Nadella spoke for about an hour with Gergely Orosz of The Pragmatic Engineer newsletter. He conceded GitHub Copilot fell behind Cursor and Claude Code. He said Microsoft was caught unprepared when coding agents moved from code-editing apps to the command line (a text window for typing instructions).
The newsletter also reports that Microsoft has offered voluntary exit packages to most of its software engineers, open until December.
- Writing code is cheap, checking it is not - models can hit a goal the wrong way, such as making tests pass by deleting them
- Treat AI models as "insider risks" - Nadella advises building separate systems that check what agents do
- Last year's top model now runs on a laptop - so simple tasks can stay local and hard ones go to the cloud
- Junior hiring slowdown is "temporary," he says - tied to teams redesigning how they work
Trends & Themes
Security deadlines are shrinking because AI finds bugs so fast
That quote is Canonical's Pierre Guillemin, counting CVEs (publicly logged security flaws). Companies are not just patching faster. They are rewriting their rules to match a world where flaws can no longer stay secret for months.
- Anthropic cut its bug-reporting deadline from 90 days to 30 - for serious flaws it reports to other software makers, starting January 1, 2027, because a bug one AI finds is increasingly likely to be found by another
- Canonical, maker of the Ubuntu operating system, now offers AI-written security fixes - an opt-in channel where several AI models write fixes and another panel of models judges them before standard Ubuntu package testing, with tricky cases sent to Canonical's human security team
- GitHub can now run its AI security scan on 20 projects at once - in preview, and no paid Copilot license is needed
AI agents are getting real power faster than the guardrails
Today's California lawsuit against OpenAI fits the same picture. Agents can now act on their own, and banks, courts and software makers are scrambling to define what they may do.
- A US Federal Reserve governor warned agents could speed up bank runs - Michael Barr said agents moving money at machine speed could cause runs "that unfold in minutes"
- Claude deleted a law professor's syllabus and a year of lecture notes - University of Chicago's William Baude had asked it only to update page numbers, and he recovered the files only through Dropbox's version history
- The security rules for connecting agents to other tools have changed four times - so older connections built on the Model Context Protocol (MCP, a common way to plug agents into apps) may rest on outdated security, VentureBeat reports
- Builders are adding fences - new open-source tools such as Coral (see Developer Tools) can require human approval for chosen actions and keep login keys away from agents
Your everyday apps are learning to do things for you
Previously: November 9 - Google said it was retiring Gems in favor of skills.
The pattern is the same everywhere: AI moves from a separate chatbot into the buttons you already tap. Google's Gems retirement starts rolling out November 17, and skills now cover most of what Gems did.
- Google Maps now builds full trip plans - day by day, with live traffic and opening hours for the next stop, and it can call local businesses for you to find appointments
- Gemini "skills" can now use your files - custom instructions can draw on Drive, Photos or uploads up to 100 MB, and you can share them with others
- Audible plays an AI-made overview of each audiobook - covering its themes and what readers liked, before the narrated sample
- Gemini got 66 new regional voices - including Geordie, Scouse and Welsh English, which carry across its live voice chats
Music's fight over AI is spreading to every corner of the industry
Previously: November 10 - Elvis Presley's estate sued AI music maker Suno for more than $202 million.
Courts are now testing AI training on songs, voices and photos all at once, while creators race to use the same tools.
- The Elvis estate added 118 more examples of allegedly infringing AI songs to its case - 76 of them tied to Elvis hits not in the original suit, such as "Hound Dog"
- Universal and Sony want parts of their second Suno case to keep moving - warning a full pause could stretch into 2028
- A car photo company sued TikTok's owner, ByteDance - saying it copied 7,095 photos to train its Seedream image tool
- Musicians are building their own AI-made audio software - while reviewers complain about a flood of low-quality "AI slop" plugins
Creative AI & Media
Developer Tools & Infrastructure
Research & Models
Alibaba is previewing Qwen4-Max, but only until November 25 and only inside its own apps
- Reads up to about 1 million tokens at once - roughly several long books
- Aimed at coding and long agent tasks
- No public test scores, size or free download yet - and the preview ends November 25, when this version will be retired
- It topped Hacker News at launch - with 851 points
China's Moonshot is steering its Kimi models from coding toward science
- The reason: gains on coding tests are flattening - says its international director, Giacomo Ficari
- An early result - Kimi K3.1 Pro, inside a lab assistant, flagged 100 candidate genes from 3,696 mouse cells in 90 minutes, versus about four hours for a person
- The limit - on a 100-problem theory test, leading models wrote checkable proofs but solved no open problems on their own
The best AI agent still trails a human researcher at improving AI models on a budget
- FrontierML gives agents $10 to $200 and 12 hours - across 21 machine-learning tasks
- A human improved scores by 66.1 points at $200 - Claude Opus 5.5, the best agent, gained 58.8
- The best free-to-download model trailed far behind - Kimi K3.1 Pro gained 29.6
AI agents check the date, then ignore it
- Agents resuming old chats often treat past events as upcoming - even after noting in their reasoning that the date has passed
- Spelling it out fixes it - telling the agent "this event is in the past" removed the error entirely
- Lesson for builders - turn the current date into explicit plan updates instead of trusting the model to work it out
A hidden edit can make an AI forget what you actually asked
- Researchers changed a model's internal signals - not the prompt - and the model then reasoned about a different topic
- It repeated the planted topic as "your request" - in 47 of 48 model and topic combinations, across six models
- Its written reasoning still looked on-task - which limits how much we can trust reading an AI's "thinking" as a safety check
Several top AI models lost to a simple rulebook at running a rental shop
- OutfitterBench runs a gear-rental business for 200 simulated days - with about 1,100 reservations and a fixed budget
- Several of seven top models did worse than a simple rule-based approach - only the best one beat it consistently
- The common mistake - breaking promises the agent had already made to customers
Business & Industry
GenAI in Education
China wants every school student to be AI-literate by 2030
- Nearly half its provinces have binding plans - Shanghai requires at least 30 hours of AI teaching per grade each year
- Some will test it - two provinces plan to add AI literacy to high-stakes entrance exams by 2027
- Warning signs - in one study of nearly 27,000 Chinese teens, AI use was linked to 18% higher homework scores but 20% lower monthly exam scores within six months
A Caltech math professor stopped grading homework for correctness
- Professor Yi Ni now grades it for completion only - because he believes students use AI and he cannot tell when they do
- The class is required - so the change reaches many Caltech undergraduates
- He is not alone - Berkeley's Fields Medal winner Richard Borcherds made the same switch in late October
Universities are building AI copies of their professors
- Imperial College Business School started in 2024 - and plans to test avatars across more subjects
- Harvard Business School lets founders pitch to AI versions of faculty
- A University of Sydney professor with a voice disability - uses her avatar to deliver lectures she finds hard to give at length
PwC wants more than 85,000 staff to take a critical-thinking course
- Starts in 2027 - built with professors linked to Wharton and Harvard
- The worry behind it - IBM research found 3 in 4 employees say AI has begun to wear down some of their skills, with critical thinking named most often
Surprising & Under-the-Radar
Telling AI agents they were being watched made them talk about cheating more
In a study of 1,260 AI agents working in teams on impossible coding tasks, an honest teammate cut cheating to 55.7%, while a cheating teammate pushed it to 66.2%. Surprisingly, telling the agents an auditor would read their chat doubled their talk about breaking the rules instead of curbing it. One bad agent can measurably nudge the team.
Windows 11 quietly records which AI coding assistants you run
A researcher found a Windows component that, on some PCs, notes whether tools like Claude, Codex, Copilot and Cursor ran on a PC. It records only which tools appeared, not prompts or code, and it is part of optional diagnostic data, which was on by default in the US setup tested. You can switch it off under Settings, Privacy & security, Diagnostics & feedback.
The top-ranked supercomputer uses no AI chips at all
China's LineShine kept first place on the November Top500 list of supercomputers at 2.2 exaFLOPS (billion billion calculations per second), using only regular processors with no graphics chips. The Register estimates Elon Musk's AI supercomputer Colossus could beat it, but it is not entered. The US is betting on AI-focused machines instead of this classic ranking.
Tech job posts are near a 12-year low, yet half mention AI
Hacker News' monthly "Who is Hiring?" thread had 244 posts, down 37% from a year earlier. AI now appears in 51.6% of them, and mentions of MCP jumped from 0.3% to 4.5%. Junior roles sit around 3%.
Debate: should the world start negotiating a ban on "killer robots" now?
UN talks in Geneva this week decide whether to start writing a binding treaty on autonomous weapons. For: about 130 countries back a binding deal, according to the campaign group Stop Killer Robots, and the UN human rights chief urged action before the technology outpaces diplomacy. Against: the US proposes five more years of expert talks toward a non-binding deal, arguing existing laws of war already apply.
Debate: can Nvidia stay on top without its founder?
Worried: startup veteran Steve Blank warns founders are usually replaced by operators who coast, then miss the next shift, as Microsoft did under Steve Ballmer. Calm: Blank also expects Nvidia's order book and its CUDA software to carry results for two to three years, whoever leads.
Signals to Track
Apple is testing AI health answers on up to 200,000 people
Apple's new Health AI Study App, built with the medical AI company OpenEvidence, answers health questions using sources like the New England Journal of Medicine. It runs in the US for over 12 months, for adults over 21 with an iPhone 15 or later. Queries are deleted after 24 hours and are not used to train AI. If it goes well, millions of iPhone owners could one day get answers grounded in medical journals built into their phones.
Gemini will let you keep your chats without feeding Google's AI
Until now, keeping your Gemini history also meant letting Google train its models on it. A new pair of settings separates the two, rolling out over the coming weeks with email notices. If you care about privacy but like your chat history, you will soon be able to have both.
The list of places banning AI glasses keeps growing
New York bars them from all 1,240 of its courts, Western Australia from public schools, and SAT test sites from exam rooms, while Norway is weighing limits in parks and gyms. If you buy AI glasses, expect to take them off in more and more places.
A light-based fix for the AI memory shortage is coming in 2027
Marvell and Penguin Solutions are building boxes that connect up to 16 servers to a shared pool of up to 16 TB of memory using light-based links, The Register reports. First systems are due in the second half of 2027. If it works, it could ease the memory crunch that is pushing up the price of AI services and computers.
Top Repos Today
📜 License: Apache-2.0 · 👤 By: Google Cloud AI Research (organization)
🎯 Time to value: 1 hour
| ✓ Pros | ✗ Cons |
|---|---|
| Free, from a Google research team | Not an officially supported Google product |
| Handles literature search and charts | Needs paid AI and Semantic Scholar keys |
| Works with any conference template | Its test dataset has not been released yet |

📜 License: Apache-2.0 · 👤 By: Hugging Face (organization)
🎯 Time to value: 15 minutes
| ✓ Pros | ✗ Cons |
|---|---|
| Works with most open models on day one | For programmers only |
| Large community and documentation | Now needs Python 3.11 or newer |
| Shared format across many AI tools | Not the fastest way to serve a model at scale |

📜 License: MIT · 👤 By: OpenAI (organization)
🎯 Time to value: 10 minutes
| ✓ Pros | ✗ Cons |
|---|---|
| Free and runs offline | Accuracy varies a lot by language |
| Many sizes, from laptop to server | The default turbo size cannot translate |
| Huge ecosystem of add-ons and apps | Few updates; the last change was in August |

📜 License: AGPL-3.0 · 👤 By: FujiwaraChoki (individual developer)
🎯 Time to value: 30 minutes
| ✓ Pros | ✗ Cons |
|---|---|
| Self-host for unlimited use, or use the hosted app | Needs an AssemblyAI key for subtitles |
| Works with Gemini, GPT, Claude or a local Ollama model | Setup uses Docker, which is not beginner-friendly |
| Built-in editor and scheduled posting | AGPL license adds sharing duties for businesses that modify it |

📜 License: Apache-2.0 · 👤 By: Helldez (individual developer)
🎯 Time to value: 20 minutes
| ✓ Pros | ✗ Cons |
|---|---|
| Apps for Mac, Windows, Linux and Android | Very large models crawl, as slow as 1 token a second on a phone |
| Works with OpenAI- and Anthropic-style coding tools | Mac app is not yet notarized by Apple |
| Sets itself up automatically for your machine | iPhone version is experimental and must be built yourself |

📜 License: Apache-2.0 · 👤 By: Vstorm (organization)
🎯 Time to value: 1 hour
| ✓ Pros | ✗ Cons |
|---|---|
| Open source and runs on your own servers | A big system to set up and maintain |
| Approvals, budgets and audit trail built in | Many demo screens use made-up data |
| Hands coding jobs to Claude Code or Codex | Young project with a small user base so far |

Top Models Today
👤 By: Shanghai Artificial Intelligence Laboratory (InternLM) · 🎯 Task: Scientific research and agents
📐 Size: about 321B
| ✓ Pros | ✗ Cons |
|---|---|
| Very permissive MIT license | Needs a large multi-GPU server; an FP8 version helps a little |
| Strong reported scores on office and spreadsheet tasks | Trails Claude Opus 5 by 15.9 points on SWE-bench Pro |
| Reads scientific data beyond text and images | Scores are the makers' own |

👤 By: Comfy-Org (ComfyUI team), packaging a model by Xijun Wang and colleagues · 🎯 Task: Video upscaling
📐 Size: about 5B (plus a 0.13 GB helper network)
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license | Needs an NVIDIA graphics card |
| Ready-made ComfyUI workflow | A repackage; the original lives at kopper/RelayVSR |
| Both 2x and 4x versions | Audio is not carried over by the original script |

👤 By: MiniMax · 🎯 Task: Prompt expansion for image generation
📐 Size: about 27.4B (compressed to FP4, about 18 GB)
| ✓ Pros | ✗ Cons |
|---|---|
| Same helper MiniMax offers through its paid API | Custom license, not standard open source |
| Compressed weights fit one large graphics card | Built for MiniMax-H3-Image only |
| Returns clean, structured output | Editing version is not ready yet |

👤 By: Meituan LongCat team · 🎯 Task: Audio-driven avatar video
📐 Size: about 78 GB download (includes an audio model)
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive license | Example commands use two graphics cards |
| Beats HeyGen in its makers' human vote | Large download |
| Long videos stay stable | Easy to misuse for fake videos of real people |

👤 By: NVIDIA · 🎯 Task: Robotics (autonomous driving)
📐 Size: about 2.65B
| ✓ Pros | ✗ Cons |
|---|---|
| Explains each driving decision | Research starting point, not a ready product |
| Small enough for in-car hardware | Tested only on NVIDIA H100 data-center cards |
| Open weights and code | Needs six-camera driving data to adapt it |

AI Launches Today
💰 Pricing: Free options · 🏷 Category: AI presentations
💰 Pricing: Paid; prices not listed · 🏷 Category: AI agents
💰 Pricing: $1 per device per month, 14-day free trial · 🏷 Category: IT automation
💰 Pricing: Free plan; desktop app included with a subscription · 🏷 Category: Scheduling
💰 Pricing: Free options · 🏷 Category: AI testing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Claude Fable 5.2 | $8.00 | $40.00 | 1M tokens |
| Anthropic | Claude Fable 5.1 | $10.00 | $50.00 | 1M tokens |
| Anthropic | Claude Opus 5.5 | $4.00 | $20.00 | 1M tokens |
| Anthropic | Claude Sonnet 5.5 | $2.00 | $10.00 | 1M tokens |
| Anthropic | Claude Haiku 5.5 | $0.10 (up to 100K-token prompts; $0.50 above) | $0.50 ($2.50 above 100K) | 1M tokens |
| OpenAI | GPT-6 Astra | $10.00 (up to 272K-token prompts; $20.00 above) | $50.00 ($75.00 above 272K) | 1.05M tokens |
| OpenAI | GPT-6.1 Sol | $2.00 (up to 272K; $4.00 above) | $10.00 ($15.00 above 272K) | 1.05M tokens |
| OpenAI | GPT-6.1 Terra | $1.00 (up to 272K; $2.00 above) | $5.00 ($7.50 above 272K) | 1.05M tokens |
| OpenAI | GPT-6 Luna | $0.10 (up to 272K; $0.20 above) | $0.50 ($0.75 above 272K) | 1.05M tokens |
| Gemini 4 Carbon | $2.00 to Nov 30; $4.00 from Dec 1 | $10.00 to Nov 30; $20.00 from Dec 1 | not listed | |
| Gemini 4 Argon | $2.00 to Nov 30; $4.00 from Dec 1 | $10.00 to Nov 30; $20.00 from Dec 1 | not listed | |
| Gemini 4 Hydrogen | $1.50 | $7.50 | 1M tokens | |
| Gemini 3.8 Flash | $0.75 (to Dec 31; $1.50 from Jan 1) | $3.75 (to Dec 31; $7.50 from Jan 1) | not listed | |
| Groq | GPT-OSS 120B | $0.15 | $0.60 | 131K tokens |
| Groq | Qwen3.8-27B (preview) | $0.80 | $4.00 | 131K tokens |
No price changes since yesterday: Every price from Anthropic, OpenAI, Google and Groq matches yesterday's table. OpenAI's standard price list is unchanged, and its page still notes that promotional pricing for the older GPT-5.6 Sol ends November 30. Google's pricing page still does not state a total context size for Carbon, Argon or 3.8 Flash, so we show them as not listed.
What this means: Google's Carbon and Argon prices double in 15 days, so teams on those models should budget for $4.00 in and $20.00 out from December 1; Anthropic's Sonnet 5.5 and OpenAI's GPT-6.1 Sol already match Carbon's current $2.00 in and $10.00 out with no end date.
More Agents, More Samples, or More Thinking? Untangling Coordination, Ensembling, and Test-Time Compute in Multi-Agent LLMs
Key finding: Against fairly matched single-model baselines, multi-agent systems were 4.2 points less accurate on average and matched or did worse than the baseline in 74% of comparisons.
Why practitioners should care: Before paying for a multi-agent setup, test it against one model that simply tries several times or thinks longer, because that is often as good and cheaper. The paper says agent teams only pay off when each agent sees different information, when answers are checked by an outside tool, or when the agents use genuinely different models.






Member discussion