Google ties school Search access to AI tools

Today's digest leads with Google's policy linking school Search access to its AI tools, plus the Gemini 4 Argon rollout and OpenAI's ChatGPT changes. It also covers Marvell's $115 billion Lumentum deal, Anthropic's IPO filing, a newspaper lawsuit against Brave, and new research on model weights.

AI lab releases

Google ties school Search access to accepting its AI tools

Google says its AI tools and Search will be off for K-12, or school-age, student accounts until administrators enable them. Schools that reject the AI features also lose Search, because Google says the two are deeply integrated. New York City Mayor Zohran Mamdani and Senator Bernie Sanders attacked the policy on X.

Why it matters: School leaders must now choose between allowing Google's AI and keeping its search engine.

Sources: The Verge, SiliconANGLE, TechCrunch, Ars Technica, The Information

Google rolls out Gemini 4 Argon and tests Gemini 4 Carbon

Google's October roundup says Gemini 4 Argon, a model for complex coding and legal work, has rolled out globally. Google calls its follow-up, Gemini 4 Carbon, its strongest model so far, which is the company's own description. Carbon is in pre-release testing with the US government and a group of trusted cyber defenders.

Why it matters: Developers and consumers can use Argon now, while wider access to Carbon depends on further safety work.

Sources: Google

OpenAI plans to merge ChatGPT's Chat and Work tabs

OpenAI told users by email that it will combine the Chat and Work tabs in ChatGPT. Select users get the merged view first, ahead of a wider rollout. ChatGPT and Codex will also share one allowance of tokens, the units that measure AI usage.

Why it matters: Some ChatGPT users may wait for usage resets more often, though OpenAI says most will not be affected.

Sources: TechCrunch, Neowin

DeepMind shows AI agent proposing drug candidates in an hour

Google DeepMind built chemistry skills for Google Antigravity that help an AI agent, software that carries out tasks by itself, handle molecular data. DeepMind says that in its own lung cancer demonstration, the agent proposed new molecules and ways to make them in an hour. The company says a single round of this work typically takes several days.

Why it matters: Drug researchers could spend less time cleaning data, if DeepMind's results hold up outside its own demonstration.

Sources: Google DeepMind

Acquisitions + deals

Marvell agrees to buy laser maker Lumentum in $115 billion deal

Marvell agreed to buy Lumentum, which makes lasers and optical network parts, paying entirely in its own shares. TechCrunch puts the enterprise value, which includes debt, at $115 billion and the equity value at about $80 billion. SiliconANGLE says Marvell shares fell almost 13% and Lumentum shares almost 19% after the news.

Why it matters: AI data center operators could end up buying networking chips and lasers from one company, if regulators and shareholders approve.

Sources: SiliconANGLE, TechCrunch, The Information

Fal raises $1.3 billion to run more AI media models

Fal sells inference, meaning it runs already trained AI models for customers, including image, video and audio models. The startup raised $1.3 billion in a round led by Bessemer Venture Partners. Bloomberg reports the deal values Fal at $16 billion, double its valuation in April.

Why it matters: Developers using hosted AI media models may see more capacity as this money turns into computing power.

Sources: TechCrunch, SiliconANGLE

Salesforce agrees to buy AI customer service startup Cresta

Salesforce agreed to buy Cresta, which makes AI software for customer service. Salesforce did not disclose a price, but The Information reports it is paying $2.7 billion. SiliconANGLE says Salesforce also agreed this year to pay $3.6 billion for Fin, formerly Intercom.

Why it matters: Salesforce customers may see more AI agents built into its customer service software.

Sources: SiliconANGLE, The Information

Anthropic filing shows $17 billion quarter and an operating loss

Anthropic's updated filing for its IPO, a first public sale of shares, shows $17 billion in third-quarter revenue. Quarterly growth slowed to 47% from 144% three months earlier. The company posted a $536 million operating loss after charges for short-notice hardware purchases and employee donation matching.

Why it matters: Investors weighing the listing can now compare Anthropic's fast sales growth with the rising cost of computing hardware.

Sources: SiliconANGLE, AP News

More funding

• Israeli startup 42AI launched with $62.5 million in seed funding for a model it says estimates how people will act. Sources: TechCrunch, SiliconANGLE
• Crunchbase News reports Chinese AI lab DeepSeek raised $15.9 billion, helping push October's global venture funding to $59 billion. Sources: Crunchbase News
• SiliconANGLE reports Traversal, which builds AI agents to diagnose and fix IT infrastructure problems, raised $120 million at a $1 billion valuation. Sources: SiliconANGLE
• SiliconANGLE reports Applied Compute, which runs a cloud platform for training and running AI models, raised $350 million at a $3.25 billion valuation. Sources: SiliconANGLE
• TechCrunch AI reports cybersecurity investor Ballistic Ventures is seeking $490 million across two new funds, based on regulatory filings. Sources: TechCrunch
• SiliconANGLE reports Athena, which automates legal work on commercial real estate deals, raised $2.5 million in pre-seed funding led by a16z Speedrun. Sources: SiliconANGLE
• TechCrunch Venture reports investor Yvonne Bajela closed a $17 million first fund, Evertrue Capital, to back early-stage startups in the UK and Europe. Sources: TechCrunch
• SiliconANGLE reports Larridin, which measures the value companies get from AI tools, raised a $30 million Series A led by Pelion Ventures. Sources: SiliconANGLE
• SiliconANGLE reports warehouse robot maker Tutor Intelligence raised a $75 million Series A led by Andreessen Horowitz. Sources: SiliconANGLE
• TechCrunch AI reports Trajectory, which helps businesses retrain AI models using feedback, raised a $45 million Series A led by Sequoia. Sources: TechCrunch

MCP + agent tooling

Bloomberg reports Goldman Sachs limits AI coding tools to engineers

Bloomberg reports Goldman Sachs has cut most staff off from advanced AI coding tools such as Claude Code. Engineers keep access under internal guidelines from October, according to Bloomberg's report. A Goldman spokesperson cited productivity and risk management, and said the experimental phase has ended.

Why it matters: Office workers hoping to build their own software with AI may find large employers pulling back access.

Sources: The Decoder, Bloomberg

Anthropic steers Claude Code away from cloning existing games

Anthropic changed the system prompt, the built-in instructions, for Claude Code on Opus 5.5 to discourage copying specific games. Anthropic's Addy Osmani said using Claude to infringe intellectual property breaks its usage policy. Original projects are meant to keep working, though The Decoder says some users report problems with them.

Why it matters: Game hobbyists using Claude Code may see refused requests, including on projects that are their own.

Sources: The Verge, The Decoder

LangSmith's bring-your-own-cloud option expands to Azure and Google Cloud

LangChain says LangSmith's Bring Your Own Cloud option is fully available on AWS and in beta on Azure and Google Cloud. Customers' traces, which are records of agent activity, stay in their own cloud account along with prompts and datasets. LangChain says it handles scaling and platform updates for those customers.

Why it matters: Teams with strict compliance rules gain more cloud choices without running the software themselves.

Sources: langchain.com

LinearB launches Agentic Studio to tackle AI code review backlogs

LinearB launched Agentic Studio, which gives engineers and AI coding agents background on why software delivery has stalled. LinearB's own report says AI-assisted pull requests, or proposed code changes, are nearly 2.6 times larger than others. The company also says those requests wait more than five times longer for review.

Why it matters: Engineering managers may need to look past code volume and check where AI-written work gets stuck.

Sources: SiliconANGLE, Business Insider

Web scraping + data tools

New York Times and other papers sue Brave over scraping, paywalls

The New York Times sued browser maker Brave in Manhattan federal court, and three other newspapers filed a similar case. The suits allege Brave scraped their articles, meaning copied them automatically, and sold access to AI firms including Anthropic. Brave did not immediately reply to requests for comment.

Why it matters: AI companies that buy web data from third parties could face closer questions about where it came from.

Sources: TechCrunch, The Verge

Research

Researchers say AI models can read and edit other models' weights

Researchers say GPT-6 Astra can read and change a model's raw weights, the numbers that set its behavior. The authors report that editing 0.016% of a small Qwen model's weights raised its math test score from 7.9% to 48.1%. Simon Willison says a model on his laptop can also interpret weights, while Matthew Honnibal calls the paper misleading.

Why it matters: Researchers may gain a new way to inspect and fix models, though the early claims are disputed.

Sources: rikiyat.github.io, Simon Willison

StateBench tests how AI models track changes in long documents

Researchers released StateBench, a benchmark, or standard test, of how language models follow changing information in long documents. The authors say recurrent and hybrid models, which carry a running memory, generalize further than transformers, another design, when trained directly on these tasks. They also say leading models, including Claude Opus 5.5, lose track on harder problems as inputs get longer.

Why it matters: Long tasks such as reading code or following a story can still trip up current models, so outputs need checking.

Sources: Hugging Face

Other notable news

Apple shares fall after Ternus's first earnings call as CEO

Apple reported 12% revenue growth to $114.48 billion on John Ternus's first earnings call as CEO. Finance chief Kevan Parekh said AI could lift revenue through third-party apps and partnerships, but TechCrunch says he gave no details. Shares fell after the company said higher memory prices would cut into holiday-quarter margins.

Why it matters: App developers get an early hint that Apple sees AI apps as a future income source.

Sources: SiliconANGLE, TechCrunch, The Verge, The Information

OpenAI hires former Army Secretary Dan Driscoll for security role

OpenAI hired Dan Driscoll, a former US Army Secretary, as a senior counselor on national security. OpenAI says he will advise on using AI for national and cyber security, especially with US allies. The Decoder says The Wall Street Journal sees the hire as a push for closer military ties and more government contracts.

Why it matters: People following military AI policy can expect OpenAI to take a more visible part in that debate.

Sources: OpenAI, The Decoder, The Information

Musk says Grok Bot limits reflect 'unsustainably high' costs

Some Grok Bot users found their weekly limits suddenly used up over the Halloween weekend. SpaceX CEO Elon Musk said on X this was no glitch, writing that serving the tool cost too much against revenue. Musk has not said how limits for subscribers will change from here.

Why it matters: Paying Grok Bot users cannot yet tell how much use their plans will include.

Sources: TechCrunch, Fortune, Business Insider

Binance says fake job interview played part in $6.3 billion hack

Binance says a bogus job interview appears to be part of a North Korean plan behind its $6.3 billion theft. The exchange says a software engineer was asked in a second-round interview to download an infected code repository. Binance says another engineer used a tampered Homebrew installation, a software setup tool, in June to get an AI coding tool.

Why it matters: Software engineers now have a concrete reason to treat interview coding tasks and install tools with more care.

Sources: Axios, The Information

That is all for today's digest, and ScrutifAI Daily will return with more AI and tech news in the next edition.

---

About ScrutifAI: ScrutifAI is an independent third-party agency that tests software for AI agents, building a single access point where agents and people can access and use tested tools without learning each one individually, at a fraction of the cost of buying them directly.